diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..80c82ecd96a3263bfc34c9dbe1ba49d3a7ac2234 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b691660646c603efbd1f01eadbd3832781f9d4ccec7ba1fcafa2932752d2d518 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..cd18793132a99580088a9b6fb72f66f54b4bca95 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6bdad8726d78dce8092878565226b6c9924de83454ca458e1aae65eda463017 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..afdf0dcd4f810914d90dbb658b297fd375c85c89 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e108c5b843c4a90e61fccbbac63455e3e037aed95847578095ed39c3fc3df1b +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..7bda1aaf14dd85e2b1364052a6bb3ec8f34769f2 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2ee0b251acf25c28a03805774955d130a6997d7abf7055aaeab20c97baa32ba +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..36dc51ae62e74bcadc48856628896392b74fb01e --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3925103ca140a8fc0ede18e311f558db1d38d53851c54dc9c706e3e60d20439d +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..bf863f55d2800600c6118815a49482abb400ee58 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05cb127b301f0153d83a82f1e6c41b03c619d0fc7268c9abf32d6ef51875e70b +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..4f78f5797a1a50aa08e332a07de87593fd431037 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f71ebe89513d3eef6d294fbf421f4ea484596e6c15f1e8909a8c2d9568c5c127 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..7da40b961a3be66f838ba3027af26ff0e5503ff2 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s1__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0053fa6f28837e28b3a47293c23495d4c98a4919cd63c3418af0308b8deda3c8 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..bac4a7e863631a15dcff26d4e9323334a2d41750 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dc1eba7005716e32dc73c97b183aebe993c62b644b687ce570b8db153ba26c0 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..1c55b540199d3c67a60dfc28c1f2d5d24ee93864 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25938b9c94b67621b22e1ccb1dca886cfed58f640741c071dcf47d8109300506 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..17ee0b6fd9f4532118dffecf0b72646cbcf036d8 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c8e90b06e7bfb8ff7604d416a0a518cd8efb008c20479f81626a1658196bb67 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..7f8584f1ee7538661387193a138adc7554831b01 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2610c4342a6f8bfd34270c6710e1aa31459448fee06ca0cd4503f0bbfc9766c8 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..70ef838ba1cde31b9864ae08054aaecd813ca012 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e64389c573a84981be884282304ba756ebf18858ea8753dba5425076984ae63 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..39779bec27a75dcc4fbab0143ec16fcb065f4db2 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aebbc4bffcaedc8c50f34bda0fdd97f1e5bb9eba50ee68ea93e6cc5e3f84f1f9 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..668abef07ec3d338b94bbaf63923b81181d63351 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43b035238334eab07c05c04c23b0e57b7fa311de3fafb9b64aab978d81b5a80d +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..40cfb818e6af7a39545da332b3173db35791881b --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s2__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76e8f632a35b0d6b5d2fd017d607f5dea3e09c8b1dd81b90cbc996511c5e941e +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..4a6cf514aca2154fbfea9e97c2081de9dcc31ce8 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbc398661b688b2f46595f0cbb00337f484a017de6004cc5d3ba269652316f0c +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..daa0078f97d843d90e9c9a56039562a2cdf70151 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9c23835a1571423cb817cdacdffd3d9b17f919c0224c6317d12fb5f05452ee7 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..624736a3b05b85ee06197cd667936e62874a8a00 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a70b7d42bdcf7d86e5971eeb7a0f899aff9379e6410c5cf19f69dc95d2f571d +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..0a0187adf664cbfbcf0196291ece78f06aa63771 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6408ac54ef9f45756476cec5fddd0b14d427561671c87133746c4105783be265 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..e7e15c22e054314fbe1e94de4d8b56b50fb56049 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3426affd3423300d95614fef29eeed5441d619d54aafa235088b2bd4fb7316d +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..39a052d48b355295cca20a21a0d84ff06c1a3c0e --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd939fc018e234eb70cb998e59a7035d990ca3c700f9683b6a2643ac5d9c4db4 +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..e014d5e964973fde89380906956e10d78b7325b1 --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e33c175a4294c5c06b07962e1ce60d6e321f368174a8175aff6e1eed4b8442ea +size 438016720 diff --git a/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..9b3bbeef2433ceb264f6622872c7d25247d08c0c --- /dev/null +++ b/outputs/trace_s1_seqft__doc__s3__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d81b5a40f27df9ff6c39f40a82fd1f18677d363b6228e2aa0cf8f93e136589e8 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..1d1452666356298b3971242255103d1fa24cd4ae --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b31fb038be8edc89a0d4af1580dcc4c39d7c57242e2b12e34c3a318e7634c5ae +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_10.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_10.pt new file mode 100644 index 0000000000000000000000000000000000000000..6caf3c2b810c64f4997752d242435660b92402e2 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_10.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:958889eecdd14f003153e73a5f6bd14ac5e09065e52a9e28067675be023f0faa +size 438022686 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..1b8a2641a4e3efee9320de9f61ee57fb98269ee2 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88780f3a9f85226cbf9e3ca94dd6d06fbed0733cc542acb5846147e78cfd9ddd +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..9ada51cbae7ba47389cb6c48cfd5f18dfcb75157 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68a0cc5658773c6394e5b1055464faa7fbb1eb7cd92e2d93f520f82c346bff0e +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..14c6942faa330dd7a8e495626fa3fe20295fe70e --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b9a09124779cd854e9f1ff74e27adf76db77a9d5e6b58b2413bb05004ae2ea7 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..eb707af7ea177efa37e6367b2b619296e6649f57 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5aaf275477602170a081cf113a18cec0aaaff03a2d0f1ea90c7be5b597efd2d +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..9fd27822ea287f2ef25f9776800d9f3388bb985d --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6db1519a683a22eb912c2412e4ee82ba27be6894ee4915bcaf0ba948efcf469 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..01b4a9d7cd807ac07292562da0be8835da47f76b --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74d1fd7fec6ba4ccadd10b683f045b4239f82860d9e1698aae99b8733e3de379 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..8158404844997c9ef00ee1b9234bec85435962de --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:581997fcee03c18516edcf8398db69e91c1261a55411e8e7e9aef47e285f7676 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_9.pt b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_9.pt new file mode 100644 index 0000000000000000000000000000000000000000..f027b0c583e7973be8db3398b7b9748d7fe1d2e5 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s1__seed12/checkpoints/task_9.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f9a8fca2e661b12833c153d72283998316baada25936ed608ee7ef6737e6f19 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..ff12344f3553edb763aaa0bce6825d7151e30b1c --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb39b65a9d0304a61f35ba603a4d1deaeec59adbac9571f28167d514e88c34d1 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_10.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_10.pt new file mode 100644 index 0000000000000000000000000000000000000000..b404e4b9721f5439e31366bf2126c913f7541461 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_10.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cddfb402a40aaeba5f561dff32d4b0269b01fdfee375fedd687d0ce300851e1 +size 438022686 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..4ba2aab59def1ae724fa71cf9b0051e0dd6a485f --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cba2ae407fab742f0502c3a87443e2c73abc560019a18a4838e8867dcc6b5b21 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..687594a91fb7a402ae81a2c4fac9666d5a7b2b9e --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2116dac130c13242b5870da24db51b8e6a8c737342a4137b61ab39784cef838 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..78c78dec5365757d28d7bd2624d8754102bf0fca --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa9bc4fe994cc57942819cefc190bcd9d288e9d74eb2a760e120e1e9ab2be05a +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..b604e5f95a56241c2df37ec7c09ca5d879a27557 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaf923cfabd3886a1b351ba322173aa38a48d698c60c2e3c4a061d5c6c1cf48a +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..628ff64d4c4cd3d122fa199776a447b1b980c1b0 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3ee0e9175c467631e523894556a4d05f8c1b8f53791f4c5cc1a580112ed7df1 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..1de0cc4948f35d3829f0dcb9e1f402814b0156f9 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e62bdc21eb37d9b975aee68be7d3b44f14c768a62c73b8d61d8d8d3360894d5 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..c87445f7d889d337d3786e8f798282e9cd0df11b --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9222083ab8c2f440d14d763c5e677b047e5a68291bf754c19dedfed401b7297 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_9.pt b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_9.pt new file mode 100644 index 0000000000000000000000000000000000000000..52f34f556244559992d214676fe8b6a118865d22 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s2__seed12/checkpoints/task_9.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b5bc90d6ec5c0fe86355adac274b3f126cbd944083f7d935442f25090bb3017 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_1.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..88997e30eaad557f7053d7e144fe2a32e08c40a5 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95ccda55d08b1e93c6b44dd80a13d809ce0bb7651ddb79a9cb5a33205da4c855 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_10.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_10.pt new file mode 100644 index 0000000000000000000000000000000000000000..8eadc63b072128f3ef8eed07a0924a6d714e9664 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_10.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2a2daf9696c9fcc805159d300ff15956180caf5e3783d4343b9e2fe202352e5 +size 438022686 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_2.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..7dd6712a771adeee181ec7b6224df0b92b1aef33 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac1f115f981c9daf4e280b209455c8ec6a18eca8573429aec8cd68638c9743e7 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_3.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..a1c0eba89a0422b74e97fcd094a266cc863c0a87 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbee04c48daed2fd821a5b98ea4922b4ee43291042497da951c3dce789513a97 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_4.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..de7b80c38267547c14c176e21d93ce740e1b9810 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19885920b5e284b754e3b485e5e4ba4a782aaa74a01827945f31ed442732badf +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_5.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..b2d5dd46662c093d0d2eac61a80a4fe81e3907ca --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a23a5cdd3dda0d20837a5cdf1fead51f9ec54753cda1583155cefa21d0d5009 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_6.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..72d11792bdaba736ef136fd2167e2040a89f0542 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dccec8aa74f30cb174f06bcd6ec1a84dbd67f71732bd72f5a2792d0c63de676f +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_7.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..b710396271defad90eabc94f12cb9b0733ca8505 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4995c7c4698717b600a7ac6d78f1c80dfe378de7460b2fc0d3db602e35737c48 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_8.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..0cd4c6feb1fe68b8ceca6d076ef4c17aec4a7689 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42a8fd6d4b6e16c7357adab99b91ea44f447579857ed69dc46d6033bdc71ff01 +size 438016720 diff --git a/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_9.pt b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_9.pt new file mode 100644 index 0000000000000000000000000000000000000000..60e3ad8cc6b7466acf6737aa195203aca0221d07 --- /dev/null +++ b/outputs/trace_s1_seqft__x__s3__seed12/checkpoints/task_9.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00bd57dca5cdac4129a9c548a257f45a4ceb4520ab0fcae1448d3ae7a69ac063 +size 438016720 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_1.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..cde55ee52333e588791acf6aed44d912d52aa469 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ac1148ad368d45c25766b5e7cfccc6c2d8fff5d71cd50bbe018d296b4b9db91 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_2.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..ccfd8448d011c01f28f3d5076d177929f3a397b5 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825d49afac3b0f83f693c7d70cef3eb91e1b7d72b8845e88cf39b1290d6330f3 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_3.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..535bde9828fef00c2fc2c7e42c44becc75ef4a1c --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c762d96594521aa2745002e15a67e465b5b7edf92eea6446a0f16da981e1c96f +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_4.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..2be4f21ab091769d3749f4f37e41b5414f2959d0 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89d5bbaaaeadf2aabda63cc9463e74b213d4d50774d265658728aee125f099fd +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_5.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..de2c6503e0315ae43c07fad60822b20ead7a7da9 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4805d954ff95b6978688b9c0416d1852ccb22fe5c945ebdb793879914b04795b +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_6.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..d1cccaad35a793f3c020ad9d175f19975fa94272 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8bee3126312a0936902e9f7f720ce6009dbd4278fed04c8d41ce09ee8212484 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_7.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..9d7f9ee9408f376a11f426f15deda99c6f5bef70 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abf2b94747a19a925b2e0a1e8acc595942b576340ec262ef8268761493892adb +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_8.pt b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..f4051b09da98265d2bf5f5513d47fdb9d872581e --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s1__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:346d793e34ba120a8d9b87330c664ec8d12c591e57ebf0a47fe76f11e998784e +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_1.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_1.pt new file mode 100644 index 0000000000000000000000000000000000000000..eba2debfa0c66f61429728ddea744dcabc97c9d4 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_1.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90fee45171d737c8d0079a011096ae2c7b7d5960b77eab04e84bf615a10c79d9 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_2.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_2.pt new file mode 100644 index 0000000000000000000000000000000000000000..132d9ecaf751cb3122e36f54ec7cd92553015b4b --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_2.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c208a1ce9f72a6754c03bb42706507c4a0e81ea1e72a70b5b22a4acf362217b9 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_3.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_3.pt new file mode 100644 index 0000000000000000000000000000000000000000..8d7b62c8be033d1c80c420e9c0d6d6bffe7456eb --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_3.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c3bbff5a90b29a01ec4c2b372f36c01b61c5c68dc001c5ba10ca42c5c175661 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_4.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_4.pt new file mode 100644 index 0000000000000000000000000000000000000000..f115793483f90f3c21765175196f92a974918855 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_4.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80eb518b57858d69c976d5e9e659621f014780315d99e7a633d1b325e7dae43c +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_5.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_5.pt new file mode 100644 index 0000000000000000000000000000000000000000..243b61ea321e75049a93424b8de56c6c3a550d0c --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_5.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:979ff27b8e3f281f066ca2094c07f73aacabb3f4682f8863c0ba67d1255b508f +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_6.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_6.pt new file mode 100644 index 0000000000000000000000000000000000000000..8e6847379b2a545b803a2700b3b93726a5359596 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_6.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e906529d92c41e843fd3a6f15104053254982c5ddce46a72501258e0e0e340c +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_7.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_7.pt new file mode 100644 index 0000000000000000000000000000000000000000..a73fae7870091f020ead0ea94eb33bcfbf4beb15 --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_7.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69d769b69d7a5410d276ac63d2242cd78a65e6fc3f1206abb9a7f0f57bf1e6c2 +size 454543166 diff --git a/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_8.pt b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_8.pt new file mode 100644 index 0000000000000000000000000000000000000000..dac88ab65564851a35553894466bcd95938d1e1a --- /dev/null +++ b/outputs/trace_s2_target_fusion__doc__s2__seed12/checkpoints/task_8.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b351d01d111d81084ae7f87dc37eec71879cc4a7a76d9ecee46313c208d358b +size 454543166 diff --git a/outputs/trace_s4_lora__x__s2__seed13/environment.json b/outputs/trace_s4_lora__x__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..369c69265c42eb62e32b424ad937bd146feb7d53 --- /dev/null +++ b/outputs/trace_s4_lora__x__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T01:33:54.437553+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s2__seed13/metrics.json b/outputs/trace_s4_lora__x__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..a9acd2fa7f2db294bb55eae99c6eb2bf096d8263 --- /dev/null +++ b/outputs/trace_s4_lora__x__s2__seed13/metrics.json @@ -0,0 +1,1887 @@ +{ + "experiment": "trace_s4_lora", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 4, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": false, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6092455875737609, + 0.6929281036423894, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5364082132459386, + 0.6100504636189624, + 0.5839510549187968, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.45649257918704594, + 0.41366955484518614, + 0.4686878158045527, + 0.7412698412698413, + null, + null, + null, + null, + null, + null + ], + [ + 0.5049960774598455, + 0.5046790240763364, + 0.4134425615897645, + 0.6038726997515304, + 0.6295829509575229, + null, + null, + null, + null, + null + ], + [ + 0.5877622377622377, + 0.47761374053167366, + 0.4261118341659034, + 0.3178880336851542, + 0.47480620155038755, + 0.6973859322777249, + null, + null, + null, + null + ], + [ + 0.6970027428042695, + 0.513759433933012, + 0.4185959893924496, + 0.5047295756515614, + 0.5126730584646634, + 0.6116721756081024, + 0.5668057397583259, + null, + null, + null + ], + [ + 0.6376623376623377, + 0.5609080728901558, + 0.5311160022825466, + 0.5703685227910457, + 0.3960975671501987, + 0.42856634621466067, + 0.5437758245060477, + 0.6990940877304513, + null, + null + ], + [ + 0.6211129197564357, + 0.45792573571370343, + 0.41064008077422365, + 0.6242720990107954, + 0.5271907366983767, + 0.19893562095396958, + 0.5360357979669855, + 0.6716591638969526, + 0.6867039964866052, + null + ], + [ + 0.5580246913580247, + 0.504625154306683, + 0.3777150116128943, + 0.5562754265883952, + 0.45749502825526217, + 0.11902439024390243, + 0.43999999999999995, + 0.6078548657137619, + 0.6832518778626563, + 0.7887368192376464 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.622317787334289, + "losses": { + "current_ce": 0.6380637633078026, + "total": 0.6380637633078026, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7079646017699115, + "macro_f1": 0.6092455875737609, + "weighted_f1": 0.7658123404856192, + "macro_precision": 0.6463220373711936, + "macro_recall": 0.7457347409294988, + "per_class_f1": { + "FAVOR": 0.7719298245614035, + "AGAINST": 0.23529411764705882, + "NONE": 0.8205128205128205 + }, + "accuracy_pct": 70.79646017699115, + "macro_f1_pct": 60.92455875737609, + "weighted_f1_pct": 76.58123404856192, + "macro_precision_pct": 64.63220373711937, + "macro_recall_pct": 74.57347409294988, + "per_class_f1_pct": { + "FAVOR": 77.19298245614034, + "AGAINST": 23.52941176470588, + "NONE": 82.05128205128204 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.6929281036423894, + "weighted_f1": 0.7276337333941021, + "macro_precision": 0.6858566800076233, + "macro_recall": 0.7038720538720539, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7788461538461539, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 69.29281036423895, + "weighted_f1_pct": 72.76337333941021, + "macro_precision_pct": 68.58566800076234, + "macro_recall_pct": 70.38720538720538, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 77.88461538461539, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7190635451505016, + "macro_f1": 0.7211968166513621, + "weighted_f1": 0.7190343692319977, + "macro_precision": 0.730749354005168, + "macro_recall": 0.719884749948549, + "per_class_f1": { + "FAVOR": 0.7111111111111111, + "AGAINST": 0.7024793388429752, + "NONE": 0.75 + }, + "accuracy_pct": 71.90635451505017, + "macro_f1_pct": 72.11968166513621, + "weighted_f1_pct": 71.90343692319978, + "macro_precision_pct": 73.07493540051681, + "macro_recall_pct": 71.9884749948549, + "per_class_f1_pct": { + "FAVOR": 71.11111111111111, + "AGAINST": 70.24793388429752, + "NONE": 75.0 + } + }, + { + "after_task": 3, + "best_epoch": 14, + "best_dev_macro_f1": 0.6524064171122994, + "losses": { + "current_ce": 0.6875484300156435, + "total": 0.6875484300156435, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.584070796460177, + "macro_f1": 0.5364082132459386, + "weighted_f1": 0.6618284182262433, + "macro_precision": 0.627104377104377, + "macro_recall": 0.7283096711564131, + "per_class_f1": { + "FAVOR": 0.6213592233009708, + "AGAINST": 0.20408163265306123, + "NONE": 0.7837837837837838 + }, + "accuracy_pct": 58.4070796460177, + "macro_f1_pct": 53.64082132459386, + "weighted_f1_pct": 66.18284182262433, + "macro_precision_pct": 62.7104377104377, + "macro_recall_pct": 72.83096711564131, + "per_class_f1_pct": { + "FAVOR": 62.13592233009708, + "AGAINST": 20.408163265306122, + "NONE": 78.37837837837837 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6100504636189624, + "weighted_f1": 0.6764582766990875, + "macro_precision": 0.689714315738906, + "macro_recall": 0.6078002244668911, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7727272727272727, + "NONE": 0.6226415094339622 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 61.00504636189624, + "weighted_f1_pct": 67.64582766990875, + "macro_precision_pct": 68.9714315738906, + "macro_recall_pct": 60.78002244668911, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 77.27272727272727, + "NONE": 62.264150943396224 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5839510549187968, + "weighted_f1": 0.5712482781996638, + "macro_precision": 0.5785434472934473, + "macro_recall": 0.6017094017094017, + "per_class_f1": { + "FAVOR": 0.6129032258064516, + "AGAINST": 0.4380952380952381, + "NONE": 0.7008547008547008 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.39510549187968, + "weighted_f1_pct": 57.124827819966384, + "macro_precision_pct": 57.85434472934473, + "macro_recall_pct": 60.17094017094017, + "per_class_f1_pct": { + "FAVOR": 61.29032258064516, + "AGAINST": 43.80952380952381, + "NONE": 70.08547008547008 + } + } + ], + "accuracy": 0.6281179138321995, + "macro_f1": 0.6253599402005684, + "weighted_f1": 0.6245423025993074, + "macro_precision": 0.6543580722059398, + "macro_recall": 0.6229289238588057, + "per_class_f1": { + "FAVOR": 0.5781990521327014, + "AGAINST": 0.6042780748663101, + "NONE": 0.6936026936026936 + }, + "accuracy_pct": 62.81179138321995, + "macro_f1_pct": 62.53599402005684, + "weighted_f1_pct": 62.454230259930746, + "macro_precision_pct": 65.43580722059397, + "macro_recall_pct": 62.29289238588057, + "per_class_f1_pct": { + "FAVOR": 57.81990521327014, + "AGAINST": 60.42780748663101, + "NONE": 69.36026936026936 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7061555460507556, + "losses": { + "current_ce": 0.6013698600772498, + "total": 0.6013698600772498, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5486725663716814, + "macro_f1": 0.45649257918704594, + "weighted_f1": 0.5326432165571765, + "macro_precision": 0.5654761904761905, + "macro_recall": 0.6197306152165999, + "per_class_f1": { + "FAVOR": 0.4418604651162791, + "AGAINST": 0.20689655172413793, + "NONE": 0.7207207207207207 + }, + "accuracy_pct": 54.86725663716814, + "macro_f1_pct": 45.64925791870459, + "weighted_f1_pct": 53.26432165571765, + "macro_precision_pct": 56.547619047619044, + "macro_recall_pct": 61.973061521659986, + "per_class_f1_pct": { + "FAVOR": 44.18604651162791, + "AGAINST": 20.689655172413794, + "NONE": 72.07207207207207 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.521505376344086, + "macro_f1": 0.41366955484518614, + "weighted_f1": 0.4941320307692576, + "macro_precision": 0.5565854519774011, + "macro_recall": 0.5035353535353536, + "per_class_f1": { + "FAVOR": 0.10810810810810811, + "AGAINST": 0.5930232558139535, + "NONE": 0.5398773006134969 + }, + "accuracy_pct": 52.1505376344086, + "macro_f1_pct": 41.36695548451861, + "weighted_f1_pct": 49.41320307692576, + "macro_precision_pct": 55.658545197740104, + "macro_recall_pct": 50.353535353535364, + "per_class_f1_pct": { + "FAVOR": 10.81081081081081, + "AGAINST": 59.30232558139535, + "NONE": 53.987730061349694 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5140845070422535, + "macro_f1": 0.4686878158045527, + "weighted_f1": 0.4544548113811797, + "macro_precision": 0.5794249775381851, + "macro_recall": 0.5102564102564102, + "per_class_f1": { + "FAVOR": 0.4888888888888889, + "AGAINST": 0.2716049382716049, + "NONE": 0.6455696202531646 + }, + "accuracy_pct": 51.40845070422535, + "macro_f1_pct": 46.86878158045527, + "weighted_f1_pct": 45.44548113811797, + "macro_precision_pct": 57.942497753818515, + "macro_recall_pct": 51.02564102564102, + "per_class_f1_pct": { + "FAVOR": 48.888888888888886, + "AGAINST": 27.160493827160494, + "NONE": 64.55696202531645 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.75, + "macro_f1": 0.7412698412698413, + "weighted_f1": 0.7476904761904762, + "macro_precision": 0.7471623536430755, + "macro_recall": 0.7400988569663269, + "per_class_f1": { + "FAVOR": 0.7238095238095238, + "AGAINST": 0.7, + "NONE": 0.8 + }, + "accuracy_pct": 75.0, + "macro_f1_pct": 74.12698412698413, + "weighted_f1_pct": 74.76904761904763, + "macro_precision_pct": 74.71623536430755, + "macro_recall_pct": 74.0098856966327, + "per_class_f1_pct": { + "FAVOR": 72.38095238095238, + "AGAINST": 70.0, + "NONE": 80.0 + } + } + ], + "accuracy": 0.5959438377535101, + "macro_f1": 0.5735372896847238, + "weighted_f1": 0.5761388054676334, + "macro_precision": 0.6509192921741892, + "macro_recall": 0.5872656755009696, + "per_class_f1": { + "FAVOR": 0.5128205128205128, + "AGAINST": 0.5323383084577115, + "NONE": 0.6754530477759473 + }, + "accuracy_pct": 59.59438377535101, + "macro_f1_pct": 57.35372896847238, + "weighted_f1_pct": 57.61388054676334, + "macro_precision_pct": 65.09192921741892, + "macro_recall_pct": 58.726567550096966, + "per_class_f1_pct": { + "FAVOR": 51.28205128205128, + "AGAINST": 53.233830845771145, + "NONE": 67.54530477759472 + } + }, + { + "after_task": 5, + "best_epoch": 8, + "best_dev_macro_f1": 0.6656256024677077, + "losses": { + "current_ce": 0.6503845500640381, + "total": 0.6503845500640381, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6017699115044248, + "macro_f1": 0.5049960774598455, + "weighted_f1": 0.6012146902874105, + "macro_precision": 0.579957356076759, + "macro_recall": 0.6495813614852567, + "per_class_f1": { + "FAVOR": 0.5434782608695652, + "AGAINST": 0.23076923076923078, + "NONE": 0.7407407407407407 + }, + "accuracy_pct": 60.17699115044248, + "macro_f1_pct": 50.499607745984555, + "weighted_f1_pct": 60.121469028741046, + "macro_precision_pct": 57.995735607675904, + "macro_recall_pct": 64.95813614852567, + "per_class_f1_pct": { + "FAVOR": 54.347826086956516, + "AGAINST": 23.076923076923077, + "NONE": 74.07407407407408 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5645161290322581, + "macro_f1": 0.5046790240763364, + "weighted_f1": 0.5575754125853304, + "macro_precision": 0.5379826635145785, + "macro_recall": 0.557351290684624, + "per_class_f1": { + "FAVOR": 0.3018867924528302, + "AGAINST": 0.6222222222222222, + "NONE": 0.5899280575539568 + }, + "accuracy_pct": 56.451612903225815, + "macro_f1_pct": 50.46790240763364, + "weighted_f1_pct": 55.75754125853304, + "macro_precision_pct": 53.79826635145785, + "macro_recall_pct": 55.7351290684624, + "per_class_f1_pct": { + "FAVOR": 30.18867924528302, + "AGAINST": 62.22222222222222, + "NONE": 58.992805755395686 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.49295774647887325, + "macro_f1": 0.4134425615897645, + "weighted_f1": 0.36887397270515704, + "macro_precision": 0.5035146133925268, + "macro_recall": 0.5324786324786325, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.03225806451612903, + "NONE": 0.6455696202531646 + }, + "accuracy_pct": 49.29577464788733, + "macro_f1_pct": 41.34425615897645, + "weighted_f1_pct": 36.887397270515706, + "macro_precision_pct": 50.35146133925268, + "macro_recall_pct": 53.24786324786325, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 3.225806451612903, + "NONE": 64.55696202531645 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.64, + "macro_f1": 0.6038726997515304, + "weighted_f1": 0.6173525127804804, + "macro_precision": 0.7136904761904761, + "macro_recall": 0.6442075996292863, + "per_class_f1": { + "FAVOR": 0.6216216216216216, + "AGAINST": 0.4235294117647059, + "NONE": 0.7664670658682635 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 60.38726997515303, + "weighted_f1_pct": 61.73525127804804, + "macro_precision_pct": 71.3690476190476, + "macro_recall_pct": 64.42075996292863, + "per_class_f1_pct": { + "FAVOR": 62.16216216216216, + "AGAINST": 42.35294117647059, + "NONE": 76.64670658682635 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.66, + "macro_f1": 0.6295829509575229, + "weighted_f1": 0.6625940341288097, + "macro_precision": 0.6255882475428952, + "macro_recall": 0.6351125681222769, + "per_class_f1": { + "FAVOR": 0.7135678391959799, + "AGAINST": 0.4318181818181818, + "NONE": 0.7433628318584071 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 62.95829509575229, + "weighted_f1_pct": 66.25940341288097, + "macro_precision_pct": 62.558824754289525, + "macro_recall_pct": 63.51125681222769, + "per_class_f1_pct": { + "FAVOR": 71.35678391959799, + "AGAINST": 43.18181818181818, + "NONE": 74.33628318584071 + } + } + ], + "accuracy": 0.5980975029726516, + "macro_f1": 0.5797054515878831, + "weighted_f1": 0.5793039239284651, + "macro_precision": 0.60143925392541, + "macro_recall": 0.599407032072936, + "per_class_f1": { + "FAVOR": 0.60431654676259, + "AGAINST": 0.4399092970521542, + "NONE": 0.6948905109489051 + }, + "accuracy_pct": 59.80975029726516, + "macro_f1_pct": 57.97054515878831, + "weighted_f1_pct": 57.93039239284651, + "macro_precision_pct": 60.143925392541, + "macro_recall_pct": 59.9407032072936, + "per_class_f1_pct": { + "FAVOR": 60.431654676258994, + "AGAINST": 43.99092970521542, + "NONE": 69.48905109489051 + } + }, + { + "after_task": 6, + "best_epoch": 4, + "best_dev_macro_f1": 0.8413387896938799, + "losses": { + "current_ce": 0.47400298946433594, + "total": 0.47400298946433594, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5877622377622377, + "weighted_f1": 0.8415465065907544, + "macro_precision": 0.5781830462681526, + "macro_recall": 0.6006552602839461, + "per_class_f1": { + "FAVOR": 0.8769230769230769, + "AGAINST": 0.0, + "NONE": 0.8863636363636364 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 58.776223776223766, + "weighted_f1_pct": 84.15465065907543, + "macro_precision_pct": 57.81830462681526, + "macro_recall_pct": 60.065526028394615, + "per_class_f1_pct": { + "FAVOR": 87.6923076923077, + "AGAINST": 0.0, + "NONE": 88.63636363636364 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.45161290322580644, + "macro_f1": 0.47761374053167366, + "weighted_f1": 0.43867629412780906, + "macro_precision": 0.5768617347564716, + "macro_recall": 0.6003928170594838, + "per_class_f1": { + "FAVOR": 0.43537414965986393, + "AGAINST": 0.3546099290780142, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 45.16129032258064, + "macro_f1_pct": 47.761374053167366, + "weighted_f1_pct": 43.86762941278091, + "macro_precision_pct": 57.68617347564716, + "macro_recall_pct": 60.039281705948376, + "per_class_f1_pct": { + "FAVOR": 43.53741496598639, + "AGAINST": 35.46099290780142, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.47183098591549294, + "macro_f1": 0.4261118341659034, + "weighted_f1": 0.39300020071157743, + "macro_precision": 0.5110233210144004, + "macro_recall": 0.5576923076923077, + "per_class_f1": { + "FAVOR": 0.5283018867924528, + "AGAINST": 0.11940298507462686, + "NONE": 0.6306306306306306 + }, + "accuracy_pct": 47.183098591549296, + "macro_f1_pct": 42.61118341659034, + "weighted_f1_pct": 39.300020071157746, + "macro_precision_pct": 51.10233210144004, + "macro_recall_pct": 55.769230769230774, + "per_class_f1_pct": { + "FAVOR": 52.83018867924528, + "AGAINST": 11.940298507462686, + "NONE": 63.06306306306306 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.39, + "macro_f1": 0.3178880336851542, + "weighted_f1": 0.3212203984102016, + "macro_precision": 0.5567460317460318, + "macro_recall": 0.4364689527340129, + "per_class_f1": { + "FAVOR": 0.4636363636363636, + "AGAINST": 0.029850746268656716, + "NONE": 0.46017699115044247 + }, + "accuracy_pct": 39.0, + "macro_f1_pct": 31.788803368515424, + "weighted_f1_pct": 32.12203984102016, + "macro_precision_pct": 55.67460317460318, + "macro_recall_pct": 43.64689527340129, + "per_class_f1_pct": { + "FAVOR": 46.36363636363636, + "AGAINST": 2.9850746268656714, + "NONE": 46.017699115044245 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.655, + "macro_f1": 0.47480620155038755, + "weighted_f1": 0.5717151162790698, + "macro_precision": 0.51301138894191, + "macro_recall": 0.5058546631362165, + "per_class_f1": { + "FAVOR": 0.75, + "AGAINST": 0.0, + "NONE": 0.6744186046511628 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 47.48062015503876, + "weighted_f1_pct": 57.17151162790698, + "macro_precision_pct": 51.301138894191, + "macro_recall_pct": 50.58546631362165, + "per_class_f1_pct": { + "FAVOR": 75.0, + "AGAINST": 0.0, + "NONE": 67.44186046511628 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6973859322777249, + "weighted_f1": 0.7374246656138342, + "macro_precision": 0.6781009339148875, + "macro_recall": 0.7653273363318341, + "per_class_f1": { + "FAVOR": 0.6129032258064516, + "AGAINST": 0.7848101265822784, + "NONE": 0.6944444444444444 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.7385932277725, + "weighted_f1_pct": 73.74246656138342, + "macro_precision_pct": 67.81009339148875, + "macro_recall_pct": 76.53273363318341, + "per_class_f1_pct": { + "FAVOR": 61.29032258064516, + "AGAINST": 78.48101265822784, + "NONE": 69.44444444444444 + } + } + ], + "accuracy": 0.569402228976697, + "macro_f1": 0.5504147701382135, + "weighted_f1": 0.5392478579869764, + "macro_precision": 0.6722204997411084, + "macro_recall": 0.5910041718079794, + "per_class_f1": { + "FAVOR": 0.6168623265741728, + "AGAINST": 0.38095238095238093, + "NONE": 0.6534296028880866 + }, + "accuracy_pct": 56.9402228976697, + "macro_f1_pct": 55.04147701382135, + "weighted_f1_pct": 53.92478579869764, + "macro_precision_pct": 67.22204997411085, + "macro_recall_pct": 59.10041718079794, + "per_class_f1_pct": { + "FAVOR": 61.686232657417285, + "AGAINST": 38.095238095238095, + "NONE": 65.34296028880865 + } + }, + { + "after_task": 7, + "best_epoch": 9, + "best_dev_macro_f1": 0.5703703703703703, + "losses": { + "current_ce": 0.5868443047458475, + "total": 0.5868443047458475, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.6970027428042695, + "weighted_f1": 0.8430653316149499, + "macro_precision": 0.6870558785529716, + "macro_recall": 0.7095983497148404, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.36363636363636365, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 69.70027428042695, + "weighted_f1_pct": 84.306533161495, + "macro_precision_pct": 68.70558785529715, + "macro_recall_pct": 70.95983497148404, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 36.36363636363637, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.4946236559139785, + "macro_f1": 0.513759433933012, + "weighted_f1": 0.510672569632613, + "macro_precision": 0.6066007566007566, + "macro_recall": 0.5997194163860831, + "per_class_f1": { + "FAVOR": 0.42028985507246375, + "AGAINST": 0.4868421052631579, + "NONE": 0.6341463414634146 + }, + "accuracy_pct": 49.46236559139785, + "macro_f1_pct": 51.375943393301206, + "weighted_f1_pct": 51.0672569632613, + "macro_precision_pct": 60.66007566007566, + "macro_recall_pct": 59.97194163860831, + "per_class_f1_pct": { + "FAVOR": 42.028985507246375, + "AGAINST": 48.68421052631579, + "NONE": 63.41463414634146 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4507042253521127, + "macro_f1": 0.4185959893924496, + "weighted_f1": 0.3910084265312983, + "macro_precision": 0.46776748237902915, + "macro_recall": 0.5226495726495727, + "per_class_f1": { + "FAVOR": 0.5050505050505051, + "AGAINST": 0.16666666666666666, + "NONE": 0.584070796460177 + }, + "accuracy_pct": 45.07042253521127, + "macro_f1_pct": 41.85959893924496, + "weighted_f1_pct": 39.10084265312983, + "macro_precision_pct": 46.776748237902915, + "macro_recall_pct": 52.26495726495727, + "per_class_f1_pct": { + "FAVOR": 50.505050505050505, + "AGAINST": 16.666666666666664, + "NONE": 58.4070796460177 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.51, + "macro_f1": 0.5047295756515614, + "weighted_f1": 0.5023351864841227, + "macro_precision": 0.6703360427986951, + "macro_recall": 0.5549274019153537, + "per_class_f1": { + "FAVOR": 0.5291005291005291, + "AGAINST": 0.48936170212765956, + "NONE": 0.49572649572649574 + }, + "accuracy_pct": 51.0, + "macro_f1_pct": 50.47295756515614, + "weighted_f1_pct": 50.233518648412264, + "macro_precision_pct": 67.03360427986951, + "macro_recall_pct": 55.492740191535376, + "per_class_f1_pct": { + "FAVOR": 52.910052910052904, + "AGAINST": 48.93617021276596, + "NONE": 49.572649572649574 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.57, + "macro_f1": 0.5126730584646634, + "weighted_f1": 0.5671445130871292, + "macro_precision": 0.5974979321753515, + "macro_recall": 0.49874753078636574, + "per_class_f1": { + "FAVOR": 0.6872246696035242, + "AGAINST": 0.3191489361702128, + "NONE": 0.5316455696202531 + }, + "accuracy_pct": 56.99999999999999, + "macro_f1_pct": 51.267305846466336, + "weighted_f1_pct": 56.714451308712924, + "macro_precision_pct": 59.74979321753515, + "macro_recall_pct": 49.87475307863657, + "per_class_f1_pct": { + "FAVOR": 68.72246696035242, + "AGAINST": 31.914893617021278, + "NONE": 53.16455696202531 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6116721756081024, + "weighted_f1": 0.6806238745563532, + "macro_precision": 0.6069048055839471, + "macro_recall": 0.6447026486756622, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7602339181286549, + "NONE": 0.64 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 61.167217560810236, + "weighted_f1_pct": 68.06238745563532, + "macro_precision_pct": 60.69048055839471, + "macro_recall_pct": 64.47026486756621, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 76.0233918128655, + "NONE": 64.0 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5668057397583259, + "weighted_f1": 0.5807030631951231, + "macro_precision": 0.5653153153153153, + "macro_recall": 0.5958605664488017, + "per_class_f1": { + "FAVOR": 0.515625, + "AGAINST": 0.6206896551724138, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 56.680573975832594, + "weighted_f1_pct": 58.07030631951231, + "macro_precision_pct": 56.531531531531535, + "macro_recall_pct": 59.58605664488017, + "per_class_f1_pct": { + "FAVOR": 51.5625, + "AGAINST": 62.06896551724138, + "NONE": 56.41025641025641 + } + } + ], + "accuracy": 0.5734919286321155, + "macro_f1": 0.5743494783378914, + "weighted_f1": 0.5691140249013457, + "macro_precision": 0.607573986219562, + "macro_recall": 0.5881024803955239, + "per_class_f1": { + "FAVOR": 0.5887265135699373, + "AGAINST": 0.5260416666666666, + "NONE": 0.60828025477707 + }, + "accuracy_pct": 57.34919286321155, + "macro_f1_pct": 57.43494783378914, + "weighted_f1_pct": 56.91140249013456, + "macro_precision_pct": 60.757398621956206, + "macro_recall_pct": 58.81024803955239, + "per_class_f1_pct": { + "FAVOR": 58.872651356993735, + "AGAINST": 52.604166666666664, + "NONE": 60.828025477707 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.6987301418628128, + "losses": { + "current_ce": 0.5270507854945732, + "total": 0.5270507854945732, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6376623376623377, + "weighted_f1": 0.7787725548787496, + "macro_precision": 0.6530559559408816, + "macro_recall": 0.7845649799781581, + "per_class_f1": { + "FAVOR": 0.7636363636363637, + "AGAINST": 0.2857142857142857, + "NONE": 0.8636363636363636 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.76623376623377, + "weighted_f1_pct": 77.87725548787496, + "macro_precision_pct": 65.30559559408816, + "macro_recall_pct": 78.45649799781582, + "per_class_f1_pct": { + "FAVOR": 76.36363636363637, + "AGAINST": 28.57142857142857, + "NONE": 86.36363636363636 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5806451612903226, + "macro_f1": 0.5609080728901558, + "weighted_f1": 0.5985169762867267, + "macro_precision": 0.5688468798527101, + "macro_recall": 0.587598204264871, + "per_class_f1": { + "FAVOR": 0.40425531914893614, + "AGAINST": 0.6421052631578947, + "NONE": 0.6363636363636364 + }, + "accuracy_pct": 58.06451612903226, + "macro_f1_pct": 56.09080728901558, + "weighted_f1_pct": 59.851697628672675, + "macro_precision_pct": 56.88468798527101, + "macro_recall_pct": 58.759820426487096, + "per_class_f1_pct": { + "FAVOR": 40.42553191489361, + "AGAINST": 64.21052631578948, + "NONE": 63.63636363636363 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5492957746478874, + "macro_f1": 0.5311160022825466, + "weighted_f1": 0.5248240728302316, + "macro_precision": 0.6276196500105419, + "macro_recall": 0.5393162393162393, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.425531914893617, + "NONE": 0.6344827586206897 + }, + "accuracy_pct": 54.929577464788736, + "macro_f1_pct": 53.111600228254666, + "weighted_f1_pct": 52.48240728302316, + "macro_precision_pct": 62.761965001054186, + "macro_recall_pct": 53.93162393162393, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 42.5531914893617, + "NONE": 63.44827586206897 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.565, + "macro_f1": 0.5703685227910457, + "weighted_f1": 0.567193053563634, + "macro_precision": 0.6135231724597817, + "macro_recall": 0.5856039542786531, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.5189873417721519, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 57.036852279104565, + "weighted_f1_pct": 56.71930535636341, + "macro_precision_pct": 61.35231724597817, + "macro_recall_pct": 58.56039542786531, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 51.89873417721519, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.38, + "macro_f1": 0.3960975671501987, + "weighted_f1": 0.39631265664160403, + "macro_precision": 0.5449415848086945, + "macro_recall": 0.4331007719357234, + "per_class_f1": { + "FAVOR": 0.3815789473684211, + "AGAINST": 0.3391812865497076, + "NONE": 0.4675324675324675 + }, + "accuracy_pct": 38.0, + "macro_f1_pct": 39.60975671501987, + "weighted_f1_pct": 39.6312656641604, + "macro_precision_pct": 54.494158480869444, + "macro_recall_pct": 43.31007719357234, + "per_class_f1_pct": { + "FAVOR": 38.15789473684211, + "AGAINST": 33.91812865497076, + "NONE": 46.75324675324675 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.5684931506849316, + "macro_f1": 0.42856634621466067, + "weighted_f1": 0.5501617512257376, + "macro_precision": 0.4379808261332702, + "macro_recall": 0.5150824587706148, + "per_class_f1": { + "FAVOR": 0.0625, + "AGAINST": 0.6871165644171779, + "NONE": 0.5360824742268041 + }, + "accuracy_pct": 56.849315068493155, + "macro_f1_pct": 42.856634621466064, + "weighted_f1_pct": 55.01617512257376, + "macro_precision_pct": 43.79808261332702, + "macro_recall_pct": 51.508245877061476, + "per_class_f1_pct": { + "FAVOR": 6.25, + "AGAINST": 68.71165644171779, + "NONE": 53.608247422680414 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5437758245060477, + "weighted_f1": 0.5875620084694495, + "macro_precision": 0.564989927974146, + "macro_recall": 0.5428467683369643, + "per_class_f1": { + "FAVOR": 0.39080459770114945, + "AGAINST": 0.7111111111111111, + "NONE": 0.5294117647058824 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 54.37758245060477, + "weighted_f1_pct": 58.75620084694495, + "macro_precision_pct": 56.4989927974146, + "macro_recall_pct": 54.28467683369643, + "per_class_f1_pct": { + "FAVOR": 39.08045977011494, + "AGAINST": 71.11111111111111, + "NONE": 52.94117647058824 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6990940877304513, + "weighted_f1": 0.7169262555626192, + "macro_precision": 0.693687343358396, + "macro_recall": 0.7203056068230457, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.7403846153846154, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.90940877304513, + "weighted_f1_pct": 71.69262555626193, + "macro_precision_pct": 69.3687343358396, + "macro_recall_pct": 72.03056068230457, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 74.03846153846155, + "NONE": 71.07438016528926 + } + } + ], + "accuracy": 0.5807860262008734, + "macro_f1": 0.5748906997485045, + "weighted_f1": 0.5774160698203054, + "macro_precision": 0.5835905468800205, + "macro_recall": 0.5738261007117292, + "per_class_f1": { + "FAVOR": 0.5049928673323824, + "AGAINST": 0.5949878738884398, + "NONE": 0.6246913580246913 + }, + "accuracy_pct": 58.07860262008734, + "macro_f1_pct": 57.48906997485045, + "weighted_f1_pct": 57.74160698203053, + "macro_precision_pct": 58.359054688002054, + "macro_recall_pct": 57.38261007117293, + "per_class_f1_pct": { + "FAVOR": 50.499286733238236, + "AGAINST": 59.49878738884398, + "NONE": 62.46913580246913 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6569698670023203, + "losses": { + "current_ce": 0.5315607939920728, + "total": 0.5315607939920728, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6211129197564357, + "weighted_f1": 0.7687350846782132, + "macro_precision": 0.6168186423505572, + "macro_recall": 0.6774420580026695, + "per_class_f1": { + "FAVOR": 0.7719298245614035, + "AGAINST": 0.26666666666666666, + "NONE": 0.8247422680412371 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 62.11129197564357, + "weighted_f1_pct": 76.87350846782131, + "macro_precision_pct": 61.68186423505572, + "macro_recall_pct": 67.74420580026695, + "per_class_f1_pct": { + "FAVOR": 77.19298245614034, + "AGAINST": 26.666666666666668, + "NONE": 82.4742268041237 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.45698924731182794, + "macro_f1": 0.45792573571370343, + "weighted_f1": 0.4297189943120409, + "macro_precision": 0.5054012345679012, + "macro_recall": 0.5668350168350168, + "per_class_f1": { + "FAVOR": 0.45977011494252873, + "AGAINST": 0.375, + "NONE": 0.5390070921985816 + }, + "accuracy_pct": 45.69892473118279, + "macro_f1_pct": 45.79257357137034, + "weighted_f1_pct": 42.97189943120409, + "macro_precision_pct": 50.54012345679012, + "macro_recall_pct": 56.68350168350168, + "per_class_f1_pct": { + "FAVOR": 45.97701149425287, + "AGAINST": 37.5, + "NONE": 53.90070921985816 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4788732394366197, + "macro_f1": 0.41064008077422365, + "weighted_f1": 0.3818494623323045, + "macro_precision": 0.530667701863354, + "macro_recall": 0.4935897435897436, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.11940298507462686, + "NONE": 0.6219512195121951 + }, + "accuracy_pct": 47.88732394366197, + "macro_f1_pct": 41.064008077422365, + "weighted_f1_pct": 38.18494623323045, + "macro_precision_pct": 53.0667701863354, + "macro_recall_pct": 49.358974358974365, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 11.940298507462686, + "NONE": 62.19512195121951 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.645, + "macro_f1": 0.6242720990107954, + "weighted_f1": 0.6381286869325625, + "macro_precision": 0.66023802514824, + "macro_recall": 0.6403614457831325, + "per_class_f1": { + "FAVOR": 0.5954198473282443, + "AGAINST": 0.52, + "NONE": 0.757396449704142 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 62.427209901079536, + "weighted_f1_pct": 63.812868693256256, + "macro_precision_pct": 66.023802514824, + "macro_recall_pct": 64.03614457831324, + "per_class_f1_pct": { + "FAVOR": 59.541984732824424, + "AGAINST": 52.0, + "NONE": 75.7396449704142 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.515, + "macro_f1": 0.5271907366983767, + "weighted_f1": 0.5325695960330935, + "macro_precision": 0.5535485413677667, + "macro_recall": 0.5440647809579848, + "per_class_f1": { + "FAVOR": 0.5029239766081871, + "AGAINST": 0.3548387096774194, + "NONE": 0.7238095238095238 + }, + "accuracy_pct": 51.5, + "macro_f1_pct": 52.71907366983767, + "weighted_f1_pct": 53.25695960330935, + "macro_precision_pct": 55.354854136776666, + "macro_recall_pct": 54.406478095798484, + "per_class_f1_pct": { + "FAVOR": 50.29239766081871, + "AGAINST": 35.483870967741936, + "NONE": 72.38095238095238 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.2808219178082192, + "macro_f1": 0.19893562095396958, + "weighted_f1": 0.21355451134262246, + "macro_precision": 0.3126274509803922, + "macro_recall": 0.3768115942028985, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.22018348623853212, + "NONE": 0.37662337662337664 + }, + "accuracy_pct": 28.08219178082192, + "macro_f1_pct": 19.89356209539696, + "weighted_f1_pct": 21.355451134262246, + "macro_precision_pct": 31.26274509803922, + "macro_recall_pct": 37.68115942028985, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 22.018348623853214, + "NONE": 37.66233766233766 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5368421052631579, + "macro_f1": 0.5360357979669855, + "weighted_f1": 0.5370724257925613, + "macro_precision": 0.564452184382525, + "macro_recall": 0.6209150326797386, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.5408805031446541, + "NONE": 0.5378151260504201 + }, + "accuracy_pct": 53.68421052631579, + "macro_f1_pct": 53.603579796698554, + "weighted_f1_pct": 53.70724257925613, + "macro_precision_pct": 56.445218438252496, + "macro_recall_pct": 62.091503267973856, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 54.088050314465406, + "NONE": 53.78151260504202 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6716591638969526, + "weighted_f1": 0.663182198029981, + "macro_precision": 0.6847045505767615, + "macro_recall": 0.7653316086423843, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6432748538011696, + "NONE": 0.7050359712230215 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 67.16591638969525, + "weighted_f1_pct": 66.3182198029981, + "macro_precision_pct": 68.47045505767615, + "macro_recall_pct": 76.53316086423843, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 64.32748538011695, + "NONE": 70.50359712230215 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6867039964866052, + "weighted_f1": 0.6717737154150197, + "macro_precision": 0.6976790716286514, + "macro_recall": 0.7098456597101582, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.6931818181818182, + "NONE": 0.7608695652173914 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.67039964866052, + "weighted_f1_pct": 67.17737154150197, + "macro_precision_pct": 69.76790716286514, + "macro_recall_pct": 70.98456597101583, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 69.31818181818183, + "NONE": 76.08695652173914 + } + } + ], + "accuracy": 0.5603557814485387, + "macro_f1": 0.5576989286483144, + "weighted_f1": 0.5452881168721021, + "macro_precision": 0.5765499096834995, + "macro_recall": 0.5942393042474742, + "per_class_f1": { + "FAVOR": 0.5625692137320044, + "AGAINST": 0.4732394366197183, + "NONE": 0.6372881355932203 + }, + "accuracy_pct": 56.03557814485387, + "macro_f1_pct": 55.76989286483144, + "weighted_f1_pct": 54.52881168721021, + "macro_precision_pct": 57.65499096834995, + "macro_recall_pct": 59.42393042474742, + "per_class_f1_pct": { + "FAVOR": 56.25692137320044, + "AGAINST": 47.32394366197183, + "NONE": 63.72881355932203 + } + }, + { + "after_task": 10, + "best_epoch": 3, + "best_dev_macro_f1": 0.6723453012127972, + "losses": { + "current_ce": 0.379657160397619, + "total": 0.379657160397619, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7345132743362832, + "macro_f1": 0.5580246913580247, + "weighted_f1": 0.7490003277613897, + "macro_precision": 0.5584656084656084, + "macro_recall": 0.5692755733527485, + "per_class_f1": { + "FAVOR": 0.8, + "AGAINST": 0.13333333333333333, + "NONE": 0.7407407407407407 + }, + "accuracy_pct": 73.45132743362832, + "macro_f1_pct": 55.80246913580247, + "weighted_f1_pct": 74.90003277613897, + "macro_precision_pct": 55.84656084656084, + "macro_recall_pct": 56.92755733527485, + "per_class_f1_pct": { + "FAVOR": 80.0, + "AGAINST": 13.333333333333334, + "NONE": 74.07407407407408 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5, + "macro_f1": 0.504625154306683, + "weighted_f1": 0.5034851513174915, + "macro_precision": 0.53758496010312, + "macro_recall": 0.5866442199775533, + "per_class_f1": { + "FAVOR": 0.4144144144144144, + "AGAINST": 0.4840764331210191, + "NONE": 0.6153846153846154 + }, + "accuracy_pct": 50.0, + "macro_f1_pct": 50.4625154306683, + "weighted_f1_pct": 50.34851513174915, + "macro_precision_pct": 53.758496010311994, + "macro_recall_pct": 58.66442199775533, + "per_class_f1_pct": { + "FAVOR": 41.44144144144144, + "AGAINST": 48.40764331210191, + "NONE": 61.53846153846154 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4084507042253521, + "macro_f1": 0.3777150116128943, + "weighted_f1": 0.34798916133116076, + "macro_precision": 0.4477270864076078, + "macro_recall": 0.41923076923076924, + "per_class_f1": { + "FAVOR": 0.4782608695652174, + "AGAINST": 0.12658227848101267, + "NONE": 0.5283018867924528 + }, + "accuracy_pct": 40.845070422535215, + "macro_f1_pct": 37.77150116128943, + "weighted_f1_pct": 34.79891613311607, + "macro_precision_pct": 44.772708640760776, + "macro_recall_pct": 41.92307692307693, + "per_class_f1_pct": { + "FAVOR": 47.82608695652174, + "AGAINST": 12.658227848101266, + "NONE": 52.83018867924528 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.565, + "macro_f1": 0.5562754265883952, + "weighted_f1": 0.5635904187359041, + "macro_precision": 0.5710874093227034, + "macro_recall": 0.577386468952734, + "per_class_f1": { + "FAVOR": 0.583941605839416, + "AGAINST": 0.42735042735042733, + "NONE": 0.6575342465753424 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 55.62754265883952, + "weighted_f1_pct": 56.35904187359041, + "macro_precision_pct": 57.108740932270344, + "macro_recall_pct": 57.738646895273405, + "per_class_f1_pct": { + "FAVOR": 58.3941605839416, + "AGAINST": 42.73504273504273, + "NONE": 65.75342465753424 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.495, + "macro_f1": 0.45749502825526217, + "weighted_f1": 0.5055099533345147, + "macro_precision": 0.5232789777739834, + "macro_recall": 0.44925538323596576, + "per_class_f1": { + "FAVOR": 0.6124401913875598, + "AGAINST": 0.2909090909090909, + "NONE": 0.4691358024691358 + }, + "accuracy_pct": 49.5, + "macro_f1_pct": 45.749502825526214, + "weighted_f1_pct": 50.55099533345147, + "macro_precision_pct": 52.327897777398334, + "macro_recall_pct": 44.92553832359658, + "per_class_f1_pct": { + "FAVOR": 61.24401913875598, + "AGAINST": 29.09090909090909, + "NONE": 46.913580246913575 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.1917808219178082, + "macro_f1": 0.11902439024390243, + "weighted_f1": 0.08818576678917474, + "macro_precision": 0.14753086419753086, + "macro_recall": 0.3060969515242379, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.04, + "NONE": 0.3170731707317073 + }, + "accuracy_pct": 19.17808219178082, + "macro_f1_pct": 11.902439024390244, + "weighted_f1_pct": 8.818576678917474, + "macro_precision_pct": 14.753086419753087, + "macro_recall_pct": 30.60969515242379, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 4.0, + "NONE": 31.70731707317073 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.4421052631578947, + "macro_f1": 0.43999999999999995, + "weighted_f1": 0.44403007518796994, + "macro_precision": 0.5303599441590029, + "macro_recall": 0.5381263616557734, + "per_class_f1": { + "FAVOR": 0.42, + "AGAINST": 0.45714285714285713, + "NONE": 0.44285714285714284 + }, + "accuracy_pct": 44.21052631578947, + "macro_f1_pct": 43.99999999999999, + "weighted_f1_pct": 44.403007518797, + "macro_precision_pct": 53.03599441590029, + "macro_recall_pct": 53.812636165577345, + "per_class_f1_pct": { + "FAVOR": 42.0, + "AGAINST": 45.714285714285715, + "NONE": 44.285714285714285 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.6078548657137619, + "weighted_f1": 0.6098523791475676, + "macro_precision": 0.6324169530355097, + "macro_recall": 0.6726820521719011, + "per_class_f1": { + "FAVOR": 0.6027397260273972, + "AGAINST": 0.6127167630057804, + "NONE": 0.6081081081081081 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 60.78548657137619, + "weighted_f1_pct": 60.985237914756766, + "macro_precision_pct": 63.24169530355097, + "macro_recall_pct": 67.26820521719011, + "per_class_f1_pct": { + "FAVOR": 60.273972602739725, + "AGAINST": 61.27167630057804, + "NONE": 60.810810810810814 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6832518778626563, + "weighted_f1": 0.6794721296667404, + "macro_precision": 0.6938951310861423, + "macro_recall": 0.6772519861950758, + "per_class_f1": { + "FAVOR": 0.6538461538461539, + "AGAINST": 0.6946107784431138, + "NONE": 0.7012987012987013 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.32518778626563, + "weighted_f1_pct": 67.94721296667404, + "macro_precision_pct": 69.38951310861424, + "macro_recall_pct": 67.72519861950758, + "per_class_f1_pct": { + "FAVOR": 65.38461538461539, + "AGAINST": 69.46107784431138, + "NONE": 70.12987012987013 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7887368192376464, + "weighted_f1": 0.8183141825838472, + "macro_precision": 0.7712210086299892, + "macro_recall": 0.8141836489162827, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7704918032786885, + "NONE": 0.8623853211009175 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 78.87368192376464, + "weighted_f1_pct": 81.83141825838473, + "macro_precision_pct": 77.12210086299892, + "macro_recall_pct": 81.41836489162827, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 77.04918032786885, + "NONE": 86.23853211009175 + } + } + ], + "accuracy": 0.5507328072153326, + "macro_f1": 0.5500747457749592, + "weighted_f1": 0.5412769395277567, + "macro_precision": 0.5593191662898689, + "macro_recall": 0.5702260600415977, + "per_class_f1": { + "FAVOR": 0.5828571428571429, + "AGAINST": 0.4694915254237288, + "NONE": 0.5978755690440061 + }, + "accuracy_pct": 55.07328072153326, + "macro_f1_pct": 55.00747457749592, + "weighted_f1_pct": 54.127693952775665, + "macro_precision_pct": 55.93191662898689, + "macro_recall_pct": 57.02260600415977, + "per_class_f1_pct": { + "FAVOR": 58.285714285714285, + "AGAINST": 46.94915254237288, + "NONE": 59.78755690440061 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.5093003265179227, + "average_forgetting": 0.1957903887336251, + "backward_transfer": -0.1957903887336251, + "final_mean_target_macro_f1_pct": 50.93003265179227, + "average_forgetting_pct": 19.57903887336251, + "backward_transfer_pct": -19.57903887336251, + "final_pooled_accuracy": 0.5507328072153326, + "final_pooled_accuracy_pct": 55.07328072153326, + "final_pooled_macro_f1": 0.5500747457749592, + "final_pooled_macro_f1_pct": 55.00747457749592, + "final_pooled_weighted_f1": 0.5412769395277567, + "final_pooled_weighted_f1_pct": 54.127693952775665, + "final_pooled_macro_precision": 0.5593191662898689, + "final_pooled_macro_precision_pct": 55.93191662898689, + "final_pooled_macro_recall": 0.5702260600415977, + "final_pooled_macro_recall_pct": 57.02260600415977, + "final_pooled_per_class_f1": { + "FAVOR": 0.5828571428571429, + "AGAINST": 0.4694915254237288, + "NONE": 0.5978755690440061 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 58.285714285714285, + "AGAINST": 46.94915254237288, + "NONE": 59.78755690440061 + } + }, + "efficiency": { + "training_wall_time_seconds": 613.9015596951358, + "training_wall_time_minutes": 10.231692661585596, + "peak_gpu_memory_allocated_gb": 5.56232213973999, + "peak_gpu_memory_reserved_gb": 6.439453125 + } +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s2__seed13/metrics.partial.json b/outputs/trace_s4_lora__x__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..72da630ecaeff91a597835594537b1aa45c3cb6d --- /dev/null +++ b/outputs/trace_s4_lora__x__s2__seed13/metrics.partial.json @@ -0,0 +1,1826 @@ +{ + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6092455875737609, + 0.6929281036423894, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5364082132459386, + 0.6100504636189624, + 0.5839510549187968, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.45649257918704594, + 0.41366955484518614, + 0.4686878158045527, + 0.7412698412698413, + null, + null, + null, + null, + null, + null + ], + [ + 0.5049960774598455, + 0.5046790240763364, + 0.4134425615897645, + 0.6038726997515304, + 0.6295829509575229, + null, + null, + null, + null, + null + ], + [ + 0.5877622377622377, + 0.47761374053167366, + 0.4261118341659034, + 0.3178880336851542, + 0.47480620155038755, + 0.6973859322777249, + null, + null, + null, + null + ], + [ + 0.6970027428042695, + 0.513759433933012, + 0.4185959893924496, + 0.5047295756515614, + 0.5126730584646634, + 0.6116721756081024, + 0.5668057397583259, + null, + null, + null + ], + [ + 0.6376623376623377, + 0.5609080728901558, + 0.5311160022825466, + 0.5703685227910457, + 0.3960975671501987, + 0.42856634621466067, + 0.5437758245060477, + 0.6990940877304513, + null, + null + ], + [ + 0.6211129197564357, + 0.45792573571370343, + 0.41064008077422365, + 0.6242720990107954, + 0.5271907366983767, + 0.19893562095396958, + 0.5360357979669855, + 0.6716591638969526, + 0.6867039964866052, + null + ], + [ + 0.5580246913580247, + 0.504625154306683, + 0.3777150116128943, + 0.5562754265883952, + 0.45749502825526217, + 0.11902439024390243, + 0.43999999999999995, + 0.6078548657137619, + 0.6832518778626563, + 0.7887368192376464 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.622317787334289, + "losses": { + "current_ce": 0.6380637633078026, + "total": 0.6380637633078026, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7079646017699115, + "macro_f1": 0.6092455875737609, + "weighted_f1": 0.7658123404856192, + "macro_precision": 0.6463220373711936, + "macro_recall": 0.7457347409294988, + "per_class_f1": { + "FAVOR": 0.7719298245614035, + "AGAINST": 0.23529411764705882, + "NONE": 0.8205128205128205 + }, + "accuracy_pct": 70.79646017699115, + "macro_f1_pct": 60.92455875737609, + "weighted_f1_pct": 76.58123404856192, + "macro_precision_pct": 64.63220373711937, + "macro_recall_pct": 74.57347409294988, + "per_class_f1_pct": { + "FAVOR": 77.19298245614034, + "AGAINST": 23.52941176470588, + "NONE": 82.05128205128204 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.6929281036423894, + "weighted_f1": 0.7276337333941021, + "macro_precision": 0.6858566800076233, + "macro_recall": 0.7038720538720539, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7788461538461539, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 69.29281036423895, + "weighted_f1_pct": 72.76337333941021, + "macro_precision_pct": 68.58566800076234, + "macro_recall_pct": 70.38720538720538, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 77.88461538461539, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7190635451505016, + "macro_f1": 0.7211968166513621, + "weighted_f1": 0.7190343692319977, + "macro_precision": 0.730749354005168, + "macro_recall": 0.719884749948549, + "per_class_f1": { + "FAVOR": 0.7111111111111111, + "AGAINST": 0.7024793388429752, + "NONE": 0.75 + }, + "accuracy_pct": 71.90635451505017, + "macro_f1_pct": 72.11968166513621, + "weighted_f1_pct": 71.90343692319978, + "macro_precision_pct": 73.07493540051681, + "macro_recall_pct": 71.9884749948549, + "per_class_f1_pct": { + "FAVOR": 71.11111111111111, + "AGAINST": 70.24793388429752, + "NONE": 75.0 + } + }, + { + "after_task": 3, + "best_epoch": 14, + "best_dev_macro_f1": 0.6524064171122994, + "losses": { + "current_ce": 0.6875484300156435, + "total": 0.6875484300156435, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.584070796460177, + "macro_f1": 0.5364082132459386, + "weighted_f1": 0.6618284182262433, + "macro_precision": 0.627104377104377, + "macro_recall": 0.7283096711564131, + "per_class_f1": { + "FAVOR": 0.6213592233009708, + "AGAINST": 0.20408163265306123, + "NONE": 0.7837837837837838 + }, + "accuracy_pct": 58.4070796460177, + "macro_f1_pct": 53.64082132459386, + "weighted_f1_pct": 66.18284182262433, + "macro_precision_pct": 62.7104377104377, + "macro_recall_pct": 72.83096711564131, + "per_class_f1_pct": { + "FAVOR": 62.13592233009708, + "AGAINST": 20.408163265306122, + "NONE": 78.37837837837837 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6100504636189624, + "weighted_f1": 0.6764582766990875, + "macro_precision": 0.689714315738906, + "macro_recall": 0.6078002244668911, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7727272727272727, + "NONE": 0.6226415094339622 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 61.00504636189624, + "weighted_f1_pct": 67.64582766990875, + "macro_precision_pct": 68.9714315738906, + "macro_recall_pct": 60.78002244668911, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 77.27272727272727, + "NONE": 62.264150943396224 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5839510549187968, + "weighted_f1": 0.5712482781996638, + "macro_precision": 0.5785434472934473, + "macro_recall": 0.6017094017094017, + "per_class_f1": { + "FAVOR": 0.6129032258064516, + "AGAINST": 0.4380952380952381, + "NONE": 0.7008547008547008 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.39510549187968, + "weighted_f1_pct": 57.124827819966384, + "macro_precision_pct": 57.85434472934473, + "macro_recall_pct": 60.17094017094017, + "per_class_f1_pct": { + "FAVOR": 61.29032258064516, + "AGAINST": 43.80952380952381, + "NONE": 70.08547008547008 + } + } + ], + "accuracy": 0.6281179138321995, + "macro_f1": 0.6253599402005684, + "weighted_f1": 0.6245423025993074, + "macro_precision": 0.6543580722059398, + "macro_recall": 0.6229289238588057, + "per_class_f1": { + "FAVOR": 0.5781990521327014, + "AGAINST": 0.6042780748663101, + "NONE": 0.6936026936026936 + }, + "accuracy_pct": 62.81179138321995, + "macro_f1_pct": 62.53599402005684, + "weighted_f1_pct": 62.454230259930746, + "macro_precision_pct": 65.43580722059397, + "macro_recall_pct": 62.29289238588057, + "per_class_f1_pct": { + "FAVOR": 57.81990521327014, + "AGAINST": 60.42780748663101, + "NONE": 69.36026936026936 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7061555460507556, + "losses": { + "current_ce": 0.6013698600772498, + "total": 0.6013698600772498, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5486725663716814, + "macro_f1": 0.45649257918704594, + "weighted_f1": 0.5326432165571765, + "macro_precision": 0.5654761904761905, + "macro_recall": 0.6197306152165999, + "per_class_f1": { + "FAVOR": 0.4418604651162791, + "AGAINST": 0.20689655172413793, + "NONE": 0.7207207207207207 + }, + "accuracy_pct": 54.86725663716814, + "macro_f1_pct": 45.64925791870459, + "weighted_f1_pct": 53.26432165571765, + "macro_precision_pct": 56.547619047619044, + "macro_recall_pct": 61.973061521659986, + "per_class_f1_pct": { + "FAVOR": 44.18604651162791, + "AGAINST": 20.689655172413794, + "NONE": 72.07207207207207 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.521505376344086, + "macro_f1": 0.41366955484518614, + "weighted_f1": 0.4941320307692576, + "macro_precision": 0.5565854519774011, + "macro_recall": 0.5035353535353536, + "per_class_f1": { + "FAVOR": 0.10810810810810811, + "AGAINST": 0.5930232558139535, + "NONE": 0.5398773006134969 + }, + "accuracy_pct": 52.1505376344086, + "macro_f1_pct": 41.36695548451861, + "weighted_f1_pct": 49.41320307692576, + "macro_precision_pct": 55.658545197740104, + "macro_recall_pct": 50.353535353535364, + "per_class_f1_pct": { + "FAVOR": 10.81081081081081, + "AGAINST": 59.30232558139535, + "NONE": 53.987730061349694 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5140845070422535, + "macro_f1": 0.4686878158045527, + "weighted_f1": 0.4544548113811797, + "macro_precision": 0.5794249775381851, + "macro_recall": 0.5102564102564102, + "per_class_f1": { + "FAVOR": 0.4888888888888889, + "AGAINST": 0.2716049382716049, + "NONE": 0.6455696202531646 + }, + "accuracy_pct": 51.40845070422535, + "macro_f1_pct": 46.86878158045527, + "weighted_f1_pct": 45.44548113811797, + "macro_precision_pct": 57.942497753818515, + "macro_recall_pct": 51.02564102564102, + "per_class_f1_pct": { + "FAVOR": 48.888888888888886, + "AGAINST": 27.160493827160494, + "NONE": 64.55696202531645 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.75, + "macro_f1": 0.7412698412698413, + "weighted_f1": 0.7476904761904762, + "macro_precision": 0.7471623536430755, + "macro_recall": 0.7400988569663269, + "per_class_f1": { + "FAVOR": 0.7238095238095238, + "AGAINST": 0.7, + "NONE": 0.8 + }, + "accuracy_pct": 75.0, + "macro_f1_pct": 74.12698412698413, + "weighted_f1_pct": 74.76904761904763, + "macro_precision_pct": 74.71623536430755, + "macro_recall_pct": 74.0098856966327, + "per_class_f1_pct": { + "FAVOR": 72.38095238095238, + "AGAINST": 70.0, + "NONE": 80.0 + } + } + ], + "accuracy": 0.5959438377535101, + "macro_f1": 0.5735372896847238, + "weighted_f1": 0.5761388054676334, + "macro_precision": 0.6509192921741892, + "macro_recall": 0.5872656755009696, + "per_class_f1": { + "FAVOR": 0.5128205128205128, + "AGAINST": 0.5323383084577115, + "NONE": 0.6754530477759473 + }, + "accuracy_pct": 59.59438377535101, + "macro_f1_pct": 57.35372896847238, + "weighted_f1_pct": 57.61388054676334, + "macro_precision_pct": 65.09192921741892, + "macro_recall_pct": 58.726567550096966, + "per_class_f1_pct": { + "FAVOR": 51.28205128205128, + "AGAINST": 53.233830845771145, + "NONE": 67.54530477759472 + } + }, + { + "after_task": 5, + "best_epoch": 8, + "best_dev_macro_f1": 0.6656256024677077, + "losses": { + "current_ce": 0.6503845500640381, + "total": 0.6503845500640381, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6017699115044248, + "macro_f1": 0.5049960774598455, + "weighted_f1": 0.6012146902874105, + "macro_precision": 0.579957356076759, + "macro_recall": 0.6495813614852567, + "per_class_f1": { + "FAVOR": 0.5434782608695652, + "AGAINST": 0.23076923076923078, + "NONE": 0.7407407407407407 + }, + "accuracy_pct": 60.17699115044248, + "macro_f1_pct": 50.499607745984555, + "weighted_f1_pct": 60.121469028741046, + "macro_precision_pct": 57.995735607675904, + "macro_recall_pct": 64.95813614852567, + "per_class_f1_pct": { + "FAVOR": 54.347826086956516, + "AGAINST": 23.076923076923077, + "NONE": 74.07407407407408 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5645161290322581, + "macro_f1": 0.5046790240763364, + "weighted_f1": 0.5575754125853304, + "macro_precision": 0.5379826635145785, + "macro_recall": 0.557351290684624, + "per_class_f1": { + "FAVOR": 0.3018867924528302, + "AGAINST": 0.6222222222222222, + "NONE": 0.5899280575539568 + }, + "accuracy_pct": 56.451612903225815, + "macro_f1_pct": 50.46790240763364, + "weighted_f1_pct": 55.75754125853304, + "macro_precision_pct": 53.79826635145785, + "macro_recall_pct": 55.7351290684624, + "per_class_f1_pct": { + "FAVOR": 30.18867924528302, + "AGAINST": 62.22222222222222, + "NONE": 58.992805755395686 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.49295774647887325, + "macro_f1": 0.4134425615897645, + "weighted_f1": 0.36887397270515704, + "macro_precision": 0.5035146133925268, + "macro_recall": 0.5324786324786325, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.03225806451612903, + "NONE": 0.6455696202531646 + }, + "accuracy_pct": 49.29577464788733, + "macro_f1_pct": 41.34425615897645, + "weighted_f1_pct": 36.887397270515706, + "macro_precision_pct": 50.35146133925268, + "macro_recall_pct": 53.24786324786325, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 3.225806451612903, + "NONE": 64.55696202531645 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.64, + "macro_f1": 0.6038726997515304, + "weighted_f1": 0.6173525127804804, + "macro_precision": 0.7136904761904761, + "macro_recall": 0.6442075996292863, + "per_class_f1": { + "FAVOR": 0.6216216216216216, + "AGAINST": 0.4235294117647059, + "NONE": 0.7664670658682635 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 60.38726997515303, + "weighted_f1_pct": 61.73525127804804, + "macro_precision_pct": 71.3690476190476, + "macro_recall_pct": 64.42075996292863, + "per_class_f1_pct": { + "FAVOR": 62.16216216216216, + "AGAINST": 42.35294117647059, + "NONE": 76.64670658682635 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.66, + "macro_f1": 0.6295829509575229, + "weighted_f1": 0.6625940341288097, + "macro_precision": 0.6255882475428952, + "macro_recall": 0.6351125681222769, + "per_class_f1": { + "FAVOR": 0.7135678391959799, + "AGAINST": 0.4318181818181818, + "NONE": 0.7433628318584071 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 62.95829509575229, + "weighted_f1_pct": 66.25940341288097, + "macro_precision_pct": 62.558824754289525, + "macro_recall_pct": 63.51125681222769, + "per_class_f1_pct": { + "FAVOR": 71.35678391959799, + "AGAINST": 43.18181818181818, + "NONE": 74.33628318584071 + } + } + ], + "accuracy": 0.5980975029726516, + "macro_f1": 0.5797054515878831, + "weighted_f1": 0.5793039239284651, + "macro_precision": 0.60143925392541, + "macro_recall": 0.599407032072936, + "per_class_f1": { + "FAVOR": 0.60431654676259, + "AGAINST": 0.4399092970521542, + "NONE": 0.6948905109489051 + }, + "accuracy_pct": 59.80975029726516, + "macro_f1_pct": 57.97054515878831, + "weighted_f1_pct": 57.93039239284651, + "macro_precision_pct": 60.143925392541, + "macro_recall_pct": 59.9407032072936, + "per_class_f1_pct": { + "FAVOR": 60.431654676258994, + "AGAINST": 43.99092970521542, + "NONE": 69.48905109489051 + } + }, + { + "after_task": 6, + "best_epoch": 4, + "best_dev_macro_f1": 0.8413387896938799, + "losses": { + "current_ce": 0.47400298946433594, + "total": 0.47400298946433594, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5877622377622377, + "weighted_f1": 0.8415465065907544, + "macro_precision": 0.5781830462681526, + "macro_recall": 0.6006552602839461, + "per_class_f1": { + "FAVOR": 0.8769230769230769, + "AGAINST": 0.0, + "NONE": 0.8863636363636364 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 58.776223776223766, + "weighted_f1_pct": 84.15465065907543, + "macro_precision_pct": 57.81830462681526, + "macro_recall_pct": 60.065526028394615, + "per_class_f1_pct": { + "FAVOR": 87.6923076923077, + "AGAINST": 0.0, + "NONE": 88.63636363636364 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.45161290322580644, + "macro_f1": 0.47761374053167366, + "weighted_f1": 0.43867629412780906, + "macro_precision": 0.5768617347564716, + "macro_recall": 0.6003928170594838, + "per_class_f1": { + "FAVOR": 0.43537414965986393, + "AGAINST": 0.3546099290780142, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 45.16129032258064, + "macro_f1_pct": 47.761374053167366, + "weighted_f1_pct": 43.86762941278091, + "macro_precision_pct": 57.68617347564716, + "macro_recall_pct": 60.039281705948376, + "per_class_f1_pct": { + "FAVOR": 43.53741496598639, + "AGAINST": 35.46099290780142, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.47183098591549294, + "macro_f1": 0.4261118341659034, + "weighted_f1": 0.39300020071157743, + "macro_precision": 0.5110233210144004, + "macro_recall": 0.5576923076923077, + "per_class_f1": { + "FAVOR": 0.5283018867924528, + "AGAINST": 0.11940298507462686, + "NONE": 0.6306306306306306 + }, + "accuracy_pct": 47.183098591549296, + "macro_f1_pct": 42.61118341659034, + "weighted_f1_pct": 39.300020071157746, + "macro_precision_pct": 51.10233210144004, + "macro_recall_pct": 55.769230769230774, + "per_class_f1_pct": { + "FAVOR": 52.83018867924528, + "AGAINST": 11.940298507462686, + "NONE": 63.06306306306306 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.39, + "macro_f1": 0.3178880336851542, + "weighted_f1": 0.3212203984102016, + "macro_precision": 0.5567460317460318, + "macro_recall": 0.4364689527340129, + "per_class_f1": { + "FAVOR": 0.4636363636363636, + "AGAINST": 0.029850746268656716, + "NONE": 0.46017699115044247 + }, + "accuracy_pct": 39.0, + "macro_f1_pct": 31.788803368515424, + "weighted_f1_pct": 32.12203984102016, + "macro_precision_pct": 55.67460317460318, + "macro_recall_pct": 43.64689527340129, + "per_class_f1_pct": { + "FAVOR": 46.36363636363636, + "AGAINST": 2.9850746268656714, + "NONE": 46.017699115044245 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.655, + "macro_f1": 0.47480620155038755, + "weighted_f1": 0.5717151162790698, + "macro_precision": 0.51301138894191, + "macro_recall": 0.5058546631362165, + "per_class_f1": { + "FAVOR": 0.75, + "AGAINST": 0.0, + "NONE": 0.6744186046511628 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 47.48062015503876, + "weighted_f1_pct": 57.17151162790698, + "macro_precision_pct": 51.301138894191, + "macro_recall_pct": 50.58546631362165, + "per_class_f1_pct": { + "FAVOR": 75.0, + "AGAINST": 0.0, + "NONE": 67.44186046511628 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6973859322777249, + "weighted_f1": 0.7374246656138342, + "macro_precision": 0.6781009339148875, + "macro_recall": 0.7653273363318341, + "per_class_f1": { + "FAVOR": 0.6129032258064516, + "AGAINST": 0.7848101265822784, + "NONE": 0.6944444444444444 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.7385932277725, + "weighted_f1_pct": 73.74246656138342, + "macro_precision_pct": 67.81009339148875, + "macro_recall_pct": 76.53273363318341, + "per_class_f1_pct": { + "FAVOR": 61.29032258064516, + "AGAINST": 78.48101265822784, + "NONE": 69.44444444444444 + } + } + ], + "accuracy": 0.569402228976697, + "macro_f1": 0.5504147701382135, + "weighted_f1": 0.5392478579869764, + "macro_precision": 0.6722204997411084, + "macro_recall": 0.5910041718079794, + "per_class_f1": { + "FAVOR": 0.6168623265741728, + "AGAINST": 0.38095238095238093, + "NONE": 0.6534296028880866 + }, + "accuracy_pct": 56.9402228976697, + "macro_f1_pct": 55.04147701382135, + "weighted_f1_pct": 53.92478579869764, + "macro_precision_pct": 67.22204997411085, + "macro_recall_pct": 59.10041718079794, + "per_class_f1_pct": { + "FAVOR": 61.686232657417285, + "AGAINST": 38.095238095238095, + "NONE": 65.34296028880865 + } + }, + { + "after_task": 7, + "best_epoch": 9, + "best_dev_macro_f1": 0.5703703703703703, + "losses": { + "current_ce": 0.5868443047458475, + "total": 0.5868443047458475, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.6970027428042695, + "weighted_f1": 0.8430653316149499, + "macro_precision": 0.6870558785529716, + "macro_recall": 0.7095983497148404, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.36363636363636365, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 69.70027428042695, + "weighted_f1_pct": 84.306533161495, + "macro_precision_pct": 68.70558785529715, + "macro_recall_pct": 70.95983497148404, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 36.36363636363637, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.4946236559139785, + "macro_f1": 0.513759433933012, + "weighted_f1": 0.510672569632613, + "macro_precision": 0.6066007566007566, + "macro_recall": 0.5997194163860831, + "per_class_f1": { + "FAVOR": 0.42028985507246375, + "AGAINST": 0.4868421052631579, + "NONE": 0.6341463414634146 + }, + "accuracy_pct": 49.46236559139785, + "macro_f1_pct": 51.375943393301206, + "weighted_f1_pct": 51.0672569632613, + "macro_precision_pct": 60.66007566007566, + "macro_recall_pct": 59.97194163860831, + "per_class_f1_pct": { + "FAVOR": 42.028985507246375, + "AGAINST": 48.68421052631579, + "NONE": 63.41463414634146 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4507042253521127, + "macro_f1": 0.4185959893924496, + "weighted_f1": 0.3910084265312983, + "macro_precision": 0.46776748237902915, + "macro_recall": 0.5226495726495727, + "per_class_f1": { + "FAVOR": 0.5050505050505051, + "AGAINST": 0.16666666666666666, + "NONE": 0.584070796460177 + }, + "accuracy_pct": 45.07042253521127, + "macro_f1_pct": 41.85959893924496, + "weighted_f1_pct": 39.10084265312983, + "macro_precision_pct": 46.776748237902915, + "macro_recall_pct": 52.26495726495727, + "per_class_f1_pct": { + "FAVOR": 50.505050505050505, + "AGAINST": 16.666666666666664, + "NONE": 58.4070796460177 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.51, + "macro_f1": 0.5047295756515614, + "weighted_f1": 0.5023351864841227, + "macro_precision": 0.6703360427986951, + "macro_recall": 0.5549274019153537, + "per_class_f1": { + "FAVOR": 0.5291005291005291, + "AGAINST": 0.48936170212765956, + "NONE": 0.49572649572649574 + }, + "accuracy_pct": 51.0, + "macro_f1_pct": 50.47295756515614, + "weighted_f1_pct": 50.233518648412264, + "macro_precision_pct": 67.03360427986951, + "macro_recall_pct": 55.492740191535376, + "per_class_f1_pct": { + "FAVOR": 52.910052910052904, + "AGAINST": 48.93617021276596, + "NONE": 49.572649572649574 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.57, + "macro_f1": 0.5126730584646634, + "weighted_f1": 0.5671445130871292, + "macro_precision": 0.5974979321753515, + "macro_recall": 0.49874753078636574, + "per_class_f1": { + "FAVOR": 0.6872246696035242, + "AGAINST": 0.3191489361702128, + "NONE": 0.5316455696202531 + }, + "accuracy_pct": 56.99999999999999, + "macro_f1_pct": 51.267305846466336, + "weighted_f1_pct": 56.714451308712924, + "macro_precision_pct": 59.74979321753515, + "macro_recall_pct": 49.87475307863657, + "per_class_f1_pct": { + "FAVOR": 68.72246696035242, + "AGAINST": 31.914893617021278, + "NONE": 53.16455696202531 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6116721756081024, + "weighted_f1": 0.6806238745563532, + "macro_precision": 0.6069048055839471, + "macro_recall": 0.6447026486756622, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7602339181286549, + "NONE": 0.64 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 61.167217560810236, + "weighted_f1_pct": 68.06238745563532, + "macro_precision_pct": 60.69048055839471, + "macro_recall_pct": 64.47026486756621, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 76.0233918128655, + "NONE": 64.0 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5668057397583259, + "weighted_f1": 0.5807030631951231, + "macro_precision": 0.5653153153153153, + "macro_recall": 0.5958605664488017, + "per_class_f1": { + "FAVOR": 0.515625, + "AGAINST": 0.6206896551724138, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 56.680573975832594, + "weighted_f1_pct": 58.07030631951231, + "macro_precision_pct": 56.531531531531535, + "macro_recall_pct": 59.58605664488017, + "per_class_f1_pct": { + "FAVOR": 51.5625, + "AGAINST": 62.06896551724138, + "NONE": 56.41025641025641 + } + } + ], + "accuracy": 0.5734919286321155, + "macro_f1": 0.5743494783378914, + "weighted_f1": 0.5691140249013457, + "macro_precision": 0.607573986219562, + "macro_recall": 0.5881024803955239, + "per_class_f1": { + "FAVOR": 0.5887265135699373, + "AGAINST": 0.5260416666666666, + "NONE": 0.60828025477707 + }, + "accuracy_pct": 57.34919286321155, + "macro_f1_pct": 57.43494783378914, + "weighted_f1_pct": 56.91140249013456, + "macro_precision_pct": 60.757398621956206, + "macro_recall_pct": 58.81024803955239, + "per_class_f1_pct": { + "FAVOR": 58.872651356993735, + "AGAINST": 52.604166666666664, + "NONE": 60.828025477707 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.6987301418628128, + "losses": { + "current_ce": 0.5270507854945732, + "total": 0.5270507854945732, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6376623376623377, + "weighted_f1": 0.7787725548787496, + "macro_precision": 0.6530559559408816, + "macro_recall": 0.7845649799781581, + "per_class_f1": { + "FAVOR": 0.7636363636363637, + "AGAINST": 0.2857142857142857, + "NONE": 0.8636363636363636 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.76623376623377, + "weighted_f1_pct": 77.87725548787496, + "macro_precision_pct": 65.30559559408816, + "macro_recall_pct": 78.45649799781582, + "per_class_f1_pct": { + "FAVOR": 76.36363636363637, + "AGAINST": 28.57142857142857, + "NONE": 86.36363636363636 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5806451612903226, + "macro_f1": 0.5609080728901558, + "weighted_f1": 0.5985169762867267, + "macro_precision": 0.5688468798527101, + "macro_recall": 0.587598204264871, + "per_class_f1": { + "FAVOR": 0.40425531914893614, + "AGAINST": 0.6421052631578947, + "NONE": 0.6363636363636364 + }, + "accuracy_pct": 58.06451612903226, + "macro_f1_pct": 56.09080728901558, + "weighted_f1_pct": 59.851697628672675, + "macro_precision_pct": 56.88468798527101, + "macro_recall_pct": 58.759820426487096, + "per_class_f1_pct": { + "FAVOR": 40.42553191489361, + "AGAINST": 64.21052631578948, + "NONE": 63.63636363636363 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5492957746478874, + "macro_f1": 0.5311160022825466, + "weighted_f1": 0.5248240728302316, + "macro_precision": 0.6276196500105419, + "macro_recall": 0.5393162393162393, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.425531914893617, + "NONE": 0.6344827586206897 + }, + "accuracy_pct": 54.929577464788736, + "macro_f1_pct": 53.111600228254666, + "weighted_f1_pct": 52.48240728302316, + "macro_precision_pct": 62.761965001054186, + "macro_recall_pct": 53.93162393162393, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 42.5531914893617, + "NONE": 63.44827586206897 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.565, + "macro_f1": 0.5703685227910457, + "weighted_f1": 0.567193053563634, + "macro_precision": 0.6135231724597817, + "macro_recall": 0.5856039542786531, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.5189873417721519, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 57.036852279104565, + "weighted_f1_pct": 56.71930535636341, + "macro_precision_pct": 61.35231724597817, + "macro_recall_pct": 58.56039542786531, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 51.89873417721519, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.38, + "macro_f1": 0.3960975671501987, + "weighted_f1": 0.39631265664160403, + "macro_precision": 0.5449415848086945, + "macro_recall": 0.4331007719357234, + "per_class_f1": { + "FAVOR": 0.3815789473684211, + "AGAINST": 0.3391812865497076, + "NONE": 0.4675324675324675 + }, + "accuracy_pct": 38.0, + "macro_f1_pct": 39.60975671501987, + "weighted_f1_pct": 39.6312656641604, + "macro_precision_pct": 54.494158480869444, + "macro_recall_pct": 43.31007719357234, + "per_class_f1_pct": { + "FAVOR": 38.15789473684211, + "AGAINST": 33.91812865497076, + "NONE": 46.75324675324675 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.5684931506849316, + "macro_f1": 0.42856634621466067, + "weighted_f1": 0.5501617512257376, + "macro_precision": 0.4379808261332702, + "macro_recall": 0.5150824587706148, + "per_class_f1": { + "FAVOR": 0.0625, + "AGAINST": 0.6871165644171779, + "NONE": 0.5360824742268041 + }, + "accuracy_pct": 56.849315068493155, + "macro_f1_pct": 42.856634621466064, + "weighted_f1_pct": 55.01617512257376, + "macro_precision_pct": 43.79808261332702, + "macro_recall_pct": 51.508245877061476, + "per_class_f1_pct": { + "FAVOR": 6.25, + "AGAINST": 68.71165644171779, + "NONE": 53.608247422680414 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5437758245060477, + "weighted_f1": 0.5875620084694495, + "macro_precision": 0.564989927974146, + "macro_recall": 0.5428467683369643, + "per_class_f1": { + "FAVOR": 0.39080459770114945, + "AGAINST": 0.7111111111111111, + "NONE": 0.5294117647058824 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 54.37758245060477, + "weighted_f1_pct": 58.75620084694495, + "macro_precision_pct": 56.4989927974146, + "macro_recall_pct": 54.28467683369643, + "per_class_f1_pct": { + "FAVOR": 39.08045977011494, + "AGAINST": 71.11111111111111, + "NONE": 52.94117647058824 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6990940877304513, + "weighted_f1": 0.7169262555626192, + "macro_precision": 0.693687343358396, + "macro_recall": 0.7203056068230457, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.7403846153846154, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.90940877304513, + "weighted_f1_pct": 71.69262555626193, + "macro_precision_pct": 69.3687343358396, + "macro_recall_pct": 72.03056068230457, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 74.03846153846155, + "NONE": 71.07438016528926 + } + } + ], + "accuracy": 0.5807860262008734, + "macro_f1": 0.5748906997485045, + "weighted_f1": 0.5774160698203054, + "macro_precision": 0.5835905468800205, + "macro_recall": 0.5738261007117292, + "per_class_f1": { + "FAVOR": 0.5049928673323824, + "AGAINST": 0.5949878738884398, + "NONE": 0.6246913580246913 + }, + "accuracy_pct": 58.07860262008734, + "macro_f1_pct": 57.48906997485045, + "weighted_f1_pct": 57.74160698203053, + "macro_precision_pct": 58.359054688002054, + "macro_recall_pct": 57.38261007117293, + "per_class_f1_pct": { + "FAVOR": 50.499286733238236, + "AGAINST": 59.49878738884398, + "NONE": 62.46913580246913 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6569698670023203, + "losses": { + "current_ce": 0.5315607939920728, + "total": 0.5315607939920728, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6211129197564357, + "weighted_f1": 0.7687350846782132, + "macro_precision": 0.6168186423505572, + "macro_recall": 0.6774420580026695, + "per_class_f1": { + "FAVOR": 0.7719298245614035, + "AGAINST": 0.26666666666666666, + "NONE": 0.8247422680412371 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 62.11129197564357, + "weighted_f1_pct": 76.87350846782131, + "macro_precision_pct": 61.68186423505572, + "macro_recall_pct": 67.74420580026695, + "per_class_f1_pct": { + "FAVOR": 77.19298245614034, + "AGAINST": 26.666666666666668, + "NONE": 82.4742268041237 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.45698924731182794, + "macro_f1": 0.45792573571370343, + "weighted_f1": 0.4297189943120409, + "macro_precision": 0.5054012345679012, + "macro_recall": 0.5668350168350168, + "per_class_f1": { + "FAVOR": 0.45977011494252873, + "AGAINST": 0.375, + "NONE": 0.5390070921985816 + }, + "accuracy_pct": 45.69892473118279, + "macro_f1_pct": 45.79257357137034, + "weighted_f1_pct": 42.97189943120409, + "macro_precision_pct": 50.54012345679012, + "macro_recall_pct": 56.68350168350168, + "per_class_f1_pct": { + "FAVOR": 45.97701149425287, + "AGAINST": 37.5, + "NONE": 53.90070921985816 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4788732394366197, + "macro_f1": 0.41064008077422365, + "weighted_f1": 0.3818494623323045, + "macro_precision": 0.530667701863354, + "macro_recall": 0.4935897435897436, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.11940298507462686, + "NONE": 0.6219512195121951 + }, + "accuracy_pct": 47.88732394366197, + "macro_f1_pct": 41.064008077422365, + "weighted_f1_pct": 38.18494623323045, + "macro_precision_pct": 53.0667701863354, + "macro_recall_pct": 49.358974358974365, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 11.940298507462686, + "NONE": 62.19512195121951 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.645, + "macro_f1": 0.6242720990107954, + "weighted_f1": 0.6381286869325625, + "macro_precision": 0.66023802514824, + "macro_recall": 0.6403614457831325, + "per_class_f1": { + "FAVOR": 0.5954198473282443, + "AGAINST": 0.52, + "NONE": 0.757396449704142 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 62.427209901079536, + "weighted_f1_pct": 63.812868693256256, + "macro_precision_pct": 66.023802514824, + "macro_recall_pct": 64.03614457831324, + "per_class_f1_pct": { + "FAVOR": 59.541984732824424, + "AGAINST": 52.0, + "NONE": 75.7396449704142 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.515, + "macro_f1": 0.5271907366983767, + "weighted_f1": 0.5325695960330935, + "macro_precision": 0.5535485413677667, + "macro_recall": 0.5440647809579848, + "per_class_f1": { + "FAVOR": 0.5029239766081871, + "AGAINST": 0.3548387096774194, + "NONE": 0.7238095238095238 + }, + "accuracy_pct": 51.5, + "macro_f1_pct": 52.71907366983767, + "weighted_f1_pct": 53.25695960330935, + "macro_precision_pct": 55.354854136776666, + "macro_recall_pct": 54.406478095798484, + "per_class_f1_pct": { + "FAVOR": 50.29239766081871, + "AGAINST": 35.483870967741936, + "NONE": 72.38095238095238 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.2808219178082192, + "macro_f1": 0.19893562095396958, + "weighted_f1": 0.21355451134262246, + "macro_precision": 0.3126274509803922, + "macro_recall": 0.3768115942028985, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.22018348623853212, + "NONE": 0.37662337662337664 + }, + "accuracy_pct": 28.08219178082192, + "macro_f1_pct": 19.89356209539696, + "weighted_f1_pct": 21.355451134262246, + "macro_precision_pct": 31.26274509803922, + "macro_recall_pct": 37.68115942028985, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 22.018348623853214, + "NONE": 37.66233766233766 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5368421052631579, + "macro_f1": 0.5360357979669855, + "weighted_f1": 0.5370724257925613, + "macro_precision": 0.564452184382525, + "macro_recall": 0.6209150326797386, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.5408805031446541, + "NONE": 0.5378151260504201 + }, + "accuracy_pct": 53.68421052631579, + "macro_f1_pct": 53.603579796698554, + "weighted_f1_pct": 53.70724257925613, + "macro_precision_pct": 56.445218438252496, + "macro_recall_pct": 62.091503267973856, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 54.088050314465406, + "NONE": 53.78151260504202 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6716591638969526, + "weighted_f1": 0.663182198029981, + "macro_precision": 0.6847045505767615, + "macro_recall": 0.7653316086423843, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6432748538011696, + "NONE": 0.7050359712230215 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 67.16591638969525, + "weighted_f1_pct": 66.3182198029981, + "macro_precision_pct": 68.47045505767615, + "macro_recall_pct": 76.53316086423843, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 64.32748538011695, + "NONE": 70.50359712230215 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6867039964866052, + "weighted_f1": 0.6717737154150197, + "macro_precision": 0.6976790716286514, + "macro_recall": 0.7098456597101582, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.6931818181818182, + "NONE": 0.7608695652173914 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.67039964866052, + "weighted_f1_pct": 67.17737154150197, + "macro_precision_pct": 69.76790716286514, + "macro_recall_pct": 70.98456597101583, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 69.31818181818183, + "NONE": 76.08695652173914 + } + } + ], + "accuracy": 0.5603557814485387, + "macro_f1": 0.5576989286483144, + "weighted_f1": 0.5452881168721021, + "macro_precision": 0.5765499096834995, + "macro_recall": 0.5942393042474742, + "per_class_f1": { + "FAVOR": 0.5625692137320044, + "AGAINST": 0.4732394366197183, + "NONE": 0.6372881355932203 + }, + "accuracy_pct": 56.03557814485387, + "macro_f1_pct": 55.76989286483144, + "weighted_f1_pct": 54.52881168721021, + "macro_precision_pct": 57.65499096834995, + "macro_recall_pct": 59.42393042474742, + "per_class_f1_pct": { + "FAVOR": 56.25692137320044, + "AGAINST": 47.32394366197183, + "NONE": 63.72881355932203 + } + }, + { + "after_task": 10, + "best_epoch": 3, + "best_dev_macro_f1": 0.6723453012127972, + "losses": { + "current_ce": 0.379657160397619, + "total": 0.379657160397619, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7345132743362832, + "macro_f1": 0.5580246913580247, + "weighted_f1": 0.7490003277613897, + "macro_precision": 0.5584656084656084, + "macro_recall": 0.5692755733527485, + "per_class_f1": { + "FAVOR": 0.8, + "AGAINST": 0.13333333333333333, + "NONE": 0.7407407407407407 + }, + "accuracy_pct": 73.45132743362832, + "macro_f1_pct": 55.80246913580247, + "weighted_f1_pct": 74.90003277613897, + "macro_precision_pct": 55.84656084656084, + "macro_recall_pct": 56.92755733527485, + "per_class_f1_pct": { + "FAVOR": 80.0, + "AGAINST": 13.333333333333334, + "NONE": 74.07407407407408 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.5, + "macro_f1": 0.504625154306683, + "weighted_f1": 0.5034851513174915, + "macro_precision": 0.53758496010312, + "macro_recall": 0.5866442199775533, + "per_class_f1": { + "FAVOR": 0.4144144144144144, + "AGAINST": 0.4840764331210191, + "NONE": 0.6153846153846154 + }, + "accuracy_pct": 50.0, + "macro_f1_pct": 50.4625154306683, + "weighted_f1_pct": 50.34851513174915, + "macro_precision_pct": 53.758496010311994, + "macro_recall_pct": 58.66442199775533, + "per_class_f1_pct": { + "FAVOR": 41.44144144144144, + "AGAINST": 48.40764331210191, + "NONE": 61.53846153846154 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.4084507042253521, + "macro_f1": 0.3777150116128943, + "weighted_f1": 0.34798916133116076, + "macro_precision": 0.4477270864076078, + "macro_recall": 0.41923076923076924, + "per_class_f1": { + "FAVOR": 0.4782608695652174, + "AGAINST": 0.12658227848101267, + "NONE": 0.5283018867924528 + }, + "accuracy_pct": 40.845070422535215, + "macro_f1_pct": 37.77150116128943, + "weighted_f1_pct": 34.79891613311607, + "macro_precision_pct": 44.772708640760776, + "macro_recall_pct": 41.92307692307693, + "per_class_f1_pct": { + "FAVOR": 47.82608695652174, + "AGAINST": 12.658227848101266, + "NONE": 52.83018867924528 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.565, + "macro_f1": 0.5562754265883952, + "weighted_f1": 0.5635904187359041, + "macro_precision": 0.5710874093227034, + "macro_recall": 0.577386468952734, + "per_class_f1": { + "FAVOR": 0.583941605839416, + "AGAINST": 0.42735042735042733, + "NONE": 0.6575342465753424 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 55.62754265883952, + "weighted_f1_pct": 56.35904187359041, + "macro_precision_pct": 57.108740932270344, + "macro_recall_pct": 57.738646895273405, + "per_class_f1_pct": { + "FAVOR": 58.3941605839416, + "AGAINST": 42.73504273504273, + "NONE": 65.75342465753424 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.495, + "macro_f1": 0.45749502825526217, + "weighted_f1": 0.5055099533345147, + "macro_precision": 0.5232789777739834, + "macro_recall": 0.44925538323596576, + "per_class_f1": { + "FAVOR": 0.6124401913875598, + "AGAINST": 0.2909090909090909, + "NONE": 0.4691358024691358 + }, + "accuracy_pct": 49.5, + "macro_f1_pct": 45.749502825526214, + "weighted_f1_pct": 50.55099533345147, + "macro_precision_pct": 52.327897777398334, + "macro_recall_pct": 44.92553832359658, + "per_class_f1_pct": { + "FAVOR": 61.24401913875598, + "AGAINST": 29.09090909090909, + "NONE": 46.913580246913575 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.1917808219178082, + "macro_f1": 0.11902439024390243, + "weighted_f1": 0.08818576678917474, + "macro_precision": 0.14753086419753086, + "macro_recall": 0.3060969515242379, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.04, + "NONE": 0.3170731707317073 + }, + "accuracy_pct": 19.17808219178082, + "macro_f1_pct": 11.902439024390244, + "weighted_f1_pct": 8.818576678917474, + "macro_precision_pct": 14.753086419753087, + "macro_recall_pct": 30.60969515242379, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 4.0, + "NONE": 31.70731707317073 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.4421052631578947, + "macro_f1": 0.43999999999999995, + "weighted_f1": 0.44403007518796994, + "macro_precision": 0.5303599441590029, + "macro_recall": 0.5381263616557734, + "per_class_f1": { + "FAVOR": 0.42, + "AGAINST": 0.45714285714285713, + "NONE": 0.44285714285714284 + }, + "accuracy_pct": 44.21052631578947, + "macro_f1_pct": 43.99999999999999, + "weighted_f1_pct": 44.403007518797, + "macro_precision_pct": 53.03599441590029, + "macro_recall_pct": 53.812636165577345, + "per_class_f1_pct": { + "FAVOR": 42.0, + "AGAINST": 45.714285714285715, + "NONE": 44.285714285714285 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.6078548657137619, + "weighted_f1": 0.6098523791475676, + "macro_precision": 0.6324169530355097, + "macro_recall": 0.6726820521719011, + "per_class_f1": { + "FAVOR": 0.6027397260273972, + "AGAINST": 0.6127167630057804, + "NONE": 0.6081081081081081 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 60.78548657137619, + "weighted_f1_pct": 60.985237914756766, + "macro_precision_pct": 63.24169530355097, + "macro_recall_pct": 67.26820521719011, + "per_class_f1_pct": { + "FAVOR": 60.273972602739725, + "AGAINST": 61.27167630057804, + "NONE": 60.810810810810814 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6832518778626563, + "weighted_f1": 0.6794721296667404, + "macro_precision": 0.6938951310861423, + "macro_recall": 0.6772519861950758, + "per_class_f1": { + "FAVOR": 0.6538461538461539, + "AGAINST": 0.6946107784431138, + "NONE": 0.7012987012987013 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.32518778626563, + "weighted_f1_pct": 67.94721296667404, + "macro_precision_pct": 69.38951310861424, + "macro_recall_pct": 67.72519861950758, + "per_class_f1_pct": { + "FAVOR": 65.38461538461539, + "AGAINST": 69.46107784431138, + "NONE": 70.12987012987013 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7887368192376464, + "weighted_f1": 0.8183141825838472, + "macro_precision": 0.7712210086299892, + "macro_recall": 0.8141836489162827, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7704918032786885, + "NONE": 0.8623853211009175 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 78.87368192376464, + "weighted_f1_pct": 81.83141825838473, + "macro_precision_pct": 77.12210086299892, + "macro_recall_pct": 81.41836489162827, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 77.04918032786885, + "NONE": 86.23853211009175 + } + } + ], + "accuracy": 0.5507328072153326, + "macro_f1": 0.5500747457749592, + "weighted_f1": 0.5412769395277567, + "macro_precision": 0.5593191662898689, + "macro_recall": 0.5702260600415977, + "per_class_f1": { + "FAVOR": 0.5828571428571429, + "AGAINST": 0.4694915254237288, + "NONE": 0.5978755690440061 + }, + "accuracy_pct": 55.07328072153326, + "macro_f1_pct": 55.00747457749592, + "weighted_f1_pct": 54.127693952775665, + "macro_precision_pct": 55.93191662898689, + "macro_recall_pct": 57.02260600415977, + "per_class_f1_pct": { + "FAVOR": 58.285714285714285, + "AGAINST": 46.94915254237288, + "NONE": 59.78755690440061 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s2__seed13/resolved_config.yaml b/outputs/trace_s4_lora__x__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..38290ee21cff63e1ea6495fad6eb2e298b637637 --- /dev/null +++ b/outputs/trace_s4_lora__x__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s4_lora + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 4 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s4_lora__x__s2__seed13/run.log b/outputs/trace_s4_lora__x__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..77dd938655222156fff31604e6fb118b6c1f4f83 --- /dev/null +++ b/outputs/trace_s4_lora__x__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 01:33:54,493 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s4_lora__x__s2__seed13 device=cuda +2026-09-02 01:44:09,834 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.5093003265179227, 'average_forgetting': 0.1957903887336251, 'backward_transfer': -0.1957903887336251, 'final_mean_target_macro_f1_pct': 50.93003265179227, 'average_forgetting_pct': 19.57903887336251, 'backward_transfer_pct': -19.57903887336251, 'final_pooled_accuracy': 0.5507328072153326, 'final_pooled_accuracy_pct': 55.07328072153326, 'final_pooled_macro_f1': 0.5500747457749592, 'final_pooled_macro_f1_pct': 55.00747457749592, 'final_pooled_weighted_f1': 0.5412769395277567, 'final_pooled_weighted_f1_pct': 54.127693952775665, 'final_pooled_macro_precision': 0.5593191662898689, 'final_pooled_macro_precision_pct': 55.93191662898689, 'final_pooled_macro_recall': 0.5702260600415977, 'final_pooled_macro_recall_pct': 57.02260600415977, 'final_pooled_per_class_f1': {'FAVOR': 0.5828571428571429, 'AGAINST': 0.4694915254237288, 'NONE': 0.5978755690440061}, 'final_pooled_per_class_f1_pct': {'FAVOR': 58.285714285714285, 'AGAINST': 46.94915254237288, 'NONE': 59.78755690440061}} +2026-09-02 01:44:09,834 [INFO] __main__: efficiency={'training_wall_time_seconds': 613.9015596951358, 'training_wall_time_minutes': 10.231692661585596, 'peak_gpu_memory_allocated_gb': 5.56232213973999, 'peak_gpu_memory_reserved_gb': 6.439453125} diff --git a/outputs/trace_s4_lora__x__s3__seed13/environment.json b/outputs/trace_s4_lora__x__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..04f7630a0d49a73cc3e91d41cbf8e491f09f2dac --- /dev/null +++ b/outputs/trace_s4_lora__x__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T01:44:17.815843+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s3__seed13/metrics.json b/outputs/trace_s4_lora__x__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..49bffeeb32cd9df79fd9892b17de892c58084f3d --- /dev/null +++ b/outputs/trace_s4_lora__x__s3__seed13/metrics.json @@ -0,0 +1,1887 @@ +{ + "experiment": "trace_s4_lora", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 4, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": false, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5708112646147389, + 0.5998266312387883, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.3744924767136375, + 0.5627755236009485, + 0.7013404231749196, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.18042791376746414, + 0.4436151051990784, + 0.4156903254351832, + 0.6032312557296334, + null, + null, + null, + null, + null, + null + ], + [ + 0.13034188034188035, + 0.2525894942561609, + 0.417463869693169, + 0.6264037385266303, + 0.8025875899187497, + null, + null, + null, + null, + null + ], + [ + 0.15785955807187227, + 0.3789677419354838, + 0.4862612132133222, + 0.6382034632034632, + 0.7360729227349351, + 0.7374031007751939, + null, + null, + null, + null + ], + [ + 0.41347926267281104, + 0.3326676447089291, + 0.32048549332545195, + 0.5780035940994712, + 0.5826284918011537, + 0.6397100464378752, + 0.6677773101384212, + null, + null, + null + ], + [ + 0.29483726150392814, + 0.4286151306896601, + 0.5461938637123898, + 0.591231114002747, + 0.6968202599270561, + 0.49066065182570034, + 0.6131044510107406, + 0.6021575885774723, + null, + null + ], + [ + 0.17617797317375994, + 0.41787906257036783, + 0.4497613277520746, + 0.608885523134481, + 0.6127814088598402, + 0.3804266072092159, + 0.5228988168094655, + 0.5849042897097817, + 0.7094223046030276, + null + ], + [ + 0.13820794590025356, + 0.3419903257931427, + 0.3781268706306187, + 0.643914676701562, + 0.5912012393947482, + 0.41830089437383594, + 0.542305087810031, + 0.4287938854959106, + 0.6059476461616938, + 0.6721045109617051 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 14, + "best_dev_macro_f1": 0.61515693996897, + "losses": { + "current_ce": 0.6929058672803821, + "total": 0.6929058672803821, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.5708112646147389, + "weighted_f1": 0.6792593028559658, + "macro_precision": 0.5797966401414677, + "macro_recall": 0.5754822588705647, + "per_class_f1": { + "FAVOR": 0.15789473684210525, + "AGAINST": 0.7959183673469388, + "NONE": 0.7586206896551724 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 57.081126461473886, + "weighted_f1_pct": 67.92593028559658, + "macro_precision_pct": 57.97966401414677, + "macro_recall_pct": 57.54822588705647, + "per_class_f1_pct": { + "FAVOR": 15.789473684210526, + "AGAINST": 79.59183673469387, + "NONE": 75.86206896551724 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5998266312387883, + "weighted_f1": 0.6187294631551585, + "macro_precision": 0.6009237085186453, + "macro_recall": 0.6256354393609296, + "per_class_f1": { + "FAVOR": 0.6165413533834586, + "AGAINST": 0.6477272727272727, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.98266312387883, + "weighted_f1_pct": 61.872946315515854, + "macro_precision_pct": 60.09237085186453, + "macro_recall_pct": 62.563543936092955, + "per_class_f1_pct": { + "FAVOR": 61.65413533834586, + "AGAINST": 64.77272727272727, + "NONE": 53.52112676056338 + } + } + ], + "accuracy": 0.6547619047619048, + "macro_f1": 0.6253617497941737, + "weighted_f1": 0.6592498035745361, + "macro_precision": 0.6192999856898244, + "macro_recall": 0.6345439882566334, + "per_class_f1": { + "FAVOR": 0.5146198830409356, + "AGAINST": 0.7258064516129032, + "NONE": 0.6356589147286822 + }, + "accuracy_pct": 65.47619047619048, + "macro_f1_pct": 62.53617497941737, + "weighted_f1_pct": 65.9249803574536, + "macro_precision_pct": 61.929998568982434, + "macro_recall_pct": 63.45439882566334, + "per_class_f1_pct": { + "FAVOR": 51.461988304093566, + "AGAINST": 72.58064516129032, + "NONE": 63.565891472868216 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7310140850150589, + "losses": { + "current_ce": 0.5811113305164106, + "total": 0.5811113305164106, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.5205479452054794, + "macro_f1": 0.3744924767136375, + "weighted_f1": 0.4962620766822291, + "macro_precision": 0.3650793650793651, + "macro_recall": 0.48000999500249875, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.6329113924050633, + "NONE": 0.49056603773584906 + }, + "accuracy_pct": 52.054794520547944, + "macro_f1_pct": 37.44924767136375, + "weighted_f1_pct": 49.62620766822291, + "macro_precision_pct": 36.50793650793651, + "macro_recall_pct": 48.000999500249875, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 63.29113924050633, + "NONE": 49.056603773584904 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5627755236009485, + "weighted_f1": 0.5817740495748138, + "macro_precision": 0.5584795321637427, + "macro_recall": 0.6042120551924474, + "per_class_f1": { + "FAVOR": 0.41304347826086957, + "AGAINST": 0.6598984771573604, + "NONE": 0.6153846153846154 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 56.27755236009485, + "weighted_f1_pct": 58.177404957481386, + "macro_precision_pct": 55.84795321637427, + "macro_recall_pct": 60.42120551924474, + "per_class_f1_pct": { + "FAVOR": 41.30434782608695, + "AGAINST": 65.98984771573603, + "NONE": 61.53846153846154 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.7013404231749196, + "weighted_f1": 0.7121090198102816, + "macro_precision": 0.6851457000710731, + "macro_recall": 0.7352231318915337, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7254901960784313, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 70.13404231749196, + "weighted_f1_pct": 71.21090198102816, + "macro_precision_pct": 68.51457000710731, + "macro_recall_pct": 73.52231318915337, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 72.54901960784314, + "NONE": 71.1864406779661 + } + } + ], + "accuracy": 0.6153846153846154, + "macro_f1": 0.5778826632024306, + "weighted_f1": 0.6139738978824454, + "macro_precision": 0.5883706467661692, + "macro_recall": 0.6138896712465202, + "per_class_f1": { + "FAVOR": 0.4479166666666667, + "AGAINST": 0.6762075134168157, + "NONE": 0.6095238095238096 + }, + "accuracy_pct": 61.53846153846154, + "macro_f1_pct": 57.78826632024307, + "weighted_f1_pct": 61.39738978824454, + "macro_precision_pct": 58.837064676616926, + "macro_recall_pct": 61.38896712465202, + "per_class_f1_pct": { + "FAVOR": 44.79166666666667, + "AGAINST": 67.62075134168157, + "NONE": 60.952380952380956 + } + }, + { + "after_task": 4, + "best_epoch": 9, + "best_dev_macro_f1": 0.6345427141476642, + "losses": { + "current_ce": 0.6060382296168615, + "total": 0.6060382296168615, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2191780821917808, + "macro_f1": 0.18042791376746414, + "weighted_f1": 0.13063285331390032, + "macro_precision": 0.31984126984126987, + "macro_recall": 0.349720139930035, + "per_class_f1": { + "FAVOR": 0.13043478260869565, + "AGAINST": 0.061855670103092786, + "NONE": 0.348993288590604 + }, + "accuracy_pct": 21.91780821917808, + "macro_f1_pct": 18.042791376746415, + "weighted_f1_pct": 13.063285331390032, + "macro_precision_pct": 31.984126984126988, + "macro_recall_pct": 34.9720139930035, + "per_class_f1_pct": { + "FAVOR": 13.043478260869565, + "AGAINST": 6.185567010309279, + "NONE": 34.899328859060404 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.45263157894736844, + "macro_f1": 0.4436151051990784, + "weighted_f1": 0.3824697118060538, + "macro_precision": 0.6047649989170457, + "macro_recall": 0.6027596223674655, + "per_class_f1": { + "FAVOR": 0.5503355704697986, + "AGAINST": 0.2413793103448276, + "NONE": 0.5391304347826087 + }, + "accuracy_pct": 45.26315789473684, + "macro_f1_pct": 44.36151051990784, + "weighted_f1_pct": 38.246971180605385, + "macro_precision_pct": 60.47649989170457, + "macro_recall_pct": 60.27596223674655, + "per_class_f1_pct": { + "FAVOR": 55.033557046979865, + "AGAINST": 24.137931034482758, + "NONE": 53.91304347826087 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.41624365482233505, + "macro_f1": 0.4156903254351832, + "weighted_f1": 0.3338268279071345, + "macro_precision": 0.5817827396774765, + "macro_recall": 0.6141687568028016, + "per_class_f1": { + "FAVOR": 0.43537414965986393, + "AGAINST": 0.16129032258064516, + "NONE": 0.6504065040650406 + }, + "accuracy_pct": 41.62436548223351, + "macro_f1_pct": 41.56903254351832, + "weighted_f1_pct": 33.38268279071345, + "macro_precision_pct": 58.17827396774765, + "macro_recall_pct": 61.41687568028016, + "per_class_f1_pct": { + "FAVOR": 43.53741496598639, + "AGAINST": 16.129032258064516, + "NONE": 65.04065040650406 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.6, + "macro_f1": 0.6032312557296334, + "weighted_f1": 0.5787851889620221, + "macro_precision": 0.6433130805223829, + "macro_recall": 0.6222446127595179, + "per_class_f1": { + "FAVOR": 0.6368159203980099, + "AGAINST": 0.43478260869565216, + "NONE": 0.7380952380952381 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.32312557296334, + "weighted_f1_pct": 57.878518896202216, + "macro_precision_pct": 64.33130805223828, + "macro_recall_pct": 62.22446127595179, + "per_class_f1_pct": { + "FAVOR": 63.681592039801, + "AGAINST": 43.47826086956522, + "NONE": 73.80952380952381 + } + } + ], + "accuracy": 0.43656207366984995, + "macro_f1": 0.429755562344462, + "weighted_f1": 0.37155734506783417, + "macro_precision": 0.5270608980414383, + "macro_recall": 0.5620866625713692, + "per_class_f1": { + "FAVOR": 0.5156537753222836, + "AGAINST": 0.23008849557522124, + "NONE": 0.5435244161358811 + }, + "accuracy_pct": 43.656207366984994, + "macro_f1_pct": 42.9755562344462, + "weighted_f1_pct": 37.15573450678342, + "macro_precision_pct": 52.70608980414383, + "macro_recall_pct": 56.20866625713692, + "per_class_f1_pct": { + "FAVOR": 51.56537753222836, + "AGAINST": 23.008849557522122, + "NONE": 54.35244161358811 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.6736155223960102, + "losses": { + "current_ce": 0.4023706999296943, + "total": 0.4023706999296943, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2054794520547945, + "macro_f1": 0.13034188034188035, + "weighted_f1": 0.10256410256410256, + "macro_precision": 0.15100250626566417, + "macro_recall": 0.3212143928035982, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.057692307692307696, + "NONE": 0.3333333333333333 + }, + "accuracy_pct": 20.54794520547945, + "macro_f1_pct": 13.034188034188036, + "weighted_f1_pct": 10.256410256410255, + "macro_precision_pct": 15.100250626566417, + "macro_recall_pct": 32.12143928035982, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 5.769230769230769, + "NONE": 33.33333333333333 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.26842105263157895, + "macro_f1": 0.2525894942561609, + "weighted_f1": 0.21613647482068535, + "macro_precision": 0.5061391416230127, + "macro_recall": 0.40777051561365285, + "per_class_f1": { + "FAVOR": 0.275, + "AGAINST": 0.14414414414414414, + "NONE": 0.3386243386243386 + }, + "accuracy_pct": 26.842105263157894, + "macro_f1_pct": 25.25894942561609, + "weighted_f1_pct": 21.613647482068536, + "macro_precision_pct": 50.61391416230126, + "macro_recall_pct": 40.77705156136528, + "per_class_f1_pct": { + "FAVOR": 27.500000000000004, + "AGAINST": 14.414414414414415, + "NONE": 33.86243386243386 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.43147208121827413, + "macro_f1": 0.417463869693169, + "weighted_f1": 0.3430121206318303, + "macro_precision": 0.5654733107563296, + "macro_recall": 0.6016489728098265, + "per_class_f1": { + "FAVOR": 0.45045045045045046, + "AGAINST": 0.19047619047619047, + "NONE": 0.6114649681528662 + }, + "accuracy_pct": 43.14720812182741, + "macro_f1_pct": 41.7463869693169, + "weighted_f1_pct": 34.30121206318303, + "macro_precision_pct": 56.547331075632954, + "macro_recall_pct": 60.16489728098266, + "per_class_f1_pct": { + "FAVOR": 45.04504504504504, + "AGAINST": 19.047619047619047, + "NONE": 61.146496815286625 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.615, + "macro_f1": 0.6264037385266303, + "weighted_f1": 0.6145566686856575, + "macro_precision": 0.6497461547704879, + "macro_recall": 0.6152495000733483, + "per_class_f1": { + "FAVOR": 0.5584415584415584, + "AGAINST": 0.6358381502890174, + "NONE": 0.684931506849315 + }, + "accuracy_pct": 61.5, + "macro_f1_pct": 62.64037385266303, + "weighted_f1_pct": 61.45566686856575, + "macro_precision_pct": 64.97461547704879, + "macro_recall_pct": 61.52495000733483, + "per_class_f1_pct": { + "FAVOR": 55.84415584415584, + "AGAINST": 63.58381502890174, + "NONE": 68.4931506849315 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8025875899187497, + "weighted_f1": 0.8367969259431766, + "macro_precision": 0.7903796653796654, + "macro_recall": 0.8200807003905455, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7894736842105263, + "NONE": 0.8849557522123894 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 80.25875899187497, + "weighted_f1_pct": 83.67969259431766, + "macro_precision_pct": 79.03796653796654, + "macro_recall_pct": 82.00807003905454, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 78.94736842105263, + "NONE": 88.49557522123894 + } + } + ], + "accuracy": 0.4887459807073955, + "macro_f1": 0.4784554201627138, + "weighted_f1": 0.4628966040143497, + "macro_precision": 0.5251892152683807, + "macro_recall": 0.5320008461844429, + "per_class_f1": { + "FAVOR": 0.46867749419953597, + "AGAINST": 0.39171974522292996, + "NONE": 0.5749690210656754 + }, + "accuracy_pct": 48.87459807073955, + "macro_f1_pct": 47.84554201627138, + "weighted_f1_pct": 46.289660401434965, + "macro_precision_pct": 52.51892152683807, + "macro_recall_pct": 53.200084618444286, + "per_class_f1_pct": { + "FAVOR": 46.8677494199536, + "AGAINST": 39.171974522293, + "NONE": 57.49690210656754 + } + }, + { + "after_task": 6, + "best_epoch": 6, + "best_dev_macro_f1": 0.7391617322651806, + "losses": { + "current_ce": 0.41747793562355495, + "total": 0.41747793562355495, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2328767123287671, + "macro_f1": 0.15785955807187227, + "weighted_f1": 0.1500030699925997, + "macro_precision": 0.21614583333333334, + "macro_recall": 0.3357071464267866, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.12962962962962962, + "NONE": 0.34394904458598724 + }, + "accuracy_pct": 23.28767123287671, + "macro_f1_pct": 15.785955807187227, + "weighted_f1_pct": 15.000306999259971, + "macro_precision_pct": 21.614583333333336, + "macro_recall_pct": 33.57071464267866, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 12.962962962962962, + "NONE": 34.394904458598724 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.37894736842105264, + "macro_f1": 0.3789677419354838, + "weighted_f1": 0.3564563667232598, + "macro_precision": 0.5156905018565098, + "macro_recall": 0.5054466230936819, + "per_class_f1": { + "FAVOR": 0.42, + "AGAINST": 0.304, + "NONE": 0.4129032258064516 + }, + "accuracy_pct": 37.89473684210527, + "macro_f1_pct": 37.89677419354838, + "weighted_f1_pct": 35.64563667232598, + "macro_precision_pct": 51.56905018565098, + "macro_recall_pct": 50.54466230936819, + "per_class_f1_pct": { + "FAVOR": 42.0, + "AGAINST": 30.4, + "NONE": 41.29032258064516 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.4873096446700508, + "macro_f1": 0.4862612132133222, + "weighted_f1": 0.4440674517042101, + "macro_precision": 0.5640479070413039, + "macro_recall": 0.6304902223694518, + "per_class_f1": { + "FAVOR": 0.48598130841121495, + "AGAINST": 0.352112676056338, + "NONE": 0.6206896551724138 + }, + "accuracy_pct": 48.73096446700508, + "macro_f1_pct": 48.62612132133222, + "weighted_f1_pct": 44.40674517042101, + "macro_precision_pct": 56.40479070413039, + "macro_recall_pct": 63.049022236945184, + "per_class_f1_pct": { + "FAVOR": 48.598130841121495, + "AGAINST": 35.2112676056338, + "NONE": 62.06896551724138 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6382034632034632, + "weighted_f1": 0.6282175324675324, + "macro_precision": 0.672889373037831, + "macro_recall": 0.6242134358665524, + "per_class_f1": { + "FAVOR": 0.5584415584415584, + "AGAINST": 0.6704545454545454, + "NONE": 0.6857142857142857 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 63.820346320346324, + "weighted_f1_pct": 62.82175324675324, + "macro_precision_pct": 67.2889373037831, + "macro_recall_pct": 62.42134358665524, + "per_class_f1_pct": { + "FAVOR": 55.84415584415584, + "AGAINST": 67.04545454545455, + "NONE": 68.57142857142857 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7360729227349351, + "weighted_f1": 0.7894011483336015, + "macro_precision": 0.7289582499666533, + "macro_recall": 0.7871508690099395, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7102803738317757, + "NONE": 0.8663594470046083 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 73.60729227349351, + "weighted_f1_pct": 78.94011483336016, + "macro_precision_pct": 72.89582499666533, + "macro_recall_pct": 78.71508690099395, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 71.02803738317756, + "NONE": 86.63594470046083 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7374031007751939, + "weighted_f1": 0.8446954105783084, + "macro_precision": 0.7155347384855583, + "macro_recall": 0.7712898920033976, + "per_class_f1": { + "FAVOR": 0.875, + "AGAINST": 0.5, + "NONE": 0.8372093023255814 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.74031007751938, + "weighted_f1_pct": 84.46954105783084, + "macro_precision_pct": 71.55347384855583, + "macro_recall_pct": 77.12898920033976, + "per_class_f1_pct": { + "FAVOR": 87.5, + "AGAINST": 50.0, + "NONE": 83.72093023255815 + } + } + ], + "accuracy": 0.5535372848948374, + "macro_f1": 0.5489191133229975, + "weighted_f1": 0.5354779997634255, + "macro_precision": 0.5782223519833679, + "macro_recall": 0.5863233532570741, + "per_class_f1": { + "FAVOR": 0.5743243243243243, + "AGAINST": 0.4507462686567164, + "NONE": 0.6216867469879518 + }, + "accuracy_pct": 55.35372848948374, + "macro_f1_pct": 54.89191133229975, + "weighted_f1_pct": 53.547799976342546, + "macro_precision_pct": 57.82223519833679, + "macro_recall_pct": 58.63233532570741, + "per_class_f1_pct": { + "FAVOR": 57.432432432432435, + "AGAINST": 45.07462686567164, + "NONE": 62.16867469879518 + } + }, + { + "after_task": 7, + "best_epoch": 15, + "best_dev_macro_f1": 0.6251282051282051, + "losses": { + "current_ce": 0.598251260952516, + "total": 0.598251260952516, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.5821917808219178, + "macro_f1": 0.41347926267281104, + "weighted_f1": 0.5581868253266838, + "macro_precision": 0.4127621597501115, + "macro_recall": 0.5362318840579711, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.7225806451612903, + "NONE": 0.5178571428571429 + }, + "accuracy_pct": 58.21917808219178, + "macro_f1_pct": 41.3479262672811, + "weighted_f1_pct": 55.81868253266838, + "macro_precision_pct": 41.27621597501115, + "macro_recall_pct": 53.62318840579711, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 72.25806451612902, + "NONE": 51.78571428571429 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.34210526315789475, + "macro_f1": 0.3326676447089291, + "weighted_f1": 0.26298148824826756, + "macro_precision": 0.43612313612313613, + "macro_recall": 0.5065359477124183, + "per_class_f1": { + "FAVOR": 0.4954128440366973, + "AGAINST": 0.09009009009009009, + "NONE": 0.4125 + }, + "accuracy_pct": 34.21052631578947, + "macro_f1_pct": 33.26676447089291, + "weighted_f1_pct": 26.298148824826757, + "macro_precision_pct": 43.61231361231361, + "macro_recall_pct": 50.65359477124183, + "per_class_f1_pct": { + "FAVOR": 49.54128440366973, + "AGAINST": 9.00900900900901, + "NONE": 41.25 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.34517766497461927, + "macro_f1": 0.32048549332545195, + "weighted_f1": 0.25645948875514873, + "macro_precision": 0.4458295459464139, + "macro_recall": 0.4585364314671967, + "per_class_f1": { + "FAVOR": 0.3582089552238806, + "AGAINST": 0.128, + "NONE": 0.4752475247524752 + }, + "accuracy_pct": 34.51776649746193, + "macro_f1_pct": 32.04854933254519, + "weighted_f1_pct": 25.645948875514872, + "macro_precision_pct": 44.582954594641386, + "macro_recall_pct": 45.85364314671967, + "per_class_f1_pct": { + "FAVOR": 35.82089552238806, + "AGAINST": 12.8, + "NONE": 47.524752475247524 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.58, + "macro_f1": 0.5780035940994712, + "weighted_f1": 0.5570422371329946, + "macro_precision": 0.5926365653294957, + "macro_recall": 0.6096402844370324, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.4067796610169492, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 57.99999999999999, + "macro_f1_pct": 57.80035940994712, + "weighted_f1_pct": 55.70422371329946, + "macro_precision_pct": 59.26365653294957, + "macro_recall_pct": 60.96402844370324, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 40.67796610169492, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.69, + "macro_f1": 0.5826284918011537, + "weighted_f1": 0.6376017159245576, + "macro_precision": 0.715089103125913, + "macro_recall": 0.5758898328613471, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.32432432432432434, + "NONE": 0.7985611510791367 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 58.262849180115374, + "weighted_f1_pct": 63.76017159245576, + "macro_precision_pct": 71.5089103125913, + "macro_recall_pct": 57.588983286134706, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 32.432432432432435, + "NONE": 79.85611510791367 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7256637168141593, + "macro_f1": 0.6397100464378752, + "weighted_f1": 0.7251645572665302, + "macro_precision": 0.687948350071736, + "macro_recall": 0.6575415604902317, + "per_class_f1": { + "FAVOR": 0.7407407407407407, + "AGAINST": 0.4444444444444444, + "NONE": 0.7339449541284404 + }, + "accuracy_pct": 72.56637168141593, + "macro_f1_pct": 63.97100464378752, + "weighted_f1_pct": 72.51645572665302, + "macro_precision_pct": 68.7948350071736, + "macro_recall_pct": 65.75415604902317, + "per_class_f1_pct": { + "FAVOR": 74.07407407407408, + "AGAINST": 44.44444444444444, + "NONE": 73.39449541284404 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6677773101384212, + "weighted_f1": 0.6953563411896745, + "macro_precision": 0.6528880070546738, + "macro_recall": 0.7086980920314253, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7395833333333334, + "NONE": 0.6464646464646465 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.77773101384213, + "weighted_f1_pct": 69.53563411896745, + "macro_precision_pct": 65.28880070546738, + "macro_recall_pct": 70.86980920314254, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 73.95833333333334, + "NONE": 64.64646464646465 + } + } + ], + "accuracy": 0.5535714285714286, + "macro_f1": 0.5473386831180124, + "weighted_f1": 0.5315355090154374, + "macro_precision": 0.6096082028477673, + "macro_recall": 0.5974800777355643, + "per_class_f1": { + "FAVOR": 0.5700636942675159, + "AGAINST": 0.45408163265306123, + "NONE": 0.6178707224334601 + }, + "accuracy_pct": 55.35714285714286, + "macro_f1_pct": 54.73386831180124, + "weighted_f1_pct": 53.153550901543746, + "macro_precision_pct": 60.96082028477673, + "macro_recall_pct": 59.74800777355643, + "per_class_f1_pct": { + "FAVOR": 57.00636942675159, + "AGAINST": 45.40816326530612, + "NONE": 61.78707224334601 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.7521164021164021, + "losses": { + "current_ce": 0.6733893568317095, + "total": 0.6733893568317095, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.4041095890410959, + "macro_f1": 0.29483726150392814, + "weighted_f1": 0.37836815645034827, + "macro_precision": 0.34638364779874214, + "macro_recall": 0.4341579210394803, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.4696969696969697, + "NONE": 0.4148148148148148 + }, + "accuracy_pct": 40.41095890410959, + "macro_f1_pct": 29.483726150392815, + "weighted_f1_pct": 37.836815645034825, + "macro_precision_pct": 34.638364779874216, + "macro_recall_pct": 43.41579210394803, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 46.96969696969697, + "NONE": 41.48148148148148 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.4631578947368421, + "macro_f1": 0.4286151306896601, + "weighted_f1": 0.447704066158481, + "macro_precision": 0.5093726060378657, + "macro_recall": 0.532316630355846, + "per_class_f1": { + "FAVOR": 0.2777777777777778, + "AGAINST": 0.5263157894736842, + "NONE": 0.48175182481751827 + }, + "accuracy_pct": 46.31578947368421, + "macro_f1_pct": 42.86151306896601, + "weighted_f1_pct": 44.7704066158481, + "macro_precision_pct": 50.93726060378657, + "macro_recall_pct": 53.231663035584596, + "per_class_f1_pct": { + "FAVOR": 27.77777777777778, + "AGAINST": 52.63157894736842, + "NONE": 48.175182481751825 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5482233502538071, + "macro_f1": 0.5461938637123898, + "weighted_f1": 0.5313678271164103, + "macro_precision": 0.5690958164642376, + "macro_recall": 0.6408699172884492, + "per_class_f1": { + "FAVOR": 0.5176470588235295, + "AGAINST": 0.49079754601226994, + "NONE": 0.6301369863013698 + }, + "accuracy_pct": 54.82233502538071, + "macro_f1_pct": 54.619386371238974, + "weighted_f1_pct": 53.136782711641025, + "macro_precision_pct": 56.90958164642376, + "macro_recall_pct": 64.08699172884492, + "per_class_f1_pct": { + "FAVOR": 51.76470588235295, + "AGAINST": 49.079754601226995, + "NONE": 63.013698630136986 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.591231114002747, + "weighted_f1": 0.5753898286705703, + "macro_precision": 0.6843616021276109, + "macro_recall": 0.5968313529289139, + "per_class_f1": { + "FAVOR": 0.46551724137931033, + "AGAINST": 0.6415094339622641, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 59.123111400274695, + "weighted_f1_pct": 57.53898286705703, + "macro_precision_pct": 68.43616021276108, + "macro_recall_pct": 59.683135292891386, + "per_class_f1_pct": { + "FAVOR": 46.55172413793103, + "AGAINST": 64.15094339622641, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.6968202599270561, + "weighted_f1": 0.7426621975651102, + "macro_precision": 0.7579075425790754, + "macro_recall": 0.6648453551002277, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5825242718446602, + "NONE": 0.8412698412698413 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 69.6820259927056, + "weighted_f1_pct": 74.26621975651102, + "macro_precision_pct": 75.79075425790755, + "macro_recall_pct": 66.48453551002277, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 58.252427184466015, + "NONE": 84.12698412698413 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5663716814159292, + "macro_f1": 0.49066065182570034, + "weighted_f1": 0.574450233063516, + "macro_precision": 0.5798771121351766, + "macro_recall": 0.6819075354932654, + "per_class_f1": { + "FAVOR": 0.5111111111111111, + "AGAINST": 0.24242424242424243, + "NONE": 0.7184466019417476 + }, + "accuracy_pct": 56.63716814159292, + "macro_f1_pct": 49.066065182570036, + "weighted_f1_pct": 57.445023306351594, + "macro_precision_pct": 57.987711213517656, + "macro_recall_pct": 68.19075354932654, + "per_class_f1_pct": { + "FAVOR": 51.11111111111111, + "AGAINST": 24.242424242424242, + "NONE": 71.84466019417476 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.6131044510107406, + "weighted_f1": 0.6668663788994947, + "macro_precision": 0.6926914355485784, + "macro_recall": 0.6318742985409652, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7378640776699029, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 61.310445101074066, + "weighted_f1_pct": 66.68663788994948, + "macro_precision_pct": 69.26914355485783, + "macro_recall_pct": 63.18742985409652, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 73.7864077669903, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6021575885774723, + "weighted_f1": 0.5863594044435837, + "macro_precision": 0.6011904761904762, + "macro_recall": 0.6264957264957265, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.44, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.215758857747225, + "weighted_f1_pct": 58.63594044435837, + "macro_precision_pct": 60.11904761904761, + "macro_recall_pct": 62.64957264957265, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 44.0, + "NONE": 72.1311475409836 + } + } + ], + "accuracy": 0.5829694323144105, + "macro_f1": 0.5641600969180139, + "weighted_f1": 0.5681749152799429, + "macro_precision": 0.6063838476263586, + "macro_recall": 0.5844317120787709, + "per_class_f1": { + "FAVOR": 0.46950092421441775, + "AGAINST": 0.5642857142857143, + "NONE": 0.6586936522539099 + }, + "accuracy_pct": 58.29694323144105, + "macro_f1_pct": 56.41600969180139, + "weighted_f1_pct": 56.81749152799429, + "macro_precision_pct": 60.63838476263585, + "macro_recall_pct": 58.443171207877086, + "per_class_f1_pct": { + "FAVOR": 46.950092421441774, + "AGAINST": 56.42857142857143, + "NONE": 65.86936522539098 + } + }, + { + "after_task": 9, + "best_epoch": 4, + "best_dev_macro_f1": 0.7328160211949003, + "losses": { + "current_ce": 0.5340694459020228, + "total": 0.5340694459020228, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.23972602739726026, + "macro_f1": 0.17617797317375994, + "weighted_f1": 0.14505014968369437, + "macro_precision": 0.2925407925407925, + "macro_recall": 0.3647826086956522, + "per_class_f1": { + "FAVOR": 0.06666666666666667, + "AGAINST": 0.0970873786407767, + "NONE": 0.36477987421383645 + }, + "accuracy_pct": 23.972602739726025, + "macro_f1_pct": 17.617797317375995, + "weighted_f1_pct": 14.505014968369437, + "macro_precision_pct": 29.25407925407925, + "macro_recall_pct": 36.47826086956522, + "per_class_f1_pct": { + "FAVOR": 6.666666666666667, + "AGAINST": 9.70873786407767, + "NONE": 36.477987421383645 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.43157894736842106, + "macro_f1": 0.41787906257036783, + "weighted_f1": 0.4272308958797451, + "macro_precision": 0.5484258166076348, + "macro_recall": 0.5279593318809005, + "per_class_f1": { + "FAVOR": 0.35135135135135137, + "AGAINST": 0.46357615894039733, + "NONE": 0.43870967741935485 + }, + "accuracy_pct": 43.15789473684211, + "macro_f1_pct": 41.78790625703678, + "weighted_f1_pct": 42.72308958797451, + "macro_precision_pct": 54.84258166076348, + "macro_recall_pct": 52.795933188090046, + "per_class_f1_pct": { + "FAVOR": 35.13513513513514, + "AGAINST": 46.35761589403973, + "NONE": 43.87096774193549 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.45685279187817257, + "macro_f1": 0.4497613277520746, + "weighted_f1": 0.3880606095172912, + "macro_precision": 0.5437356760886173, + "macro_recall": 0.6021169529758806, + "per_class_f1": { + "FAVOR": 0.4943820224719101, + "AGAINST": 0.26666666666666666, + "NONE": 0.5882352941176471 + }, + "accuracy_pct": 45.68527918781726, + "macro_f1_pct": 44.97613277520746, + "weighted_f1_pct": 38.80606095172912, + "macro_precision_pct": 54.37356760886173, + "macro_recall_pct": 60.211695297588065, + "per_class_f1_pct": { + "FAVOR": 49.43820224719101, + "AGAINST": 26.666666666666668, + "NONE": 58.82352941176471 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.605, + "macro_f1": 0.608885523134481, + "weighted_f1": 0.5934994209829213, + "macro_precision": 0.6468870827744022, + "macro_recall": 0.6161567028775701, + "per_class_f1": { + "FAVOR": 0.5039370078740157, + "AGAINST": 0.6391752577319587, + "NONE": 0.6835443037974683 + }, + "accuracy_pct": 60.5, + "macro_f1_pct": 60.888552313448095, + "weighted_f1_pct": 59.34994209829213, + "macro_precision_pct": 64.68870827744023, + "macro_recall_pct": 61.615670287757005, + "per_class_f1_pct": { + "FAVOR": 50.39370078740157, + "AGAINST": 63.91752577319587, + "NONE": 68.35443037974683 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.695, + "macro_f1": 0.6127814088598402, + "weighted_f1": 0.6625642701525055, + "macro_precision": 0.7115890083632018, + "macro_recall": 0.5864327095711404, + "per_class_f1": { + "FAVOR": 0.6222222222222222, + "AGAINST": 0.4235294117647059, + "NONE": 0.7925925925925926 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 61.27814088598402, + "weighted_f1_pct": 66.25642701525055, + "macro_precision_pct": 71.15890083632019, + "macro_recall_pct": 58.643270957114034, + "per_class_f1_pct": { + "FAVOR": 62.22222222222222, + "AGAINST": 42.35294117647059, + "NONE": 79.25925925925927 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.46017699115044247, + "macro_f1": 0.3804266072092159, + "weighted_f1": 0.3899808407572933, + "macro_precision": 0.5133477633477633, + "macro_recall": 0.5668244145128019, + "per_class_f1": { + "FAVOR": 0.2564102564102564, + "AGAINST": 0.2608695652173913, + "NONE": 0.624 + }, + "accuracy_pct": 46.017699115044245, + "macro_f1_pct": 38.04266072092159, + "weighted_f1_pct": 38.99808407572933, + "macro_precision_pct": 51.334776334776336, + "macro_recall_pct": 56.682441451280184, + "per_class_f1_pct": { + "FAVOR": 25.64102564102564, + "AGAINST": 26.08695652173913, + "NONE": 62.4 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.5698924731182796, + "macro_f1": 0.5228988168094655, + "weighted_f1": 0.5680079559363524, + "macro_precision": 0.6152063862928349, + "macro_recall": 0.5760942760942761, + "per_class_f1": { + "FAVOR": 0.375, + "AGAINST": 0.627906976744186, + "NONE": 0.5657894736842105 + }, + "accuracy_pct": 56.98924731182796, + "macro_f1_pct": 52.28988168094655, + "weighted_f1_pct": 56.80079559363524, + "macro_precision_pct": 61.52063862928349, + "macro_recall_pct": 57.60942760942761, + "per_class_f1_pct": { + "FAVOR": 37.5, + "AGAINST": 62.7906976744186, + "NONE": 56.57894736842105 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.5849042897097817, + "weighted_f1": 0.5496978821824924, + "macro_precision": 0.6521317829457365, + "macro_recall": 0.6649572649572649, + "per_class_f1": { + "FAVOR": 0.7142857142857143, + "AGAINST": 0.3157894736842105, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 58.49042897097817, + "weighted_f1_pct": 54.969788218249235, + "macro_precision_pct": 65.21317829457365, + "macro_recall_pct": 66.49572649572649, + "per_class_f1_pct": { + "FAVOR": 71.42857142857143, + "AGAINST": 31.57894736842105, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7094223046030276, + "weighted_f1": 0.715, + "macro_precision": 0.7094223046030276, + "macro_recall": 0.7094223046030276, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.676923076923077, + "NONE": 0.7590361445783133 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.94223046030275, + "weighted_f1_pct": 71.5, + "macro_precision_pct": 70.94223046030275, + "macro_recall_pct": 70.94223046030275, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 67.6923076923077, + "NONE": 75.90361445783132 + } + } + ], + "accuracy": 0.5432020330368488, + "macro_f1": 0.5273308678061798, + "weighted_f1": 0.5226299464400128, + "macro_precision": 0.5805833458935049, + "macro_recall": 0.5572547882322926, + "per_class_f1": { + "FAVOR": 0.48721804511278194, + "AGAINST": 0.4695977549111319, + "NONE": 0.6251768033946252 + }, + "accuracy_pct": 54.32020330368488, + "macro_f1_pct": 52.733086780617974, + "weighted_f1_pct": 52.26299464400128, + "macro_precision_pct": 58.058334589350494, + "macro_recall_pct": 55.72547882322926, + "per_class_f1_pct": { + "FAVOR": 48.721804511278194, + "AGAINST": 46.95977549111319, + "NONE": 62.51768033946252 + } + }, + { + "after_task": 10, + "best_epoch": 4, + "best_dev_macro_f1": 0.6480304456999372, + "losses": { + "current_ce": 0.629376731851162, + "total": 0.629376731851162, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2054794520547945, + "macro_f1": 0.13820794590025356, + "weighted_f1": 0.08039984251786149, + "macro_precision": 0.18015873015873016, + "macro_recall": 0.3466666666666667, + "per_class_f1": { + "FAVOR": 0.07142857142857142, + "AGAINST": 0.0, + "NONE": 0.3431952662721893 + }, + "accuracy_pct": 20.54794520547945, + "macro_f1_pct": 13.820794590025356, + "weighted_f1_pct": 8.03998425178615, + "macro_precision_pct": 18.015873015873016, + "macro_recall_pct": 34.66666666666667, + "per_class_f1_pct": { + "FAVOR": 7.142857142857142, + "AGAINST": 0.0, + "NONE": 34.319526627218934 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.3736842105263158, + "macro_f1": 0.3419903257931427, + "weighted_f1": 0.35347372913718356, + "macro_precision": 0.4618641349669639, + "macro_recall": 0.4745824255628177, + "per_class_f1": { + "FAVOR": 0.19718309859154928, + "AGAINST": 0.4166666666666667, + "NONE": 0.4121212121212121 + }, + "accuracy_pct": 37.368421052631575, + "macro_f1_pct": 34.19903257931427, + "weighted_f1_pct": 35.34737291371835, + "macro_precision_pct": 46.18641349669639, + "macro_recall_pct": 47.458242556281775, + "per_class_f1_pct": { + "FAVOR": 19.718309859154928, + "AGAINST": 41.66666666666667, + "NONE": 41.21212121212121 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.40609137055837563, + "macro_f1": 0.3781268706306187, + "weighted_f1": 0.26777977386048984, + "macro_precision": 0.591991341991342, + "macro_recall": 0.6083742158703117, + "per_class_f1": { + "FAVOR": 0.46153846153846156, + "AGAINST": 0.05172413793103448, + "NONE": 0.6211180124223602 + }, + "accuracy_pct": 40.609137055837564, + "macro_f1_pct": 37.81268706306187, + "weighted_f1_pct": 26.777977386048985, + "macro_precision_pct": 59.19913419913419, + "macro_recall_pct": 60.83742158703117, + "per_class_f1_pct": { + "FAVOR": 46.15384615384615, + "AGAINST": 5.172413793103448, + "NONE": 62.11180124223602 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.643914676701562, + "weighted_f1": 0.6248105446466102, + "macro_precision": 0.6555882352941177, + "macro_recall": 0.6645202634362525, + "per_class_f1": { + "FAVOR": 0.644808743169399, + "AGAINST": 0.5396825396825397, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 64.3914676701562, + "weighted_f1_pct": 62.48105446466102, + "macro_precision_pct": 65.55882352941177, + "macro_recall_pct": 66.45202634362525, + "per_class_f1_pct": { + "FAVOR": 64.48087431693989, + "AGAINST": 53.96825396825397, + "NONE": 74.72527472527473 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.695, + "macro_f1": 0.5912012393947482, + "weighted_f1": 0.6397486535713724, + "macro_precision": 0.7262277951933124, + "macro_recall": 0.658115386751069, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.2898550724637681, + "NONE": 0.8076923076923077 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 59.12012393947482, + "weighted_f1_pct": 63.97486535713725, + "macro_precision_pct": 72.62277951933123, + "macro_recall_pct": 65.8115386751069, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 28.985507246376812, + "NONE": 80.76923076923077 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5486725663716814, + "macro_f1": 0.41830089437383594, + "weighted_f1": 0.5162513563399008, + "macro_precision": 0.5050879161990273, + "macro_recall": 0.499502487562189, + "per_class_f1": { + "FAVOR": 0.449438202247191, + "AGAINST": 0.13333333333333333, + "NONE": 0.6721311475409836 + }, + "accuracy_pct": 54.86725663716814, + "macro_f1_pct": 41.83008943738359, + "weighted_f1_pct": 51.625135633990084, + "macro_precision_pct": 50.50879161990272, + "macro_recall_pct": 49.950248756218905, + "per_class_f1_pct": { + "FAVOR": 44.9438202247191, + "AGAINST": 13.333333333333334, + "NONE": 67.21311475409836 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.553763440860215, + "macro_f1": 0.542305087810031, + "weighted_f1": 0.543518645860433, + "macro_precision": 0.5721607378129118, + "macro_recall": 0.62996632996633, + "per_class_f1": { + "FAVOR": 0.4675324675324675, + "AGAINST": 0.5316455696202531, + "NONE": 0.6277372262773723 + }, + "accuracy_pct": 55.376344086021504, + "macro_f1_pct": 54.230508781003095, + "weighted_f1_pct": 54.3518645860433, + "macro_precision_pct": 57.21607378129118, + "macro_recall_pct": 62.996632996633, + "per_class_f1_pct": { + "FAVOR": 46.75324675324675, + "AGAINST": 53.16455696202531, + "NONE": 62.77372262773723 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5070422535211268, + "macro_f1": 0.4287938854959106, + "weighted_f1": 0.379958740981192, + "macro_precision": 0.6733498349834983, + "macro_recall": 0.5594017094017094, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.03278688524590164, + "NONE": 0.6535947712418301 + }, + "accuracy_pct": 50.70422535211267, + "macro_f1_pct": 42.879388549591056, + "weighted_f1_pct": 37.995874098119195, + "macro_precision_pct": 67.33498349834983, + "macro_recall_pct": 55.940170940170944, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 3.278688524590164, + "NONE": 65.359477124183 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.65, + "macro_f1": 0.6059476461616938, + "weighted_f1": 0.6207997609393086, + "macro_precision": 0.7269206502624407, + "macro_recall": 0.646339202965709, + "per_class_f1": { + "FAVOR": 0.624113475177305, + "AGAINST": 0.40963855421686746, + "NONE": 0.7840909090909091 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 60.594764616169385, + "weighted_f1_pct": 62.07997609393085, + "macro_precision_pct": 72.69206502624407, + "macro_recall_pct": 64.6339202965709, + "per_class_f1_pct": { + "FAVOR": 62.4113475177305, + "AGAINST": 40.963855421686745, + "NONE": 78.4090909090909 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.695, + "macro_f1": 0.6721045109617051, + "weighted_f1": 0.6981399158468471, + "macro_precision": 0.6667176163691204, + "macro_recall": 0.6803401560683114, + "per_class_f1": { + "FAVOR": 0.7309644670050761, + "AGAINST": 0.4888888888888889, + "NONE": 0.7964601769911505 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 67.21045109617052, + "weighted_f1_pct": 69.81399158468471, + "macro_precision_pct": 66.67176163691204, + "macro_recall_pct": 68.03401560683115, + "per_class_f1_pct": { + "FAVOR": 73.09644670050761, + "AGAINST": 48.888888888888886, + "NONE": 79.64601769911505 + } + } + ], + "accuracy": 0.5372040586245772, + "macro_f1": 0.514031119507793, + "weighted_f1": 0.49501382005080913, + "macro_precision": 0.5756201470487184, + "macro_recall": 0.5724163579551608, + "per_class_f1": { + "FAVOR": 0.5613026819923371, + "AGAINST": 0.3343782654127482, + "NONE": 0.6464124111182935 + }, + "accuracy_pct": 53.720405862457724, + "macro_f1_pct": 51.4031119507793, + "weighted_f1_pct": 49.50138200508091, + "macro_precision_pct": 57.56201470487184, + "macro_recall_pct": 57.241635795516075, + "per_class_f1_pct": { + "FAVOR": 56.13026819923371, + "AGAINST": 33.437826541274816, + "NONE": 64.64124111182934 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.47608930832235014, + "average_forgetting": 0.23169045893874543, + "backward_transfer": -0.22780465810831987, + "final_mean_target_macro_f1_pct": 47.60893083223501, + "average_forgetting_pct": 23.169045893874543, + "backward_transfer_pct": -22.780465810831988, + "final_pooled_accuracy": 0.5372040586245772, + "final_pooled_accuracy_pct": 53.720405862457724, + "final_pooled_macro_f1": 0.514031119507793, + "final_pooled_macro_f1_pct": 51.4031119507793, + "final_pooled_weighted_f1": 0.49501382005080913, + "final_pooled_weighted_f1_pct": 49.50138200508091, + "final_pooled_macro_precision": 0.5756201470487184, + "final_pooled_macro_precision_pct": 57.56201470487184, + "final_pooled_macro_recall": 0.5724163579551608, + "final_pooled_macro_recall_pct": 57.241635795516075, + "final_pooled_per_class_f1": { + "FAVOR": 0.5613026819923371, + "AGAINST": 0.3343782654127482, + "NONE": 0.6464124111182935 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 56.13026819923371, + "AGAINST": 33.437826541274816, + "NONE": 64.64124111182934 + } + }, + "efficiency": { + "training_wall_time_seconds": 611.4533863409888, + "training_wall_time_minutes": 10.190889772349813, + "peak_gpu_memory_allocated_gb": 5.56232213973999, + "peak_gpu_memory_reserved_gb": 6.103515625 + } +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s3__seed13/metrics.partial.json b/outputs/trace_s4_lora__x__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..0e2a3504934d9ceb5dde5952fab8ab44ba5399d6 --- /dev/null +++ b/outputs/trace_s4_lora__x__s3__seed13/metrics.partial.json @@ -0,0 +1,1826 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5708112646147389, + 0.5998266312387883, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.3744924767136375, + 0.5627755236009485, + 0.7013404231749196, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.18042791376746414, + 0.4436151051990784, + 0.4156903254351832, + 0.6032312557296334, + null, + null, + null, + null, + null, + null + ], + [ + 0.13034188034188035, + 0.2525894942561609, + 0.417463869693169, + 0.6264037385266303, + 0.8025875899187497, + null, + null, + null, + null, + null + ], + [ + 0.15785955807187227, + 0.3789677419354838, + 0.4862612132133222, + 0.6382034632034632, + 0.7360729227349351, + 0.7374031007751939, + null, + null, + null, + null + ], + [ + 0.41347926267281104, + 0.3326676447089291, + 0.32048549332545195, + 0.5780035940994712, + 0.5826284918011537, + 0.6397100464378752, + 0.6677773101384212, + null, + null, + null + ], + [ + 0.29483726150392814, + 0.4286151306896601, + 0.5461938637123898, + 0.591231114002747, + 0.6968202599270561, + 0.49066065182570034, + 0.6131044510107406, + 0.6021575885774723, + null, + null + ], + [ + 0.17617797317375994, + 0.41787906257036783, + 0.4497613277520746, + 0.608885523134481, + 0.6127814088598402, + 0.3804266072092159, + 0.5228988168094655, + 0.5849042897097817, + 0.7094223046030276, + null + ], + [ + 0.13820794590025356, + 0.3419903257931427, + 0.3781268706306187, + 0.643914676701562, + 0.5912012393947482, + 0.41830089437383594, + 0.542305087810031, + 0.4287938854959106, + 0.6059476461616938, + 0.6721045109617051 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 14, + "best_dev_macro_f1": 0.61515693996897, + "losses": { + "current_ce": 0.6929058672803821, + "total": 0.6929058672803821, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.5708112646147389, + "weighted_f1": 0.6792593028559658, + "macro_precision": 0.5797966401414677, + "macro_recall": 0.5754822588705647, + "per_class_f1": { + "FAVOR": 0.15789473684210525, + "AGAINST": 0.7959183673469388, + "NONE": 0.7586206896551724 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 57.081126461473886, + "weighted_f1_pct": 67.92593028559658, + "macro_precision_pct": 57.97966401414677, + "macro_recall_pct": 57.54822588705647, + "per_class_f1_pct": { + "FAVOR": 15.789473684210526, + "AGAINST": 79.59183673469387, + "NONE": 75.86206896551724 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5998266312387883, + "weighted_f1": 0.6187294631551585, + "macro_precision": 0.6009237085186453, + "macro_recall": 0.6256354393609296, + "per_class_f1": { + "FAVOR": 0.6165413533834586, + "AGAINST": 0.6477272727272727, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.98266312387883, + "weighted_f1_pct": 61.872946315515854, + "macro_precision_pct": 60.09237085186453, + "macro_recall_pct": 62.563543936092955, + "per_class_f1_pct": { + "FAVOR": 61.65413533834586, + "AGAINST": 64.77272727272727, + "NONE": 53.52112676056338 + } + } + ], + "accuracy": 0.6547619047619048, + "macro_f1": 0.6253617497941737, + "weighted_f1": 0.6592498035745361, + "macro_precision": 0.6192999856898244, + "macro_recall": 0.6345439882566334, + "per_class_f1": { + "FAVOR": 0.5146198830409356, + "AGAINST": 0.7258064516129032, + "NONE": 0.6356589147286822 + }, + "accuracy_pct": 65.47619047619048, + "macro_f1_pct": 62.53617497941737, + "weighted_f1_pct": 65.9249803574536, + "macro_precision_pct": 61.929998568982434, + "macro_recall_pct": 63.45439882566334, + "per_class_f1_pct": { + "FAVOR": 51.461988304093566, + "AGAINST": 72.58064516129032, + "NONE": 63.565891472868216 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7310140850150589, + "losses": { + "current_ce": 0.5811113305164106, + "total": 0.5811113305164106, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.5205479452054794, + "macro_f1": 0.3744924767136375, + "weighted_f1": 0.4962620766822291, + "macro_precision": 0.3650793650793651, + "macro_recall": 0.48000999500249875, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.6329113924050633, + "NONE": 0.49056603773584906 + }, + "accuracy_pct": 52.054794520547944, + "macro_f1_pct": 37.44924767136375, + "weighted_f1_pct": 49.62620766822291, + "macro_precision_pct": 36.50793650793651, + "macro_recall_pct": 48.000999500249875, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 63.29113924050633, + "NONE": 49.056603773584904 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5627755236009485, + "weighted_f1": 0.5817740495748138, + "macro_precision": 0.5584795321637427, + "macro_recall": 0.6042120551924474, + "per_class_f1": { + "FAVOR": 0.41304347826086957, + "AGAINST": 0.6598984771573604, + "NONE": 0.6153846153846154 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 56.27755236009485, + "weighted_f1_pct": 58.177404957481386, + "macro_precision_pct": 55.84795321637427, + "macro_recall_pct": 60.42120551924474, + "per_class_f1_pct": { + "FAVOR": 41.30434782608695, + "AGAINST": 65.98984771573603, + "NONE": 61.53846153846154 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.7013404231749196, + "weighted_f1": 0.7121090198102816, + "macro_precision": 0.6851457000710731, + "macro_recall": 0.7352231318915337, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7254901960784313, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 70.13404231749196, + "weighted_f1_pct": 71.21090198102816, + "macro_precision_pct": 68.51457000710731, + "macro_recall_pct": 73.52231318915337, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 72.54901960784314, + "NONE": 71.1864406779661 + } + } + ], + "accuracy": 0.6153846153846154, + "macro_f1": 0.5778826632024306, + "weighted_f1": 0.6139738978824454, + "macro_precision": 0.5883706467661692, + "macro_recall": 0.6138896712465202, + "per_class_f1": { + "FAVOR": 0.4479166666666667, + "AGAINST": 0.6762075134168157, + "NONE": 0.6095238095238096 + }, + "accuracy_pct": 61.53846153846154, + "macro_f1_pct": 57.78826632024307, + "weighted_f1_pct": 61.39738978824454, + "macro_precision_pct": 58.837064676616926, + "macro_recall_pct": 61.38896712465202, + "per_class_f1_pct": { + "FAVOR": 44.79166666666667, + "AGAINST": 67.62075134168157, + "NONE": 60.952380952380956 + } + }, + { + "after_task": 4, + "best_epoch": 9, + "best_dev_macro_f1": 0.6345427141476642, + "losses": { + "current_ce": 0.6060382296168615, + "total": 0.6060382296168615, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2191780821917808, + "macro_f1": 0.18042791376746414, + "weighted_f1": 0.13063285331390032, + "macro_precision": 0.31984126984126987, + "macro_recall": 0.349720139930035, + "per_class_f1": { + "FAVOR": 0.13043478260869565, + "AGAINST": 0.061855670103092786, + "NONE": 0.348993288590604 + }, + "accuracy_pct": 21.91780821917808, + "macro_f1_pct": 18.042791376746415, + "weighted_f1_pct": 13.063285331390032, + "macro_precision_pct": 31.984126984126988, + "macro_recall_pct": 34.9720139930035, + "per_class_f1_pct": { + "FAVOR": 13.043478260869565, + "AGAINST": 6.185567010309279, + "NONE": 34.899328859060404 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.45263157894736844, + "macro_f1": 0.4436151051990784, + "weighted_f1": 0.3824697118060538, + "macro_precision": 0.6047649989170457, + "macro_recall": 0.6027596223674655, + "per_class_f1": { + "FAVOR": 0.5503355704697986, + "AGAINST": 0.2413793103448276, + "NONE": 0.5391304347826087 + }, + "accuracy_pct": 45.26315789473684, + "macro_f1_pct": 44.36151051990784, + "weighted_f1_pct": 38.246971180605385, + "macro_precision_pct": 60.47649989170457, + "macro_recall_pct": 60.27596223674655, + "per_class_f1_pct": { + "FAVOR": 55.033557046979865, + "AGAINST": 24.137931034482758, + "NONE": 53.91304347826087 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.41624365482233505, + "macro_f1": 0.4156903254351832, + "weighted_f1": 0.3338268279071345, + "macro_precision": 0.5817827396774765, + "macro_recall": 0.6141687568028016, + "per_class_f1": { + "FAVOR": 0.43537414965986393, + "AGAINST": 0.16129032258064516, + "NONE": 0.6504065040650406 + }, + "accuracy_pct": 41.62436548223351, + "macro_f1_pct": 41.56903254351832, + "weighted_f1_pct": 33.38268279071345, + "macro_precision_pct": 58.17827396774765, + "macro_recall_pct": 61.41687568028016, + "per_class_f1_pct": { + "FAVOR": 43.53741496598639, + "AGAINST": 16.129032258064516, + "NONE": 65.04065040650406 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.6, + "macro_f1": 0.6032312557296334, + "weighted_f1": 0.5787851889620221, + "macro_precision": 0.6433130805223829, + "macro_recall": 0.6222446127595179, + "per_class_f1": { + "FAVOR": 0.6368159203980099, + "AGAINST": 0.43478260869565216, + "NONE": 0.7380952380952381 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.32312557296334, + "weighted_f1_pct": 57.878518896202216, + "macro_precision_pct": 64.33130805223828, + "macro_recall_pct": 62.22446127595179, + "per_class_f1_pct": { + "FAVOR": 63.681592039801, + "AGAINST": 43.47826086956522, + "NONE": 73.80952380952381 + } + } + ], + "accuracy": 0.43656207366984995, + "macro_f1": 0.429755562344462, + "weighted_f1": 0.37155734506783417, + "macro_precision": 0.5270608980414383, + "macro_recall": 0.5620866625713692, + "per_class_f1": { + "FAVOR": 0.5156537753222836, + "AGAINST": 0.23008849557522124, + "NONE": 0.5435244161358811 + }, + "accuracy_pct": 43.656207366984994, + "macro_f1_pct": 42.9755562344462, + "weighted_f1_pct": 37.15573450678342, + "macro_precision_pct": 52.70608980414383, + "macro_recall_pct": 56.20866625713692, + "per_class_f1_pct": { + "FAVOR": 51.56537753222836, + "AGAINST": 23.008849557522122, + "NONE": 54.35244161358811 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.6736155223960102, + "losses": { + "current_ce": 0.4023706999296943, + "total": 0.4023706999296943, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2054794520547945, + "macro_f1": 0.13034188034188035, + "weighted_f1": 0.10256410256410256, + "macro_precision": 0.15100250626566417, + "macro_recall": 0.3212143928035982, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.057692307692307696, + "NONE": 0.3333333333333333 + }, + "accuracy_pct": 20.54794520547945, + "macro_f1_pct": 13.034188034188036, + "weighted_f1_pct": 10.256410256410255, + "macro_precision_pct": 15.100250626566417, + "macro_recall_pct": 32.12143928035982, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 5.769230769230769, + "NONE": 33.33333333333333 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.26842105263157895, + "macro_f1": 0.2525894942561609, + "weighted_f1": 0.21613647482068535, + "macro_precision": 0.5061391416230127, + "macro_recall": 0.40777051561365285, + "per_class_f1": { + "FAVOR": 0.275, + "AGAINST": 0.14414414414414414, + "NONE": 0.3386243386243386 + }, + "accuracy_pct": 26.842105263157894, + "macro_f1_pct": 25.25894942561609, + "weighted_f1_pct": 21.613647482068536, + "macro_precision_pct": 50.61391416230126, + "macro_recall_pct": 40.77705156136528, + "per_class_f1_pct": { + "FAVOR": 27.500000000000004, + "AGAINST": 14.414414414414415, + "NONE": 33.86243386243386 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.43147208121827413, + "macro_f1": 0.417463869693169, + "weighted_f1": 0.3430121206318303, + "macro_precision": 0.5654733107563296, + "macro_recall": 0.6016489728098265, + "per_class_f1": { + "FAVOR": 0.45045045045045046, + "AGAINST": 0.19047619047619047, + "NONE": 0.6114649681528662 + }, + "accuracy_pct": 43.14720812182741, + "macro_f1_pct": 41.7463869693169, + "weighted_f1_pct": 34.30121206318303, + "macro_precision_pct": 56.547331075632954, + "macro_recall_pct": 60.16489728098266, + "per_class_f1_pct": { + "FAVOR": 45.04504504504504, + "AGAINST": 19.047619047619047, + "NONE": 61.146496815286625 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.615, + "macro_f1": 0.6264037385266303, + "weighted_f1": 0.6145566686856575, + "macro_precision": 0.6497461547704879, + "macro_recall": 0.6152495000733483, + "per_class_f1": { + "FAVOR": 0.5584415584415584, + "AGAINST": 0.6358381502890174, + "NONE": 0.684931506849315 + }, + "accuracy_pct": 61.5, + "macro_f1_pct": 62.64037385266303, + "weighted_f1_pct": 61.45566686856575, + "macro_precision_pct": 64.97461547704879, + "macro_recall_pct": 61.52495000733483, + "per_class_f1_pct": { + "FAVOR": 55.84415584415584, + "AGAINST": 63.58381502890174, + "NONE": 68.4931506849315 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8025875899187497, + "weighted_f1": 0.8367969259431766, + "macro_precision": 0.7903796653796654, + "macro_recall": 0.8200807003905455, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7894736842105263, + "NONE": 0.8849557522123894 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 80.25875899187497, + "weighted_f1_pct": 83.67969259431766, + "macro_precision_pct": 79.03796653796654, + "macro_recall_pct": 82.00807003905454, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 78.94736842105263, + "NONE": 88.49557522123894 + } + } + ], + "accuracy": 0.4887459807073955, + "macro_f1": 0.4784554201627138, + "weighted_f1": 0.4628966040143497, + "macro_precision": 0.5251892152683807, + "macro_recall": 0.5320008461844429, + "per_class_f1": { + "FAVOR": 0.46867749419953597, + "AGAINST": 0.39171974522292996, + "NONE": 0.5749690210656754 + }, + "accuracy_pct": 48.87459807073955, + "macro_f1_pct": 47.84554201627138, + "weighted_f1_pct": 46.289660401434965, + "macro_precision_pct": 52.51892152683807, + "macro_recall_pct": 53.200084618444286, + "per_class_f1_pct": { + "FAVOR": 46.8677494199536, + "AGAINST": 39.171974522293, + "NONE": 57.49690210656754 + } + }, + { + "after_task": 6, + "best_epoch": 6, + "best_dev_macro_f1": 0.7391617322651806, + "losses": { + "current_ce": 0.41747793562355495, + "total": 0.41747793562355495, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2328767123287671, + "macro_f1": 0.15785955807187227, + "weighted_f1": 0.1500030699925997, + "macro_precision": 0.21614583333333334, + "macro_recall": 0.3357071464267866, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.12962962962962962, + "NONE": 0.34394904458598724 + }, + "accuracy_pct": 23.28767123287671, + "macro_f1_pct": 15.785955807187227, + "weighted_f1_pct": 15.000306999259971, + "macro_precision_pct": 21.614583333333336, + "macro_recall_pct": 33.57071464267866, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 12.962962962962962, + "NONE": 34.394904458598724 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.37894736842105264, + "macro_f1": 0.3789677419354838, + "weighted_f1": 0.3564563667232598, + "macro_precision": 0.5156905018565098, + "macro_recall": 0.5054466230936819, + "per_class_f1": { + "FAVOR": 0.42, + "AGAINST": 0.304, + "NONE": 0.4129032258064516 + }, + "accuracy_pct": 37.89473684210527, + "macro_f1_pct": 37.89677419354838, + "weighted_f1_pct": 35.64563667232598, + "macro_precision_pct": 51.56905018565098, + "macro_recall_pct": 50.54466230936819, + "per_class_f1_pct": { + "FAVOR": 42.0, + "AGAINST": 30.4, + "NONE": 41.29032258064516 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.4873096446700508, + "macro_f1": 0.4862612132133222, + "weighted_f1": 0.4440674517042101, + "macro_precision": 0.5640479070413039, + "macro_recall": 0.6304902223694518, + "per_class_f1": { + "FAVOR": 0.48598130841121495, + "AGAINST": 0.352112676056338, + "NONE": 0.6206896551724138 + }, + "accuracy_pct": 48.73096446700508, + "macro_f1_pct": 48.62612132133222, + "weighted_f1_pct": 44.40674517042101, + "macro_precision_pct": 56.40479070413039, + "macro_recall_pct": 63.049022236945184, + "per_class_f1_pct": { + "FAVOR": 48.598130841121495, + "AGAINST": 35.2112676056338, + "NONE": 62.06896551724138 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6382034632034632, + "weighted_f1": 0.6282175324675324, + "macro_precision": 0.672889373037831, + "macro_recall": 0.6242134358665524, + "per_class_f1": { + "FAVOR": 0.5584415584415584, + "AGAINST": 0.6704545454545454, + "NONE": 0.6857142857142857 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 63.820346320346324, + "weighted_f1_pct": 62.82175324675324, + "macro_precision_pct": 67.2889373037831, + "macro_recall_pct": 62.42134358665524, + "per_class_f1_pct": { + "FAVOR": 55.84415584415584, + "AGAINST": 67.04545454545455, + "NONE": 68.57142857142857 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7360729227349351, + "weighted_f1": 0.7894011483336015, + "macro_precision": 0.7289582499666533, + "macro_recall": 0.7871508690099395, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7102803738317757, + "NONE": 0.8663594470046083 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 73.60729227349351, + "weighted_f1_pct": 78.94011483336016, + "macro_precision_pct": 72.89582499666533, + "macro_recall_pct": 78.71508690099395, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 71.02803738317756, + "NONE": 86.63594470046083 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7374031007751939, + "weighted_f1": 0.8446954105783084, + "macro_precision": 0.7155347384855583, + "macro_recall": 0.7712898920033976, + "per_class_f1": { + "FAVOR": 0.875, + "AGAINST": 0.5, + "NONE": 0.8372093023255814 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.74031007751938, + "weighted_f1_pct": 84.46954105783084, + "macro_precision_pct": 71.55347384855583, + "macro_recall_pct": 77.12898920033976, + "per_class_f1_pct": { + "FAVOR": 87.5, + "AGAINST": 50.0, + "NONE": 83.72093023255815 + } + } + ], + "accuracy": 0.5535372848948374, + "macro_f1": 0.5489191133229975, + "weighted_f1": 0.5354779997634255, + "macro_precision": 0.5782223519833679, + "macro_recall": 0.5863233532570741, + "per_class_f1": { + "FAVOR": 0.5743243243243243, + "AGAINST": 0.4507462686567164, + "NONE": 0.6216867469879518 + }, + "accuracy_pct": 55.35372848948374, + "macro_f1_pct": 54.89191133229975, + "weighted_f1_pct": 53.547799976342546, + "macro_precision_pct": 57.82223519833679, + "macro_recall_pct": 58.63233532570741, + "per_class_f1_pct": { + "FAVOR": 57.432432432432435, + "AGAINST": 45.07462686567164, + "NONE": 62.16867469879518 + } + }, + { + "after_task": 7, + "best_epoch": 15, + "best_dev_macro_f1": 0.6251282051282051, + "losses": { + "current_ce": 0.598251260952516, + "total": 0.598251260952516, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.5821917808219178, + "macro_f1": 0.41347926267281104, + "weighted_f1": 0.5581868253266838, + "macro_precision": 0.4127621597501115, + "macro_recall": 0.5362318840579711, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.7225806451612903, + "NONE": 0.5178571428571429 + }, + "accuracy_pct": 58.21917808219178, + "macro_f1_pct": 41.3479262672811, + "weighted_f1_pct": 55.81868253266838, + "macro_precision_pct": 41.27621597501115, + "macro_recall_pct": 53.62318840579711, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 72.25806451612902, + "NONE": 51.78571428571429 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.34210526315789475, + "macro_f1": 0.3326676447089291, + "weighted_f1": 0.26298148824826756, + "macro_precision": 0.43612313612313613, + "macro_recall": 0.5065359477124183, + "per_class_f1": { + "FAVOR": 0.4954128440366973, + "AGAINST": 0.09009009009009009, + "NONE": 0.4125 + }, + "accuracy_pct": 34.21052631578947, + "macro_f1_pct": 33.26676447089291, + "weighted_f1_pct": 26.298148824826757, + "macro_precision_pct": 43.61231361231361, + "macro_recall_pct": 50.65359477124183, + "per_class_f1_pct": { + "FAVOR": 49.54128440366973, + "AGAINST": 9.00900900900901, + "NONE": 41.25 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.34517766497461927, + "macro_f1": 0.32048549332545195, + "weighted_f1": 0.25645948875514873, + "macro_precision": 0.4458295459464139, + "macro_recall": 0.4585364314671967, + "per_class_f1": { + "FAVOR": 0.3582089552238806, + "AGAINST": 0.128, + "NONE": 0.4752475247524752 + }, + "accuracy_pct": 34.51776649746193, + "macro_f1_pct": 32.04854933254519, + "weighted_f1_pct": 25.645948875514872, + "macro_precision_pct": 44.582954594641386, + "macro_recall_pct": 45.85364314671967, + "per_class_f1_pct": { + "FAVOR": 35.82089552238806, + "AGAINST": 12.8, + "NONE": 47.524752475247524 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.58, + "macro_f1": 0.5780035940994712, + "weighted_f1": 0.5570422371329946, + "macro_precision": 0.5926365653294957, + "macro_recall": 0.6096402844370324, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.4067796610169492, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 57.99999999999999, + "macro_f1_pct": 57.80035940994712, + "weighted_f1_pct": 55.70422371329946, + "macro_precision_pct": 59.26365653294957, + "macro_recall_pct": 60.96402844370324, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 40.67796610169492, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.69, + "macro_f1": 0.5826284918011537, + "weighted_f1": 0.6376017159245576, + "macro_precision": 0.715089103125913, + "macro_recall": 0.5758898328613471, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.32432432432432434, + "NONE": 0.7985611510791367 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 58.262849180115374, + "weighted_f1_pct": 63.76017159245576, + "macro_precision_pct": 71.5089103125913, + "macro_recall_pct": 57.588983286134706, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 32.432432432432435, + "NONE": 79.85611510791367 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7256637168141593, + "macro_f1": 0.6397100464378752, + "weighted_f1": 0.7251645572665302, + "macro_precision": 0.687948350071736, + "macro_recall": 0.6575415604902317, + "per_class_f1": { + "FAVOR": 0.7407407407407407, + "AGAINST": 0.4444444444444444, + "NONE": 0.7339449541284404 + }, + "accuracy_pct": 72.56637168141593, + "macro_f1_pct": 63.97100464378752, + "weighted_f1_pct": 72.51645572665302, + "macro_precision_pct": 68.7948350071736, + "macro_recall_pct": 65.75415604902317, + "per_class_f1_pct": { + "FAVOR": 74.07407407407408, + "AGAINST": 44.44444444444444, + "NONE": 73.39449541284404 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6677773101384212, + "weighted_f1": 0.6953563411896745, + "macro_precision": 0.6528880070546738, + "macro_recall": 0.7086980920314253, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7395833333333334, + "NONE": 0.6464646464646465 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.77773101384213, + "weighted_f1_pct": 69.53563411896745, + "macro_precision_pct": 65.28880070546738, + "macro_recall_pct": 70.86980920314254, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 73.95833333333334, + "NONE": 64.64646464646465 + } + } + ], + "accuracy": 0.5535714285714286, + "macro_f1": 0.5473386831180124, + "weighted_f1": 0.5315355090154374, + "macro_precision": 0.6096082028477673, + "macro_recall": 0.5974800777355643, + "per_class_f1": { + "FAVOR": 0.5700636942675159, + "AGAINST": 0.45408163265306123, + "NONE": 0.6178707224334601 + }, + "accuracy_pct": 55.35714285714286, + "macro_f1_pct": 54.73386831180124, + "weighted_f1_pct": 53.153550901543746, + "macro_precision_pct": 60.96082028477673, + "macro_recall_pct": 59.74800777355643, + "per_class_f1_pct": { + "FAVOR": 57.00636942675159, + "AGAINST": 45.40816326530612, + "NONE": 61.78707224334601 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.7521164021164021, + "losses": { + "current_ce": 0.6733893568317095, + "total": 0.6733893568317095, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.4041095890410959, + "macro_f1": 0.29483726150392814, + "weighted_f1": 0.37836815645034827, + "macro_precision": 0.34638364779874214, + "macro_recall": 0.4341579210394803, + "per_class_f1": { + "FAVOR": 0.0, + "AGAINST": 0.4696969696969697, + "NONE": 0.4148148148148148 + }, + "accuracy_pct": 40.41095890410959, + "macro_f1_pct": 29.483726150392815, + "weighted_f1_pct": 37.836815645034825, + "macro_precision_pct": 34.638364779874216, + "macro_recall_pct": 43.41579210394803, + "per_class_f1_pct": { + "FAVOR": 0.0, + "AGAINST": 46.96969696969697, + "NONE": 41.48148148148148 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.4631578947368421, + "macro_f1": 0.4286151306896601, + "weighted_f1": 0.447704066158481, + "macro_precision": 0.5093726060378657, + "macro_recall": 0.532316630355846, + "per_class_f1": { + "FAVOR": 0.2777777777777778, + "AGAINST": 0.5263157894736842, + "NONE": 0.48175182481751827 + }, + "accuracy_pct": 46.31578947368421, + "macro_f1_pct": 42.86151306896601, + "weighted_f1_pct": 44.7704066158481, + "macro_precision_pct": 50.93726060378657, + "macro_recall_pct": 53.231663035584596, + "per_class_f1_pct": { + "FAVOR": 27.77777777777778, + "AGAINST": 52.63157894736842, + "NONE": 48.175182481751825 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5482233502538071, + "macro_f1": 0.5461938637123898, + "weighted_f1": 0.5313678271164103, + "macro_precision": 0.5690958164642376, + "macro_recall": 0.6408699172884492, + "per_class_f1": { + "FAVOR": 0.5176470588235295, + "AGAINST": 0.49079754601226994, + "NONE": 0.6301369863013698 + }, + "accuracy_pct": 54.82233502538071, + "macro_f1_pct": 54.619386371238974, + "weighted_f1_pct": 53.136782711641025, + "macro_precision_pct": 56.90958164642376, + "macro_recall_pct": 64.08699172884492, + "per_class_f1_pct": { + "FAVOR": 51.76470588235295, + "AGAINST": 49.079754601226995, + "NONE": 63.013698630136986 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.591231114002747, + "weighted_f1": 0.5753898286705703, + "macro_precision": 0.6843616021276109, + "macro_recall": 0.5968313529289139, + "per_class_f1": { + "FAVOR": 0.46551724137931033, + "AGAINST": 0.6415094339622641, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 59.123111400274695, + "weighted_f1_pct": 57.53898286705703, + "macro_precision_pct": 68.43616021276108, + "macro_recall_pct": 59.683135292891386, + "per_class_f1_pct": { + "FAVOR": 46.55172413793103, + "AGAINST": 64.15094339622641, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.6968202599270561, + "weighted_f1": 0.7426621975651102, + "macro_precision": 0.7579075425790754, + "macro_recall": 0.6648453551002277, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5825242718446602, + "NONE": 0.8412698412698413 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 69.6820259927056, + "weighted_f1_pct": 74.26621975651102, + "macro_precision_pct": 75.79075425790755, + "macro_recall_pct": 66.48453551002277, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 58.252427184466015, + "NONE": 84.12698412698413 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5663716814159292, + "macro_f1": 0.49066065182570034, + "weighted_f1": 0.574450233063516, + "macro_precision": 0.5798771121351766, + "macro_recall": 0.6819075354932654, + "per_class_f1": { + "FAVOR": 0.5111111111111111, + "AGAINST": 0.24242424242424243, + "NONE": 0.7184466019417476 + }, + "accuracy_pct": 56.63716814159292, + "macro_f1_pct": 49.066065182570036, + "weighted_f1_pct": 57.445023306351594, + "macro_precision_pct": 57.987711213517656, + "macro_recall_pct": 68.19075354932654, + "per_class_f1_pct": { + "FAVOR": 51.11111111111111, + "AGAINST": 24.242424242424242, + "NONE": 71.84466019417476 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.6131044510107406, + "weighted_f1": 0.6668663788994947, + "macro_precision": 0.6926914355485784, + "macro_recall": 0.6318742985409652, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.7378640776699029, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 61.310445101074066, + "weighted_f1_pct": 66.68663788994948, + "macro_precision_pct": 69.26914355485783, + "macro_recall_pct": 63.18742985409652, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 73.7864077669903, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6021575885774723, + "weighted_f1": 0.5863594044435837, + "macro_precision": 0.6011904761904762, + "macro_recall": 0.6264957264957265, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.44, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.215758857747225, + "weighted_f1_pct": 58.63594044435837, + "macro_precision_pct": 60.11904761904761, + "macro_recall_pct": 62.64957264957265, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 44.0, + "NONE": 72.1311475409836 + } + } + ], + "accuracy": 0.5829694323144105, + "macro_f1": 0.5641600969180139, + "weighted_f1": 0.5681749152799429, + "macro_precision": 0.6063838476263586, + "macro_recall": 0.5844317120787709, + "per_class_f1": { + "FAVOR": 0.46950092421441775, + "AGAINST": 0.5642857142857143, + "NONE": 0.6586936522539099 + }, + "accuracy_pct": 58.29694323144105, + "macro_f1_pct": 56.41600969180139, + "weighted_f1_pct": 56.81749152799429, + "macro_precision_pct": 60.63838476263585, + "macro_recall_pct": 58.443171207877086, + "per_class_f1_pct": { + "FAVOR": 46.950092421441774, + "AGAINST": 56.42857142857143, + "NONE": 65.86936522539098 + } + }, + { + "after_task": 9, + "best_epoch": 4, + "best_dev_macro_f1": 0.7328160211949003, + "losses": { + "current_ce": 0.5340694459020228, + "total": 0.5340694459020228, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.23972602739726026, + "macro_f1": 0.17617797317375994, + "weighted_f1": 0.14505014968369437, + "macro_precision": 0.2925407925407925, + "macro_recall": 0.3647826086956522, + "per_class_f1": { + "FAVOR": 0.06666666666666667, + "AGAINST": 0.0970873786407767, + "NONE": 0.36477987421383645 + }, + "accuracy_pct": 23.972602739726025, + "macro_f1_pct": 17.617797317375995, + "weighted_f1_pct": 14.505014968369437, + "macro_precision_pct": 29.25407925407925, + "macro_recall_pct": 36.47826086956522, + "per_class_f1_pct": { + "FAVOR": 6.666666666666667, + "AGAINST": 9.70873786407767, + "NONE": 36.477987421383645 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.43157894736842106, + "macro_f1": 0.41787906257036783, + "weighted_f1": 0.4272308958797451, + "macro_precision": 0.5484258166076348, + "macro_recall": 0.5279593318809005, + "per_class_f1": { + "FAVOR": 0.35135135135135137, + "AGAINST": 0.46357615894039733, + "NONE": 0.43870967741935485 + }, + "accuracy_pct": 43.15789473684211, + "macro_f1_pct": 41.78790625703678, + "weighted_f1_pct": 42.72308958797451, + "macro_precision_pct": 54.84258166076348, + "macro_recall_pct": 52.795933188090046, + "per_class_f1_pct": { + "FAVOR": 35.13513513513514, + "AGAINST": 46.35761589403973, + "NONE": 43.87096774193549 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.45685279187817257, + "macro_f1": 0.4497613277520746, + "weighted_f1": 0.3880606095172912, + "macro_precision": 0.5437356760886173, + "macro_recall": 0.6021169529758806, + "per_class_f1": { + "FAVOR": 0.4943820224719101, + "AGAINST": 0.26666666666666666, + "NONE": 0.5882352941176471 + }, + "accuracy_pct": 45.68527918781726, + "macro_f1_pct": 44.97613277520746, + "weighted_f1_pct": 38.80606095172912, + "macro_precision_pct": 54.37356760886173, + "macro_recall_pct": 60.211695297588065, + "per_class_f1_pct": { + "FAVOR": 49.43820224719101, + "AGAINST": 26.666666666666668, + "NONE": 58.82352941176471 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.605, + "macro_f1": 0.608885523134481, + "weighted_f1": 0.5934994209829213, + "macro_precision": 0.6468870827744022, + "macro_recall": 0.6161567028775701, + "per_class_f1": { + "FAVOR": 0.5039370078740157, + "AGAINST": 0.6391752577319587, + "NONE": 0.6835443037974683 + }, + "accuracy_pct": 60.5, + "macro_f1_pct": 60.888552313448095, + "weighted_f1_pct": 59.34994209829213, + "macro_precision_pct": 64.68870827744023, + "macro_recall_pct": 61.615670287757005, + "per_class_f1_pct": { + "FAVOR": 50.39370078740157, + "AGAINST": 63.91752577319587, + "NONE": 68.35443037974683 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.695, + "macro_f1": 0.6127814088598402, + "weighted_f1": 0.6625642701525055, + "macro_precision": 0.7115890083632018, + "macro_recall": 0.5864327095711404, + "per_class_f1": { + "FAVOR": 0.6222222222222222, + "AGAINST": 0.4235294117647059, + "NONE": 0.7925925925925926 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 61.27814088598402, + "weighted_f1_pct": 66.25642701525055, + "macro_precision_pct": 71.15890083632019, + "macro_recall_pct": 58.643270957114034, + "per_class_f1_pct": { + "FAVOR": 62.22222222222222, + "AGAINST": 42.35294117647059, + "NONE": 79.25925925925927 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.46017699115044247, + "macro_f1": 0.3804266072092159, + "weighted_f1": 0.3899808407572933, + "macro_precision": 0.5133477633477633, + "macro_recall": 0.5668244145128019, + "per_class_f1": { + "FAVOR": 0.2564102564102564, + "AGAINST": 0.2608695652173913, + "NONE": 0.624 + }, + "accuracy_pct": 46.017699115044245, + "macro_f1_pct": 38.04266072092159, + "weighted_f1_pct": 38.99808407572933, + "macro_precision_pct": 51.334776334776336, + "macro_recall_pct": 56.682441451280184, + "per_class_f1_pct": { + "FAVOR": 25.64102564102564, + "AGAINST": 26.08695652173913, + "NONE": 62.4 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.5698924731182796, + "macro_f1": 0.5228988168094655, + "weighted_f1": 0.5680079559363524, + "macro_precision": 0.6152063862928349, + "macro_recall": 0.5760942760942761, + "per_class_f1": { + "FAVOR": 0.375, + "AGAINST": 0.627906976744186, + "NONE": 0.5657894736842105 + }, + "accuracy_pct": 56.98924731182796, + "macro_f1_pct": 52.28988168094655, + "weighted_f1_pct": 56.80079559363524, + "macro_precision_pct": 61.52063862928349, + "macro_recall_pct": 57.60942760942761, + "per_class_f1_pct": { + "FAVOR": 37.5, + "AGAINST": 62.7906976744186, + "NONE": 56.57894736842105 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.5849042897097817, + "weighted_f1": 0.5496978821824924, + "macro_precision": 0.6521317829457365, + "macro_recall": 0.6649572649572649, + "per_class_f1": { + "FAVOR": 0.7142857142857143, + "AGAINST": 0.3157894736842105, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 58.49042897097817, + "weighted_f1_pct": 54.969788218249235, + "macro_precision_pct": 65.21317829457365, + "macro_recall_pct": 66.49572649572649, + "per_class_f1_pct": { + "FAVOR": 71.42857142857143, + "AGAINST": 31.57894736842105, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7094223046030276, + "weighted_f1": 0.715, + "macro_precision": 0.7094223046030276, + "macro_recall": 0.7094223046030276, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.676923076923077, + "NONE": 0.7590361445783133 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.94223046030275, + "weighted_f1_pct": 71.5, + "macro_precision_pct": 70.94223046030275, + "macro_recall_pct": 70.94223046030275, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 67.6923076923077, + "NONE": 75.90361445783132 + } + } + ], + "accuracy": 0.5432020330368488, + "macro_f1": 0.5273308678061798, + "weighted_f1": 0.5226299464400128, + "macro_precision": 0.5805833458935049, + "macro_recall": 0.5572547882322926, + "per_class_f1": { + "FAVOR": 0.48721804511278194, + "AGAINST": 0.4695977549111319, + "NONE": 0.6251768033946252 + }, + "accuracy_pct": 54.32020330368488, + "macro_f1_pct": 52.733086780617974, + "weighted_f1_pct": 52.26299464400128, + "macro_precision_pct": 58.058334589350494, + "macro_recall_pct": 55.72547882322926, + "per_class_f1_pct": { + "FAVOR": 48.721804511278194, + "AGAINST": 46.95977549111319, + "NONE": 62.51768033946252 + } + }, + { + "after_task": 10, + "best_epoch": 4, + "best_dev_macro_f1": 0.6480304456999372, + "losses": { + "current_ce": 0.629376731851162, + "total": 0.629376731851162, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 0, + "samples_per_task": 64, + "label_counts": {}, + "task_label_counts": {} + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.2054794520547945, + "macro_f1": 0.13820794590025356, + "weighted_f1": 0.08039984251786149, + "macro_precision": 0.18015873015873016, + "macro_recall": 0.3466666666666667, + "per_class_f1": { + "FAVOR": 0.07142857142857142, + "AGAINST": 0.0, + "NONE": 0.3431952662721893 + }, + "accuracy_pct": 20.54794520547945, + "macro_f1_pct": 13.820794590025356, + "weighted_f1_pct": 8.03998425178615, + "macro_precision_pct": 18.015873015873016, + "macro_recall_pct": 34.66666666666667, + "per_class_f1_pct": { + "FAVOR": 7.142857142857142, + "AGAINST": 0.0, + "NONE": 34.319526627218934 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.3736842105263158, + "macro_f1": 0.3419903257931427, + "weighted_f1": 0.35347372913718356, + "macro_precision": 0.4618641349669639, + "macro_recall": 0.4745824255628177, + "per_class_f1": { + "FAVOR": 0.19718309859154928, + "AGAINST": 0.4166666666666667, + "NONE": 0.4121212121212121 + }, + "accuracy_pct": 37.368421052631575, + "macro_f1_pct": 34.19903257931427, + "weighted_f1_pct": 35.34737291371835, + "macro_precision_pct": 46.18641349669639, + "macro_recall_pct": 47.458242556281775, + "per_class_f1_pct": { + "FAVOR": 19.718309859154928, + "AGAINST": 41.66666666666667, + "NONE": 41.21212121212121 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.40609137055837563, + "macro_f1": 0.3781268706306187, + "weighted_f1": 0.26777977386048984, + "macro_precision": 0.591991341991342, + "macro_recall": 0.6083742158703117, + "per_class_f1": { + "FAVOR": 0.46153846153846156, + "AGAINST": 0.05172413793103448, + "NONE": 0.6211180124223602 + }, + "accuracy_pct": 40.609137055837564, + "macro_f1_pct": 37.81268706306187, + "weighted_f1_pct": 26.777977386048985, + "macro_precision_pct": 59.19913419913419, + "macro_recall_pct": 60.83742158703117, + "per_class_f1_pct": { + "FAVOR": 46.15384615384615, + "AGAINST": 5.172413793103448, + "NONE": 62.11180124223602 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.643914676701562, + "weighted_f1": 0.6248105446466102, + "macro_precision": 0.6555882352941177, + "macro_recall": 0.6645202634362525, + "per_class_f1": { + "FAVOR": 0.644808743169399, + "AGAINST": 0.5396825396825397, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 64.3914676701562, + "weighted_f1_pct": 62.48105446466102, + "macro_precision_pct": 65.55882352941177, + "macro_recall_pct": 66.45202634362525, + "per_class_f1_pct": { + "FAVOR": 64.48087431693989, + "AGAINST": 53.96825396825397, + "NONE": 74.72527472527473 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.695, + "macro_f1": 0.5912012393947482, + "weighted_f1": 0.6397486535713724, + "macro_precision": 0.7262277951933124, + "macro_recall": 0.658115386751069, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.2898550724637681, + "NONE": 0.8076923076923077 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 59.12012393947482, + "weighted_f1_pct": 63.97486535713725, + "macro_precision_pct": 72.62277951933123, + "macro_recall_pct": 65.8115386751069, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 28.985507246376812, + "NONE": 80.76923076923077 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.5486725663716814, + "macro_f1": 0.41830089437383594, + "weighted_f1": 0.5162513563399008, + "macro_precision": 0.5050879161990273, + "macro_recall": 0.499502487562189, + "per_class_f1": { + "FAVOR": 0.449438202247191, + "AGAINST": 0.13333333333333333, + "NONE": 0.6721311475409836 + }, + "accuracy_pct": 54.86725663716814, + "macro_f1_pct": 41.83008943738359, + "weighted_f1_pct": 51.625135633990084, + "macro_precision_pct": 50.50879161990272, + "macro_recall_pct": 49.950248756218905, + "per_class_f1_pct": { + "FAVOR": 44.9438202247191, + "AGAINST": 13.333333333333334, + "NONE": 67.21311475409836 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.553763440860215, + "macro_f1": 0.542305087810031, + "weighted_f1": 0.543518645860433, + "macro_precision": 0.5721607378129118, + "macro_recall": 0.62996632996633, + "per_class_f1": { + "FAVOR": 0.4675324675324675, + "AGAINST": 0.5316455696202531, + "NONE": 0.6277372262773723 + }, + "accuracy_pct": 55.376344086021504, + "macro_f1_pct": 54.230508781003095, + "weighted_f1_pct": 54.3518645860433, + "macro_precision_pct": 57.21607378129118, + "macro_recall_pct": 62.996632996633, + "per_class_f1_pct": { + "FAVOR": 46.75324675324675, + "AGAINST": 53.16455696202531, + "NONE": 62.77372262773723 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5070422535211268, + "macro_f1": 0.4287938854959106, + "weighted_f1": 0.379958740981192, + "macro_precision": 0.6733498349834983, + "macro_recall": 0.5594017094017094, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.03278688524590164, + "NONE": 0.6535947712418301 + }, + "accuracy_pct": 50.70422535211267, + "macro_f1_pct": 42.879388549591056, + "weighted_f1_pct": 37.995874098119195, + "macro_precision_pct": 67.33498349834983, + "macro_recall_pct": 55.940170940170944, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 3.278688524590164, + "NONE": 65.359477124183 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.65, + "macro_f1": 0.6059476461616938, + "weighted_f1": 0.6207997609393086, + "macro_precision": 0.7269206502624407, + "macro_recall": 0.646339202965709, + "per_class_f1": { + "FAVOR": 0.624113475177305, + "AGAINST": 0.40963855421686746, + "NONE": 0.7840909090909091 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 60.594764616169385, + "weighted_f1_pct": 62.07997609393085, + "macro_precision_pct": 72.69206502624407, + "macro_recall_pct": 64.6339202965709, + "per_class_f1_pct": { + "FAVOR": 62.4113475177305, + "AGAINST": 40.963855421686745, + "NONE": 78.4090909090909 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.695, + "macro_f1": 0.6721045109617051, + "weighted_f1": 0.6981399158468471, + "macro_precision": 0.6667176163691204, + "macro_recall": 0.6803401560683114, + "per_class_f1": { + "FAVOR": 0.7309644670050761, + "AGAINST": 0.4888888888888889, + "NONE": 0.7964601769911505 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 67.21045109617052, + "weighted_f1_pct": 69.81399158468471, + "macro_precision_pct": 66.67176163691204, + "macro_recall_pct": 68.03401560683115, + "per_class_f1_pct": { + "FAVOR": 73.09644670050761, + "AGAINST": 48.888888888888886, + "NONE": 79.64601769911505 + } + } + ], + "accuracy": 0.5372040586245772, + "macro_f1": 0.514031119507793, + "weighted_f1": 0.49501382005080913, + "macro_precision": 0.5756201470487184, + "macro_recall": 0.5724163579551608, + "per_class_f1": { + "FAVOR": 0.5613026819923371, + "AGAINST": 0.3343782654127482, + "NONE": 0.6464124111182935 + }, + "accuracy_pct": 53.720405862457724, + "macro_f1_pct": 51.4031119507793, + "weighted_f1_pct": 49.50138200508091, + "macro_precision_pct": 57.56201470487184, + "macro_recall_pct": 57.241635795516075, + "per_class_f1_pct": { + "FAVOR": 56.13026819923371, + "AGAINST": 33.437826541274816, + "NONE": 64.64124111182934 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s4_lora__x__s3__seed13/resolved_config.yaml b/outputs/trace_s4_lora__x__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f50e3ad7a3163eee286cbfaf41e8dcd3e56173fa --- /dev/null +++ b/outputs/trace_s4_lora__x__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s4_lora + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 4 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s4_lora__x__s3__seed13/run.log b/outputs/trace_s4_lora__x__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..12a1178807e2a14d260494160474a6d8add032d5 --- /dev/null +++ b/outputs/trace_s4_lora__x__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 01:44:17,869 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s4_lora__x__s3__seed13 device=cuda +2026-09-02 01:54:30,637 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.47608930832235014, 'average_forgetting': 0.23169045893874543, 'backward_transfer': -0.22780465810831987, 'final_mean_target_macro_f1_pct': 47.60893083223501, 'average_forgetting_pct': 23.169045893874543, 'backward_transfer_pct': -22.780465810831988, 'final_pooled_accuracy': 0.5372040586245772, 'final_pooled_accuracy_pct': 53.720405862457724, 'final_pooled_macro_f1': 0.514031119507793, 'final_pooled_macro_f1_pct': 51.4031119507793, 'final_pooled_weighted_f1': 0.49501382005080913, 'final_pooled_weighted_f1_pct': 49.50138200508091, 'final_pooled_macro_precision': 0.5756201470487184, 'final_pooled_macro_precision_pct': 57.56201470487184, 'final_pooled_macro_recall': 0.5724163579551608, 'final_pooled_macro_recall_pct': 57.241635795516075, 'final_pooled_per_class_f1': {'FAVOR': 0.5613026819923371, 'AGAINST': 0.3343782654127482, 'NONE': 0.6464124111182935}, 'final_pooled_per_class_f1_pct': {'FAVOR': 56.13026819923371, 'AGAINST': 33.437826541274816, 'NONE': 64.64124111182934}} +2026-09-02 01:54:30,637 [INFO] __main__: efficiency={'training_wall_time_seconds': 611.4533863409888, 'training_wall_time_minutes': 10.190889772349813, 'peak_gpu_memory_allocated_gb': 5.56232213973999, 'peak_gpu_memory_reserved_gb': 6.103515625} diff --git a/outputs/trace_s5_balanced_replay__doc__s1__seed13/environment.json b/outputs/trace_s5_balanced_replay__doc__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..38b6fc23c0272bee48ee7403cbbd68e2b191e5bb --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T01:54:38.814364+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.json b/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..4aad1f5da803cc28c06ca1feb901c84ff4e6f499 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.json @@ -0,0 +1,1587 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.6155173412078051, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5694161861580516, + 0.6956361701518025, + null, + null, + null, + null, + null, + null + ], + [ + 0.5292705369293579, + 0.6688695179351244, + 0.569517700185404, + null, + null, + null, + null, + null + ], + [ + 0.5217190618274629, + 0.613353638755958, + 0.5125331533353226, + 0.5730053319396435, + null, + null, + null, + null + ], + [ + 0.5517885838908566, + 0.6508294594071905, + 0.4968656936302855, + 0.5570733785888508, + 0.6546257001019442, + null, + null, + null + ], + [ + 0.5185760867396362, + 0.6351920141919599, + 0.4997972354580364, + 0.5446644351512296, + 0.634896023500377, + 0.6691383965626724, + null, + null + ], + [ + 0.528340831926188, + 0.6726239875674654, + 0.5287480209475094, + 0.5872269180724953, + 0.6264762020792722, + 0.6814555771644994, + 0.670982707354219, + null + ], + [ + 0.5433127677439283, + 0.6661248665952351, + 0.5425058409447767, + 0.5840808825115622, + 0.6298615112135373, + 0.6833801657343669, + 0.6519210169391337, + 0.707316090361901 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 13, + "best_dev_macro_f1": 0.6697781414802692, + "losses": { + "current_ce": 0.6033996852018215, + "total": 0.6033996852018215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6899618805590851, + "macro_f1": 0.6155173412078051, + "weighted_f1": 0.7021062475404232, + "macro_precision": 0.6083931545026292, + "macro_recall": 0.6320453995290597, + "per_class_f1": { + "FAVOR": 0.512280701754386, + "AGAINST": 0.5157894736842106, + "NONE": 0.8184818481848185 + }, + "accuracy_pct": 68.99618805590852, + "macro_f1_pct": 61.55173412078051, + "weighted_f1_pct": 70.21062475404231, + "macro_precision_pct": 60.83931545026292, + "macro_recall_pct": 63.20453995290597, + "per_class_f1_pct": { + "FAVOR": 51.2280701754386, + "AGAINST": 51.578947368421055, + "NONE": 81.84818481848185 + } + } + ], + "accuracy": 0.6899618805590851, + "macro_f1": 0.6155173412078051, + "weighted_f1": 0.7021062475404232, + "macro_precision": 0.6083931545026292, + "macro_recall": 0.6320453995290597, + "per_class_f1": { + "FAVOR": 0.512280701754386, + "AGAINST": 0.5157894736842106, + "NONE": 0.8184818481848185 + }, + "accuracy_pct": 68.99618805590852, + "macro_f1_pct": 61.55173412078051, + "weighted_f1_pct": 70.21062475404231, + "macro_precision_pct": 60.83931545026292, + "macro_recall_pct": 63.20453995290597, + "per_class_f1_pct": { + "FAVOR": 51.2280701754386, + "AGAINST": 51.578947368421055, + "NONE": 81.84818481848185 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7311337628642974, + "losses": { + "current_ce": 0.5307365087951933, + "replay_ce": 0.02806483572348952, + "total": 0.5588013440086728, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6531130876747141, + "macro_f1": 0.5694161861580516, + "weighted_f1": 0.6618331942814528, + "macro_precision": 0.5650006999308553, + "macro_recall": 0.5910686384542594, + "per_class_f1": { + "FAVOR": 0.5029239766081871, + "AGAINST": 0.41533546325878595, + "NONE": 0.7899891186071817 + }, + "accuracy_pct": 65.31130876747142, + "macro_f1_pct": 56.941618615805154, + "weighted_f1_pct": 66.18331942814528, + "macro_precision_pct": 56.50006999308553, + "macro_recall_pct": 59.10686384542594, + "per_class_f1_pct": { + "FAVOR": 50.29239766081871, + "AGAINST": 41.533546325878596, + "NONE": 78.99891186071817 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7110016420361248, + "macro_f1": 0.6956361701518025, + "weighted_f1": 0.7188045338776238, + "macro_precision": 0.6999122917984398, + "macro_recall": 0.7098670619031916, + "per_class_f1": { + "FAVOR": 0.6127167630057804, + "AGAINST": 0.6873065015479877, + "NONE": 0.7868852459016393 + }, + "accuracy_pct": 71.10016420361248, + "macro_f1_pct": 69.56361701518024, + "weighted_f1_pct": 71.88045338776237, + "macro_precision_pct": 69.99122917984398, + "macro_recall_pct": 70.98670619031915, + "per_class_f1_pct": { + "FAVOR": 61.27167630057804, + "AGAINST": 68.73065015479877, + "NONE": 78.68852459016394 + } + } + ], + "accuracy": 0.6783667621776505, + "macro_f1": 0.6334756640850786, + "weighted_f1": 0.6867442139840507, + "macro_precision": 0.6322937896760786, + "macro_recall": 0.6522518761257706, + "per_class_f1": { + "FAVOR": 0.5581395348837209, + "AGAINST": 0.5534591194968553, + "NONE": 0.7888283378746594 + }, + "accuracy_pct": 67.83667621776505, + "macro_f1_pct": 63.34756640850786, + "weighted_f1_pct": 68.67442139840507, + "macro_precision_pct": 63.229378967607865, + "macro_recall_pct": 65.22518761257706, + "per_class_f1_pct": { + "FAVOR": 55.81395348837209, + "AGAINST": 55.34591194968554, + "NONE": 78.88283378746594 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.6837498989719051, + "losses": { + "current_ce": 0.546360798870095, + "replay_ce": 0.017297741932386073, + "total": 0.5636585399234992, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6302414231257941, + "macro_f1": 0.5292705369293579, + "weighted_f1": 0.6329911674399076, + "macro_precision": 0.5269146331751605, + "macro_recall": 0.533691673452022, + "per_class_f1": { + "FAVOR": 0.42033898305084744, + "AGAINST": 0.39375, + "NONE": 0.7737226277372263 + }, + "accuracy_pct": 63.024142312579414, + "macro_f1_pct": 52.927053692935786, + "weighted_f1_pct": 63.29911674399076, + "macro_precision_pct": 52.69146331751605, + "macro_recall_pct": 53.36916734520219, + "per_class_f1_pct": { + "FAVOR": 42.03389830508474, + "AGAINST": 39.375, + "NONE": 77.37226277372264 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6765188834154351, + "macro_f1": 0.6688695179351244, + "weighted_f1": 0.6832681716948531, + "macro_precision": 0.6755026482606038, + "macro_recall": 0.6954341246288583, + "per_class_f1": { + "FAVOR": 0.6276923076923077, + "AGAINST": 0.6512820512820513, + "NONE": 0.7276341948310139 + }, + "accuracy_pct": 67.65188834154351, + "macro_f1_pct": 66.88695179351244, + "weighted_f1_pct": 68.32681716948531, + "macro_precision_pct": 67.55026482606038, + "macro_recall_pct": 69.54341246288584, + "per_class_f1_pct": { + "FAVOR": 62.76923076923077, + "AGAINST": 65.12820512820512, + "NONE": 72.76341948310139 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6101231190150479, + "macro_f1": 0.569517700185404, + "weighted_f1": 0.6234517578663603, + "macro_precision": 0.5720758270427145, + "macro_recall": 0.5939638545932486, + "per_class_f1": { + "FAVOR": 0.43089430894308944, + "AGAINST": 0.5767790262172284, + "NONE": 0.7008797653958945 + }, + "accuracy_pct": 61.01231190150479, + "macro_f1_pct": 56.9517700185404, + "weighted_f1_pct": 62.345175786636034, + "macro_precision_pct": 57.20758270427145, + "macro_recall_pct": 59.39638545932486, + "per_class_f1_pct": { + "FAVOR": 43.08943089430895, + "AGAINST": 57.67790262172284, + "NONE": 70.08797653958945 + } + } + ], + "accuracy": 0.6365773389750823, + "macro_f1": 0.5979827342535521, + "weighted_f1": 0.6474129609335886, + "macro_precision": 0.5925061339263092, + "macro_recall": 0.6166227411947646, + "per_class_f1": { + "FAVOR": 0.5011547344110855, + "AGAINST": 0.5530546623794212, + "NONE": 0.7397388059701493 + }, + "accuracy_pct": 63.65773389750823, + "macro_f1_pct": 59.798273425355205, + "weighted_f1_pct": 64.74129609335886, + "macro_precision_pct": 59.250613392630925, + "macro_recall_pct": 61.66227411947646, + "per_class_f1_pct": { + "FAVOR": 50.11547344110855, + "AGAINST": 55.30546623794213, + "NONE": 73.97388059701493 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.685362115437692, + "losses": { + "current_ce": 0.5138460865930508, + "replay_ce": 0.020093058473675658, + "total": 0.533939144407448, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5959339263024143, + "macro_f1": 0.5217190618274629, + "weighted_f1": 0.6157188732089818, + "macro_precision": 0.5309637871422258, + "macro_recall": 0.5583553398912876, + "per_class_f1": { + "FAVOR": 0.44938271604938274, + "AGAINST": 0.37012987012987014, + "NONE": 0.7456445993031359 + }, + "accuracy_pct": 59.593392630241425, + "macro_f1_pct": 52.171906182746284, + "weighted_f1_pct": 61.57188732089818, + "macro_precision_pct": 53.09637871422258, + "macro_recall_pct": 55.83553398912876, + "per_class_f1_pct": { + "FAVOR": 44.93827160493827, + "AGAINST": 37.01298701298701, + "NONE": 74.56445993031359 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6190476190476191, + "macro_f1": 0.613353638755958, + "weighted_f1": 0.6272319375336906, + "macro_precision": 0.6348633536898273, + "macro_recall": 0.6450364206794089, + "per_class_f1": { + "FAVOR": 0.563953488372093, + "AGAINST": 0.6080402010050251, + "NONE": 0.6680672268907563 + }, + "accuracy_pct": 61.904761904761905, + "macro_f1_pct": 61.3353638755958, + "weighted_f1_pct": 62.72319375336905, + "macro_precision_pct": 63.48633536898273, + "macro_recall_pct": 64.50364206794089, + "per_class_f1_pct": { + "FAVOR": 56.395348837209305, + "AGAINST": 60.80402010050251, + "NONE": 66.80672268907563 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5417236662106704, + "macro_f1": 0.5125331533353226, + "weighted_f1": 0.5654020911298463, + "macro_precision": 0.5357188508210813, + "macro_recall": 0.5549233265587468, + "per_class_f1": { + "FAVOR": 0.3569230769230769, + "AGAINST": 0.5548902195608783, + "NONE": 0.6257861635220126 + }, + "accuracy_pct": 54.172366621067034, + "macro_f1_pct": 51.25331533353226, + "weighted_f1_pct": 56.54020911298463, + "macro_precision_pct": 53.571885082108125, + "macro_recall_pct": 55.49233265587468, + "per_class_f1_pct": { + "FAVOR": 35.69230769230769, + "AGAINST": 55.489021956087825, + "NONE": 62.57861635220125 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5730053319396435, + "weighted_f1": 0.6249112125865773, + "macro_precision": 0.5923607358033588, + "macro_recall": 0.6094249923919034, + "per_class_f1": { + "FAVOR": 0.5317073170731708, + "AGAINST": 0.46835443037974683, + "NONE": 0.7189542483660131 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 57.300533193964355, + "weighted_f1_pct": 62.491121258657735, + "macro_precision_pct": 59.236073580335876, + "macro_recall_pct": 60.94249923919034, + "per_class_f1_pct": { + "FAVOR": 53.170731707317074, + "AGAINST": 46.835443037974684, + "NONE": 71.89542483660131 + } + } + ], + "accuracy": 0.5883404864091559, + "macro_f1": 0.5624827302356098, + "weighted_f1": 0.6082392850633415, + "macro_precision": 0.5752730310625047, + "macro_recall": 0.5984835689256194, + "per_class_f1": { + "FAVOR": 0.4784366576819407, + "AGAINST": 0.5134602757715037, + "NONE": 0.695551257253385 + }, + "accuracy_pct": 58.83404864091559, + "macro_f1_pct": 56.24827302356098, + "weighted_f1_pct": 60.82392850633415, + "macro_precision_pct": 57.52730310625047, + "macro_recall_pct": 59.84835689256194, + "per_class_f1_pct": { + "FAVOR": 47.84366576819407, + "AGAINST": 51.346027577150366, + "NONE": 69.5551257253385 + } + }, + { + "after_task": 5, + "best_epoch": 2, + "best_dev_macro_f1": 0.6720908926179384, + "losses": { + "current_ce": 0.4921029800283057, + "replay_ce": 0.029563177262233304, + "total": 0.521666157653644, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.5517885838908566, + "weighted_f1": 0.6479521895550832, + "macro_precision": 0.5522860929049819, + "macro_recall": 0.5761337045106109, + "per_class_f1": { + "FAVOR": 0.4573002754820937, + "AGAINST": 0.41904761904761906, + "NONE": 0.7790178571428571 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 55.17885838908566, + "weighted_f1_pct": 64.79521895550832, + "macro_precision_pct": 55.228609290498184, + "macro_recall_pct": 57.61337045106109, + "per_class_f1_pct": { + "FAVOR": 45.73002754820937, + "AGAINST": 41.904761904761905, + "NONE": 77.90178571428571 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6633825944170771, + "macro_f1": 0.6508294594071905, + "weighted_f1": 0.6701214266900546, + "macro_precision": 0.6551864165674209, + "macro_recall": 0.669622528930551, + "per_class_f1": { + "FAVOR": 0.5894039735099338, + "AGAINST": 0.6347607052896725, + "NONE": 0.7283236994219653 + }, + "accuracy_pct": 66.33825944170772, + "macro_f1_pct": 65.08294594071906, + "weighted_f1_pct": 67.01214266900546, + "macro_precision_pct": 65.51864165674209, + "macro_recall_pct": 66.9622528930551, + "per_class_f1_pct": { + "FAVOR": 58.94039735099338, + "AGAINST": 63.476070528967256, + "NONE": 72.83236994219652 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5143638850889193, + "macro_f1": 0.4968656936302855, + "weighted_f1": 0.5432809051046937, + "macro_precision": 0.5420035501451835, + "macro_recall": 0.5585061923779701, + "per_class_f1": { + "FAVOR": 0.36787564766839376, + "AGAINST": 0.5204301075268817, + "NONE": 0.602291325695581 + }, + "accuracy_pct": 51.43638850889193, + "macro_f1_pct": 49.686569363028546, + "weighted_f1_pct": 54.32809051046937, + "macro_precision_pct": 54.20035501451835, + "macro_recall_pct": 55.850619237797005, + "per_class_f1_pct": { + "FAVOR": 36.787564766839374, + "AGAINST": 52.04301075268817, + "NONE": 60.2291325695581 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.585949177877429, + "macro_f1": 0.5570733785888508, + "weighted_f1": 0.6081637701000742, + "macro_precision": 0.5915731300609096, + "macro_recall": 0.5957592057969584, + "per_class_f1": { + "FAVOR": 0.5219298245614035, + "AGAINST": 0.44912280701754387, + "NONE": 0.7001675041876047 + }, + "accuracy_pct": 58.594917787742894, + "macro_f1_pct": 55.707337858885076, + "weighted_f1_pct": 60.816377010007415, + "macro_precision_pct": 59.15731300609096, + "macro_recall_pct": 59.57592057969584, + "per_class_f1_pct": { + "FAVOR": 52.19298245614035, + "AGAINST": 44.91228070175438, + "NONE": 70.01675041876047 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6881287726358148, + "macro_f1": 0.6546257001019442, + "weighted_f1": 0.6956462080302849, + "macro_precision": 0.6518476621417798, + "macro_recall": 0.6628828198467108, + "per_class_f1": { + "FAVOR": 0.5748031496062992, + "AGAINST": 0.5789473684210527, + "NONE": 0.810126582278481 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 65.46257001019443, + "weighted_f1_pct": 69.56462080302849, + "macro_precision_pct": 65.18476621417798, + "macro_recall_pct": 66.28828198467109, + "per_class_f1_pct": { + "FAVOR": 57.48031496062992, + "AGAINST": 57.89473684210527, + "NONE": 81.0126582278481 + } + } + ], + "accuracy": 0.6109930154873975, + "macro_f1": 0.5815727981752827, + "weighted_f1": 0.6291099409562191, + "macro_precision": 0.5900546965334866, + "macro_recall": 0.6109428349334495, + "per_class_f1": { + "FAVOR": 0.4940374787052811, + "AGAINST": 0.5254629629629629, + "NONE": 0.7252179528576042 + }, + "accuracy_pct": 61.099301548739746, + "macro_f1_pct": 58.15727981752828, + "weighted_f1_pct": 62.91099409562191, + "macro_precision_pct": 59.00546965334866, + "macro_recall_pct": 61.094283493344946, + "per_class_f1_pct": { + "FAVOR": 49.40374787052811, + "AGAINST": 52.54629629629629, + "NONE": 72.52179528576042 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7184062850729518, + "losses": { + "current_ce": 0.4117540765198923, + "replay_ce": 0.04057449822846268, + "total": 0.452328574799356, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5933926302414231, + "macro_f1": 0.5185760867396362, + "weighted_f1": 0.6162278908487981, + "macro_precision": 0.5207216559626828, + "macro_recall": 0.5434194890077243, + "per_class_f1": { + "FAVOR": 0.410958904109589, + "AGAINST": 0.3965014577259475, + "NONE": 0.7482678983833718 + }, + "accuracy_pct": 59.339263024142305, + "macro_f1_pct": 51.85760867396362, + "weighted_f1_pct": 61.62278908487981, + "macro_precision_pct": 52.072165596268285, + "macro_recall_pct": 54.34194890077243, + "per_class_f1_pct": { + "FAVOR": 41.0958904109589, + "AGAINST": 39.650145772594755, + "NONE": 74.82678983833718 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6403940886699507, + "macro_f1": 0.6351920141919599, + "weighted_f1": 0.6462805337939896, + "macro_precision": 0.6468095399701064, + "macro_recall": 0.6661981364216514, + "per_class_f1": { + "FAVOR": 0.5958702064896755, + "AGAINST": 0.6307692307692307, + "NONE": 0.6789366053169734 + }, + "accuracy_pct": 64.03940886699507, + "macro_f1_pct": 63.519201419195994, + "weighted_f1_pct": 64.62805337939896, + "macro_precision_pct": 64.68095399701063, + "macro_recall_pct": 66.61981364216514, + "per_class_f1_pct": { + "FAVOR": 59.587020648967545, + "AGAINST": 63.07692307692307, + "NONE": 67.89366053169734 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5225718194254446, + "macro_f1": 0.4997972354580364, + "weighted_f1": 0.5467535084905882, + "macro_precision": 0.5342033218593144, + "macro_recall": 0.550931702923668, + "per_class_f1": { + "FAVOR": 0.36151603498542273, + "AGAINST": 0.5375494071146245, + "NONE": 0.600326264274062 + }, + "accuracy_pct": 52.25718194254446, + "macro_f1_pct": 49.97972354580364, + "weighted_f1_pct": 54.67535084905882, + "macro_precision_pct": 53.42033218593144, + "macro_recall_pct": 55.0931702923668, + "per_class_f1_pct": { + "FAVOR": 36.15160349854227, + "AGAINST": 53.75494071146245, + "NONE": 60.0326264274062 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.57847533632287, + "macro_f1": 0.5446644351512296, + "weighted_f1": 0.6041335106943165, + "macro_precision": 0.5625404272346487, + "macro_recall": 0.5728174162551244, + "per_class_f1": { + "FAVOR": 0.467005076142132, + "AGAINST": 0.4520123839009288, + "NONE": 0.714975845410628 + }, + "accuracy_pct": 57.847533632286996, + "macro_f1_pct": 54.46644351512296, + "weighted_f1_pct": 60.413351069431656, + "macro_precision_pct": 56.25404272346487, + "macro_recall_pct": 57.28174162551244, + "per_class_f1_pct": { + "FAVOR": 46.7005076142132, + "AGAINST": 45.20123839009288, + "NONE": 71.49758454106279 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6639839034205232, + "macro_f1": 0.634896023500377, + "weighted_f1": 0.671621705336156, + "macro_precision": 0.6337234625111974, + "macro_recall": 0.6446684269086628, + "per_class_f1": { + "FAVOR": 0.5537190082644629, + "AGAINST": 0.5774647887323944, + "NONE": 0.7735042735042735 + }, + "accuracy_pct": 66.39839034205231, + "macro_f1_pct": 63.4896023500377, + "weighted_f1_pct": 67.1621705336156, + "macro_precision_pct": 63.37234625111974, + "macro_recall_pct": 64.46684269086627, + "per_class_f1_pct": { + "FAVOR": 55.371900826446286, + "AGAINST": 57.74647887323944, + "NONE": 77.35042735042735 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6941649899396378, + "macro_f1": 0.6691383965626724, + "weighted_f1": 0.7066846132031092, + "macro_precision": 0.666905200027261, + "macro_recall": 0.6973991232611922, + "per_class_f1": { + "FAVOR": 0.5639097744360902, + "AGAINST": 0.6536964980544747, + "NONE": 0.7898089171974523 + }, + "accuracy_pct": 69.41649899396378, + "macro_f1_pct": 66.91383965626724, + "weighted_f1_pct": 70.66846132031091, + "macro_precision_pct": 66.69052000272609, + "macro_recall_pct": 69.73991232611922, + "per_class_f1_pct": { + "FAVOR": 56.390977443609025, + "AGAINST": 65.36964980544747, + "NONE": 78.98089171974523 + } + } + ], + "accuracy": 0.6071240105540897, + "macro_f1": 0.5795386644369359, + "weighted_f1": 0.6249017135969034, + "macro_precision": 0.5864645229053557, + "macro_recall": 0.6075279521187976, + "per_class_f1": { + "FAVOR": 0.4843509492047204, + "AGAINST": 0.5382786495482644, + "NONE": 0.7159863945578231 + }, + "accuracy_pct": 60.71240105540897, + "macro_f1_pct": 57.95386644369359, + "weighted_f1_pct": 62.490171359690336, + "macro_precision_pct": 58.646452290535564, + "macro_recall_pct": 60.752795211879764, + "per_class_f1_pct": { + "FAVOR": 48.43509492047204, + "AGAINST": 53.827864954826445, + "NONE": 71.5986394557823 + } + }, + { + "after_task": 7, + "best_epoch": 13, + "best_dev_macro_f1": 0.7332904506331527, + "losses": { + "current_ce": 0.36610839086456026, + "replay_ce": 0.017698088319351275, + "total": 0.38380647870462115, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6200762388818297, + "macro_f1": 0.528340831926188, + "weighted_f1": 0.6359720262459907, + "macro_precision": 0.5306727590900252, + "macro_recall": 0.5467173551051546, + "per_class_f1": { + "FAVOR": 0.4134078212290503, + "AGAINST": 0.38977635782747605, + "NONE": 0.7818383167220376 + }, + "accuracy_pct": 62.00762388818297, + "macro_f1_pct": 52.8340831926188, + "weighted_f1_pct": 63.59720262459907, + "macro_precision_pct": 53.06727590900252, + "macro_recall_pct": 54.67173551051546, + "per_class_f1_pct": { + "FAVOR": 41.340782122905026, + "AGAINST": 38.977635782747605, + "NONE": 78.18383167220377 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6847290640394089, + "macro_f1": 0.6726239875674654, + "weighted_f1": 0.6900788317368352, + "macro_precision": 0.6698032506415741, + "macro_recall": 0.6887602089684147, + "per_class_f1": { + "FAVOR": 0.627831715210356, + "AGAINST": 0.6451612903225806, + "NONE": 0.74487895716946 + }, + "accuracy_pct": 68.4729064039409, + "macro_f1_pct": 67.26239875674655, + "weighted_f1_pct": 69.00788317368351, + "macro_precision_pct": 66.98032506415741, + "macro_recall_pct": 68.87602089684147, + "per_class_f1_pct": { + "FAVOR": 62.7831715210356, + "AGAINST": 64.51612903225806, + "NONE": 74.487895716946 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5649794801641587, + "macro_f1": 0.5287480209475094, + "weighted_f1": 0.5879003192307674, + "macro_precision": 0.5409336979982918, + "macro_recall": 0.5675054867388313, + "per_class_f1": { + "FAVOR": 0.38509316770186336, + "AGAINST": 0.5217391304347826, + "NONE": 0.6794117647058824 + }, + "accuracy_pct": 56.49794801641587, + "macro_f1_pct": 52.87480209475094, + "weighted_f1_pct": 58.79003192307673, + "macro_precision_pct": 54.09336979982918, + "macro_recall_pct": 56.75054867388313, + "per_class_f1_pct": { + "FAVOR": 38.50931677018634, + "AGAINST": 52.17391304347826, + "NONE": 67.94117647058823 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6322869955156951, + "macro_f1": 0.5872269180724953, + "weighted_f1": 0.6495787862138902, + "macro_precision": 0.5883134571857637, + "macro_recall": 0.6074539372340838, + "per_class_f1": { + "FAVOR": 0.5173333333333333, + "AGAINST": 0.4797297297297297, + "NONE": 0.7646176911544228 + }, + "accuracy_pct": 63.22869955156951, + "macro_f1_pct": 58.722691807249525, + "weighted_f1_pct": 64.95787862138901, + "macro_precision_pct": 58.831345718576365, + "macro_recall_pct": 60.745393723408384, + "per_class_f1_pct": { + "FAVOR": 51.733333333333334, + "AGAINST": 47.97297297297297, + "NONE": 76.46176911544228 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6264762020792722, + "weighted_f1": 0.6721173287287074, + "macro_precision": 0.6292235801581595, + "macro_recall": 0.6247689372228797, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.546875, + "NONE": 0.7992202729044834 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 62.647620207927226, + "weighted_f1_pct": 67.21173287287074, + "macro_precision_pct": 62.92235801581595, + "macro_recall_pct": 62.476893722287976, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 54.6875, + "NONE": 79.92202729044834 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6814555771644994, + "weighted_f1": 0.7251352466440751, + "macro_precision": 0.6817443297515938, + "macro_recall": 0.7051885218551885, + "per_class_f1": { + "FAVOR": 0.6148409893992933, + "AGAINST": 0.6094420600858369, + "NONE": 0.8200836820083682 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 68.14555771644994, + "weighted_f1_pct": 72.51352466440751, + "macro_precision_pct": 68.17443297515938, + "macro_recall_pct": 70.51885218551885, + "per_class_f1_pct": { + "FAVOR": 61.48409893992933, + "AGAINST": 60.94420600858369, + "NONE": 82.00836820083683 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7168758716875872, + "macro_f1": 0.670982707354219, + "weighted_f1": 0.7270188618515017, + "macro_precision": 0.6596710854829722, + "macro_recall": 0.7008511759517018, + "per_class_f1": { + "FAVOR": 0.5387453874538746, + "AGAINST": 0.6700251889168766, + "NONE": 0.804177545691906 + }, + "accuracy_pct": 71.68758716875871, + "macro_f1_pct": 67.0982707354219, + "weighted_f1_pct": 72.70188618515017, + "macro_precision_pct": 65.96710854829722, + "macro_recall_pct": 70.08511759517017, + "per_class_f1_pct": { + "FAVOR": 53.874538745387454, + "AGAINST": 67.00251889168766, + "NONE": 80.4177545691906 + } + } + ], + "accuracy": 0.6532061238074107, + "macro_f1": 0.6126288751911302, + "weighted_f1": 0.665966090750988, + "macro_precision": 0.6091610960968175, + "macro_recall": 0.6320459588969595, + "per_class_f1": { + "FAVOR": 0.5132991133924405, + "AGAINST": 0.5552213149978513, + "NONE": 0.7693661971830986 + }, + "accuracy_pct": 65.32061238074107, + "macro_f1_pct": 61.26288751911302, + "weighted_f1_pct": 66.5966090750988, + "macro_precision_pct": 60.91610960968175, + "macro_recall_pct": 63.204595889695945, + "per_class_f1_pct": { + "FAVOR": 51.32991133924405, + "AGAINST": 55.52213149978513, + "NONE": 76.93661971830986 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.7875357278487338, + "losses": { + "current_ce": 0.34310466447005084, + "replay_ce": 0.014769438629988215, + "total": 0.35787410408112347, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.5433127677439283, + "weighted_f1": 0.6476966032743156, + "macro_precision": 0.541347418875975, + "macro_recall": 0.5585246069995525, + "per_class_f1": { + "FAVOR": 0.41543026706231456, + "AGAINST": 0.4268292682926829, + "NONE": 0.7876787678767877 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 54.33127677439283, + "weighted_f1_pct": 64.76966032743155, + "macro_precision_pct": 54.13474188759751, + "macro_recall_pct": 55.85246069995525, + "per_class_f1_pct": { + "FAVOR": 41.54302670623146, + "AGAINST": 42.68292682926829, + "NONE": 78.76787678767877 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6748768472906403, + "macro_f1": 0.6661248665952351, + "weighted_f1": 0.6803848335247603, + "macro_precision": 0.6649800254315382, + "macro_recall": 0.6880914808808257, + "per_class_f1": { + "FAVOR": 0.6227544910179641, + "AGAINST": 0.6518105849582173, + "NONE": 0.7238095238095238 + }, + "accuracy_pct": 67.48768472906403, + "macro_f1_pct": 66.61248665952351, + "weighted_f1_pct": 68.03848335247604, + "macro_precision_pct": 66.49800254315382, + "macro_recall_pct": 68.80914808808257, + "per_class_f1_pct": { + "FAVOR": 62.27544910179641, + "AGAINST": 65.18105849582173, + "NONE": 72.38095238095238 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5425058409447767, + "weighted_f1": 0.593102252081552, + "macro_precision": 0.5623688595358503, + "macro_recall": 0.5945419770746185, + "per_class_f1": { + "FAVOR": 0.4046242774566474, + "AGAINST": 0.5633187772925764, + "NONE": 0.6595744680851063 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 54.250584094477674, + "weighted_f1_pct": 59.310225208155195, + "macro_precision_pct": 56.23688595358502, + "macro_recall_pct": 59.45419770746185, + "per_class_f1_pct": { + "FAVOR": 40.46242774566474, + "AGAINST": 56.33187772925764, + "NONE": 65.95744680851064 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6292974588938715, + "macro_f1": 0.5840808825115622, + "weighted_f1": 0.6456598031385071, + "macro_precision": 0.5875843454790823, + "macro_recall": 0.6057666435192529, + "per_class_f1": { + "FAVOR": 0.5357142857142857, + "AGAINST": 0.45936395759717313, + "NONE": 0.7571644042232277 + }, + "accuracy_pct": 62.92974588938714, + "macro_f1_pct": 58.40808825115622, + "weighted_f1_pct": 64.56598031385072, + "macro_precision_pct": 58.75843454790824, + "macro_recall_pct": 60.57666435192529, + "per_class_f1_pct": { + "FAVOR": 53.57142857142857, + "AGAINST": 45.936395759717314, + "NONE": 75.71644042232278 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6298615112135373, + "weighted_f1": 0.6736840735856147, + "macro_precision": 0.630582342226508, + "macro_recall": 0.6294703004356653, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.5609756097560976, + "NONE": 0.7952755905511811 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 62.98615112135373, + "weighted_f1_pct": 67.36840735856147, + "macro_precision_pct": 63.058234222650796, + "macro_recall_pct": 62.947030043566535, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 56.09756097560976, + "NONE": 79.52755905511812 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.716297786720322, + "macro_f1": 0.6833801657343669, + "weighted_f1": 0.7252650330273138, + "macro_precision": 0.6777713699999537, + "macro_recall": 0.699851000425713, + "per_class_f1": { + "FAVOR": 0.6007604562737643, + "AGAINST": 0.6324786324786325, + "NONE": 0.8169014084507042 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.33801657343669, + "weighted_f1_pct": 72.52650330273138, + "macro_precision_pct": 67.77713699999536, + "macro_recall_pct": 69.9851000425713, + "per_class_f1_pct": { + "FAVOR": 60.07604562737643, + "AGAINST": 63.24786324786324, + "NONE": 81.69014084507043 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6519210169391337, + "weighted_f1": 0.7023819175543574, + "macro_precision": 0.6454454381523106, + "macro_recall": 0.6928929943341359, + "per_class_f1": { + "FAVOR": 0.5289855072463768, + "AGAINST": 0.6556603773584906, + "NONE": 0.771117166212534 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 65.19210169391337, + "weighted_f1_pct": 70.23819175543574, + "macro_precision_pct": 64.54454381523107, + "macro_recall_pct": 69.28929943341359, + "per_class_f1_pct": { + "FAVOR": 52.89855072463768, + "AGAINST": 65.56603773584906, + "NONE": 77.11171662125341 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7541528239202658, + "macro_f1": 0.707316090361901, + "weighted_f1": 0.7578762023181934, + "macro_precision": 0.6993793705131774, + "macro_recall": 0.718292377726244, + "per_class_f1": { + "FAVOR": 0.6271186440677966, + "AGAINST": 0.6508474576271186, + "NONE": 0.8439821693907875 + }, + "accuracy_pct": 75.41528239202658, + "macro_f1_pct": 70.73160903619011, + "weighted_f1_pct": 75.78762023181935, + "macro_precision_pct": 69.93793705131775, + "macro_recall_pct": 71.8292377726244, + "per_class_f1_pct": { + "FAVOR": 62.71186440677966, + "AGAINST": 65.08474576271186, + "NONE": 84.39821693907875 + } + } + ], + "accuracy": 0.6627520062634567, + "macro_f1": 0.6248573052912697, + "weighted_f1": 0.6749224993922747, + "macro_precision": 0.6203728032267358, + "macro_recall": 0.6464981967532092, + "per_class_f1": { + "FAVOR": 0.5272277227722773, + "AGAINST": 0.5778454510848877, + "NONE": 0.769498742016644 + }, + "accuracy_pct": 66.27520062634567, + "macro_f1_pct": 62.48573052912697, + "weighted_f1_pct": 67.49224993922746, + "macro_precision_pct": 62.03728032267358, + "macro_recall_pct": 64.64981967532091, + "per_class_f1_pct": { + "FAVOR": 52.722772277227726, + "AGAINST": 57.78454510848877, + "NONE": 76.9498742016644 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6260628927555552, + "average_forgetting": 0.02482500893651848, + "backward_transfer": -0.021033756545850078, + "final_mean_target_macro_f1_pct": 62.60628927555552, + "average_forgetting_pct": 2.482500893651848, + "backward_transfer_pct": -2.1033756545850077, + "final_pooled_accuracy": 0.6627520062634567, + "final_pooled_accuracy_pct": 66.27520062634567, + "final_pooled_macro_f1": 0.6248573052912697, + "final_pooled_macro_f1_pct": 62.48573052912697, + "final_pooled_weighted_f1": 0.6749224993922747, + "final_pooled_weighted_f1_pct": 67.49224993922746, + "final_pooled_macro_precision": 0.6203728032267358, + "final_pooled_macro_precision_pct": 62.03728032267358, + "final_pooled_macro_recall": 0.6464981967532092, + "final_pooled_macro_recall_pct": 64.64981967532091, + "final_pooled_per_class_f1": { + "FAVOR": 0.5272277227722773, + "AGAINST": 0.5778454510848877, + "NONE": 0.769498742016644 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 52.722772277227726, + "AGAINST": 57.78454510848877, + "NONE": 76.9498742016644 + } + }, + "efficiency": { + "training_wall_time_seconds": 1774.8458118368872, + "training_wall_time_minutes": 29.580763530614785, + "peak_gpu_memory_allocated_gb": 10.39700698852539, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..6378669755b2494adec186dc6f6a4b66ad530aee --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s1__seed13/metrics.partial.json @@ -0,0 +1,1481 @@ +{ + "matrix": [ + [ + 0.6155173412078051, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5694161861580516, + 0.6956361701518025, + null, + null, + null, + null, + null, + null + ], + [ + 0.5292705369293579, + 0.6688695179351244, + 0.569517700185404, + null, + null, + null, + null, + null + ], + [ + 0.5217190618274629, + 0.613353638755958, + 0.5125331533353226, + 0.5730053319396435, + null, + null, + null, + null + ], + [ + 0.5517885838908566, + 0.6508294594071905, + 0.4968656936302855, + 0.5570733785888508, + 0.6546257001019442, + null, + null, + null + ], + [ + 0.5185760867396362, + 0.6351920141919599, + 0.4997972354580364, + 0.5446644351512296, + 0.634896023500377, + 0.6691383965626724, + null, + null + ], + [ + 0.528340831926188, + 0.6726239875674654, + 0.5287480209475094, + 0.5872269180724953, + 0.6264762020792722, + 0.6814555771644994, + 0.670982707354219, + null + ], + [ + 0.5433127677439283, + 0.6661248665952351, + 0.5425058409447767, + 0.5840808825115622, + 0.6298615112135373, + 0.6833801657343669, + 0.6519210169391337, + 0.707316090361901 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 13, + "best_dev_macro_f1": 0.6697781414802692, + "losses": { + "current_ce": 0.6033996852018215, + "total": 0.6033996852018215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6899618805590851, + "macro_f1": 0.6155173412078051, + "weighted_f1": 0.7021062475404232, + "macro_precision": 0.6083931545026292, + "macro_recall": 0.6320453995290597, + "per_class_f1": { + "FAVOR": 0.512280701754386, + "AGAINST": 0.5157894736842106, + "NONE": 0.8184818481848185 + }, + "accuracy_pct": 68.99618805590852, + "macro_f1_pct": 61.55173412078051, + "weighted_f1_pct": 70.21062475404231, + "macro_precision_pct": 60.83931545026292, + "macro_recall_pct": 63.20453995290597, + "per_class_f1_pct": { + "FAVOR": 51.2280701754386, + "AGAINST": 51.578947368421055, + "NONE": 81.84818481848185 + } + } + ], + "accuracy": 0.6899618805590851, + "macro_f1": 0.6155173412078051, + "weighted_f1": 0.7021062475404232, + "macro_precision": 0.6083931545026292, + "macro_recall": 0.6320453995290597, + "per_class_f1": { + "FAVOR": 0.512280701754386, + "AGAINST": 0.5157894736842106, + "NONE": 0.8184818481848185 + }, + "accuracy_pct": 68.99618805590852, + "macro_f1_pct": 61.55173412078051, + "weighted_f1_pct": 70.21062475404231, + "macro_precision_pct": 60.83931545026292, + "macro_recall_pct": 63.20453995290597, + "per_class_f1_pct": { + "FAVOR": 51.2280701754386, + "AGAINST": 51.578947368421055, + "NONE": 81.84818481848185 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7311337628642974, + "losses": { + "current_ce": 0.5307365087951933, + "replay_ce": 0.02806483572348952, + "total": 0.5588013440086728, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6531130876747141, + "macro_f1": 0.5694161861580516, + "weighted_f1": 0.6618331942814528, + "macro_precision": 0.5650006999308553, + "macro_recall": 0.5910686384542594, + "per_class_f1": { + "FAVOR": 0.5029239766081871, + "AGAINST": 0.41533546325878595, + "NONE": 0.7899891186071817 + }, + "accuracy_pct": 65.31130876747142, + "macro_f1_pct": 56.941618615805154, + "weighted_f1_pct": 66.18331942814528, + "macro_precision_pct": 56.50006999308553, + "macro_recall_pct": 59.10686384542594, + "per_class_f1_pct": { + "FAVOR": 50.29239766081871, + "AGAINST": 41.533546325878596, + "NONE": 78.99891186071817 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7110016420361248, + "macro_f1": 0.6956361701518025, + "weighted_f1": 0.7188045338776238, + "macro_precision": 0.6999122917984398, + "macro_recall": 0.7098670619031916, + "per_class_f1": { + "FAVOR": 0.6127167630057804, + "AGAINST": 0.6873065015479877, + "NONE": 0.7868852459016393 + }, + "accuracy_pct": 71.10016420361248, + "macro_f1_pct": 69.56361701518024, + "weighted_f1_pct": 71.88045338776237, + "macro_precision_pct": 69.99122917984398, + "macro_recall_pct": 70.98670619031915, + "per_class_f1_pct": { + "FAVOR": 61.27167630057804, + "AGAINST": 68.73065015479877, + "NONE": 78.68852459016394 + } + } + ], + "accuracy": 0.6783667621776505, + "macro_f1": 0.6334756640850786, + "weighted_f1": 0.6867442139840507, + "macro_precision": 0.6322937896760786, + "macro_recall": 0.6522518761257706, + "per_class_f1": { + "FAVOR": 0.5581395348837209, + "AGAINST": 0.5534591194968553, + "NONE": 0.7888283378746594 + }, + "accuracy_pct": 67.83667621776505, + "macro_f1_pct": 63.34756640850786, + "weighted_f1_pct": 68.67442139840507, + "macro_precision_pct": 63.229378967607865, + "macro_recall_pct": 65.22518761257706, + "per_class_f1_pct": { + "FAVOR": 55.81395348837209, + "AGAINST": 55.34591194968554, + "NONE": 78.88283378746594 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.6837498989719051, + "losses": { + "current_ce": 0.546360798870095, + "replay_ce": 0.017297741932386073, + "total": 0.5636585399234992, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6302414231257941, + "macro_f1": 0.5292705369293579, + "weighted_f1": 0.6329911674399076, + "macro_precision": 0.5269146331751605, + "macro_recall": 0.533691673452022, + "per_class_f1": { + "FAVOR": 0.42033898305084744, + "AGAINST": 0.39375, + "NONE": 0.7737226277372263 + }, + "accuracy_pct": 63.024142312579414, + "macro_f1_pct": 52.927053692935786, + "weighted_f1_pct": 63.29911674399076, + "macro_precision_pct": 52.69146331751605, + "macro_recall_pct": 53.36916734520219, + "per_class_f1_pct": { + "FAVOR": 42.03389830508474, + "AGAINST": 39.375, + "NONE": 77.37226277372264 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6765188834154351, + "macro_f1": 0.6688695179351244, + "weighted_f1": 0.6832681716948531, + "macro_precision": 0.6755026482606038, + "macro_recall": 0.6954341246288583, + "per_class_f1": { + "FAVOR": 0.6276923076923077, + "AGAINST": 0.6512820512820513, + "NONE": 0.7276341948310139 + }, + "accuracy_pct": 67.65188834154351, + "macro_f1_pct": 66.88695179351244, + "weighted_f1_pct": 68.32681716948531, + "macro_precision_pct": 67.55026482606038, + "macro_recall_pct": 69.54341246288584, + "per_class_f1_pct": { + "FAVOR": 62.76923076923077, + "AGAINST": 65.12820512820512, + "NONE": 72.76341948310139 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6101231190150479, + "macro_f1": 0.569517700185404, + "weighted_f1": 0.6234517578663603, + "macro_precision": 0.5720758270427145, + "macro_recall": 0.5939638545932486, + "per_class_f1": { + "FAVOR": 0.43089430894308944, + "AGAINST": 0.5767790262172284, + "NONE": 0.7008797653958945 + }, + "accuracy_pct": 61.01231190150479, + "macro_f1_pct": 56.9517700185404, + "weighted_f1_pct": 62.345175786636034, + "macro_precision_pct": 57.20758270427145, + "macro_recall_pct": 59.39638545932486, + "per_class_f1_pct": { + "FAVOR": 43.08943089430895, + "AGAINST": 57.67790262172284, + "NONE": 70.08797653958945 + } + } + ], + "accuracy": 0.6365773389750823, + "macro_f1": 0.5979827342535521, + "weighted_f1": 0.6474129609335886, + "macro_precision": 0.5925061339263092, + "macro_recall": 0.6166227411947646, + "per_class_f1": { + "FAVOR": 0.5011547344110855, + "AGAINST": 0.5530546623794212, + "NONE": 0.7397388059701493 + }, + "accuracy_pct": 63.65773389750823, + "macro_f1_pct": 59.798273425355205, + "weighted_f1_pct": 64.74129609335886, + "macro_precision_pct": 59.250613392630925, + "macro_recall_pct": 61.66227411947646, + "per_class_f1_pct": { + "FAVOR": 50.11547344110855, + "AGAINST": 55.30546623794213, + "NONE": 73.97388059701493 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.685362115437692, + "losses": { + "current_ce": 0.5138460865930508, + "replay_ce": 0.020093058473675658, + "total": 0.533939144407448, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5959339263024143, + "macro_f1": 0.5217190618274629, + "weighted_f1": 0.6157188732089818, + "macro_precision": 0.5309637871422258, + "macro_recall": 0.5583553398912876, + "per_class_f1": { + "FAVOR": 0.44938271604938274, + "AGAINST": 0.37012987012987014, + "NONE": 0.7456445993031359 + }, + "accuracy_pct": 59.593392630241425, + "macro_f1_pct": 52.171906182746284, + "weighted_f1_pct": 61.57188732089818, + "macro_precision_pct": 53.09637871422258, + "macro_recall_pct": 55.83553398912876, + "per_class_f1_pct": { + "FAVOR": 44.93827160493827, + "AGAINST": 37.01298701298701, + "NONE": 74.56445993031359 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6190476190476191, + "macro_f1": 0.613353638755958, + "weighted_f1": 0.6272319375336906, + "macro_precision": 0.6348633536898273, + "macro_recall": 0.6450364206794089, + "per_class_f1": { + "FAVOR": 0.563953488372093, + "AGAINST": 0.6080402010050251, + "NONE": 0.6680672268907563 + }, + "accuracy_pct": 61.904761904761905, + "macro_f1_pct": 61.3353638755958, + "weighted_f1_pct": 62.72319375336905, + "macro_precision_pct": 63.48633536898273, + "macro_recall_pct": 64.50364206794089, + "per_class_f1_pct": { + "FAVOR": 56.395348837209305, + "AGAINST": 60.80402010050251, + "NONE": 66.80672268907563 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5417236662106704, + "macro_f1": 0.5125331533353226, + "weighted_f1": 0.5654020911298463, + "macro_precision": 0.5357188508210813, + "macro_recall": 0.5549233265587468, + "per_class_f1": { + "FAVOR": 0.3569230769230769, + "AGAINST": 0.5548902195608783, + "NONE": 0.6257861635220126 + }, + "accuracy_pct": 54.172366621067034, + "macro_f1_pct": 51.25331533353226, + "weighted_f1_pct": 56.54020911298463, + "macro_precision_pct": 53.571885082108125, + "macro_recall_pct": 55.49233265587468, + "per_class_f1_pct": { + "FAVOR": 35.69230769230769, + "AGAINST": 55.489021956087825, + "NONE": 62.57861635220125 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5730053319396435, + "weighted_f1": 0.6249112125865773, + "macro_precision": 0.5923607358033588, + "macro_recall": 0.6094249923919034, + "per_class_f1": { + "FAVOR": 0.5317073170731708, + "AGAINST": 0.46835443037974683, + "NONE": 0.7189542483660131 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 57.300533193964355, + "weighted_f1_pct": 62.491121258657735, + "macro_precision_pct": 59.236073580335876, + "macro_recall_pct": 60.94249923919034, + "per_class_f1_pct": { + "FAVOR": 53.170731707317074, + "AGAINST": 46.835443037974684, + "NONE": 71.89542483660131 + } + } + ], + "accuracy": 0.5883404864091559, + "macro_f1": 0.5624827302356098, + "weighted_f1": 0.6082392850633415, + "macro_precision": 0.5752730310625047, + "macro_recall": 0.5984835689256194, + "per_class_f1": { + "FAVOR": 0.4784366576819407, + "AGAINST": 0.5134602757715037, + "NONE": 0.695551257253385 + }, + "accuracy_pct": 58.83404864091559, + "macro_f1_pct": 56.24827302356098, + "weighted_f1_pct": 60.82392850633415, + "macro_precision_pct": 57.52730310625047, + "macro_recall_pct": 59.84835689256194, + "per_class_f1_pct": { + "FAVOR": 47.84366576819407, + "AGAINST": 51.346027577150366, + "NONE": 69.5551257253385 + } + }, + { + "after_task": 5, + "best_epoch": 2, + "best_dev_macro_f1": 0.6720908926179384, + "losses": { + "current_ce": 0.4921029800283057, + "replay_ce": 0.029563177262233304, + "total": 0.521666157653644, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.5517885838908566, + "weighted_f1": 0.6479521895550832, + "macro_precision": 0.5522860929049819, + "macro_recall": 0.5761337045106109, + "per_class_f1": { + "FAVOR": 0.4573002754820937, + "AGAINST": 0.41904761904761906, + "NONE": 0.7790178571428571 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 55.17885838908566, + "weighted_f1_pct": 64.79521895550832, + "macro_precision_pct": 55.228609290498184, + "macro_recall_pct": 57.61337045106109, + "per_class_f1_pct": { + "FAVOR": 45.73002754820937, + "AGAINST": 41.904761904761905, + "NONE": 77.90178571428571 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6633825944170771, + "macro_f1": 0.6508294594071905, + "weighted_f1": 0.6701214266900546, + "macro_precision": 0.6551864165674209, + "macro_recall": 0.669622528930551, + "per_class_f1": { + "FAVOR": 0.5894039735099338, + "AGAINST": 0.6347607052896725, + "NONE": 0.7283236994219653 + }, + "accuracy_pct": 66.33825944170772, + "macro_f1_pct": 65.08294594071906, + "weighted_f1_pct": 67.01214266900546, + "macro_precision_pct": 65.51864165674209, + "macro_recall_pct": 66.9622528930551, + "per_class_f1_pct": { + "FAVOR": 58.94039735099338, + "AGAINST": 63.476070528967256, + "NONE": 72.83236994219652 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5143638850889193, + "macro_f1": 0.4968656936302855, + "weighted_f1": 0.5432809051046937, + "macro_precision": 0.5420035501451835, + "macro_recall": 0.5585061923779701, + "per_class_f1": { + "FAVOR": 0.36787564766839376, + "AGAINST": 0.5204301075268817, + "NONE": 0.602291325695581 + }, + "accuracy_pct": 51.43638850889193, + "macro_f1_pct": 49.686569363028546, + "weighted_f1_pct": 54.32809051046937, + "macro_precision_pct": 54.20035501451835, + "macro_recall_pct": 55.850619237797005, + "per_class_f1_pct": { + "FAVOR": 36.787564766839374, + "AGAINST": 52.04301075268817, + "NONE": 60.2291325695581 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.585949177877429, + "macro_f1": 0.5570733785888508, + "weighted_f1": 0.6081637701000742, + "macro_precision": 0.5915731300609096, + "macro_recall": 0.5957592057969584, + "per_class_f1": { + "FAVOR": 0.5219298245614035, + "AGAINST": 0.44912280701754387, + "NONE": 0.7001675041876047 + }, + "accuracy_pct": 58.594917787742894, + "macro_f1_pct": 55.707337858885076, + "weighted_f1_pct": 60.816377010007415, + "macro_precision_pct": 59.15731300609096, + "macro_recall_pct": 59.57592057969584, + "per_class_f1_pct": { + "FAVOR": 52.19298245614035, + "AGAINST": 44.91228070175438, + "NONE": 70.01675041876047 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6881287726358148, + "macro_f1": 0.6546257001019442, + "weighted_f1": 0.6956462080302849, + "macro_precision": 0.6518476621417798, + "macro_recall": 0.6628828198467108, + "per_class_f1": { + "FAVOR": 0.5748031496062992, + "AGAINST": 0.5789473684210527, + "NONE": 0.810126582278481 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 65.46257001019443, + "weighted_f1_pct": 69.56462080302849, + "macro_precision_pct": 65.18476621417798, + "macro_recall_pct": 66.28828198467109, + "per_class_f1_pct": { + "FAVOR": 57.48031496062992, + "AGAINST": 57.89473684210527, + "NONE": 81.0126582278481 + } + } + ], + "accuracy": 0.6109930154873975, + "macro_f1": 0.5815727981752827, + "weighted_f1": 0.6291099409562191, + "macro_precision": 0.5900546965334866, + "macro_recall": 0.6109428349334495, + "per_class_f1": { + "FAVOR": 0.4940374787052811, + "AGAINST": 0.5254629629629629, + "NONE": 0.7252179528576042 + }, + "accuracy_pct": 61.099301548739746, + "macro_f1_pct": 58.15727981752828, + "weighted_f1_pct": 62.91099409562191, + "macro_precision_pct": 59.00546965334866, + "macro_recall_pct": 61.094283493344946, + "per_class_f1_pct": { + "FAVOR": 49.40374787052811, + "AGAINST": 52.54629629629629, + "NONE": 72.52179528576042 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7184062850729518, + "losses": { + "current_ce": 0.4117540765198923, + "replay_ce": 0.04057449822846268, + "total": 0.452328574799356, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5933926302414231, + "macro_f1": 0.5185760867396362, + "weighted_f1": 0.6162278908487981, + "macro_precision": 0.5207216559626828, + "macro_recall": 0.5434194890077243, + "per_class_f1": { + "FAVOR": 0.410958904109589, + "AGAINST": 0.3965014577259475, + "NONE": 0.7482678983833718 + }, + "accuracy_pct": 59.339263024142305, + "macro_f1_pct": 51.85760867396362, + "weighted_f1_pct": 61.62278908487981, + "macro_precision_pct": 52.072165596268285, + "macro_recall_pct": 54.34194890077243, + "per_class_f1_pct": { + "FAVOR": 41.0958904109589, + "AGAINST": 39.650145772594755, + "NONE": 74.82678983833718 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6403940886699507, + "macro_f1": 0.6351920141919599, + "weighted_f1": 0.6462805337939896, + "macro_precision": 0.6468095399701064, + "macro_recall": 0.6661981364216514, + "per_class_f1": { + "FAVOR": 0.5958702064896755, + "AGAINST": 0.6307692307692307, + "NONE": 0.6789366053169734 + }, + "accuracy_pct": 64.03940886699507, + "macro_f1_pct": 63.519201419195994, + "weighted_f1_pct": 64.62805337939896, + "macro_precision_pct": 64.68095399701063, + "macro_recall_pct": 66.61981364216514, + "per_class_f1_pct": { + "FAVOR": 59.587020648967545, + "AGAINST": 63.07692307692307, + "NONE": 67.89366053169734 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5225718194254446, + "macro_f1": 0.4997972354580364, + "weighted_f1": 0.5467535084905882, + "macro_precision": 0.5342033218593144, + "macro_recall": 0.550931702923668, + "per_class_f1": { + "FAVOR": 0.36151603498542273, + "AGAINST": 0.5375494071146245, + "NONE": 0.600326264274062 + }, + "accuracy_pct": 52.25718194254446, + "macro_f1_pct": 49.97972354580364, + "weighted_f1_pct": 54.67535084905882, + "macro_precision_pct": 53.42033218593144, + "macro_recall_pct": 55.0931702923668, + "per_class_f1_pct": { + "FAVOR": 36.15160349854227, + "AGAINST": 53.75494071146245, + "NONE": 60.0326264274062 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.57847533632287, + "macro_f1": 0.5446644351512296, + "weighted_f1": 0.6041335106943165, + "macro_precision": 0.5625404272346487, + "macro_recall": 0.5728174162551244, + "per_class_f1": { + "FAVOR": 0.467005076142132, + "AGAINST": 0.4520123839009288, + "NONE": 0.714975845410628 + }, + "accuracy_pct": 57.847533632286996, + "macro_f1_pct": 54.46644351512296, + "weighted_f1_pct": 60.413351069431656, + "macro_precision_pct": 56.25404272346487, + "macro_recall_pct": 57.28174162551244, + "per_class_f1_pct": { + "FAVOR": 46.7005076142132, + "AGAINST": 45.20123839009288, + "NONE": 71.49758454106279 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6639839034205232, + "macro_f1": 0.634896023500377, + "weighted_f1": 0.671621705336156, + "macro_precision": 0.6337234625111974, + "macro_recall": 0.6446684269086628, + "per_class_f1": { + "FAVOR": 0.5537190082644629, + "AGAINST": 0.5774647887323944, + "NONE": 0.7735042735042735 + }, + "accuracy_pct": 66.39839034205231, + "macro_f1_pct": 63.4896023500377, + "weighted_f1_pct": 67.1621705336156, + "macro_precision_pct": 63.37234625111974, + "macro_recall_pct": 64.46684269086627, + "per_class_f1_pct": { + "FAVOR": 55.371900826446286, + "AGAINST": 57.74647887323944, + "NONE": 77.35042735042735 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6941649899396378, + "macro_f1": 0.6691383965626724, + "weighted_f1": 0.7066846132031092, + "macro_precision": 0.666905200027261, + "macro_recall": 0.6973991232611922, + "per_class_f1": { + "FAVOR": 0.5639097744360902, + "AGAINST": 0.6536964980544747, + "NONE": 0.7898089171974523 + }, + "accuracy_pct": 69.41649899396378, + "macro_f1_pct": 66.91383965626724, + "weighted_f1_pct": 70.66846132031091, + "macro_precision_pct": 66.69052000272609, + "macro_recall_pct": 69.73991232611922, + "per_class_f1_pct": { + "FAVOR": 56.390977443609025, + "AGAINST": 65.36964980544747, + "NONE": 78.98089171974523 + } + } + ], + "accuracy": 0.6071240105540897, + "macro_f1": 0.5795386644369359, + "weighted_f1": 0.6249017135969034, + "macro_precision": 0.5864645229053557, + "macro_recall": 0.6075279521187976, + "per_class_f1": { + "FAVOR": 0.4843509492047204, + "AGAINST": 0.5382786495482644, + "NONE": 0.7159863945578231 + }, + "accuracy_pct": 60.71240105540897, + "macro_f1_pct": 57.95386644369359, + "weighted_f1_pct": 62.490171359690336, + "macro_precision_pct": 58.646452290535564, + "macro_recall_pct": 60.752795211879764, + "per_class_f1_pct": { + "FAVOR": 48.43509492047204, + "AGAINST": 53.827864954826445, + "NONE": 71.5986394557823 + } + }, + { + "after_task": 7, + "best_epoch": 13, + "best_dev_macro_f1": 0.7332904506331527, + "losses": { + "current_ce": 0.36610839086456026, + "replay_ce": 0.017698088319351275, + "total": 0.38380647870462115, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6200762388818297, + "macro_f1": 0.528340831926188, + "weighted_f1": 0.6359720262459907, + "macro_precision": 0.5306727590900252, + "macro_recall": 0.5467173551051546, + "per_class_f1": { + "FAVOR": 0.4134078212290503, + "AGAINST": 0.38977635782747605, + "NONE": 0.7818383167220376 + }, + "accuracy_pct": 62.00762388818297, + "macro_f1_pct": 52.8340831926188, + "weighted_f1_pct": 63.59720262459907, + "macro_precision_pct": 53.06727590900252, + "macro_recall_pct": 54.67173551051546, + "per_class_f1_pct": { + "FAVOR": 41.340782122905026, + "AGAINST": 38.977635782747605, + "NONE": 78.18383167220377 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6847290640394089, + "macro_f1": 0.6726239875674654, + "weighted_f1": 0.6900788317368352, + "macro_precision": 0.6698032506415741, + "macro_recall": 0.6887602089684147, + "per_class_f1": { + "FAVOR": 0.627831715210356, + "AGAINST": 0.6451612903225806, + "NONE": 0.74487895716946 + }, + "accuracy_pct": 68.4729064039409, + "macro_f1_pct": 67.26239875674655, + "weighted_f1_pct": 69.00788317368351, + "macro_precision_pct": 66.98032506415741, + "macro_recall_pct": 68.87602089684147, + "per_class_f1_pct": { + "FAVOR": 62.7831715210356, + "AGAINST": 64.51612903225806, + "NONE": 74.487895716946 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5649794801641587, + "macro_f1": 0.5287480209475094, + "weighted_f1": 0.5879003192307674, + "macro_precision": 0.5409336979982918, + "macro_recall": 0.5675054867388313, + "per_class_f1": { + "FAVOR": 0.38509316770186336, + "AGAINST": 0.5217391304347826, + "NONE": 0.6794117647058824 + }, + "accuracy_pct": 56.49794801641587, + "macro_f1_pct": 52.87480209475094, + "weighted_f1_pct": 58.79003192307673, + "macro_precision_pct": 54.09336979982918, + "macro_recall_pct": 56.75054867388313, + "per_class_f1_pct": { + "FAVOR": 38.50931677018634, + "AGAINST": 52.17391304347826, + "NONE": 67.94117647058823 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6322869955156951, + "macro_f1": 0.5872269180724953, + "weighted_f1": 0.6495787862138902, + "macro_precision": 0.5883134571857637, + "macro_recall": 0.6074539372340838, + "per_class_f1": { + "FAVOR": 0.5173333333333333, + "AGAINST": 0.4797297297297297, + "NONE": 0.7646176911544228 + }, + "accuracy_pct": 63.22869955156951, + "macro_f1_pct": 58.722691807249525, + "weighted_f1_pct": 64.95787862138901, + "macro_precision_pct": 58.831345718576365, + "macro_recall_pct": 60.745393723408384, + "per_class_f1_pct": { + "FAVOR": 51.733333333333334, + "AGAINST": 47.97297297297297, + "NONE": 76.46176911544228 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6264762020792722, + "weighted_f1": 0.6721173287287074, + "macro_precision": 0.6292235801581595, + "macro_recall": 0.6247689372228797, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.546875, + "NONE": 0.7992202729044834 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 62.647620207927226, + "weighted_f1_pct": 67.21173287287074, + "macro_precision_pct": 62.92235801581595, + "macro_recall_pct": 62.476893722287976, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 54.6875, + "NONE": 79.92202729044834 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6814555771644994, + "weighted_f1": 0.7251352466440751, + "macro_precision": 0.6817443297515938, + "macro_recall": 0.7051885218551885, + "per_class_f1": { + "FAVOR": 0.6148409893992933, + "AGAINST": 0.6094420600858369, + "NONE": 0.8200836820083682 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 68.14555771644994, + "weighted_f1_pct": 72.51352466440751, + "macro_precision_pct": 68.17443297515938, + "macro_recall_pct": 70.51885218551885, + "per_class_f1_pct": { + "FAVOR": 61.48409893992933, + "AGAINST": 60.94420600858369, + "NONE": 82.00836820083683 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7168758716875872, + "macro_f1": 0.670982707354219, + "weighted_f1": 0.7270188618515017, + "macro_precision": 0.6596710854829722, + "macro_recall": 0.7008511759517018, + "per_class_f1": { + "FAVOR": 0.5387453874538746, + "AGAINST": 0.6700251889168766, + "NONE": 0.804177545691906 + }, + "accuracy_pct": 71.68758716875871, + "macro_f1_pct": 67.0982707354219, + "weighted_f1_pct": 72.70188618515017, + "macro_precision_pct": 65.96710854829722, + "macro_recall_pct": 70.08511759517017, + "per_class_f1_pct": { + "FAVOR": 53.874538745387454, + "AGAINST": 67.00251889168766, + "NONE": 80.4177545691906 + } + } + ], + "accuracy": 0.6532061238074107, + "macro_f1": 0.6126288751911302, + "weighted_f1": 0.665966090750988, + "macro_precision": 0.6091610960968175, + "macro_recall": 0.6320459588969595, + "per_class_f1": { + "FAVOR": 0.5132991133924405, + "AGAINST": 0.5552213149978513, + "NONE": 0.7693661971830986 + }, + "accuracy_pct": 65.32061238074107, + "macro_f1_pct": 61.26288751911302, + "weighted_f1_pct": 66.5966090750988, + "macro_precision_pct": 60.91610960968175, + "macro_recall_pct": 63.204595889695945, + "per_class_f1_pct": { + "FAVOR": 51.32991133924405, + "AGAINST": 55.52213149978513, + "NONE": 76.93661971830986 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.7875357278487338, + "losses": { + "current_ce": 0.34310466447005084, + "replay_ce": 0.014769438629988215, + "total": 0.35787410408112347, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.5433127677439283, + "weighted_f1": 0.6476966032743156, + "macro_precision": 0.541347418875975, + "macro_recall": 0.5585246069995525, + "per_class_f1": { + "FAVOR": 0.41543026706231456, + "AGAINST": 0.4268292682926829, + "NONE": 0.7876787678767877 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 54.33127677439283, + "weighted_f1_pct": 64.76966032743155, + "macro_precision_pct": 54.13474188759751, + "macro_recall_pct": 55.85246069995525, + "per_class_f1_pct": { + "FAVOR": 41.54302670623146, + "AGAINST": 42.68292682926829, + "NONE": 78.76787678767877 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6748768472906403, + "macro_f1": 0.6661248665952351, + "weighted_f1": 0.6803848335247603, + "macro_precision": 0.6649800254315382, + "macro_recall": 0.6880914808808257, + "per_class_f1": { + "FAVOR": 0.6227544910179641, + "AGAINST": 0.6518105849582173, + "NONE": 0.7238095238095238 + }, + "accuracy_pct": 67.48768472906403, + "macro_f1_pct": 66.61248665952351, + "weighted_f1_pct": 68.03848335247604, + "macro_precision_pct": 66.49800254315382, + "macro_recall_pct": 68.80914808808257, + "per_class_f1_pct": { + "FAVOR": 62.27544910179641, + "AGAINST": 65.18105849582173, + "NONE": 72.38095238095238 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5425058409447767, + "weighted_f1": 0.593102252081552, + "macro_precision": 0.5623688595358503, + "macro_recall": 0.5945419770746185, + "per_class_f1": { + "FAVOR": 0.4046242774566474, + "AGAINST": 0.5633187772925764, + "NONE": 0.6595744680851063 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 54.250584094477674, + "weighted_f1_pct": 59.310225208155195, + "macro_precision_pct": 56.23688595358502, + "macro_recall_pct": 59.45419770746185, + "per_class_f1_pct": { + "FAVOR": 40.46242774566474, + "AGAINST": 56.33187772925764, + "NONE": 65.95744680851064 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6292974588938715, + "macro_f1": 0.5840808825115622, + "weighted_f1": 0.6456598031385071, + "macro_precision": 0.5875843454790823, + "macro_recall": 0.6057666435192529, + "per_class_f1": { + "FAVOR": 0.5357142857142857, + "AGAINST": 0.45936395759717313, + "NONE": 0.7571644042232277 + }, + "accuracy_pct": 62.92974588938714, + "macro_f1_pct": 58.40808825115622, + "weighted_f1_pct": 64.56598031385072, + "macro_precision_pct": 58.75843454790824, + "macro_recall_pct": 60.57666435192529, + "per_class_f1_pct": { + "FAVOR": 53.57142857142857, + "AGAINST": 45.936395759717314, + "NONE": 75.71644042232278 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6298615112135373, + "weighted_f1": 0.6736840735856147, + "macro_precision": 0.630582342226508, + "macro_recall": 0.6294703004356653, + "per_class_f1": { + "FAVOR": 0.5333333333333333, + "AGAINST": 0.5609756097560976, + "NONE": 0.7952755905511811 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 62.98615112135373, + "weighted_f1_pct": 67.36840735856147, + "macro_precision_pct": 63.058234222650796, + "macro_recall_pct": 62.947030043566535, + "per_class_f1_pct": { + "FAVOR": 53.333333333333336, + "AGAINST": 56.09756097560976, + "NONE": 79.52755905511812 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.716297786720322, + "macro_f1": 0.6833801657343669, + "weighted_f1": 0.7252650330273138, + "macro_precision": 0.6777713699999537, + "macro_recall": 0.699851000425713, + "per_class_f1": { + "FAVOR": 0.6007604562737643, + "AGAINST": 0.6324786324786325, + "NONE": 0.8169014084507042 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.33801657343669, + "weighted_f1_pct": 72.52650330273138, + "macro_precision_pct": 67.77713699999536, + "macro_recall_pct": 69.9851000425713, + "per_class_f1_pct": { + "FAVOR": 60.07604562737643, + "AGAINST": 63.24786324786324, + "NONE": 81.69014084507043 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6519210169391337, + "weighted_f1": 0.7023819175543574, + "macro_precision": 0.6454454381523106, + "macro_recall": 0.6928929943341359, + "per_class_f1": { + "FAVOR": 0.5289855072463768, + "AGAINST": 0.6556603773584906, + "NONE": 0.771117166212534 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 65.19210169391337, + "weighted_f1_pct": 70.23819175543574, + "macro_precision_pct": 64.54454381523107, + "macro_recall_pct": 69.28929943341359, + "per_class_f1_pct": { + "FAVOR": 52.89855072463768, + "AGAINST": 65.56603773584906, + "NONE": 77.11171662125341 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7541528239202658, + "macro_f1": 0.707316090361901, + "weighted_f1": 0.7578762023181934, + "macro_precision": 0.6993793705131774, + "macro_recall": 0.718292377726244, + "per_class_f1": { + "FAVOR": 0.6271186440677966, + "AGAINST": 0.6508474576271186, + "NONE": 0.8439821693907875 + }, + "accuracy_pct": 75.41528239202658, + "macro_f1_pct": 70.73160903619011, + "weighted_f1_pct": 75.78762023181935, + "macro_precision_pct": 69.93793705131775, + "macro_recall_pct": 71.8292377726244, + "per_class_f1_pct": { + "FAVOR": 62.71186440677966, + "AGAINST": 65.08474576271186, + "NONE": 84.39821693907875 + } + } + ], + "accuracy": 0.6627520062634567, + "macro_f1": 0.6248573052912697, + "weighted_f1": 0.6749224993922747, + "macro_precision": 0.6203728032267358, + "macro_recall": 0.6464981967532092, + "per_class_f1": { + "FAVOR": 0.5272277227722773, + "AGAINST": 0.5778454510848877, + "NONE": 0.769498742016644 + }, + "accuracy_pct": 66.27520062634567, + "macro_f1_pct": 62.48573052912697, + "weighted_f1_pct": 67.49224993922746, + "macro_precision_pct": 62.03728032267358, + "macro_recall_pct": 64.64981967532091, + "per_class_f1_pct": { + "FAVOR": 52.722772277227726, + "AGAINST": 57.78454510848877, + "NONE": 76.9498742016644 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s1__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__doc__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be95fd9b215e1213f7c975dad0ba5ef7e1dfa0cf --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__doc__s1__seed13/run.log b/outputs/trace_s5_balanced_replay__doc__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..6b3acb91409db61c407d4bb51387305dc9dac05d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 01:54:38,864 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__doc__s1__seed13 device=cuda +2026-09-02 02:24:15,061 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6260628927555552, 'average_forgetting': 0.02482500893651848, 'backward_transfer': -0.021033756545850078, 'final_mean_target_macro_f1_pct': 62.60628927555552, 'average_forgetting_pct': 2.482500893651848, 'backward_transfer_pct': -2.1033756545850077, 'final_pooled_accuracy': 0.6627520062634567, 'final_pooled_accuracy_pct': 66.27520062634567, 'final_pooled_macro_f1': 0.6248573052912697, 'final_pooled_macro_f1_pct': 62.48573052912697, 'final_pooled_weighted_f1': 0.6749224993922747, 'final_pooled_weighted_f1_pct': 67.49224993922746, 'final_pooled_macro_precision': 0.6203728032267358, 'final_pooled_macro_precision_pct': 62.03728032267358, 'final_pooled_macro_recall': 0.6464981967532092, 'final_pooled_macro_recall_pct': 64.64981967532091, 'final_pooled_per_class_f1': {'FAVOR': 0.5272277227722773, 'AGAINST': 0.5778454510848877, 'NONE': 0.769498742016644}, 'final_pooled_per_class_f1_pct': {'FAVOR': 52.722772277227726, 'AGAINST': 57.78454510848877, 'NONE': 76.9498742016644}} +2026-09-02 02:24:15,061 [INFO] __main__: efficiency={'training_wall_time_seconds': 1774.8458118368872, 'training_wall_time_minutes': 29.580763530614785, 'peak_gpu_memory_allocated_gb': 10.39700698852539, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s5_balanced_replay__doc__s2__seed13/environment.json b/outputs/trace_s5_balanced_replay__doc__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..8b59206da3feffe12e8b0f887de27f6bb7ec7922 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T02:24:23.211154+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.json b/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..abdf2f811c85ed0c29b7d27f96d4f9e30c5e2c03 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.json @@ -0,0 +1,1587 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.650326937695683, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6560366170561073, + 0.5713458795975188, + null, + null, + null, + null, + null, + null + ], + [ + 0.6456917552053215, + 0.5689984253505641, + 0.701203582607083, + null, + null, + null, + null, + null + ], + [ + 0.6472234304770189, + 0.5451148117814785, + 0.6215575361395841, + 0.6800942707642245, + null, + null, + null, + null + ], + [ + 0.6610539356453522, + 0.5284507378925155, + 0.621670468122081, + 0.6800097883070263, + 0.6505803433514277, + null, + null, + null + ], + [ + 0.6384527324926988, + 0.5136705760569336, + 0.5987870466649188, + 0.6336632888317335, + 0.5532676774291256, + 0.5722906604966338, + null, + null + ], + [ + 0.6417109973480675, + 0.5087635146504589, + 0.6053263636674339, + 0.6305624913683209, + 0.5359445325812328, + 0.508902244212873, + 0.6665093852204661, + null + ], + [ + 0.6574152952597495, + 0.5232417608267447, + 0.6329253517295949, + 0.6484062030938297, + 0.5402147308600167, + 0.5140735542576033, + 0.6505110120533816, + 0.6560230199087248 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6952735888792025, + "losses": { + "current_ce": 0.5369577404998598, + "total": 0.5369577404998598, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6584564860426929, + "macro_f1": 0.650326937695683, + "weighted_f1": 0.6683429957246021, + "macro_precision": 0.6615466592530813, + "macro_recall": 0.6760739090806451, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6535211267605634, + "NONE": 0.7196819085487077 + }, + "accuracy_pct": 65.84564860426929, + "macro_f1_pct": 65.0326937695683, + "weighted_f1_pct": 66.83429957246021, + "macro_precision_pct": 66.15466592530814, + "macro_recall_pct": 67.60739090806452, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 65.35211267605634, + "NONE": 71.96819085487077 + } + } + ], + "accuracy": 0.6584564860426929, + "macro_f1": 0.650326937695683, + "weighted_f1": 0.6683429957246021, + "macro_precision": 0.6615466592530813, + "macro_recall": 0.6760739090806451, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6535211267605634, + "NONE": 0.7196819085487077 + }, + "accuracy_pct": 65.84564860426929, + "macro_f1_pct": 65.0326937695683, + "weighted_f1_pct": 66.83429957246021, + "macro_precision_pct": 66.15466592530814, + "macro_recall_pct": 67.60739090806452, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 65.35211267605634, + "NONE": 71.96819085487077 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6814705645763666, + "losses": { + "current_ce": 0.5659300363377521, + "replay_ce": 0.013906713662334114, + "total": 0.5798367498998056, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6650246305418719, + "macro_f1": 0.6560366170561073, + "weighted_f1": 0.6742270808858206, + "macro_precision": 0.6640638385648335, + "macro_recall": 0.6804553631712235, + "per_class_f1": { + "FAVOR": 0.5689655172413793, + "AGAINST": 0.6758241758241759, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.50246305418719, + "macro_f1_pct": 65.60366170561073, + "weighted_f1_pct": 67.42270808858206, + "macro_precision_pct": 66.40638385648336, + "macro_recall_pct": 68.04553631712234, + "per_class_f1_pct": { + "FAVOR": 56.896551724137936, + "AGAINST": 67.58241758241759, + "NONE": 72.33201581027669 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.624813153961136, + "macro_f1": 0.5713458795975188, + "weighted_f1": 0.6390692492648636, + "macro_precision": 0.5827844497199336, + "macro_recall": 0.5967379771510352, + "per_class_f1": { + "FAVOR": 0.5545023696682464, + "AGAINST": 0.4015444015444015, + "NONE": 0.7579908675799086 + }, + "accuracy_pct": 62.4813153961136, + "macro_f1_pct": 57.13458795975188, + "weighted_f1_pct": 63.90692492648636, + "macro_precision_pct": 58.27844497199336, + "macro_recall_pct": 59.67379771510352, + "per_class_f1_pct": { + "FAVOR": 55.45023696682464, + "AGAINST": 40.15444015444015, + "NONE": 75.79908675799086 + } + } + ], + "accuracy": 0.6439749608763693, + "macro_f1": 0.6219143302393711, + "weighted_f1": 0.6575589681911539, + "macro_precision": 0.6306472059478535, + "macro_recall": 0.6465015515326394, + "per_class_f1": { + "FAVOR": 0.561038961038961, + "AGAINST": 0.5617977528089888, + "NONE": 0.7429062768701634 + }, + "accuracy_pct": 64.39749608763692, + "macro_f1_pct": 62.191433023937115, + "weighted_f1_pct": 65.75589681911539, + "macro_precision_pct": 63.06472059478535, + "macro_recall_pct": 64.65015515326394, + "per_class_f1_pct": { + "FAVOR": 56.1038961038961, + "AGAINST": 56.17977528089888, + "NONE": 74.29062768701634 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7286487162648153, + "losses": { + "current_ce": 0.47335088935991126, + "replay_ce": 0.030791237960303468, + "total": 0.5041421271328416, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6456917552053215, + "weighted_f1": 0.6556286228195528, + "macro_precision": 0.6693890779452277, + "macro_recall": 0.6805181703375212, + "per_class_f1": { + "FAVOR": 0.5677083333333334, + "AGAINST": 0.6929577464788732, + "NONE": 0.6764091858037579 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.56917552053216, + "weighted_f1_pct": 65.56286228195528, + "macro_precision_pct": 66.93890779452278, + "macro_recall_pct": 68.05181703375212, + "per_class_f1_pct": { + "FAVOR": 56.770833333333336, + "AGAINST": 69.29577464788733, + "NONE": 67.64091858037578 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6053811659192825, + "macro_f1": 0.5689984253505641, + "weighted_f1": 0.624063906020043, + "macro_precision": 0.596054558754073, + "macro_recall": 0.6027845046277203, + "per_class_f1": { + "FAVOR": 0.5374449339207048, + "AGAINST": 0.44696969696969696, + "NONE": 0.7225806451612903 + }, + "accuracy_pct": 60.53811659192825, + "macro_f1_pct": 56.89984253505641, + "weighted_f1_pct": 62.4063906020043, + "macro_precision_pct": 59.6054558754073, + "macro_recall_pct": 60.27845046277203, + "per_class_f1_pct": { + "FAVOR": 53.74449339207048, + "AGAINST": 44.696969696969695, + "NONE": 72.25806451612902 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7364185110663984, + "macro_f1": 0.701203582607083, + "weighted_f1": 0.7441807379017134, + "macro_precision": 0.7002499720990216, + "macro_recall": 0.7119377998688344, + "per_class_f1": { + "FAVOR": 0.6044776119402985, + "AGAINST": 0.6605504587155964, + "NONE": 0.8385826771653543 + }, + "accuracy_pct": 73.64185110663985, + "macro_f1_pct": 70.1203582607083, + "weighted_f1_pct": 74.41807379017133, + "macro_precision_pct": 70.02499720990217, + "macro_recall_pct": 71.19377998688344, + "per_class_f1_pct": { + "FAVOR": 60.447761194029844, + "AGAINST": 66.05504587155964, + "NONE": 83.85826771653542 + } + } + ], + "accuracy": 0.6563380281690141, + "macro_f1": 0.6388710987978061, + "weighted_f1": 0.6708381973609397, + "macro_precision": 0.6524655764809218, + "macro_recall": 0.6663428712028788, + "per_class_f1": { + "FAVOR": 0.5641952983725136, + "AGAINST": 0.6069295101553166, + "NONE": 0.7454884878655881 + }, + "accuracy_pct": 65.63380281690141, + "macro_f1_pct": 63.88710987978061, + "weighted_f1_pct": 67.08381973609397, + "macro_precision_pct": 65.24655764809218, + "macro_recall_pct": 66.63428712028788, + "per_class_f1_pct": { + "FAVOR": 56.41952983725136, + "AGAINST": 60.69295101553166, + "NONE": 74.5488487865588 + } + }, + { + "after_task": 4, + "best_epoch": 4, + "best_dev_macro_f1": 0.7765688493201192, + "losses": { + "current_ce": 0.4071689939674209, + "replay_ce": 0.016352145824188768, + "total": 0.42352114005708225, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6502463054187192, + "macro_f1": 0.6472234304770189, + "weighted_f1": 0.6558487983995327, + "macro_precision": 0.6673761865215805, + "macro_recall": 0.6847718604454673, + "per_class_f1": { + "FAVOR": 0.5785123966942148, + "AGAINST": 0.6894736842105263, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 65.02463054187191, + "macro_f1_pct": 64.7223430477019, + "weighted_f1_pct": 65.58487983995327, + "macro_precision_pct": 66.73761865215805, + "macro_recall_pct": 68.47718604454673, + "per_class_f1_pct": { + "FAVOR": 57.85123966942148, + "AGAINST": 68.94736842105263, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5680119581464873, + "macro_f1": 0.5451148117814785, + "weighted_f1": 0.5940077052184676, + "macro_precision": 0.5754913914671476, + "macro_recall": 0.5861377541368659, + "per_class_f1": { + "FAVOR": 0.4972972972972973, + "AGAINST": 0.45454545454545453, + "NONE": 0.6835016835016835 + }, + "accuracy_pct": 56.80119581464873, + "macro_f1_pct": 54.51148117814785, + "weighted_f1_pct": 59.40077052184676, + "macro_precision_pct": 57.54913914671476, + "macro_recall_pct": 58.61377541368659, + "per_class_f1_pct": { + "FAVOR": 49.72972972972973, + "AGAINST": 45.45454545454545, + "NONE": 68.35016835016835 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6378269617706237, + "macro_f1": 0.6215575361395841, + "weighted_f1": 0.6518170459755632, + "macro_precision": 0.6367075768273373, + "macro_recall": 0.6649879764822294, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.6006825938566553, + "NONE": 0.7185354691075515 + }, + "accuracy_pct": 63.78269617706237, + "macro_f1_pct": 62.15575361395841, + "weighted_f1_pct": 65.18170459755632, + "macro_precision_pct": 63.67075768273372, + "macro_recall_pct": 66.49879764822295, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 60.068259385665534, + "NONE": 71.85354691075516 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7043189368770764, + "macro_f1": 0.6800942707642245, + "weighted_f1": 0.7160951884645446, + "macro_precision": 0.6927192846274982, + "macro_recall": 0.7012413142350026, + "per_class_f1": { + "FAVOR": 0.6504854368932039, + "AGAINST": 0.6073298429319371, + "NONE": 0.7824675324675324 + }, + "accuracy_pct": 70.43189368770764, + "macro_f1_pct": 68.00942707642244, + "weighted_f1_pct": 71.60951884645445, + "macro_precision_pct": 69.27192846274983, + "macro_recall_pct": 70.12413142350026, + "per_class_f1_pct": { + "FAVOR": 65.0485436893204, + "AGAINST": 60.73298429319372, + "NONE": 78.24675324675324 + } + } + ], + "accuracy": 0.638199411022297, + "macro_f1": 0.6212250078450148, + "weighted_f1": 0.6518263303024311, + "macro_precision": 0.6322872365480602, + "macro_recall": 0.6602623633754909, + "per_class_f1": { + "FAVOR": 0.5586034912718204, + "AGAINST": 0.5878236529041287, + "NONE": 0.7172478793590952 + }, + "accuracy_pct": 63.819941102229706, + "macro_f1_pct": 62.122500784501476, + "weighted_f1_pct": 65.18263303024311, + "macro_precision_pct": 63.22872365480602, + "macro_recall_pct": 66.0262363375491, + "per_class_f1_pct": { + "FAVOR": 55.86034912718204, + "AGAINST": 58.782365290412876, + "NONE": 71.72478793590952 + } + }, + { + "after_task": 5, + "best_epoch": 6, + "best_dev_macro_f1": 0.6707087421943244, + "losses": { + "current_ce": 0.5392686740998869, + "replay_ce": 0.02804929068579373, + "total": 0.567317964942367, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6699507389162561, + "macro_f1": 0.6610539356453522, + "weighted_f1": 0.6810092012257247, + "macro_precision": 0.6746804909292684, + "macro_recall": 0.6835922633534391, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6646525679758308, + "NONE": 0.7378640776699029 + }, + "accuracy_pct": 66.99507389162561, + "macro_f1_pct": 66.10539356453522, + "weighted_f1_pct": 68.10092012257248, + "macro_precision_pct": 67.46804909292685, + "macro_recall_pct": 68.35922633534392, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 66.46525679758308, + "NONE": 73.7864077669903 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5560538116591929, + "macro_f1": 0.5284507378925155, + "weighted_f1": 0.5792316117663827, + "macro_precision": 0.5568831388466092, + "macro_recall": 0.5650125078284083, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.41533546325878595, + "NONE": 0.6700167504187605 + }, + "accuracy_pct": 55.60538116591929, + "macro_f1_pct": 52.84507378925155, + "weighted_f1_pct": 57.923161176638274, + "macro_precision_pct": 55.68831388466092, + "macro_recall_pct": 56.50125078284083, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 41.533546325878596, + "NONE": 67.00167504187606 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.635814889336016, + "macro_f1": 0.621670468122081, + "weighted_f1": 0.6487204228957977, + "macro_precision": 0.6320083442624393, + "macro_recall": 0.6634945290117703, + "per_class_f1": { + "FAVOR": 0.5304659498207885, + "AGAINST": 0.6254545454545455, + "NONE": 0.7090909090909091 + }, + "accuracy_pct": 63.581488933601605, + "macro_f1_pct": 62.167046812208106, + "weighted_f1_pct": 64.87204228957977, + "macro_precision_pct": 63.20083442624394, + "macro_recall_pct": 66.34945290117703, + "per_class_f1_pct": { + "FAVOR": 53.04659498207885, + "AGAINST": 62.54545454545455, + "NONE": 70.9090909090909 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7059800664451827, + "macro_f1": 0.6800097883070263, + "weighted_f1": 0.7160229482480209, + "macro_precision": 0.6681104936253678, + "macro_recall": 0.718569382959826, + "per_class_f1": { + "FAVOR": 0.6360153256704981, + "AGAINST": 0.6242424242424243, + "NONE": 0.7797716150081566 + }, + "accuracy_pct": 70.59800664451828, + "macro_f1_pct": 68.00097883070264, + "weighted_f1_pct": 71.60229482480209, + "macro_precision_pct": 66.81104936253678, + "macro_recall_pct": 71.8569382959826, + "per_class_f1_pct": { + "FAVOR": 63.601532567049816, + "AGAINST": 62.42424242424243, + "NONE": 77.97716150081567 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.7102922490470139, + "macro_f1": 0.6505803433514277, + "weighted_f1": 0.7186789802717299, + "macro_precision": 0.6385932409343807, + "macro_recall": 0.678668309822994, + "per_class_f1": { + "FAVOR": 0.6024096385542169, + "AGAINST": 0.536144578313253, + "NONE": 0.8131868131868132 + }, + "accuracy_pct": 71.0292249047014, + "macro_f1_pct": 65.05803433514276, + "weighted_f1_pct": 71.867898027173, + "macro_precision_pct": 63.859324093438076, + "macro_recall_pct": 67.8668309822994, + "per_class_f1_pct": { + "FAVOR": 60.24096385542169, + "AGAINST": 53.6144578313253, + "NONE": 81.31868131868131 + } + } + ], + "accuracy": 0.6573957016434893, + "macro_f1": 0.6296226140311755, + "weighted_f1": 0.671518865213296, + "macro_precision": 0.628909470155761, + "macro_recall": 0.66316733753996, + "per_class_f1": { + "FAVOR": 0.5645933014354066, + "AGAINST": 0.573055028462998, + "NONE": 0.751219512195122 + }, + "accuracy_pct": 65.73957016434892, + "macro_f1_pct": 62.962261403117544, + "weighted_f1_pct": 67.1518865213296, + "macro_precision_pct": 62.8909470155761, + "macro_recall_pct": 66.316733753996, + "per_class_f1_pct": { + "FAVOR": 56.45933014354066, + "AGAINST": 57.3055028462998, + "NONE": 75.1219512195122 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.706558393776937, + "losses": { + "current_ce": 0.5090020271816424, + "replay_ce": 0.028507245545938892, + "total": 0.5375092730044373, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.6384527324926988, + "weighted_f1": 0.6524882237107746, + "macro_precision": 0.654497559020172, + "macro_recall": 0.6687910816814675, + "per_class_f1": { + "FAVOR": 0.5619834710743802, + "AGAINST": 0.6648501362397821, + "NONE": 0.6885245901639344 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 63.84527324926989, + "weighted_f1_pct": 65.24882237107747, + "macro_precision_pct": 65.4497559020172, + "macro_recall_pct": 66.87910816814676, + "per_class_f1_pct": { + "FAVOR": 56.19834710743802, + "AGAINST": 66.4850136239782, + "NONE": 68.85245901639344 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5381165919282511, + "macro_f1": 0.5136705760569336, + "weighted_f1": 0.561725689571272, + "macro_precision": 0.5500778895323242, + "macro_recall": 0.5540339362235875, + "per_class_f1": { + "FAVOR": 0.5023255813953489, + "AGAINST": 0.39263803680981596, + "NONE": 0.6460481099656358 + }, + "accuracy_pct": 53.81165919282511, + "macro_f1_pct": 51.367057605693354, + "weighted_f1_pct": 56.1725689571272, + "macro_precision_pct": 55.00778895323242, + "macro_recall_pct": 55.403393622358756, + "per_class_f1_pct": { + "FAVOR": 50.23255813953489, + "AGAINST": 39.263803680981596, + "NONE": 64.60481099656357 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.613682092555332, + "macro_f1": 0.5987870466649188, + "weighted_f1": 0.627822053380213, + "macro_precision": 0.6109262663297585, + "macro_recall": 0.6379235557396478, + "per_class_f1": { + "FAVOR": 0.505338078291815, + "AGAINST": 0.5985401459854015, + "NONE": 0.6924829157175398 + }, + "accuracy_pct": 61.3682092555332, + "macro_f1_pct": 59.87870466649188, + "weighted_f1_pct": 62.7822053380213, + "macro_precision_pct": 61.09262663297586, + "macro_recall_pct": 63.79235557396478, + "per_class_f1_pct": { + "FAVOR": 50.5338078291815, + "AGAINST": 59.854014598540154, + "NONE": 69.24829157175398 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.659468438538206, + "macro_f1": 0.6336632888317335, + "weighted_f1": 0.6720157439024904, + "macro_precision": 0.6300948851786172, + "macro_recall": 0.6765347291161655, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.615819209039548, + "NONE": 0.733446519524618 + }, + "accuracy_pct": 65.9468438538206, + "macro_f1_pct": 63.36632888317335, + "weighted_f1_pct": 67.20157439024904, + "macro_precision_pct": 63.009488517861726, + "macro_recall_pct": 67.65347291161655, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 61.5819209039548, + "NONE": 73.3446519524618 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5984752223634053, + "macro_f1": 0.5532676774291256, + "weighted_f1": 0.6262648371933532, + "macro_precision": 0.5657435994223377, + "macro_recall": 0.603718007966374, + "per_class_f1": { + "FAVOR": 0.46078431372549017, + "AGAINST": 0.47513812154696133, + "NONE": 0.7238805970149254 + }, + "accuracy_pct": 59.84752223634053, + "macro_f1_pct": 55.326767742912565, + "weighted_f1_pct": 62.626483719335326, + "macro_precision_pct": 56.57435994223376, + "macro_recall_pct": 60.3718007966374, + "per_class_f1_pct": { + "FAVOR": 46.07843137254902, + "AGAINST": 47.51381215469613, + "NONE": 72.38805970149254 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6101231190150479, + "macro_f1": 0.5722906604966338, + "weighted_f1": 0.6253506513169128, + "macro_precision": 0.5763600927416497, + "macro_recall": 0.6047318965799046, + "per_class_f1": { + "FAVOR": 0.42696629213483145, + "AGAINST": 0.5954198473282443, + "NONE": 0.6944858420268256 + }, + "accuracy_pct": 61.01231190150479, + "macro_f1_pct": 57.22906604966338, + "weighted_f1_pct": 62.53506513169128, + "macro_precision_pct": 57.63600927416497, + "macro_recall_pct": 60.47318965799045, + "per_class_f1_pct": { + "FAVOR": 42.69662921348314, + "AGAINST": 59.541984732824424, + "NONE": 69.44858420268257 + } + } + ], + "accuracy": 0.6087291399229782, + "macro_f1": 0.587002540393159, + "weighted_f1": 0.6261705731555443, + "macro_precision": 0.597415401768073, + "macro_recall": 0.627127604968689, + "per_class_f1": { + "FAVOR": 0.5014925373134328, + "AGAINST": 0.5609424558223833, + "NONE": 0.6985726280436608 + }, + "accuracy_pct": 60.87291399229782, + "macro_f1_pct": 58.7002540393159, + "weighted_f1_pct": 62.617057315554426, + "macro_precision_pct": 59.7415401768073, + "macro_recall_pct": 62.7127604968689, + "per_class_f1_pct": { + "FAVOR": 50.14925373134328, + "AGAINST": 56.09424558223834, + "NONE": 69.85726280436609 + } + }, + { + "after_task": 7, + "best_epoch": 13, + "best_dev_macro_f1": 0.6420317815331883, + "losses": { + "current_ce": 0.49913866385107947, + "replay_ce": 0.027286833096739082, + "total": 0.5264254988304207, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.645320197044335, + "macro_f1": 0.6417109973480675, + "weighted_f1": 0.6478900637019774, + "macro_precision": 0.6600809037541315, + "macro_recall": 0.680843515197465, + "per_class_f1": { + "FAVOR": 0.5872093023255814, + "AGAINST": 0.6783042394014963, + "NONE": 0.6596194503171248 + }, + "accuracy_pct": 64.5320197044335, + "macro_f1_pct": 64.17109973480675, + "weighted_f1_pct": 64.78900637019774, + "macro_precision_pct": 66.00809037541315, + "macro_recall_pct": 68.0843515197465, + "per_class_f1_pct": { + "FAVOR": 58.720930232558146, + "AGAINST": 67.83042394014963, + "NONE": 65.96194503171247 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5381165919282511, + "macro_f1": 0.5087635146504589, + "weighted_f1": 0.5607106249177384, + "macro_precision": 0.5527262968884489, + "macro_recall": 0.5492370165921132, + "per_class_f1": { + "FAVOR": 0.4978540772532189, + "AGAINST": 0.3767123287671233, + "NONE": 0.6517241379310345 + }, + "accuracy_pct": 53.81165919282511, + "macro_f1_pct": 50.876351465045886, + "weighted_f1_pct": 56.07106249177384, + "macro_precision_pct": 55.27262968884489, + "macro_recall_pct": 54.92370165921132, + "per_class_f1_pct": { + "FAVOR": 49.78540772532189, + "AGAINST": 37.67123287671233, + "NONE": 65.17241379310344 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6053263636674339, + "weighted_f1": 0.6333736335307105, + "macro_precision": 0.6205305113022228, + "macro_recall": 0.6487009998504251, + "per_class_f1": { + "FAVOR": 0.5144927536231884, + "AGAINST": 0.6056338028169014, + "NONE": 0.695852534562212 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.53263636674339, + "weighted_f1_pct": 63.337363353071055, + "macro_precision_pct": 62.05305113022228, + "macro_recall_pct": 64.8700999850425, + "per_class_f1_pct": { + "FAVOR": 51.449275362318836, + "AGAINST": 60.56338028169014, + "NONE": 69.5852534562212 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.659468438538206, + "macro_f1": 0.6305624913683209, + "weighted_f1": 0.6723173308051257, + "macro_precision": 0.6346952089179763, + "macro_recall": 0.6650447765770797, + "per_class_f1": { + "FAVOR": 0.5565217391304348, + "AGAINST": 0.5931758530183727, + "NONE": 0.7419898819561551 + }, + "accuracy_pct": 65.9468438538206, + "macro_f1_pct": 63.056249136832086, + "weighted_f1_pct": 67.23173308051257, + "macro_precision_pct": 63.46952089179763, + "macro_recall_pct": 66.50447765770797, + "per_class_f1_pct": { + "FAVOR": 55.65217391304348, + "AGAINST": 59.31758530183727, + "NONE": 74.19898819561551 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5359445325812328, + "weighted_f1": 0.6001677712322866, + "macro_precision": 0.5575479738539114, + "macro_recall": 0.5963751531289657, + "per_class_f1": { + "FAVOR": 0.45901639344262296, + "AGAINST": 0.46236559139784944, + "NONE": 0.6864516129032258 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 53.59445325812327, + "weighted_f1_pct": 60.01677712322866, + "macro_precision_pct": 55.75479738539114, + "macro_recall_pct": 59.637515312896575, + "per_class_f1_pct": { + "FAVOR": 45.90163934426229, + "AGAINST": 46.236559139784944, + "NONE": 68.64516129032258 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5280437756497948, + "macro_f1": 0.508902244212873, + "weighted_f1": 0.5481641270907472, + "macro_precision": 0.5370851370851372, + "macro_recall": 0.5662855538985435, + "per_class_f1": { + "FAVOR": 0.40718562874251496, + "AGAINST": 0.515625, + "NONE": 0.6038961038961039 + }, + "accuracy_pct": 52.804377564979475, + "macro_f1_pct": 50.8902244212873, + "weighted_f1_pct": 54.816412709074726, + "macro_precision_pct": 53.70851370851371, + "macro_recall_pct": 56.62855538985435, + "per_class_f1_pct": { + "FAVOR": 40.7185628742515, + "AGAINST": 51.5625, + "NONE": 60.3896103896104 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6665093852204661, + "weighted_f1": 0.6913206813836135, + "macro_precision": 0.6645412426602632, + "macro_recall": 0.6834683643821448, + "per_class_f1": { + "FAVOR": 0.6064981949458483, + "AGAINST": 0.6332046332046332, + "NONE": 0.759825327510917 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.65093852204662, + "weighted_f1_pct": 69.13206813836135, + "macro_precision_pct": 66.45412426602631, + "macro_recall_pct": 68.34683643821448, + "per_class_f1_pct": { + "FAVOR": 60.64981949458483, + "AGAINST": 63.32046332046332, + "NONE": 75.9825327510917 + } + } + ], + "accuracy": 0.5997267759562842, + "macro_f1": 0.5821248680259985, + "weighted_f1": 0.6163254991862058, + "macro_precision": 0.5965165169762537, + "macro_recall": 0.6244427639362952, + "per_class_f1": { + "FAVOR": 0.5114698385726423, + "AGAINST": 0.5517792882846861, + "NONE": 0.6831254772206669 + }, + "accuracy_pct": 59.97267759562842, + "macro_f1_pct": 58.21248680259985, + "weighted_f1_pct": 61.63254991862058, + "macro_precision_pct": 59.651651697625375, + "macro_recall_pct": 62.444276393629515, + "per_class_f1_pct": { + "FAVOR": 51.146983857264225, + "AGAINST": 55.177928828468616, + "NONE": 68.31254772206668 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7447044649702456, + "losses": { + "current_ce": 0.3590776797293163, + "replay_ce": 0.017706449472201548, + "total": 0.37678412891379215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6633825944170771, + "macro_f1": 0.6574152952597495, + "weighted_f1": 0.6705844150660821, + "macro_precision": 0.6650662443663294, + "macro_recall": 0.6853607710925531, + "per_class_f1": { + "FAVOR": 0.5955056179775281, + "AGAINST": 0.6703910614525139, + "NONE": 0.7063492063492064 + }, + "accuracy_pct": 66.33825944170772, + "macro_f1_pct": 65.74152952597495, + "weighted_f1_pct": 67.05844150660822, + "macro_precision_pct": 66.50662443663295, + "macro_recall_pct": 68.53607710925532, + "per_class_f1_pct": { + "FAVOR": 59.55056179775281, + "AGAINST": 67.0391061452514, + "NONE": 70.63492063492063 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5500747384155455, + "macro_f1": 0.5232417608267447, + "weighted_f1": 0.5755753652670312, + "macro_precision": 0.55141400304414, + "macro_recall": 0.5581599333542483, + "per_class_f1": { + "FAVOR": 0.4852941176470588, + "AGAINST": 0.4144144144144144, + "NONE": 0.6700167504187605 + }, + "accuracy_pct": 55.00747384155456, + "macro_f1_pct": 52.32417608267447, + "weighted_f1_pct": 57.557536526703124, + "macro_precision_pct": 55.141400304414, + "macro_recall_pct": 55.81599333542483, + "per_class_f1_pct": { + "FAVOR": 48.529411764705884, + "AGAINST": 41.44144144144144, + "NONE": 67.00167504187606 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6458752515090543, + "macro_f1": 0.6329253517295949, + "weighted_f1": 0.659914445503971, + "macro_precision": 0.6460895296964608, + "macro_recall": 0.6716509613061338, + "per_class_f1": { + "FAVOR": 0.5639344262295082, + "AGAINST": 0.6153846153846154, + "NONE": 0.7194570135746606 + }, + "accuracy_pct": 64.58752515090544, + "macro_f1_pct": 63.29253517295948, + "weighted_f1_pct": 65.9914445503971, + "macro_precision_pct": 64.60895296964608, + "macro_recall_pct": 67.16509613061338, + "per_class_f1_pct": { + "FAVOR": 56.393442622950815, + "AGAINST": 61.53846153846154, + "NONE": 71.94570135746606 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.6794019933554817, + "macro_f1": 0.6484062030938297, + "weighted_f1": 0.6902211238486534, + "macro_precision": 0.6401485590467154, + "macro_recall": 0.6793521663088958, + "per_class_f1": { + "FAVOR": 0.5795918367346938, + "AGAINST": 0.6046511627906976, + "NONE": 0.7609756097560976 + }, + "accuracy_pct": 67.94019933554817, + "macro_f1_pct": 64.84062030938297, + "weighted_f1_pct": 69.02211238486534, + "macro_precision_pct": 64.01485590467153, + "macro_recall_pct": 67.93521663088958, + "per_class_f1_pct": { + "FAVOR": 57.95918367346938, + "AGAINST": 60.46511627906976, + "NONE": 76.09756097560975 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5844980940279543, + "macro_f1": 0.5402147308600167, + "weighted_f1": 0.6123158027516609, + "macro_precision": 0.5595762262428928, + "macro_recall": 0.596520213023481, + "per_class_f1": { + "FAVOR": 0.4665127020785219, + "AGAINST": 0.4438040345821326, + "NONE": 0.7103274559193955 + }, + "accuracy_pct": 58.449809402795424, + "macro_f1_pct": 54.02147308600167, + "weighted_f1_pct": 61.23158027516609, + "macro_precision_pct": 55.957622624289286, + "macro_recall_pct": 59.6520213023481, + "per_class_f1_pct": { + "FAVOR": 46.65127020785219, + "AGAINST": 44.38040345821326, + "NONE": 71.03274559193954 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5348837209302325, + "macro_f1": 0.5140735542576033, + "weighted_f1": 0.5549465951201534, + "macro_precision": 0.542068008183479, + "macro_recall": 0.5686804717186371, + "per_class_f1": { + "FAVOR": 0.4049079754601227, + "AGAINST": 0.5269230769230769, + "NONE": 0.6103896103896104 + }, + "accuracy_pct": 53.48837209302325, + "macro_f1_pct": 51.40735542576033, + "weighted_f1_pct": 55.494659512015346, + "macro_precision_pct": 54.2068008183479, + "macro_recall_pct": 56.86804717186371, + "per_class_f1_pct": { + "FAVOR": 40.49079754601227, + "AGAINST": 52.69230769230769, + "NONE": 61.038961038961034 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6505110120533816, + "weighted_f1": 0.6781761626783075, + "macro_precision": 0.647541084130488, + "macro_recall": 0.662714629184784, + "per_class_f1": { + "FAVOR": 0.587360594795539, + "AGAINST": 0.609375, + "NONE": 0.7547974413646056 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 65.05110120533816, + "weighted_f1_pct": 67.81761626783074, + "macro_precision_pct": 64.75410841304881, + "macro_recall_pct": 66.27146291847839, + "per_class_f1_pct": { + "FAVOR": 58.7360594795539, + "AGAINST": 60.9375, + "NONE": 75.47974413646055 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.697350069735007, + "macro_f1": 0.6560230199087248, + "weighted_f1": 0.7122525009160348, + "macro_precision": 0.6473944105114856, + "macro_recall": 0.6941621913153161, + "per_class_f1": { + "FAVOR": 0.5405405405405406, + "AGAINST": 0.6345177664974619, + "NONE": 0.793010752688172 + }, + "accuracy_pct": 69.7350069735007, + "macro_f1_pct": 65.60230199087248, + "weighted_f1_pct": 71.22525009160347, + "macro_precision_pct": 64.73944105114856, + "macro_recall_pct": 69.4162191315316, + "per_class_f1_pct": { + "FAVOR": 54.054054054054056, + "AGAINST": 63.45177664974619, + "NONE": 79.3010752688172 + } + } + ], + "accuracy": 0.6238011352515169, + "macro_f1": 0.5999409732308351, + "weighted_f1": 0.6402964685909629, + "macro_precision": 0.6058727546321164, + "macro_recall": 0.6378250337281027, + "per_class_f1": { + "FAVOR": 0.5216072782410918, + "AGAINST": 0.5616291532690246, + "NONE": 0.7165864881823886 + }, + "accuracy_pct": 62.38011352515169, + "macro_f1_pct": 59.99409732308351, + "weighted_f1_pct": 64.02964685909629, + "macro_precision_pct": 60.58727546321164, + "macro_recall_pct": 63.782503372810275, + "per_class_f1_pct": { + "FAVOR": 52.16072782410917, + "AGAINST": 56.162915326902464, + "NONE": 71.65864881823886 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6028513659987056, + "average_forgetting": 0.04804144994311226, + "backward_transfer": -0.046509021664588084, + "final_mean_target_macro_f1_pct": 60.28513659987056, + "average_forgetting_pct": 4.804144994311226, + "backward_transfer_pct": -4.650902166458808, + "final_pooled_accuracy": 0.6238011352515169, + "final_pooled_accuracy_pct": 62.38011352515169, + "final_pooled_macro_f1": 0.5999409732308351, + "final_pooled_macro_f1_pct": 59.99409732308351, + "final_pooled_weighted_f1": 0.6402964685909629, + "final_pooled_weighted_f1_pct": 64.02964685909629, + "final_pooled_macro_precision": 0.6058727546321164, + "final_pooled_macro_precision_pct": 60.58727546321164, + "final_pooled_macro_recall": 0.6378250337281027, + "final_pooled_macro_recall_pct": 63.782503372810275, + "final_pooled_per_class_f1": { + "FAVOR": 0.5216072782410918, + "AGAINST": 0.5616291532690246, + "NONE": 0.7165864881823886 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 52.16072782410917, + "AGAINST": 56.162915326902464, + "NONE": 71.65864881823886 + } + }, + "efficiency": { + "training_wall_time_seconds": 1798.0146911861375, + "training_wall_time_minutes": 29.96691151976896, + "peak_gpu_memory_allocated_gb": 10.39700698852539, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb9c841d0fbd85ba4ae8fe72ff2962a7c8b0bb8 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s2__seed13/metrics.partial.json @@ -0,0 +1,1481 @@ +{ + "matrix": [ + [ + 0.650326937695683, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6560366170561073, + 0.5713458795975188, + null, + null, + null, + null, + null, + null + ], + [ + 0.6456917552053215, + 0.5689984253505641, + 0.701203582607083, + null, + null, + null, + null, + null + ], + [ + 0.6472234304770189, + 0.5451148117814785, + 0.6215575361395841, + 0.6800942707642245, + null, + null, + null, + null + ], + [ + 0.6610539356453522, + 0.5284507378925155, + 0.621670468122081, + 0.6800097883070263, + 0.6505803433514277, + null, + null, + null + ], + [ + 0.6384527324926988, + 0.5136705760569336, + 0.5987870466649188, + 0.6336632888317335, + 0.5532676774291256, + 0.5722906604966338, + null, + null + ], + [ + 0.6417109973480675, + 0.5087635146504589, + 0.6053263636674339, + 0.6305624913683209, + 0.5359445325812328, + 0.508902244212873, + 0.6665093852204661, + null + ], + [ + 0.6574152952597495, + 0.5232417608267447, + 0.6329253517295949, + 0.6484062030938297, + 0.5402147308600167, + 0.5140735542576033, + 0.6505110120533816, + 0.6560230199087248 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6952735888792025, + "losses": { + "current_ce": 0.5369577404998598, + "total": 0.5369577404998598, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6584564860426929, + "macro_f1": 0.650326937695683, + "weighted_f1": 0.6683429957246021, + "macro_precision": 0.6615466592530813, + "macro_recall": 0.6760739090806451, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6535211267605634, + "NONE": 0.7196819085487077 + }, + "accuracy_pct": 65.84564860426929, + "macro_f1_pct": 65.0326937695683, + "weighted_f1_pct": 66.83429957246021, + "macro_precision_pct": 66.15466592530814, + "macro_recall_pct": 67.60739090806452, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 65.35211267605634, + "NONE": 71.96819085487077 + } + } + ], + "accuracy": 0.6584564860426929, + "macro_f1": 0.650326937695683, + "weighted_f1": 0.6683429957246021, + "macro_precision": 0.6615466592530813, + "macro_recall": 0.6760739090806451, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6535211267605634, + "NONE": 0.7196819085487077 + }, + "accuracy_pct": 65.84564860426929, + "macro_f1_pct": 65.0326937695683, + "weighted_f1_pct": 66.83429957246021, + "macro_precision_pct": 66.15466592530814, + "macro_recall_pct": 67.60739090806452, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 65.35211267605634, + "NONE": 71.96819085487077 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6814705645763666, + "losses": { + "current_ce": 0.5659300363377521, + "replay_ce": 0.013906713662334114, + "total": 0.5798367498998056, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6650246305418719, + "macro_f1": 0.6560366170561073, + "weighted_f1": 0.6742270808858206, + "macro_precision": 0.6640638385648335, + "macro_recall": 0.6804553631712235, + "per_class_f1": { + "FAVOR": 0.5689655172413793, + "AGAINST": 0.6758241758241759, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.50246305418719, + "macro_f1_pct": 65.60366170561073, + "weighted_f1_pct": 67.42270808858206, + "macro_precision_pct": 66.40638385648336, + "macro_recall_pct": 68.04553631712234, + "per_class_f1_pct": { + "FAVOR": 56.896551724137936, + "AGAINST": 67.58241758241759, + "NONE": 72.33201581027669 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.624813153961136, + "macro_f1": 0.5713458795975188, + "weighted_f1": 0.6390692492648636, + "macro_precision": 0.5827844497199336, + "macro_recall": 0.5967379771510352, + "per_class_f1": { + "FAVOR": 0.5545023696682464, + "AGAINST": 0.4015444015444015, + "NONE": 0.7579908675799086 + }, + "accuracy_pct": 62.4813153961136, + "macro_f1_pct": 57.13458795975188, + "weighted_f1_pct": 63.90692492648636, + "macro_precision_pct": 58.27844497199336, + "macro_recall_pct": 59.67379771510352, + "per_class_f1_pct": { + "FAVOR": 55.45023696682464, + "AGAINST": 40.15444015444015, + "NONE": 75.79908675799086 + } + } + ], + "accuracy": 0.6439749608763693, + "macro_f1": 0.6219143302393711, + "weighted_f1": 0.6575589681911539, + "macro_precision": 0.6306472059478535, + "macro_recall": 0.6465015515326394, + "per_class_f1": { + "FAVOR": 0.561038961038961, + "AGAINST": 0.5617977528089888, + "NONE": 0.7429062768701634 + }, + "accuracy_pct": 64.39749608763692, + "macro_f1_pct": 62.191433023937115, + "weighted_f1_pct": 65.75589681911539, + "macro_precision_pct": 63.06472059478535, + "macro_recall_pct": 64.65015515326394, + "per_class_f1_pct": { + "FAVOR": 56.1038961038961, + "AGAINST": 56.17977528089888, + "NONE": 74.29062768701634 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7286487162648153, + "losses": { + "current_ce": 0.47335088935991126, + "replay_ce": 0.030791237960303468, + "total": 0.5041421271328416, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6456917552053215, + "weighted_f1": 0.6556286228195528, + "macro_precision": 0.6693890779452277, + "macro_recall": 0.6805181703375212, + "per_class_f1": { + "FAVOR": 0.5677083333333334, + "AGAINST": 0.6929577464788732, + "NONE": 0.6764091858037579 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.56917552053216, + "weighted_f1_pct": 65.56286228195528, + "macro_precision_pct": 66.93890779452278, + "macro_recall_pct": 68.05181703375212, + "per_class_f1_pct": { + "FAVOR": 56.770833333333336, + "AGAINST": 69.29577464788733, + "NONE": 67.64091858037578 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6053811659192825, + "macro_f1": 0.5689984253505641, + "weighted_f1": 0.624063906020043, + "macro_precision": 0.596054558754073, + "macro_recall": 0.6027845046277203, + "per_class_f1": { + "FAVOR": 0.5374449339207048, + "AGAINST": 0.44696969696969696, + "NONE": 0.7225806451612903 + }, + "accuracy_pct": 60.53811659192825, + "macro_f1_pct": 56.89984253505641, + "weighted_f1_pct": 62.4063906020043, + "macro_precision_pct": 59.6054558754073, + "macro_recall_pct": 60.27845046277203, + "per_class_f1_pct": { + "FAVOR": 53.74449339207048, + "AGAINST": 44.696969696969695, + "NONE": 72.25806451612902 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7364185110663984, + "macro_f1": 0.701203582607083, + "weighted_f1": 0.7441807379017134, + "macro_precision": 0.7002499720990216, + "macro_recall": 0.7119377998688344, + "per_class_f1": { + "FAVOR": 0.6044776119402985, + "AGAINST": 0.6605504587155964, + "NONE": 0.8385826771653543 + }, + "accuracy_pct": 73.64185110663985, + "macro_f1_pct": 70.1203582607083, + "weighted_f1_pct": 74.41807379017133, + "macro_precision_pct": 70.02499720990217, + "macro_recall_pct": 71.19377998688344, + "per_class_f1_pct": { + "FAVOR": 60.447761194029844, + "AGAINST": 66.05504587155964, + "NONE": 83.85826771653542 + } + } + ], + "accuracy": 0.6563380281690141, + "macro_f1": 0.6388710987978061, + "weighted_f1": 0.6708381973609397, + "macro_precision": 0.6524655764809218, + "macro_recall": 0.6663428712028788, + "per_class_f1": { + "FAVOR": 0.5641952983725136, + "AGAINST": 0.6069295101553166, + "NONE": 0.7454884878655881 + }, + "accuracy_pct": 65.63380281690141, + "macro_f1_pct": 63.88710987978061, + "weighted_f1_pct": 67.08381973609397, + "macro_precision_pct": 65.24655764809218, + "macro_recall_pct": 66.63428712028788, + "per_class_f1_pct": { + "FAVOR": 56.41952983725136, + "AGAINST": 60.69295101553166, + "NONE": 74.5488487865588 + } + }, + { + "after_task": 4, + "best_epoch": 4, + "best_dev_macro_f1": 0.7765688493201192, + "losses": { + "current_ce": 0.4071689939674209, + "replay_ce": 0.016352145824188768, + "total": 0.42352114005708225, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6502463054187192, + "macro_f1": 0.6472234304770189, + "weighted_f1": 0.6558487983995327, + "macro_precision": 0.6673761865215805, + "macro_recall": 0.6847718604454673, + "per_class_f1": { + "FAVOR": 0.5785123966942148, + "AGAINST": 0.6894736842105263, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 65.02463054187191, + "macro_f1_pct": 64.7223430477019, + "weighted_f1_pct": 65.58487983995327, + "macro_precision_pct": 66.73761865215805, + "macro_recall_pct": 68.47718604454673, + "per_class_f1_pct": { + "FAVOR": 57.85123966942148, + "AGAINST": 68.94736842105263, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5680119581464873, + "macro_f1": 0.5451148117814785, + "weighted_f1": 0.5940077052184676, + "macro_precision": 0.5754913914671476, + "macro_recall": 0.5861377541368659, + "per_class_f1": { + "FAVOR": 0.4972972972972973, + "AGAINST": 0.45454545454545453, + "NONE": 0.6835016835016835 + }, + "accuracy_pct": 56.80119581464873, + "macro_f1_pct": 54.51148117814785, + "weighted_f1_pct": 59.40077052184676, + "macro_precision_pct": 57.54913914671476, + "macro_recall_pct": 58.61377541368659, + "per_class_f1_pct": { + "FAVOR": 49.72972972972973, + "AGAINST": 45.45454545454545, + "NONE": 68.35016835016835 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6378269617706237, + "macro_f1": 0.6215575361395841, + "weighted_f1": 0.6518170459755632, + "macro_precision": 0.6367075768273373, + "macro_recall": 0.6649879764822294, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.6006825938566553, + "NONE": 0.7185354691075515 + }, + "accuracy_pct": 63.78269617706237, + "macro_f1_pct": 62.15575361395841, + "weighted_f1_pct": 65.18170459755632, + "macro_precision_pct": 63.67075768273372, + "macro_recall_pct": 66.49879764822295, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 60.068259385665534, + "NONE": 71.85354691075516 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7043189368770764, + "macro_f1": 0.6800942707642245, + "weighted_f1": 0.7160951884645446, + "macro_precision": 0.6927192846274982, + "macro_recall": 0.7012413142350026, + "per_class_f1": { + "FAVOR": 0.6504854368932039, + "AGAINST": 0.6073298429319371, + "NONE": 0.7824675324675324 + }, + "accuracy_pct": 70.43189368770764, + "macro_f1_pct": 68.00942707642244, + "weighted_f1_pct": 71.60951884645445, + "macro_precision_pct": 69.27192846274983, + "macro_recall_pct": 70.12413142350026, + "per_class_f1_pct": { + "FAVOR": 65.0485436893204, + "AGAINST": 60.73298429319372, + "NONE": 78.24675324675324 + } + } + ], + "accuracy": 0.638199411022297, + "macro_f1": 0.6212250078450148, + "weighted_f1": 0.6518263303024311, + "macro_precision": 0.6322872365480602, + "macro_recall": 0.6602623633754909, + "per_class_f1": { + "FAVOR": 0.5586034912718204, + "AGAINST": 0.5878236529041287, + "NONE": 0.7172478793590952 + }, + "accuracy_pct": 63.819941102229706, + "macro_f1_pct": 62.122500784501476, + "weighted_f1_pct": 65.18263303024311, + "macro_precision_pct": 63.22872365480602, + "macro_recall_pct": 66.0262363375491, + "per_class_f1_pct": { + "FAVOR": 55.86034912718204, + "AGAINST": 58.782365290412876, + "NONE": 71.72478793590952 + } + }, + { + "after_task": 5, + "best_epoch": 6, + "best_dev_macro_f1": 0.6707087421943244, + "losses": { + "current_ce": 0.5392686740998869, + "replay_ce": 0.02804929068579373, + "total": 0.567317964942367, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6699507389162561, + "macro_f1": 0.6610539356453522, + "weighted_f1": 0.6810092012257247, + "macro_precision": 0.6746804909292684, + "macro_recall": 0.6835922633534391, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6646525679758308, + "NONE": 0.7378640776699029 + }, + "accuracy_pct": 66.99507389162561, + "macro_f1_pct": 66.10539356453522, + "weighted_f1_pct": 68.10092012257248, + "macro_precision_pct": 67.46804909292685, + "macro_recall_pct": 68.35922633534392, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 66.46525679758308, + "NONE": 73.7864077669903 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5560538116591929, + "macro_f1": 0.5284507378925155, + "weighted_f1": 0.5792316117663827, + "macro_precision": 0.5568831388466092, + "macro_recall": 0.5650125078284083, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.41533546325878595, + "NONE": 0.6700167504187605 + }, + "accuracy_pct": 55.60538116591929, + "macro_f1_pct": 52.84507378925155, + "weighted_f1_pct": 57.923161176638274, + "macro_precision_pct": 55.68831388466092, + "macro_recall_pct": 56.50125078284083, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 41.533546325878596, + "NONE": 67.00167504187606 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.635814889336016, + "macro_f1": 0.621670468122081, + "weighted_f1": 0.6487204228957977, + "macro_precision": 0.6320083442624393, + "macro_recall": 0.6634945290117703, + "per_class_f1": { + "FAVOR": 0.5304659498207885, + "AGAINST": 0.6254545454545455, + "NONE": 0.7090909090909091 + }, + "accuracy_pct": 63.581488933601605, + "macro_f1_pct": 62.167046812208106, + "weighted_f1_pct": 64.87204228957977, + "macro_precision_pct": 63.20083442624394, + "macro_recall_pct": 66.34945290117703, + "per_class_f1_pct": { + "FAVOR": 53.04659498207885, + "AGAINST": 62.54545454545455, + "NONE": 70.9090909090909 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7059800664451827, + "macro_f1": 0.6800097883070263, + "weighted_f1": 0.7160229482480209, + "macro_precision": 0.6681104936253678, + "macro_recall": 0.718569382959826, + "per_class_f1": { + "FAVOR": 0.6360153256704981, + "AGAINST": 0.6242424242424243, + "NONE": 0.7797716150081566 + }, + "accuracy_pct": 70.59800664451828, + "macro_f1_pct": 68.00097883070264, + "weighted_f1_pct": 71.60229482480209, + "macro_precision_pct": 66.81104936253678, + "macro_recall_pct": 71.8569382959826, + "per_class_f1_pct": { + "FAVOR": 63.601532567049816, + "AGAINST": 62.42424242424243, + "NONE": 77.97716150081567 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.7102922490470139, + "macro_f1": 0.6505803433514277, + "weighted_f1": 0.7186789802717299, + "macro_precision": 0.6385932409343807, + "macro_recall": 0.678668309822994, + "per_class_f1": { + "FAVOR": 0.6024096385542169, + "AGAINST": 0.536144578313253, + "NONE": 0.8131868131868132 + }, + "accuracy_pct": 71.0292249047014, + "macro_f1_pct": 65.05803433514276, + "weighted_f1_pct": 71.867898027173, + "macro_precision_pct": 63.859324093438076, + "macro_recall_pct": 67.8668309822994, + "per_class_f1_pct": { + "FAVOR": 60.24096385542169, + "AGAINST": 53.6144578313253, + "NONE": 81.31868131868131 + } + } + ], + "accuracy": 0.6573957016434893, + "macro_f1": 0.6296226140311755, + "weighted_f1": 0.671518865213296, + "macro_precision": 0.628909470155761, + "macro_recall": 0.66316733753996, + "per_class_f1": { + "FAVOR": 0.5645933014354066, + "AGAINST": 0.573055028462998, + "NONE": 0.751219512195122 + }, + "accuracy_pct": 65.73957016434892, + "macro_f1_pct": 62.962261403117544, + "weighted_f1_pct": 67.1518865213296, + "macro_precision_pct": 62.8909470155761, + "macro_recall_pct": 66.316733753996, + "per_class_f1_pct": { + "FAVOR": 56.45933014354066, + "AGAINST": 57.3055028462998, + "NONE": 75.1219512195122 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.706558393776937, + "losses": { + "current_ce": 0.5090020271816424, + "replay_ce": 0.028507245545938892, + "total": 0.5375092730044373, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.6384527324926988, + "weighted_f1": 0.6524882237107746, + "macro_precision": 0.654497559020172, + "macro_recall": 0.6687910816814675, + "per_class_f1": { + "FAVOR": 0.5619834710743802, + "AGAINST": 0.6648501362397821, + "NONE": 0.6885245901639344 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 63.84527324926989, + "weighted_f1_pct": 65.24882237107747, + "macro_precision_pct": 65.4497559020172, + "macro_recall_pct": 66.87910816814676, + "per_class_f1_pct": { + "FAVOR": 56.19834710743802, + "AGAINST": 66.4850136239782, + "NONE": 68.85245901639344 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5381165919282511, + "macro_f1": 0.5136705760569336, + "weighted_f1": 0.561725689571272, + "macro_precision": 0.5500778895323242, + "macro_recall": 0.5540339362235875, + "per_class_f1": { + "FAVOR": 0.5023255813953489, + "AGAINST": 0.39263803680981596, + "NONE": 0.6460481099656358 + }, + "accuracy_pct": 53.81165919282511, + "macro_f1_pct": 51.367057605693354, + "weighted_f1_pct": 56.1725689571272, + "macro_precision_pct": 55.00778895323242, + "macro_recall_pct": 55.403393622358756, + "per_class_f1_pct": { + "FAVOR": 50.23255813953489, + "AGAINST": 39.263803680981596, + "NONE": 64.60481099656357 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.613682092555332, + "macro_f1": 0.5987870466649188, + "weighted_f1": 0.627822053380213, + "macro_precision": 0.6109262663297585, + "macro_recall": 0.6379235557396478, + "per_class_f1": { + "FAVOR": 0.505338078291815, + "AGAINST": 0.5985401459854015, + "NONE": 0.6924829157175398 + }, + "accuracy_pct": 61.3682092555332, + "macro_f1_pct": 59.87870466649188, + "weighted_f1_pct": 62.7822053380213, + "macro_precision_pct": 61.09262663297586, + "macro_recall_pct": 63.79235557396478, + "per_class_f1_pct": { + "FAVOR": 50.5338078291815, + "AGAINST": 59.854014598540154, + "NONE": 69.24829157175398 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.659468438538206, + "macro_f1": 0.6336632888317335, + "weighted_f1": 0.6720157439024904, + "macro_precision": 0.6300948851786172, + "macro_recall": 0.6765347291161655, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.615819209039548, + "NONE": 0.733446519524618 + }, + "accuracy_pct": 65.9468438538206, + "macro_f1_pct": 63.36632888317335, + "weighted_f1_pct": 67.20157439024904, + "macro_precision_pct": 63.009488517861726, + "macro_recall_pct": 67.65347291161655, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 61.5819209039548, + "NONE": 73.3446519524618 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5984752223634053, + "macro_f1": 0.5532676774291256, + "weighted_f1": 0.6262648371933532, + "macro_precision": 0.5657435994223377, + "macro_recall": 0.603718007966374, + "per_class_f1": { + "FAVOR": 0.46078431372549017, + "AGAINST": 0.47513812154696133, + "NONE": 0.7238805970149254 + }, + "accuracy_pct": 59.84752223634053, + "macro_f1_pct": 55.326767742912565, + "weighted_f1_pct": 62.626483719335326, + "macro_precision_pct": 56.57435994223376, + "macro_recall_pct": 60.3718007966374, + "per_class_f1_pct": { + "FAVOR": 46.07843137254902, + "AGAINST": 47.51381215469613, + "NONE": 72.38805970149254 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6101231190150479, + "macro_f1": 0.5722906604966338, + "weighted_f1": 0.6253506513169128, + "macro_precision": 0.5763600927416497, + "macro_recall": 0.6047318965799046, + "per_class_f1": { + "FAVOR": 0.42696629213483145, + "AGAINST": 0.5954198473282443, + "NONE": 0.6944858420268256 + }, + "accuracy_pct": 61.01231190150479, + "macro_f1_pct": 57.22906604966338, + "weighted_f1_pct": 62.53506513169128, + "macro_precision_pct": 57.63600927416497, + "macro_recall_pct": 60.47318965799045, + "per_class_f1_pct": { + "FAVOR": 42.69662921348314, + "AGAINST": 59.541984732824424, + "NONE": 69.44858420268257 + } + } + ], + "accuracy": 0.6087291399229782, + "macro_f1": 0.587002540393159, + "weighted_f1": 0.6261705731555443, + "macro_precision": 0.597415401768073, + "macro_recall": 0.627127604968689, + "per_class_f1": { + "FAVOR": 0.5014925373134328, + "AGAINST": 0.5609424558223833, + "NONE": 0.6985726280436608 + }, + "accuracy_pct": 60.87291399229782, + "macro_f1_pct": 58.7002540393159, + "weighted_f1_pct": 62.617057315554426, + "macro_precision_pct": 59.7415401768073, + "macro_recall_pct": 62.7127604968689, + "per_class_f1_pct": { + "FAVOR": 50.14925373134328, + "AGAINST": 56.09424558223834, + "NONE": 69.85726280436609 + } + }, + { + "after_task": 7, + "best_epoch": 13, + "best_dev_macro_f1": 0.6420317815331883, + "losses": { + "current_ce": 0.49913866385107947, + "replay_ce": 0.027286833096739082, + "total": 0.5264254988304207, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.645320197044335, + "macro_f1": 0.6417109973480675, + "weighted_f1": 0.6478900637019774, + "macro_precision": 0.6600809037541315, + "macro_recall": 0.680843515197465, + "per_class_f1": { + "FAVOR": 0.5872093023255814, + "AGAINST": 0.6783042394014963, + "NONE": 0.6596194503171248 + }, + "accuracy_pct": 64.5320197044335, + "macro_f1_pct": 64.17109973480675, + "weighted_f1_pct": 64.78900637019774, + "macro_precision_pct": 66.00809037541315, + "macro_recall_pct": 68.0843515197465, + "per_class_f1_pct": { + "FAVOR": 58.720930232558146, + "AGAINST": 67.83042394014963, + "NONE": 65.96194503171247 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5381165919282511, + "macro_f1": 0.5087635146504589, + "weighted_f1": 0.5607106249177384, + "macro_precision": 0.5527262968884489, + "macro_recall": 0.5492370165921132, + "per_class_f1": { + "FAVOR": 0.4978540772532189, + "AGAINST": 0.3767123287671233, + "NONE": 0.6517241379310345 + }, + "accuracy_pct": 53.81165919282511, + "macro_f1_pct": 50.876351465045886, + "weighted_f1_pct": 56.07106249177384, + "macro_precision_pct": 55.27262968884489, + "macro_recall_pct": 54.92370165921132, + "per_class_f1_pct": { + "FAVOR": 49.78540772532189, + "AGAINST": 37.67123287671233, + "NONE": 65.17241379310344 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6053263636674339, + "weighted_f1": 0.6333736335307105, + "macro_precision": 0.6205305113022228, + "macro_recall": 0.6487009998504251, + "per_class_f1": { + "FAVOR": 0.5144927536231884, + "AGAINST": 0.6056338028169014, + "NONE": 0.695852534562212 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.53263636674339, + "weighted_f1_pct": 63.337363353071055, + "macro_precision_pct": 62.05305113022228, + "macro_recall_pct": 64.8700999850425, + "per_class_f1_pct": { + "FAVOR": 51.449275362318836, + "AGAINST": 60.56338028169014, + "NONE": 69.5852534562212 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.659468438538206, + "macro_f1": 0.6305624913683209, + "weighted_f1": 0.6723173308051257, + "macro_precision": 0.6346952089179763, + "macro_recall": 0.6650447765770797, + "per_class_f1": { + "FAVOR": 0.5565217391304348, + "AGAINST": 0.5931758530183727, + "NONE": 0.7419898819561551 + }, + "accuracy_pct": 65.9468438538206, + "macro_f1_pct": 63.056249136832086, + "weighted_f1_pct": 67.23173308051257, + "macro_precision_pct": 63.46952089179763, + "macro_recall_pct": 66.50447765770797, + "per_class_f1_pct": { + "FAVOR": 55.65217391304348, + "AGAINST": 59.31758530183727, + "NONE": 74.19898819561551 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5359445325812328, + "weighted_f1": 0.6001677712322866, + "macro_precision": 0.5575479738539114, + "macro_recall": 0.5963751531289657, + "per_class_f1": { + "FAVOR": 0.45901639344262296, + "AGAINST": 0.46236559139784944, + "NONE": 0.6864516129032258 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 53.59445325812327, + "weighted_f1_pct": 60.01677712322866, + "macro_precision_pct": 55.75479738539114, + "macro_recall_pct": 59.637515312896575, + "per_class_f1_pct": { + "FAVOR": 45.90163934426229, + "AGAINST": 46.236559139784944, + "NONE": 68.64516129032258 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5280437756497948, + "macro_f1": 0.508902244212873, + "weighted_f1": 0.5481641270907472, + "macro_precision": 0.5370851370851372, + "macro_recall": 0.5662855538985435, + "per_class_f1": { + "FAVOR": 0.40718562874251496, + "AGAINST": 0.515625, + "NONE": 0.6038961038961039 + }, + "accuracy_pct": 52.804377564979475, + "macro_f1_pct": 50.8902244212873, + "weighted_f1_pct": 54.816412709074726, + "macro_precision_pct": 53.70851370851371, + "macro_recall_pct": 56.62855538985435, + "per_class_f1_pct": { + "FAVOR": 40.7185628742515, + "AGAINST": 51.5625, + "NONE": 60.3896103896104 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6665093852204661, + "weighted_f1": 0.6913206813836135, + "macro_precision": 0.6645412426602632, + "macro_recall": 0.6834683643821448, + "per_class_f1": { + "FAVOR": 0.6064981949458483, + "AGAINST": 0.6332046332046332, + "NONE": 0.759825327510917 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.65093852204662, + "weighted_f1_pct": 69.13206813836135, + "macro_precision_pct": 66.45412426602631, + "macro_recall_pct": 68.34683643821448, + "per_class_f1_pct": { + "FAVOR": 60.64981949458483, + "AGAINST": 63.32046332046332, + "NONE": 75.9825327510917 + } + } + ], + "accuracy": 0.5997267759562842, + "macro_f1": 0.5821248680259985, + "weighted_f1": 0.6163254991862058, + "macro_precision": 0.5965165169762537, + "macro_recall": 0.6244427639362952, + "per_class_f1": { + "FAVOR": 0.5114698385726423, + "AGAINST": 0.5517792882846861, + "NONE": 0.6831254772206669 + }, + "accuracy_pct": 59.97267759562842, + "macro_f1_pct": 58.21248680259985, + "weighted_f1_pct": 61.63254991862058, + "macro_precision_pct": 59.651651697625375, + "macro_recall_pct": 62.444276393629515, + "per_class_f1_pct": { + "FAVOR": 51.146983857264225, + "AGAINST": 55.177928828468616, + "NONE": 68.31254772206668 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7447044649702456, + "losses": { + "current_ce": 0.3590776797293163, + "replay_ce": 0.017706449472201548, + "total": 0.37678412891379215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6633825944170771, + "macro_f1": 0.6574152952597495, + "weighted_f1": 0.6705844150660821, + "macro_precision": 0.6650662443663294, + "macro_recall": 0.6853607710925531, + "per_class_f1": { + "FAVOR": 0.5955056179775281, + "AGAINST": 0.6703910614525139, + "NONE": 0.7063492063492064 + }, + "accuracy_pct": 66.33825944170772, + "macro_f1_pct": 65.74152952597495, + "weighted_f1_pct": 67.05844150660822, + "macro_precision_pct": 66.50662443663295, + "macro_recall_pct": 68.53607710925532, + "per_class_f1_pct": { + "FAVOR": 59.55056179775281, + "AGAINST": 67.0391061452514, + "NONE": 70.63492063492063 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5500747384155455, + "macro_f1": 0.5232417608267447, + "weighted_f1": 0.5755753652670312, + "macro_precision": 0.55141400304414, + "macro_recall": 0.5581599333542483, + "per_class_f1": { + "FAVOR": 0.4852941176470588, + "AGAINST": 0.4144144144144144, + "NONE": 0.6700167504187605 + }, + "accuracy_pct": 55.00747384155456, + "macro_f1_pct": 52.32417608267447, + "weighted_f1_pct": 57.557536526703124, + "macro_precision_pct": 55.141400304414, + "macro_recall_pct": 55.81599333542483, + "per_class_f1_pct": { + "FAVOR": 48.529411764705884, + "AGAINST": 41.44144144144144, + "NONE": 67.00167504187606 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6458752515090543, + "macro_f1": 0.6329253517295949, + "weighted_f1": 0.659914445503971, + "macro_precision": 0.6460895296964608, + "macro_recall": 0.6716509613061338, + "per_class_f1": { + "FAVOR": 0.5639344262295082, + "AGAINST": 0.6153846153846154, + "NONE": 0.7194570135746606 + }, + "accuracy_pct": 64.58752515090544, + "macro_f1_pct": 63.29253517295948, + "weighted_f1_pct": 65.9914445503971, + "macro_precision_pct": 64.60895296964608, + "macro_recall_pct": 67.16509613061338, + "per_class_f1_pct": { + "FAVOR": 56.393442622950815, + "AGAINST": 61.53846153846154, + "NONE": 71.94570135746606 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.6794019933554817, + "macro_f1": 0.6484062030938297, + "weighted_f1": 0.6902211238486534, + "macro_precision": 0.6401485590467154, + "macro_recall": 0.6793521663088958, + "per_class_f1": { + "FAVOR": 0.5795918367346938, + "AGAINST": 0.6046511627906976, + "NONE": 0.7609756097560976 + }, + "accuracy_pct": 67.94019933554817, + "macro_f1_pct": 64.84062030938297, + "weighted_f1_pct": 69.02211238486534, + "macro_precision_pct": 64.01485590467153, + "macro_recall_pct": 67.93521663088958, + "per_class_f1_pct": { + "FAVOR": 57.95918367346938, + "AGAINST": 60.46511627906976, + "NONE": 76.09756097560975 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.5844980940279543, + "macro_f1": 0.5402147308600167, + "weighted_f1": 0.6123158027516609, + "macro_precision": 0.5595762262428928, + "macro_recall": 0.596520213023481, + "per_class_f1": { + "FAVOR": 0.4665127020785219, + "AGAINST": 0.4438040345821326, + "NONE": 0.7103274559193955 + }, + "accuracy_pct": 58.449809402795424, + "macro_f1_pct": 54.02147308600167, + "weighted_f1_pct": 61.23158027516609, + "macro_precision_pct": 55.957622624289286, + "macro_recall_pct": 59.6520213023481, + "per_class_f1_pct": { + "FAVOR": 46.65127020785219, + "AGAINST": 44.38040345821326, + "NONE": 71.03274559193954 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5348837209302325, + "macro_f1": 0.5140735542576033, + "weighted_f1": 0.5549465951201534, + "macro_precision": 0.542068008183479, + "macro_recall": 0.5686804717186371, + "per_class_f1": { + "FAVOR": 0.4049079754601227, + "AGAINST": 0.5269230769230769, + "NONE": 0.6103896103896104 + }, + "accuracy_pct": 53.48837209302325, + "macro_f1_pct": 51.40735542576033, + "weighted_f1_pct": 55.494659512015346, + "macro_precision_pct": 54.2068008183479, + "macro_recall_pct": 56.86804717186371, + "per_class_f1_pct": { + "FAVOR": 40.49079754601227, + "AGAINST": 52.69230769230769, + "NONE": 61.038961038961034 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6505110120533816, + "weighted_f1": 0.6781761626783075, + "macro_precision": 0.647541084130488, + "macro_recall": 0.662714629184784, + "per_class_f1": { + "FAVOR": 0.587360594795539, + "AGAINST": 0.609375, + "NONE": 0.7547974413646056 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 65.05110120533816, + "weighted_f1_pct": 67.81761626783074, + "macro_precision_pct": 64.75410841304881, + "macro_recall_pct": 66.27146291847839, + "per_class_f1_pct": { + "FAVOR": 58.7360594795539, + "AGAINST": 60.9375, + "NONE": 75.47974413646055 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.697350069735007, + "macro_f1": 0.6560230199087248, + "weighted_f1": 0.7122525009160348, + "macro_precision": 0.6473944105114856, + "macro_recall": 0.6941621913153161, + "per_class_f1": { + "FAVOR": 0.5405405405405406, + "AGAINST": 0.6345177664974619, + "NONE": 0.793010752688172 + }, + "accuracy_pct": 69.7350069735007, + "macro_f1_pct": 65.60230199087248, + "weighted_f1_pct": 71.22525009160347, + "macro_precision_pct": 64.73944105114856, + "macro_recall_pct": 69.4162191315316, + "per_class_f1_pct": { + "FAVOR": 54.054054054054056, + "AGAINST": 63.45177664974619, + "NONE": 79.3010752688172 + } + } + ], + "accuracy": 0.6238011352515169, + "macro_f1": 0.5999409732308351, + "weighted_f1": 0.6402964685909629, + "macro_precision": 0.6058727546321164, + "macro_recall": 0.6378250337281027, + "per_class_f1": { + "FAVOR": 0.5216072782410918, + "AGAINST": 0.5616291532690246, + "NONE": 0.7165864881823886 + }, + "accuracy_pct": 62.38011352515169, + "macro_f1_pct": 59.99409732308351, + "weighted_f1_pct": 64.02964685909629, + "macro_precision_pct": 60.58727546321164, + "macro_recall_pct": 63.782503372810275, + "per_class_f1_pct": { + "FAVOR": 52.16072782410917, + "AGAINST": 56.162915326902464, + "NONE": 71.65864881823886 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s2__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__doc__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6d2aa2c511b34c9b9f5203632fff443c32765d4b --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__doc__s2__seed13/run.log b/outputs/trace_s5_balanced_replay__doc__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..c3deca4732f5ab692a9bed306192c15a7c04496d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 02:24:23,261 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__doc__s2__seed13 device=cuda +2026-09-02 02:54:22,635 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6028513659987056, 'average_forgetting': 0.04804144994311226, 'backward_transfer': -0.046509021664588084, 'final_mean_target_macro_f1_pct': 60.28513659987056, 'average_forgetting_pct': 4.804144994311226, 'backward_transfer_pct': -4.650902166458808, 'final_pooled_accuracy': 0.6238011352515169, 'final_pooled_accuracy_pct': 62.38011352515169, 'final_pooled_macro_f1': 0.5999409732308351, 'final_pooled_macro_f1_pct': 59.99409732308351, 'final_pooled_weighted_f1': 0.6402964685909629, 'final_pooled_weighted_f1_pct': 64.02964685909629, 'final_pooled_macro_precision': 0.6058727546321164, 'final_pooled_macro_precision_pct': 60.58727546321164, 'final_pooled_macro_recall': 0.6378250337281027, 'final_pooled_macro_recall_pct': 63.782503372810275, 'final_pooled_per_class_f1': {'FAVOR': 0.5216072782410918, 'AGAINST': 0.5616291532690246, 'NONE': 0.7165864881823886}, 'final_pooled_per_class_f1_pct': {'FAVOR': 52.16072782410917, 'AGAINST': 56.162915326902464, 'NONE': 71.65864881823886}} +2026-09-02 02:54:22,636 [INFO] __main__: efficiency={'training_wall_time_seconds': 1798.0146911861375, 'training_wall_time_minutes': 29.96691151976896, 'peak_gpu_memory_allocated_gb': 10.39700698852539, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s5_balanced_replay__doc__s3__seed13/environment.json b/outputs/trace_s5_balanced_replay__doc__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..d70808cfe5dac6f98a3ebff0b13489014dd1e94c --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T02:54:30.820733+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.json b/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..1a28182a92c1e3492110355f1157a3fb0ba88c10 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.json @@ -0,0 +1,1587 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.6140498605309744, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5363954515332227, + 0.5870608057216559, + null, + null, + null, + null, + null, + null + ], + [ + 0.5286377111292535, + 0.5065905314986573, + 0.6807757552032361, + null, + null, + null, + null, + null + ], + [ + 0.5156055940577761, + 0.4952486934947032, + 0.6415107605771583, + 0.6381071896022883, + null, + null, + null, + null + ], + [ + 0.5475358661997051, + 0.5361536536226933, + 0.670237324230575, + 0.6377855201836102, + 0.6862555761909755, + null, + null, + null + ], + [ + 0.5553693503797965, + 0.4963349869938531, + 0.6374063276566542, + 0.6327629101468054, + 0.6514130999184037, + 0.5857708678738013, + null, + null + ], + [ + 0.5696546606513216, + 0.5067232724535081, + 0.6942918156925938, + 0.6536090422295241, + 0.6331795420658066, + 0.5664794010324349, + 0.6734492111640747, + null + ], + [ + 0.5577715410531044, + 0.5387373665576193, + 0.6525145343404303, + 0.6649113986463385, + 0.6441436520453835, + 0.586595530073791, + 0.6942757318686733, + 0.6972985861874751 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 13, + "best_dev_macro_f1": 0.6784533458895279, + "losses": { + "current_ce": 0.6031591256680312, + "total": 0.6031591256680312, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6874205844980941, + "macro_f1": 0.6140498605309744, + "weighted_f1": 0.6993946656967844, + "macro_precision": 0.6060267540850498, + "macro_recall": 0.6311044357885316, + "per_class_f1": { + "FAVOR": 0.5121107266435986, + "AGAINST": 0.5159574468085106, + "NONE": 0.8140814081408141 + }, + "accuracy_pct": 68.7420584498094, + "macro_f1_pct": 61.404986053097446, + "weighted_f1_pct": 69.93946656967844, + "macro_precision_pct": 60.60267540850498, + "macro_recall_pct": 63.11044357885316, + "per_class_f1_pct": { + "FAVOR": 51.21107266435986, + "AGAINST": 51.59574468085106, + "NONE": 81.4081408140814 + } + } + ], + "accuracy": 0.6874205844980941, + "macro_f1": 0.6140498605309744, + "weighted_f1": 0.6993946656967844, + "macro_precision": 0.6060267540850498, + "macro_recall": 0.6311044357885316, + "per_class_f1": { + "FAVOR": 0.5121107266435986, + "AGAINST": 0.5159574468085106, + "NONE": 0.8140814081408141 + }, + "accuracy_pct": 68.7420584498094, + "macro_f1_pct": 61.404986053097446, + "weighted_f1_pct": 69.93946656967844, + "macro_precision_pct": 60.60267540850498, + "macro_recall_pct": 63.11044357885316, + "per_class_f1_pct": { + "FAVOR": 51.21107266435986, + "AGAINST": 51.59574468085106, + "NONE": 81.4081408140814 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.6877257003839282, + "losses": { + "current_ce": 0.5780828830859964, + "replay_ce": 0.02472864391295136, + "total": 0.6028115268264498, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6175349428208387, + "macro_f1": 0.5363954515332227, + "weighted_f1": 0.6292442901267782, + "macro_precision": 0.5303341457399642, + "macro_recall": 0.548571040836836, + "per_class_f1": { + "FAVOR": 0.4414715719063545, + "AGAINST": 0.41225626740947074, + "NONE": 0.7554585152838428 + }, + "accuracy_pct": 61.753494282083864, + "macro_f1_pct": 53.63954515332227, + "weighted_f1_pct": 62.92442901267782, + "macro_precision_pct": 53.03341457399642, + "macro_recall_pct": 54.85710408368361, + "per_class_f1_pct": { + "FAVOR": 44.14715719063545, + "AGAINST": 41.22562674094708, + "NONE": 75.54585152838428 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.6306429548563611, + "macro_f1": 0.5870608057216559, + "weighted_f1": 0.6408878604535333, + "macro_precision": 0.581583176080208, + "macro_recall": 0.6066188613058382, + "per_class_f1": { + "FAVOR": 0.45643153526970953, + "AGAINST": 0.5805168986083499, + "NONE": 0.724233983286908 + }, + "accuracy_pct": 63.064295485636116, + "macro_f1_pct": 58.70608057216559, + "weighted_f1_pct": 64.08878604535333, + "macro_precision_pct": 58.1583176080208, + "macro_recall_pct": 60.66188613058382, + "per_class_f1_pct": { + "FAVOR": 45.643153526970956, + "AGAINST": 58.05168986083499, + "NONE": 72.42339832869081 + } + } + ], + "accuracy": 0.6238471673254282, + "macro_f1": 0.5667756831701469, + "weighted_f1": 0.635023315747392, + "macro_precision": 0.5603208996271358, + "macro_recall": 0.5825023839480742, + "per_class_f1": { + "FAVOR": 0.44814814814814813, + "AGAINST": 0.5104408352668214, + "NONE": 0.7417380660954712 + }, + "accuracy_pct": 62.38471673254282, + "macro_f1_pct": 56.67756831701469, + "weighted_f1_pct": 63.5023315747392, + "macro_precision_pct": 56.032089962713584, + "macro_recall_pct": 58.25023839480742, + "per_class_f1_pct": { + "FAVOR": 44.81481481481481, + "AGAINST": 51.04408352668214, + "NONE": 74.17380660954713 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6990467322869559, + "losses": { + "current_ce": 0.5617383919656277, + "replay_ce": 0.019168897760322406, + "total": 0.5809072901095663, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5286377111292535, + "weighted_f1": 0.6201406100751903, + "macro_precision": 0.5262362489050426, + "macro_recall": 0.5541643743168798, + "per_class_f1": { + "FAVOR": 0.44192634560906513, + "AGAINST": 0.3988439306358382, + "NONE": 0.7451428571428571 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 52.86377111292535, + "weighted_f1_pct": 62.014061007519025, + "macro_precision_pct": 52.62362489050426, + "macro_recall_pct": 55.41643743168798, + "per_class_f1_pct": { + "FAVOR": 44.19263456090651, + "AGAINST": 39.884393063583815, + "NONE": 74.5142857142857 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5362517099863201, + "macro_f1": 0.5065905314986573, + "weighted_f1": 0.554172457946242, + "macro_precision": 0.5276892759369395, + "macro_recall": 0.5425590014642576, + "per_class_f1": { + "FAVOR": 0.37992831541218636, + "AGAINST": 0.5207956600361664, + "NONE": 0.6190476190476191 + }, + "accuracy_pct": 53.62517099863201, + "macro_f1_pct": 50.65905314986573, + "weighted_f1_pct": 55.4172457946242, + "macro_precision_pct": 52.768927593693945, + "macro_recall_pct": 54.25590014642576, + "per_class_f1_pct": { + "FAVOR": 37.992831541218635, + "AGAINST": 52.07956600361664, + "NONE": 61.904761904761905 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.704225352112676, + "macro_f1": 0.6807757552032361, + "weighted_f1": 0.7114995709232887, + "macro_precision": 0.6774064738988937, + "macro_recall": 0.6927321429900649, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6641221374045801, + "NONE": 0.7948717948717948 + }, + "accuracy_pct": 70.4225352112676, + "macro_f1_pct": 68.0775755203236, + "weighted_f1_pct": 71.14995709232888, + "macro_precision_pct": 67.74064738988938, + "macro_recall_pct": 69.2732142990065, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 66.41221374045801, + "NONE": 79.48717948717949 + } + } + ], + "accuracy": 0.6029776674937966, + "macro_f1": 0.5659188979861076, + "weighted_f1": 0.620079648715121, + "macro_precision": 0.5665982226230061, + "macro_recall": 0.5930514363128246, + "per_class_f1": { + "FAVOR": 0.4642857142857143, + "AGAINST": 0.5167958656330749, + "NONE": 0.7166751140395337 + }, + "accuracy_pct": 60.297766749379655, + "macro_f1_pct": 56.591889798610765, + "weighted_f1_pct": 62.0079648715121, + "macro_precision_pct": 56.65982226230061, + "macro_recall_pct": 59.30514363128246, + "per_class_f1_pct": { + "FAVOR": 46.42857142857143, + "AGAINST": 51.67958656330749, + "NONE": 71.66751140395337 + } + }, + { + "after_task": 4, + "best_epoch": 13, + "best_dev_macro_f1": 0.7325086141692579, + "losses": { + "current_ce": 0.39830680891750303, + "replay_ce": 0.0163246866183098, + "total": 0.4146314956066085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5997458703939009, + "macro_f1": 0.5156055940577761, + "weighted_f1": 0.6181459128006215, + "macro_precision": 0.5132743530276737, + "macro_recall": 0.5344510830894272, + "per_class_f1": { + "FAVOR": 0.4176470588235294, + "AGAINST": 0.3710144927536232, + "NONE": 0.7581552305961755 + }, + "accuracy_pct": 59.974587039390094, + "macro_f1_pct": 51.5605594057776, + "weighted_f1_pct": 61.814591280062146, + "macro_precision_pct": 51.32743530276736, + "macro_recall_pct": 53.44510830894272, + "per_class_f1_pct": { + "FAVOR": 41.76470588235294, + "AGAINST": 37.10144927536232, + "NONE": 75.81552305961755 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5417236662106704, + "macro_f1": 0.4952486934947032, + "weighted_f1": 0.5558189909709718, + "macro_precision": 0.5214951668552047, + "macro_recall": 0.5153494830009727, + "per_class_f1": { + "FAVOR": 0.3274336283185841, + "AGAINST": 0.5250836120401338, + "NONE": 0.6332288401253918 + }, + "accuracy_pct": 54.172366621067034, + "macro_f1_pct": 49.52486934947032, + "weighted_f1_pct": 55.58189909709718, + "macro_precision_pct": 52.14951668552047, + "macro_recall_pct": 51.53494830009727, + "per_class_f1_pct": { + "FAVOR": 32.743362831858406, + "AGAINST": 52.508361204013376, + "NONE": 63.322884012539184 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6780684104627767, + "macro_f1": 0.6415107605771583, + "weighted_f1": 0.682806849657495, + "macro_precision": 0.6404564019333322, + "macro_recall": 0.6488310128693326, + "per_class_f1": { + "FAVOR": 0.5258620689655172, + "AGAINST": 0.6028368794326241, + "NONE": 0.7958333333333333 + }, + "accuracy_pct": 67.80684104627767, + "macro_f1_pct": 64.15107605771583, + "weighted_f1_pct": 68.2806849657495, + "macro_precision_pct": 64.04564019333321, + "macro_recall_pct": 64.88310128693327, + "per_class_f1_pct": { + "FAVOR": 52.58620689655172, + "AGAINST": 60.28368794326241, + "NONE": 79.58333333333333 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6381071896022883, + "weighted_f1": 0.6993159882973538, + "macro_precision": 0.6403376087774426, + "macro_recall": 0.6718439201243994, + "per_class_f1": { + "FAVOR": 0.48132780082987553, + "AGAINST": 0.6517857142857143, + "NONE": 0.7812080536912752 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 63.810718960228826, + "weighted_f1_pct": 69.93159882973538, + "macro_precision_pct": 64.03376087774426, + "macro_recall_pct": 67.18439201243994, + "per_class_f1_pct": { + "FAVOR": 48.13278008298755, + "AGAINST": 65.17857142857143, + "NONE": 78.12080536912752 + } + } + ], + "accuracy": 0.6222547584187409, + "macro_f1": 0.5731037035371861, + "weighted_f1": 0.6372953286284827, + "macro_precision": 0.5742928000809887, + "macro_recall": 0.5933692627431935, + "per_class_f1": { + "FAVOR": 0.43695861405197306, + "AGAINST": 0.5403466826060969, + "NONE": 0.7420058139534884 + }, + "accuracy_pct": 62.225475841874086, + "macro_f1_pct": 57.31037035371861, + "weighted_f1_pct": 63.729532862848274, + "macro_precision_pct": 57.42928000809887, + "macro_recall_pct": 59.336926274319346, + "per_class_f1_pct": { + "FAVOR": 43.695861405197306, + "AGAINST": 54.034668260609685, + "NONE": 74.20058139534885 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.7252836000300503, + "losses": { + "current_ce": 0.43025804015142577, + "replay_ce": 0.017140992184036544, + "total": 0.44739903286809013, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6315120711562897, + "macro_f1": 0.5475358661997051, + "weighted_f1": 0.6455073835241486, + "macro_precision": 0.5490845491785312, + "macro_recall": 0.5740715066423127, + "per_class_f1": { + "FAVOR": 0.46612466124661245, + "AGAINST": 0.3961038961038961, + "NONE": 0.7803790412486065 + }, + "accuracy_pct": 63.15120711562897, + "macro_f1_pct": 54.753586619970505, + "weighted_f1_pct": 64.55073835241485, + "macro_precision_pct": 54.90845491785312, + "macro_recall_pct": 57.40715066423127, + "per_class_f1_pct": { + "FAVOR": 46.612466124661246, + "AGAINST": 39.61038961038961, + "NONE": 78.03790412486065 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5827633378932968, + "macro_f1": 0.5361536536226933, + "weighted_f1": 0.6010036346139527, + "macro_precision": 0.5382056814095649, + "macro_recall": 0.5622818126584447, + "per_class_f1": { + "FAVOR": 0.38869257950530034, + "AGAINST": 0.510548523206751, + "NONE": 0.7092198581560284 + }, + "accuracy_pct": 58.276333789329684, + "macro_f1_pct": 53.61536536226933, + "weighted_f1_pct": 60.10036346139527, + "macro_precision_pct": 53.820568140956496, + "macro_recall_pct": 56.22818126584447, + "per_class_f1_pct": { + "FAVOR": 38.869257950530034, + "AGAINST": 51.0548523206751, + "NONE": 70.92198581560284 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7082494969818913, + "macro_f1": 0.670237324230575, + "weighted_f1": 0.71078702112509, + "macro_precision": 0.6679052854059564, + "macro_recall": 0.6734020912355474, + "per_class_f1": { + "FAVOR": 0.5959183673469388, + "AGAINST": 0.5905511811023622, + "NONE": 0.8242424242424242 + }, + "accuracy_pct": 70.82494969818913, + "macro_f1_pct": 67.02373242305751, + "weighted_f1_pct": 71.07870211250899, + "macro_precision_pct": 66.79052854059564, + "macro_recall_pct": 67.34020912355474, + "per_class_f1_pct": { + "FAVOR": 59.591836734693885, + "AGAINST": 59.055118110236215, + "NONE": 82.42424242424242 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6680613668061367, + "macro_f1": 0.6377855201836102, + "weighted_f1": 0.6834087472056228, + "macro_precision": 0.6325216510042381, + "macro_recall": 0.6858172965320474, + "per_class_f1": { + "FAVOR": 0.5139318885448917, + "AGAINST": 0.6563307493540051, + "NONE": 0.7430939226519337 + }, + "accuracy_pct": 66.80613668061368, + "macro_f1_pct": 63.778552018361026, + "weighted_f1_pct": 68.34087472056228, + "macro_precision_pct": 63.25216510042381, + "macro_recall_pct": 68.58172965320473, + "per_class_f1_pct": { + "FAVOR": 51.39318885448917, + "AGAINST": 65.63307493540051, + "NONE": 74.30939226519338 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7011494252873564, + "macro_f1": 0.6862555761909755, + "weighted_f1": 0.7088302092621667, + "macro_precision": 0.6854044734812396, + "macro_recall": 0.7025163803332818, + "per_class_f1": { + "FAVOR": 0.5914634146341463, + "AGAINST": 0.6949152542372882, + "NONE": 0.7723880597014925 + }, + "accuracy_pct": 70.11494252873564, + "macro_f1_pct": 68.62555761909755, + "weighted_f1_pct": 70.88302092621667, + "macro_precision_pct": 68.54044734812396, + "macro_recall_pct": 70.25163803332818, + "per_class_f1_pct": { + "FAVOR": 59.14634146341463, + "AGAINST": 69.49152542372882, + "NONE": 77.23880597014924 + } + } + ], + "accuracy": 0.6527985633043999, + "macro_f1": 0.6140847392914711, + "weighted_f1": 0.6659320109604544, + "macro_precision": 0.6110093103688388, + "macro_recall": 0.637180136780282, + "per_class_f1": { + "FAVOR": 0.5090439276485789, + "AGAINST": 0.5706246482836241, + "NONE": 0.7625856419422103 + }, + "accuracy_pct": 65.27985633043998, + "macro_f1_pct": 61.408473929147114, + "weighted_f1_pct": 66.59320109604543, + "macro_precision_pct": 61.10093103688388, + "macro_recall_pct": 63.71801367802819, + "per_class_f1_pct": { + "FAVOR": 50.90439276485789, + "AGAINST": 57.062464828362415, + "NONE": 76.25856419422104 + } + }, + { + "after_task": 6, + "best_epoch": 13, + "best_dev_macro_f1": 0.7084182131987915, + "losses": { + "current_ce": 0.4469472507635752, + "replay_ce": 0.018813061178253408, + "total": 0.46576031156276404, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6289707750952986, + "macro_f1": 0.5553693503797965, + "weighted_f1": 0.6498522493871518, + "macro_precision": 0.5646460504489225, + "macro_recall": 0.5877661876572552, + "per_class_f1": { + "FAVOR": 0.45454545454545453, + "AGAINST": 0.4336569579288026, + "NONE": 0.7779056386651323 + }, + "accuracy_pct": 62.897077509529865, + "macro_f1_pct": 55.53693503797965, + "weighted_f1_pct": 64.98522493871518, + "macro_precision_pct": 56.464605044892245, + "macro_recall_pct": 58.776618765725516, + "per_class_f1_pct": { + "FAVOR": 45.45454545454545, + "AGAINST": 43.36569579288026, + "NONE": 77.79056386651322 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5321477428180574, + "macro_f1": 0.4963349869938531, + "weighted_f1": 0.5621340308136051, + "macro_precision": 0.5213793588199122, + "macro_recall": 0.5317608063674352, + "per_class_f1": { + "FAVOR": 0.3413173652694611, + "AGAINST": 0.48, + "NONE": 0.667687595712098 + }, + "accuracy_pct": 53.214774281805745, + "macro_f1_pct": 49.63349869938531, + "weighted_f1_pct": 56.21340308136051, + "macro_precision_pct": 52.13793588199122, + "macro_recall_pct": 53.176080636743514, + "per_class_f1_pct": { + "FAVOR": 34.13173652694611, + "AGAINST": 48.0, + "NONE": 66.7687595712098 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6374063276566542, + "weighted_f1": 0.6800389356971274, + "macro_precision": 0.6374932903918411, + "macro_recall": 0.6465651355113403, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.5853658536585366, + "NONE": 0.7974413646055437 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 63.74063276566542, + "weighted_f1_pct": 68.00389356971273, + "macro_precision_pct": 63.74932903918411, + "macro_recall_pct": 64.65651355113403, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 58.536585365853654, + "NONE": 79.74413646055437 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6624825662482566, + "macro_f1": 0.6327629101468054, + "weighted_f1": 0.6793448472479968, + "macro_precision": 0.6315750343348167, + "macro_recall": 0.6849989508775547, + "per_class_f1": { + "FAVOR": 0.50920245398773, + "AGAINST": 0.6482412060301508, + "NONE": 0.7408450704225352 + }, + "accuracy_pct": 66.24825662482566, + "macro_f1_pct": 63.27629101468054, + "weighted_f1_pct": 67.93448472479969, + "macro_precision_pct": 63.157503433481665, + "macro_recall_pct": 68.49989508775548, + "per_class_f1_pct": { + "FAVOR": 50.920245398773, + "AGAINST": 64.82412060301507, + "NONE": 74.08450704225352 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6551724137931034, + "macro_f1": 0.6514130999184037, + "weighted_f1": 0.665774161993362, + "macro_precision": 0.675569024531265, + "macro_recall": 0.6835622617159666, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.6647058823529411, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 65.51724137931035, + "macro_f1_pct": 65.14130999184037, + "weighted_f1_pct": 66.5774161993362, + "macro_precision_pct": 67.5569024531265, + "macro_recall_pct": 68.35622617159666, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 66.47058823529412, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6113602391629297, + "macro_f1": 0.5857708678738013, + "weighted_f1": 0.6309448859845098, + "macro_precision": 0.6056863920472862, + "macro_recall": 0.6249355221151801, + "per_class_f1": { + "FAVOR": 0.5420560747663551, + "AGAINST": 0.5016722408026756, + "NONE": 0.7135842880523732 + }, + "accuracy_pct": 61.13602391629297, + "macro_f1_pct": 58.577086787380125, + "weighted_f1_pct": 63.09448859845098, + "macro_precision_pct": 60.568639204728626, + "macro_recall_pct": 62.49355221151801, + "per_class_f1_pct": { + "FAVOR": 54.20560747663551, + "AGAINST": 50.16722408026756, + "NONE": 71.35842880523732 + } + } + ], + "accuracy": 0.6236907730673317, + "macro_f1": 0.5940613856027568, + "weighted_f1": 0.6429187990896728, + "macro_precision": 0.6026140861464323, + "macro_recall": 0.6268757793897599, + "per_class_f1": { + "FAVOR": 0.4952651515151515, + "AGAINST": 0.5521821631878557, + "NONE": 0.7347368421052631 + }, + "accuracy_pct": 62.36907730673317, + "macro_f1_pct": 59.40613856027568, + "weighted_f1_pct": 64.29187990896727, + "macro_precision_pct": 60.26140861464323, + "macro_recall_pct": 62.68757793897599, + "per_class_f1_pct": { + "FAVOR": 49.52651515151515, + "AGAINST": 55.21821631878557, + "NONE": 73.47368421052632 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.7345031712473573, + "losses": { + "current_ce": 0.38242701980329696, + "replay_ce": 0.015249845888354771, + "total": 0.3976768662354776, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6340533672172808, + "macro_f1": 0.5696546606513216, + "weighted_f1": 0.655328123081914, + "macro_precision": 0.571356902780798, + "macro_recall": 0.6029914247779172, + "per_class_f1": { + "FAVOR": 0.4656084656084656, + "AGAINST": 0.47305389221556887, + "NONE": 0.7703016241299304 + }, + "accuracy_pct": 63.40533672172808, + "macro_f1_pct": 56.96546606513216, + "weighted_f1_pct": 65.5328123081914, + "macro_precision_pct": 57.1356902780798, + "macro_recall_pct": 60.299142477791726, + "per_class_f1_pct": { + "FAVOR": 46.56084656084656, + "AGAINST": 47.30538922155689, + "NONE": 77.03016241299304 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5444596443228454, + "macro_f1": 0.5067232724535081, + "weighted_f1": 0.5693682514906114, + "macro_precision": 0.521663456621161, + "macro_recall": 0.5417310590561344, + "per_class_f1": { + "FAVOR": 0.3613707165109034, + "AGAINST": 0.48717948717948717, + "NONE": 0.6716196136701337 + }, + "accuracy_pct": 54.44596443228454, + "macro_f1_pct": 50.672327245350814, + "weighted_f1_pct": 56.93682514906114, + "macro_precision_pct": 52.1663456621161, + "macro_recall_pct": 54.17310590561344, + "per_class_f1_pct": { + "FAVOR": 36.13707165109034, + "AGAINST": 48.717948717948715, + "NONE": 67.16196136701336 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7243460764587525, + "macro_f1": 0.6942918156925938, + "weighted_f1": 0.7264628370782807, + "macro_precision": 0.6937458867331897, + "macro_recall": 0.6986569204697428, + "per_class_f1": { + "FAVOR": 0.6303501945525292, + "AGAINST": 0.6363636363636364, + "NONE": 0.8161616161616162 + }, + "accuracy_pct": 72.43460764587525, + "macro_f1_pct": 69.42918156925938, + "weighted_f1_pct": 72.64628370782808, + "macro_precision_pct": 69.37458867331897, + "macro_recall_pct": 69.86569204697429, + "per_class_f1_pct": { + "FAVOR": 63.035019455252915, + "AGAINST": 63.63636363636363, + "NONE": 81.61616161616162 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6536090422295241, + "weighted_f1": 0.7078227679323863, + "macro_precision": 0.6510546950027681, + "macro_recall": 0.6910367560651638, + "per_class_f1": { + "FAVOR": 0.5060240963855421, + "AGAINST": 0.6761363636363636, + "NONE": 0.7786666666666666 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 65.36090422295241, + "weighted_f1_pct": 70.78227679323862, + "macro_precision_pct": 65.10546950027681, + "macro_recall_pct": 69.10367560651638, + "per_class_f1_pct": { + "FAVOR": 50.602409638554214, + "AGAINST": 67.61363636363636, + "NONE": 77.86666666666666 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6420361247947455, + "macro_f1": 0.6331795420658066, + "weighted_f1": 0.6522141118656085, + "macro_precision": 0.642970425213416, + "macro_recall": 0.6547794758143809, + "per_class_f1": { + "FAVOR": 0.5524861878453039, + "AGAINST": 0.641399416909621, + "NONE": 0.7056530214424951 + }, + "accuracy_pct": 64.20361247947454, + "macro_f1_pct": 63.317954206580666, + "weighted_f1_pct": 65.22141118656086, + "macro_precision_pct": 64.2970425213416, + "macro_recall_pct": 65.47794758143809, + "per_class_f1_pct": { + "FAVOR": 55.24861878453039, + "AGAINST": 64.1399416909621, + "NONE": 70.56530214424951 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6083707025411061, + "macro_f1": 0.5664794010324349, + "weighted_f1": 0.6267179216069773, + "macro_precision": 0.5940905290174878, + "macro_recall": 0.5968225768359013, + "per_class_f1": { + "FAVOR": 0.5318681318681319, + "AGAINST": 0.4330708661417323, + "NONE": 0.7344992050874404 + }, + "accuracy_pct": 60.837070254110614, + "macro_f1_pct": 56.647940103243485, + "weighted_f1_pct": 62.67179216069773, + "macro_precision_pct": 59.40905290174878, + "macro_recall_pct": 59.68225768359014, + "per_class_f1_pct": { + "FAVOR": 53.18681318681319, + "AGAINST": 43.30708661417323, + "NONE": 73.44992050874404 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.6981891348088531, + "macro_f1": 0.6734492111640747, + "weighted_f1": 0.7101950719193101, + "macro_precision": 0.684854124290744, + "macro_recall": 0.6927507967737853, + "per_class_f1": { + "FAVOR": 0.5913621262458472, + "AGAINST": 0.638095238095238, + "NONE": 0.7908902691511387 + }, + "accuracy_pct": 69.81891348088531, + "macro_f1_pct": 67.34492111640748, + "weighted_f1_pct": 71.01950719193101, + "macro_precision_pct": 68.4854124290744, + "macro_recall_pct": 69.27507967737853, + "per_class_f1_pct": { + "FAVOR": 59.136212624584715, + "AGAINST": 63.8095238095238, + "NONE": 79.08902691511386 + } + } + ], + "accuracy": 0.6427779010428223, + "macro_f1": 0.6104683289045756, + "weighted_f1": 0.6586643210381135, + "macro_precision": 0.6161955860721057, + "macro_recall": 0.6385841911900889, + "per_class_f1": { + "FAVOR": 0.516209476309227, + "AGAINST": 0.5637766681797549, + "NONE": 0.7514188422247446 + }, + "accuracy_pct": 64.27779010428223, + "macro_f1_pct": 61.04683289045756, + "weighted_f1_pct": 65.86643210381135, + "macro_precision_pct": 61.619558607210564, + "macro_recall_pct": 63.85841911900889, + "per_class_f1_pct": { + "FAVOR": 51.6209476309227, + "AGAINST": 56.377666817975495, + "NONE": 75.14188422247446 + } + }, + { + "after_task": 8, + "best_epoch": 5, + "best_dev_macro_f1": 0.7696866480590473, + "losses": { + "current_ce": 0.32281980286626255, + "replay_ce": 0.013316215624274942, + "total": 0.3361360181013451, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6353240152477764, + "macro_f1": 0.5577715410531044, + "weighted_f1": 0.6539453006223833, + "macro_precision": 0.5556293663710544, + "macro_recall": 0.5791757443718227, + "per_class_f1": { + "FAVOR": 0.4068965517241379, + "AGAINST": 0.4864864864864865, + "NONE": 0.7799315849486887 + }, + "accuracy_pct": 63.53240152477764, + "macro_f1_pct": 55.777154105310444, + "weighted_f1_pct": 65.39453006223833, + "macro_precision_pct": 55.562936637105445, + "macro_recall_pct": 57.917574437182274, + "per_class_f1_pct": { + "FAVOR": 40.689655172413794, + "AGAINST": 48.64864864864865, + "NONE": 77.99315849486888 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5718194254445964, + "macro_f1": 0.5387373665576193, + "weighted_f1": 0.5836247248375711, + "macro_precision": 0.5516861126920888, + "macro_recall": 0.569176025403009, + "per_class_f1": { + "FAVOR": 0.41295546558704455, + "AGAINST": 0.5633802816901409, + "NONE": 0.6398763523956723 + }, + "accuracy_pct": 57.18194254445964, + "macro_f1_pct": 53.87373665576193, + "weighted_f1_pct": 58.36247248375711, + "macro_precision_pct": 55.16861126920888, + "macro_recall_pct": 56.917602540300905, + "per_class_f1_pct": { + "FAVOR": 41.29554655870445, + "AGAINST": 56.33802816901409, + "NONE": 63.98763523956723 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6525145343404303, + "weighted_f1": 0.6957425311307706, + "macro_precision": 0.6643920299665041, + "macro_recall": 0.6581052480094485, + "per_class_f1": { + "FAVOR": 0.526829268292683, + "AGAINST": 0.6169491525423729, + "NONE": 0.8137651821862348 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 65.25145343404303, + "weighted_f1_pct": 69.57425311307705, + "macro_precision_pct": 66.43920299665041, + "macro_recall_pct": 65.81052480094485, + "per_class_f1_pct": { + "FAVOR": 52.6829268292683, + "AGAINST": 61.694915254237294, + "NONE": 81.37651821862349 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.708507670850767, + "macro_f1": 0.6649113986463385, + "weighted_f1": 0.72114037178311, + "macro_precision": 0.6557420786718238, + "macro_recall": 0.7015697343851087, + "per_class_f1": { + "FAVOR": 0.5285714285714286, + "AGAINST": 0.6683046683046683, + "NONE": 0.7978580990629184 + }, + "accuracy_pct": 70.85076708507671, + "macro_f1_pct": 66.49113986463385, + "weighted_f1_pct": 72.114037178311, + "macro_precision_pct": 65.57420786718238, + "macro_recall_pct": 70.15697343851087, + "per_class_f1_pct": { + "FAVOR": 52.85714285714286, + "AGAINST": 66.83046683046683, + "NONE": 79.78580990629183 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6551724137931034, + "macro_f1": 0.6441436520453835, + "weighted_f1": 0.6593887576595523, + "macro_precision": 0.6616971006892021, + "macro_recall": 0.6684597551866384, + "per_class_f1": { + "FAVOR": 0.5815602836879432, + "AGAINST": 0.6482758620689655, + "NONE": 0.7025948103792415 + }, + "accuracy_pct": 65.51724137931035, + "macro_f1_pct": 64.41436520453834, + "weighted_f1_pct": 65.93887576595523, + "macro_precision_pct": 66.16971006892021, + "macro_recall_pct": 66.84597551866383, + "per_class_f1_pct": { + "FAVOR": 58.156028368794324, + "AGAINST": 64.82758620689654, + "NONE": 70.25948103792416 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.624813153961136, + "macro_f1": 0.586595530073791, + "weighted_f1": 0.6463444832717596, + "macro_precision": 0.5933749780295695, + "macro_recall": 0.6140192252783859, + "per_class_f1": { + "FAVOR": 0.5384615384615384, + "AGAINST": 0.4666666666666667, + "NONE": 0.7546583850931677 + }, + "accuracy_pct": 62.4813153961136, + "macro_f1_pct": 58.659553007379095, + "weighted_f1_pct": 64.63444832717596, + "macro_precision_pct": 59.33749780295695, + "macro_recall_pct": 61.40192252783859, + "per_class_f1_pct": { + "FAVOR": 53.84615384615385, + "AGAINST": 46.666666666666664, + "NONE": 75.46583850931677 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.6942757318686733, + "weighted_f1": 0.7313290311471726, + "macro_precision": 0.6872030458300458, + "macro_recall": 0.7155454304879592, + "per_class_f1": { + "FAVOR": 0.6259541984732825, + "AGAINST": 0.6446280991735537, + "NONE": 0.8122448979591836 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 69.42757318686733, + "weighted_f1_pct": 73.13290311471727, + "macro_precision_pct": 68.72030458300458, + "macro_recall_pct": 71.55454304879592, + "per_class_f1_pct": { + "FAVOR": 62.59541984732825, + "AGAINST": 64.46280991735537, + "NONE": 81.22448979591836 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7358803986710963, + "macro_f1": 0.6972985861874751, + "weighted_f1": 0.7436047607697664, + "macro_precision": 0.6879262549142448, + "macro_recall": 0.7151130650976182, + "per_class_f1": { + "FAVOR": 0.6493506493506493, + "AGAINST": 0.6153846153846154, + "NONE": 0.8271604938271605 + }, + "accuracy_pct": 73.58803986710963, + "macro_f1_pct": 69.72985861874751, + "weighted_f1_pct": 74.36047607697664, + "macro_precision_pct": 68.79262549142447, + "macro_recall_pct": 71.51130650976182, + "per_class_f1_pct": { + "FAVOR": 64.93506493506493, + "AGAINST": 61.53846153846154, + "NONE": 82.71604938271605 + } + } + ], + "accuracy": 0.6637306713642591, + "macro_f1": 0.6282154565936443, + "weighted_f1": 0.6761319344764697, + "macro_precision": 0.6266867870423803, + "macro_recall": 0.6516883123486513, + "per_class_f1": { + "FAVOR": 0.5321610365571494, + "AGAINST": 0.5862412761714856, + "NONE": 0.7662440570522979 + }, + "accuracy_pct": 66.37306713642592, + "macro_f1_pct": 62.82154565936443, + "weighted_f1_pct": 67.61319344764696, + "macro_precision_pct": 62.66867870423803, + "macro_recall_pct": 65.16883123486514, + "per_class_f1_pct": { + "FAVOR": 53.21610365571494, + "AGAINST": 58.62412761714856, + "NONE": 76.62440570522979 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6295310425966019, + "average_forgetting": 0.022219632116894213, + "backward_transfer": -0.018074215957380853, + "final_mean_target_macro_f1_pct": 62.953104259660186, + "average_forgetting_pct": 2.2219632116894212, + "backward_transfer_pct": -1.8074215957380853, + "final_pooled_accuracy": 0.6637306713642591, + "final_pooled_accuracy_pct": 66.37306713642592, + "final_pooled_macro_f1": 0.6282154565936443, + "final_pooled_macro_f1_pct": 62.82154565936443, + "final_pooled_weighted_f1": 0.6761319344764697, + "final_pooled_weighted_f1_pct": 67.61319344764696, + "final_pooled_macro_precision": 0.6266867870423803, + "final_pooled_macro_precision_pct": 62.66867870423803, + "final_pooled_macro_recall": 0.6516883123486513, + "final_pooled_macro_recall_pct": 65.16883123486514, + "final_pooled_per_class_f1": { + "FAVOR": 0.5321610365571494, + "AGAINST": 0.5862412761714856, + "NONE": 0.7662440570522979 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 53.21610365571494, + "AGAINST": 58.62412761714856, + "NONE": 76.62440570522979 + } + }, + "efficiency": { + "training_wall_time_seconds": 1786.4002451079432, + "training_wall_time_minutes": 29.77333741846572, + "peak_gpu_memory_allocated_gb": 10.39700698852539, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..7cacb80b76557c812577479613b35de4c339bf75 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s3__seed13/metrics.partial.json @@ -0,0 +1,1481 @@ +{ + "matrix": [ + [ + 0.6140498605309744, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5363954515332227, + 0.5870608057216559, + null, + null, + null, + null, + null, + null + ], + [ + 0.5286377111292535, + 0.5065905314986573, + 0.6807757552032361, + null, + null, + null, + null, + null + ], + [ + 0.5156055940577761, + 0.4952486934947032, + 0.6415107605771583, + 0.6381071896022883, + null, + null, + null, + null + ], + [ + 0.5475358661997051, + 0.5361536536226933, + 0.670237324230575, + 0.6377855201836102, + 0.6862555761909755, + null, + null, + null + ], + [ + 0.5553693503797965, + 0.4963349869938531, + 0.6374063276566542, + 0.6327629101468054, + 0.6514130999184037, + 0.5857708678738013, + null, + null + ], + [ + 0.5696546606513216, + 0.5067232724535081, + 0.6942918156925938, + 0.6536090422295241, + 0.6331795420658066, + 0.5664794010324349, + 0.6734492111640747, + null + ], + [ + 0.5577715410531044, + 0.5387373665576193, + 0.6525145343404303, + 0.6649113986463385, + 0.6441436520453835, + 0.586595530073791, + 0.6942757318686733, + 0.6972985861874751 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 13, + "best_dev_macro_f1": 0.6784533458895279, + "losses": { + "current_ce": 0.6031591256680312, + "total": 0.6031591256680312, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6874205844980941, + "macro_f1": 0.6140498605309744, + "weighted_f1": 0.6993946656967844, + "macro_precision": 0.6060267540850498, + "macro_recall": 0.6311044357885316, + "per_class_f1": { + "FAVOR": 0.5121107266435986, + "AGAINST": 0.5159574468085106, + "NONE": 0.8140814081408141 + }, + "accuracy_pct": 68.7420584498094, + "macro_f1_pct": 61.404986053097446, + "weighted_f1_pct": 69.93946656967844, + "macro_precision_pct": 60.60267540850498, + "macro_recall_pct": 63.11044357885316, + "per_class_f1_pct": { + "FAVOR": 51.21107266435986, + "AGAINST": 51.59574468085106, + "NONE": 81.4081408140814 + } + } + ], + "accuracy": 0.6874205844980941, + "macro_f1": 0.6140498605309744, + "weighted_f1": 0.6993946656967844, + "macro_precision": 0.6060267540850498, + "macro_recall": 0.6311044357885316, + "per_class_f1": { + "FAVOR": 0.5121107266435986, + "AGAINST": 0.5159574468085106, + "NONE": 0.8140814081408141 + }, + "accuracy_pct": 68.7420584498094, + "macro_f1_pct": 61.404986053097446, + "weighted_f1_pct": 69.93946656967844, + "macro_precision_pct": 60.60267540850498, + "macro_recall_pct": 63.11044357885316, + "per_class_f1_pct": { + "FAVOR": 51.21107266435986, + "AGAINST": 51.59574468085106, + "NONE": 81.4081408140814 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.6877257003839282, + "losses": { + "current_ce": 0.5780828830859964, + "replay_ce": 0.02472864391295136, + "total": 0.6028115268264498, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6175349428208387, + "macro_f1": 0.5363954515332227, + "weighted_f1": 0.6292442901267782, + "macro_precision": 0.5303341457399642, + "macro_recall": 0.548571040836836, + "per_class_f1": { + "FAVOR": 0.4414715719063545, + "AGAINST": 0.41225626740947074, + "NONE": 0.7554585152838428 + }, + "accuracy_pct": 61.753494282083864, + "macro_f1_pct": 53.63954515332227, + "weighted_f1_pct": 62.92442901267782, + "macro_precision_pct": 53.03341457399642, + "macro_recall_pct": 54.85710408368361, + "per_class_f1_pct": { + "FAVOR": 44.14715719063545, + "AGAINST": 41.22562674094708, + "NONE": 75.54585152838428 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.6306429548563611, + "macro_f1": 0.5870608057216559, + "weighted_f1": 0.6408878604535333, + "macro_precision": 0.581583176080208, + "macro_recall": 0.6066188613058382, + "per_class_f1": { + "FAVOR": 0.45643153526970953, + "AGAINST": 0.5805168986083499, + "NONE": 0.724233983286908 + }, + "accuracy_pct": 63.064295485636116, + "macro_f1_pct": 58.70608057216559, + "weighted_f1_pct": 64.08878604535333, + "macro_precision_pct": 58.1583176080208, + "macro_recall_pct": 60.66188613058382, + "per_class_f1_pct": { + "FAVOR": 45.643153526970956, + "AGAINST": 58.05168986083499, + "NONE": 72.42339832869081 + } + } + ], + "accuracy": 0.6238471673254282, + "macro_f1": 0.5667756831701469, + "weighted_f1": 0.635023315747392, + "macro_precision": 0.5603208996271358, + "macro_recall": 0.5825023839480742, + "per_class_f1": { + "FAVOR": 0.44814814814814813, + "AGAINST": 0.5104408352668214, + "NONE": 0.7417380660954712 + }, + "accuracy_pct": 62.38471673254282, + "macro_f1_pct": 56.67756831701469, + "weighted_f1_pct": 63.5023315747392, + "macro_precision_pct": 56.032089962713584, + "macro_recall_pct": 58.25023839480742, + "per_class_f1_pct": { + "FAVOR": 44.81481481481481, + "AGAINST": 51.04408352668214, + "NONE": 74.17380660954713 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6990467322869559, + "losses": { + "current_ce": 0.5617383919656277, + "replay_ce": 0.019168897760322406, + "total": 0.5809072901095663, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5286377111292535, + "weighted_f1": 0.6201406100751903, + "macro_precision": 0.5262362489050426, + "macro_recall": 0.5541643743168798, + "per_class_f1": { + "FAVOR": 0.44192634560906513, + "AGAINST": 0.3988439306358382, + "NONE": 0.7451428571428571 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 52.86377111292535, + "weighted_f1_pct": 62.014061007519025, + "macro_precision_pct": 52.62362489050426, + "macro_recall_pct": 55.41643743168798, + "per_class_f1_pct": { + "FAVOR": 44.19263456090651, + "AGAINST": 39.884393063583815, + "NONE": 74.5142857142857 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5362517099863201, + "macro_f1": 0.5065905314986573, + "weighted_f1": 0.554172457946242, + "macro_precision": 0.5276892759369395, + "macro_recall": 0.5425590014642576, + "per_class_f1": { + "FAVOR": 0.37992831541218636, + "AGAINST": 0.5207956600361664, + "NONE": 0.6190476190476191 + }, + "accuracy_pct": 53.62517099863201, + "macro_f1_pct": 50.65905314986573, + "weighted_f1_pct": 55.4172457946242, + "macro_precision_pct": 52.768927593693945, + "macro_recall_pct": 54.25590014642576, + "per_class_f1_pct": { + "FAVOR": 37.992831541218635, + "AGAINST": 52.07956600361664, + "NONE": 61.904761904761905 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.704225352112676, + "macro_f1": 0.6807757552032361, + "weighted_f1": 0.7114995709232887, + "macro_precision": 0.6774064738988937, + "macro_recall": 0.6927321429900649, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6641221374045801, + "NONE": 0.7948717948717948 + }, + "accuracy_pct": 70.4225352112676, + "macro_f1_pct": 68.0775755203236, + "weighted_f1_pct": 71.14995709232888, + "macro_precision_pct": 67.74064738988938, + "macro_recall_pct": 69.2732142990065, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 66.41221374045801, + "NONE": 79.48717948717949 + } + } + ], + "accuracy": 0.6029776674937966, + "macro_f1": 0.5659188979861076, + "weighted_f1": 0.620079648715121, + "macro_precision": 0.5665982226230061, + "macro_recall": 0.5930514363128246, + "per_class_f1": { + "FAVOR": 0.4642857142857143, + "AGAINST": 0.5167958656330749, + "NONE": 0.7166751140395337 + }, + "accuracy_pct": 60.297766749379655, + "macro_f1_pct": 56.591889798610765, + "weighted_f1_pct": 62.0079648715121, + "macro_precision_pct": 56.65982226230061, + "macro_recall_pct": 59.30514363128246, + "per_class_f1_pct": { + "FAVOR": 46.42857142857143, + "AGAINST": 51.67958656330749, + "NONE": 71.66751140395337 + } + }, + { + "after_task": 4, + "best_epoch": 13, + "best_dev_macro_f1": 0.7325086141692579, + "losses": { + "current_ce": 0.39830680891750303, + "replay_ce": 0.0163246866183098, + "total": 0.4146314956066085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5997458703939009, + "macro_f1": 0.5156055940577761, + "weighted_f1": 0.6181459128006215, + "macro_precision": 0.5132743530276737, + "macro_recall": 0.5344510830894272, + "per_class_f1": { + "FAVOR": 0.4176470588235294, + "AGAINST": 0.3710144927536232, + "NONE": 0.7581552305961755 + }, + "accuracy_pct": 59.974587039390094, + "macro_f1_pct": 51.5605594057776, + "weighted_f1_pct": 61.814591280062146, + "macro_precision_pct": 51.32743530276736, + "macro_recall_pct": 53.44510830894272, + "per_class_f1_pct": { + "FAVOR": 41.76470588235294, + "AGAINST": 37.10144927536232, + "NONE": 75.81552305961755 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5417236662106704, + "macro_f1": 0.4952486934947032, + "weighted_f1": 0.5558189909709718, + "macro_precision": 0.5214951668552047, + "macro_recall": 0.5153494830009727, + "per_class_f1": { + "FAVOR": 0.3274336283185841, + "AGAINST": 0.5250836120401338, + "NONE": 0.6332288401253918 + }, + "accuracy_pct": 54.172366621067034, + "macro_f1_pct": 49.52486934947032, + "weighted_f1_pct": 55.58189909709718, + "macro_precision_pct": 52.14951668552047, + "macro_recall_pct": 51.53494830009727, + "per_class_f1_pct": { + "FAVOR": 32.743362831858406, + "AGAINST": 52.508361204013376, + "NONE": 63.322884012539184 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6780684104627767, + "macro_f1": 0.6415107605771583, + "weighted_f1": 0.682806849657495, + "macro_precision": 0.6404564019333322, + "macro_recall": 0.6488310128693326, + "per_class_f1": { + "FAVOR": 0.5258620689655172, + "AGAINST": 0.6028368794326241, + "NONE": 0.7958333333333333 + }, + "accuracy_pct": 67.80684104627767, + "macro_f1_pct": 64.15107605771583, + "weighted_f1_pct": 68.2806849657495, + "macro_precision_pct": 64.04564019333321, + "macro_recall_pct": 64.88310128693327, + "per_class_f1_pct": { + "FAVOR": 52.58620689655172, + "AGAINST": 60.28368794326241, + "NONE": 79.58333333333333 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6381071896022883, + "weighted_f1": 0.6993159882973538, + "macro_precision": 0.6403376087774426, + "macro_recall": 0.6718439201243994, + "per_class_f1": { + "FAVOR": 0.48132780082987553, + "AGAINST": 0.6517857142857143, + "NONE": 0.7812080536912752 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 63.810718960228826, + "weighted_f1_pct": 69.93159882973538, + "macro_precision_pct": 64.03376087774426, + "macro_recall_pct": 67.18439201243994, + "per_class_f1_pct": { + "FAVOR": 48.13278008298755, + "AGAINST": 65.17857142857143, + "NONE": 78.12080536912752 + } + } + ], + "accuracy": 0.6222547584187409, + "macro_f1": 0.5731037035371861, + "weighted_f1": 0.6372953286284827, + "macro_precision": 0.5742928000809887, + "macro_recall": 0.5933692627431935, + "per_class_f1": { + "FAVOR": 0.43695861405197306, + "AGAINST": 0.5403466826060969, + "NONE": 0.7420058139534884 + }, + "accuracy_pct": 62.225475841874086, + "macro_f1_pct": 57.31037035371861, + "weighted_f1_pct": 63.729532862848274, + "macro_precision_pct": 57.42928000809887, + "macro_recall_pct": 59.336926274319346, + "per_class_f1_pct": { + "FAVOR": 43.695861405197306, + "AGAINST": 54.034668260609685, + "NONE": 74.20058139534885 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.7252836000300503, + "losses": { + "current_ce": 0.43025804015142577, + "replay_ce": 0.017140992184036544, + "total": 0.44739903286809013, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6315120711562897, + "macro_f1": 0.5475358661997051, + "weighted_f1": 0.6455073835241486, + "macro_precision": 0.5490845491785312, + "macro_recall": 0.5740715066423127, + "per_class_f1": { + "FAVOR": 0.46612466124661245, + "AGAINST": 0.3961038961038961, + "NONE": 0.7803790412486065 + }, + "accuracy_pct": 63.15120711562897, + "macro_f1_pct": 54.753586619970505, + "weighted_f1_pct": 64.55073835241485, + "macro_precision_pct": 54.90845491785312, + "macro_recall_pct": 57.40715066423127, + "per_class_f1_pct": { + "FAVOR": 46.612466124661246, + "AGAINST": 39.61038961038961, + "NONE": 78.03790412486065 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5827633378932968, + "macro_f1": 0.5361536536226933, + "weighted_f1": 0.6010036346139527, + "macro_precision": 0.5382056814095649, + "macro_recall": 0.5622818126584447, + "per_class_f1": { + "FAVOR": 0.38869257950530034, + "AGAINST": 0.510548523206751, + "NONE": 0.7092198581560284 + }, + "accuracy_pct": 58.276333789329684, + "macro_f1_pct": 53.61536536226933, + "weighted_f1_pct": 60.10036346139527, + "macro_precision_pct": 53.820568140956496, + "macro_recall_pct": 56.22818126584447, + "per_class_f1_pct": { + "FAVOR": 38.869257950530034, + "AGAINST": 51.0548523206751, + "NONE": 70.92198581560284 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7082494969818913, + "macro_f1": 0.670237324230575, + "weighted_f1": 0.71078702112509, + "macro_precision": 0.6679052854059564, + "macro_recall": 0.6734020912355474, + "per_class_f1": { + "FAVOR": 0.5959183673469388, + "AGAINST": 0.5905511811023622, + "NONE": 0.8242424242424242 + }, + "accuracy_pct": 70.82494969818913, + "macro_f1_pct": 67.02373242305751, + "weighted_f1_pct": 71.07870211250899, + "macro_precision_pct": 66.79052854059564, + "macro_recall_pct": 67.34020912355474, + "per_class_f1_pct": { + "FAVOR": 59.591836734693885, + "AGAINST": 59.055118110236215, + "NONE": 82.42424242424242 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6680613668061367, + "macro_f1": 0.6377855201836102, + "weighted_f1": 0.6834087472056228, + "macro_precision": 0.6325216510042381, + "macro_recall": 0.6858172965320474, + "per_class_f1": { + "FAVOR": 0.5139318885448917, + "AGAINST": 0.6563307493540051, + "NONE": 0.7430939226519337 + }, + "accuracy_pct": 66.80613668061368, + "macro_f1_pct": 63.778552018361026, + "weighted_f1_pct": 68.34087472056228, + "macro_precision_pct": 63.25216510042381, + "macro_recall_pct": 68.58172965320473, + "per_class_f1_pct": { + "FAVOR": 51.39318885448917, + "AGAINST": 65.63307493540051, + "NONE": 74.30939226519338 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7011494252873564, + "macro_f1": 0.6862555761909755, + "weighted_f1": 0.7088302092621667, + "macro_precision": 0.6854044734812396, + "macro_recall": 0.7025163803332818, + "per_class_f1": { + "FAVOR": 0.5914634146341463, + "AGAINST": 0.6949152542372882, + "NONE": 0.7723880597014925 + }, + "accuracy_pct": 70.11494252873564, + "macro_f1_pct": 68.62555761909755, + "weighted_f1_pct": 70.88302092621667, + "macro_precision_pct": 68.54044734812396, + "macro_recall_pct": 70.25163803332818, + "per_class_f1_pct": { + "FAVOR": 59.14634146341463, + "AGAINST": 69.49152542372882, + "NONE": 77.23880597014924 + } + } + ], + "accuracy": 0.6527985633043999, + "macro_f1": 0.6140847392914711, + "weighted_f1": 0.6659320109604544, + "macro_precision": 0.6110093103688388, + "macro_recall": 0.637180136780282, + "per_class_f1": { + "FAVOR": 0.5090439276485789, + "AGAINST": 0.5706246482836241, + "NONE": 0.7625856419422103 + }, + "accuracy_pct": 65.27985633043998, + "macro_f1_pct": 61.408473929147114, + "weighted_f1_pct": 66.59320109604543, + "macro_precision_pct": 61.10093103688388, + "macro_recall_pct": 63.71801367802819, + "per_class_f1_pct": { + "FAVOR": 50.90439276485789, + "AGAINST": 57.062464828362415, + "NONE": 76.25856419422104 + } + }, + { + "after_task": 6, + "best_epoch": 13, + "best_dev_macro_f1": 0.7084182131987915, + "losses": { + "current_ce": 0.4469472507635752, + "replay_ce": 0.018813061178253408, + "total": 0.46576031156276404, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6289707750952986, + "macro_f1": 0.5553693503797965, + "weighted_f1": 0.6498522493871518, + "macro_precision": 0.5646460504489225, + "macro_recall": 0.5877661876572552, + "per_class_f1": { + "FAVOR": 0.45454545454545453, + "AGAINST": 0.4336569579288026, + "NONE": 0.7779056386651323 + }, + "accuracy_pct": 62.897077509529865, + "macro_f1_pct": 55.53693503797965, + "weighted_f1_pct": 64.98522493871518, + "macro_precision_pct": 56.464605044892245, + "macro_recall_pct": 58.776618765725516, + "per_class_f1_pct": { + "FAVOR": 45.45454545454545, + "AGAINST": 43.36569579288026, + "NONE": 77.79056386651322 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5321477428180574, + "macro_f1": 0.4963349869938531, + "weighted_f1": 0.5621340308136051, + "macro_precision": 0.5213793588199122, + "macro_recall": 0.5317608063674352, + "per_class_f1": { + "FAVOR": 0.3413173652694611, + "AGAINST": 0.48, + "NONE": 0.667687595712098 + }, + "accuracy_pct": 53.214774281805745, + "macro_f1_pct": 49.63349869938531, + "weighted_f1_pct": 56.21340308136051, + "macro_precision_pct": 52.13793588199122, + "macro_recall_pct": 53.176080636743514, + "per_class_f1_pct": { + "FAVOR": 34.13173652694611, + "AGAINST": 48.0, + "NONE": 66.7687595712098 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6374063276566542, + "weighted_f1": 0.6800389356971274, + "macro_precision": 0.6374932903918411, + "macro_recall": 0.6465651355113403, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.5853658536585366, + "NONE": 0.7974413646055437 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 63.74063276566542, + "weighted_f1_pct": 68.00389356971273, + "macro_precision_pct": 63.74932903918411, + "macro_recall_pct": 64.65651355113403, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 58.536585365853654, + "NONE": 79.74413646055437 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6624825662482566, + "macro_f1": 0.6327629101468054, + "weighted_f1": 0.6793448472479968, + "macro_precision": 0.6315750343348167, + "macro_recall": 0.6849989508775547, + "per_class_f1": { + "FAVOR": 0.50920245398773, + "AGAINST": 0.6482412060301508, + "NONE": 0.7408450704225352 + }, + "accuracy_pct": 66.24825662482566, + "macro_f1_pct": 63.27629101468054, + "weighted_f1_pct": 67.93448472479969, + "macro_precision_pct": 63.157503433481665, + "macro_recall_pct": 68.49989508775548, + "per_class_f1_pct": { + "FAVOR": 50.920245398773, + "AGAINST": 64.82412060301507, + "NONE": 74.08450704225352 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6551724137931034, + "macro_f1": 0.6514130999184037, + "weighted_f1": 0.665774161993362, + "macro_precision": 0.675569024531265, + "macro_recall": 0.6835622617159666, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.6647058823529411, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 65.51724137931035, + "macro_f1_pct": 65.14130999184037, + "weighted_f1_pct": 66.5774161993362, + "macro_precision_pct": 67.5569024531265, + "macro_recall_pct": 68.35622617159666, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 66.47058823529412, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6113602391629297, + "macro_f1": 0.5857708678738013, + "weighted_f1": 0.6309448859845098, + "macro_precision": 0.6056863920472862, + "macro_recall": 0.6249355221151801, + "per_class_f1": { + "FAVOR": 0.5420560747663551, + "AGAINST": 0.5016722408026756, + "NONE": 0.7135842880523732 + }, + "accuracy_pct": 61.13602391629297, + "macro_f1_pct": 58.577086787380125, + "weighted_f1_pct": 63.09448859845098, + "macro_precision_pct": 60.568639204728626, + "macro_recall_pct": 62.49355221151801, + "per_class_f1_pct": { + "FAVOR": 54.20560747663551, + "AGAINST": 50.16722408026756, + "NONE": 71.35842880523732 + } + } + ], + "accuracy": 0.6236907730673317, + "macro_f1": 0.5940613856027568, + "weighted_f1": 0.6429187990896728, + "macro_precision": 0.6026140861464323, + "macro_recall": 0.6268757793897599, + "per_class_f1": { + "FAVOR": 0.4952651515151515, + "AGAINST": 0.5521821631878557, + "NONE": 0.7347368421052631 + }, + "accuracy_pct": 62.36907730673317, + "macro_f1_pct": 59.40613856027568, + "weighted_f1_pct": 64.29187990896727, + "macro_precision_pct": 60.26140861464323, + "macro_recall_pct": 62.68757793897599, + "per_class_f1_pct": { + "FAVOR": 49.52651515151515, + "AGAINST": 55.21821631878557, + "NONE": 73.47368421052632 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.7345031712473573, + "losses": { + "current_ce": 0.38242701980329696, + "replay_ce": 0.015249845888354771, + "total": 0.3976768662354776, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6340533672172808, + "macro_f1": 0.5696546606513216, + "weighted_f1": 0.655328123081914, + "macro_precision": 0.571356902780798, + "macro_recall": 0.6029914247779172, + "per_class_f1": { + "FAVOR": 0.4656084656084656, + "AGAINST": 0.47305389221556887, + "NONE": 0.7703016241299304 + }, + "accuracy_pct": 63.40533672172808, + "macro_f1_pct": 56.96546606513216, + "weighted_f1_pct": 65.5328123081914, + "macro_precision_pct": 57.1356902780798, + "macro_recall_pct": 60.299142477791726, + "per_class_f1_pct": { + "FAVOR": 46.56084656084656, + "AGAINST": 47.30538922155689, + "NONE": 77.03016241299304 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5444596443228454, + "macro_f1": 0.5067232724535081, + "weighted_f1": 0.5693682514906114, + "macro_precision": 0.521663456621161, + "macro_recall": 0.5417310590561344, + "per_class_f1": { + "FAVOR": 0.3613707165109034, + "AGAINST": 0.48717948717948717, + "NONE": 0.6716196136701337 + }, + "accuracy_pct": 54.44596443228454, + "macro_f1_pct": 50.672327245350814, + "weighted_f1_pct": 56.93682514906114, + "macro_precision_pct": 52.1663456621161, + "macro_recall_pct": 54.17310590561344, + "per_class_f1_pct": { + "FAVOR": 36.13707165109034, + "AGAINST": 48.717948717948715, + "NONE": 67.16196136701336 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7243460764587525, + "macro_f1": 0.6942918156925938, + "weighted_f1": 0.7264628370782807, + "macro_precision": 0.6937458867331897, + "macro_recall": 0.6986569204697428, + "per_class_f1": { + "FAVOR": 0.6303501945525292, + "AGAINST": 0.6363636363636364, + "NONE": 0.8161616161616162 + }, + "accuracy_pct": 72.43460764587525, + "macro_f1_pct": 69.42918156925938, + "weighted_f1_pct": 72.64628370782808, + "macro_precision_pct": 69.37458867331897, + "macro_recall_pct": 69.86569204697429, + "per_class_f1_pct": { + "FAVOR": 63.035019455252915, + "AGAINST": 63.63636363636363, + "NONE": 81.61616161616162 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6903765690376569, + "macro_f1": 0.6536090422295241, + "weighted_f1": 0.7078227679323863, + "macro_precision": 0.6510546950027681, + "macro_recall": 0.6910367560651638, + "per_class_f1": { + "FAVOR": 0.5060240963855421, + "AGAINST": 0.6761363636363636, + "NONE": 0.7786666666666666 + }, + "accuracy_pct": 69.0376569037657, + "macro_f1_pct": 65.36090422295241, + "weighted_f1_pct": 70.78227679323862, + "macro_precision_pct": 65.10546950027681, + "macro_recall_pct": 69.10367560651638, + "per_class_f1_pct": { + "FAVOR": 50.602409638554214, + "AGAINST": 67.61363636363636, + "NONE": 77.86666666666666 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6420361247947455, + "macro_f1": 0.6331795420658066, + "weighted_f1": 0.6522141118656085, + "macro_precision": 0.642970425213416, + "macro_recall": 0.6547794758143809, + "per_class_f1": { + "FAVOR": 0.5524861878453039, + "AGAINST": 0.641399416909621, + "NONE": 0.7056530214424951 + }, + "accuracy_pct": 64.20361247947454, + "macro_f1_pct": 63.317954206580666, + "weighted_f1_pct": 65.22141118656086, + "macro_precision_pct": 64.2970425213416, + "macro_recall_pct": 65.47794758143809, + "per_class_f1_pct": { + "FAVOR": 55.24861878453039, + "AGAINST": 64.1399416909621, + "NONE": 70.56530214424951 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6083707025411061, + "macro_f1": 0.5664794010324349, + "weighted_f1": 0.6267179216069773, + "macro_precision": 0.5940905290174878, + "macro_recall": 0.5968225768359013, + "per_class_f1": { + "FAVOR": 0.5318681318681319, + "AGAINST": 0.4330708661417323, + "NONE": 0.7344992050874404 + }, + "accuracy_pct": 60.837070254110614, + "macro_f1_pct": 56.647940103243485, + "weighted_f1_pct": 62.67179216069773, + "macro_precision_pct": 59.40905290174878, + "macro_recall_pct": 59.68225768359014, + "per_class_f1_pct": { + "FAVOR": 53.18681318681319, + "AGAINST": 43.30708661417323, + "NONE": 73.44992050874404 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.6981891348088531, + "macro_f1": 0.6734492111640747, + "weighted_f1": 0.7101950719193101, + "macro_precision": 0.684854124290744, + "macro_recall": 0.6927507967737853, + "per_class_f1": { + "FAVOR": 0.5913621262458472, + "AGAINST": 0.638095238095238, + "NONE": 0.7908902691511387 + }, + "accuracy_pct": 69.81891348088531, + "macro_f1_pct": 67.34492111640748, + "weighted_f1_pct": 71.01950719193101, + "macro_precision_pct": 68.4854124290744, + "macro_recall_pct": 69.27507967737853, + "per_class_f1_pct": { + "FAVOR": 59.136212624584715, + "AGAINST": 63.8095238095238, + "NONE": 79.08902691511386 + } + } + ], + "accuracy": 0.6427779010428223, + "macro_f1": 0.6104683289045756, + "weighted_f1": 0.6586643210381135, + "macro_precision": 0.6161955860721057, + "macro_recall": 0.6385841911900889, + "per_class_f1": { + "FAVOR": 0.516209476309227, + "AGAINST": 0.5637766681797549, + "NONE": 0.7514188422247446 + }, + "accuracy_pct": 64.27779010428223, + "macro_f1_pct": 61.04683289045756, + "weighted_f1_pct": 65.86643210381135, + "macro_precision_pct": 61.619558607210564, + "macro_recall_pct": 63.85841911900889, + "per_class_f1_pct": { + "FAVOR": 51.6209476309227, + "AGAINST": 56.377666817975495, + "NONE": 75.14188422247446 + } + }, + { + "after_task": 8, + "best_epoch": 5, + "best_dev_macro_f1": 0.7696866480590473, + "losses": { + "current_ce": 0.32281980286626255, + "replay_ce": 0.013316215624274942, + "total": 0.3361360181013451, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6353240152477764, + "macro_f1": 0.5577715410531044, + "weighted_f1": 0.6539453006223833, + "macro_precision": 0.5556293663710544, + "macro_recall": 0.5791757443718227, + "per_class_f1": { + "FAVOR": 0.4068965517241379, + "AGAINST": 0.4864864864864865, + "NONE": 0.7799315849486887 + }, + "accuracy_pct": 63.53240152477764, + "macro_f1_pct": 55.777154105310444, + "weighted_f1_pct": 65.39453006223833, + "macro_precision_pct": 55.562936637105445, + "macro_recall_pct": 57.917574437182274, + "per_class_f1_pct": { + "FAVOR": 40.689655172413794, + "AGAINST": 48.64864864864865, + "NONE": 77.99315849486888 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5718194254445964, + "macro_f1": 0.5387373665576193, + "weighted_f1": 0.5836247248375711, + "macro_precision": 0.5516861126920888, + "macro_recall": 0.569176025403009, + "per_class_f1": { + "FAVOR": 0.41295546558704455, + "AGAINST": 0.5633802816901409, + "NONE": 0.6398763523956723 + }, + "accuracy_pct": 57.18194254445964, + "macro_f1_pct": 53.87373665576193, + "weighted_f1_pct": 58.36247248375711, + "macro_precision_pct": 55.16861126920888, + "macro_recall_pct": 56.917602540300905, + "per_class_f1_pct": { + "FAVOR": 41.29554655870445, + "AGAINST": 56.33802816901409, + "NONE": 63.98763523956723 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6525145343404303, + "weighted_f1": 0.6957425311307706, + "macro_precision": 0.6643920299665041, + "macro_recall": 0.6581052480094485, + "per_class_f1": { + "FAVOR": 0.526829268292683, + "AGAINST": 0.6169491525423729, + "NONE": 0.8137651821862348 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 65.25145343404303, + "weighted_f1_pct": 69.57425311307705, + "macro_precision_pct": 66.43920299665041, + "macro_recall_pct": 65.81052480094485, + "per_class_f1_pct": { + "FAVOR": 52.6829268292683, + "AGAINST": 61.694915254237294, + "NONE": 81.37651821862349 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.708507670850767, + "macro_f1": 0.6649113986463385, + "weighted_f1": 0.72114037178311, + "macro_precision": 0.6557420786718238, + "macro_recall": 0.7015697343851087, + "per_class_f1": { + "FAVOR": 0.5285714285714286, + "AGAINST": 0.6683046683046683, + "NONE": 0.7978580990629184 + }, + "accuracy_pct": 70.85076708507671, + "macro_f1_pct": 66.49113986463385, + "weighted_f1_pct": 72.114037178311, + "macro_precision_pct": 65.57420786718238, + "macro_recall_pct": 70.15697343851087, + "per_class_f1_pct": { + "FAVOR": 52.85714285714286, + "AGAINST": 66.83046683046683, + "NONE": 79.78580990629183 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6551724137931034, + "macro_f1": 0.6441436520453835, + "weighted_f1": 0.6593887576595523, + "macro_precision": 0.6616971006892021, + "macro_recall": 0.6684597551866384, + "per_class_f1": { + "FAVOR": 0.5815602836879432, + "AGAINST": 0.6482758620689655, + "NONE": 0.7025948103792415 + }, + "accuracy_pct": 65.51724137931035, + "macro_f1_pct": 64.41436520453834, + "weighted_f1_pct": 65.93887576595523, + "macro_precision_pct": 66.16971006892021, + "macro_recall_pct": 66.84597551866383, + "per_class_f1_pct": { + "FAVOR": 58.156028368794324, + "AGAINST": 64.82758620689654, + "NONE": 70.25948103792416 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.624813153961136, + "macro_f1": 0.586595530073791, + "weighted_f1": 0.6463444832717596, + "macro_precision": 0.5933749780295695, + "macro_recall": 0.6140192252783859, + "per_class_f1": { + "FAVOR": 0.5384615384615384, + "AGAINST": 0.4666666666666667, + "NONE": 0.7546583850931677 + }, + "accuracy_pct": 62.4813153961136, + "macro_f1_pct": 58.659553007379095, + "weighted_f1_pct": 64.63444832717596, + "macro_precision_pct": 59.33749780295695, + "macro_recall_pct": 61.40192252783859, + "per_class_f1_pct": { + "FAVOR": 53.84615384615385, + "AGAINST": 46.666666666666664, + "NONE": 75.46583850931677 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.6942757318686733, + "weighted_f1": 0.7313290311471726, + "macro_precision": 0.6872030458300458, + "macro_recall": 0.7155454304879592, + "per_class_f1": { + "FAVOR": 0.6259541984732825, + "AGAINST": 0.6446280991735537, + "NONE": 0.8122448979591836 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 69.42757318686733, + "weighted_f1_pct": 73.13290311471727, + "macro_precision_pct": 68.72030458300458, + "macro_recall_pct": 71.55454304879592, + "per_class_f1_pct": { + "FAVOR": 62.59541984732825, + "AGAINST": 64.46280991735537, + "NONE": 81.22448979591836 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7358803986710963, + "macro_f1": 0.6972985861874751, + "weighted_f1": 0.7436047607697664, + "macro_precision": 0.6879262549142448, + "macro_recall": 0.7151130650976182, + "per_class_f1": { + "FAVOR": 0.6493506493506493, + "AGAINST": 0.6153846153846154, + "NONE": 0.8271604938271605 + }, + "accuracy_pct": 73.58803986710963, + "macro_f1_pct": 69.72985861874751, + "weighted_f1_pct": 74.36047607697664, + "macro_precision_pct": 68.79262549142447, + "macro_recall_pct": 71.51130650976182, + "per_class_f1_pct": { + "FAVOR": 64.93506493506493, + "AGAINST": 61.53846153846154, + "NONE": 82.71604938271605 + } + } + ], + "accuracy": 0.6637306713642591, + "macro_f1": 0.6282154565936443, + "weighted_f1": 0.6761319344764697, + "macro_precision": 0.6266867870423803, + "macro_recall": 0.6516883123486513, + "per_class_f1": { + "FAVOR": 0.5321610365571494, + "AGAINST": 0.5862412761714856, + "NONE": 0.7662440570522979 + }, + "accuracy_pct": 66.37306713642592, + "macro_f1_pct": 62.82154565936443, + "weighted_f1_pct": 67.61319344764696, + "macro_precision_pct": 62.66867870423803, + "macro_recall_pct": 65.16883123486514, + "per_class_f1_pct": { + "FAVOR": 53.21610365571494, + "AGAINST": 58.62412761714856, + "NONE": 76.62440570522979 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__doc__s3__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__doc__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..88d080cb1292330cc6bd1ade138b2d3f962c8345 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__doc__s3__seed13/run.log b/outputs/trace_s5_balanced_replay__doc__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..1c7923d7f9e1edafb530ca3eea9bf6dcfc4ab569 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__doc__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 02:54:30,853 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__doc__s3__seed13 device=cuda +2026-09-02 03:24:18,627 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6295310425966019, 'average_forgetting': 0.022219632116894213, 'backward_transfer': -0.018074215957380853, 'final_mean_target_macro_f1_pct': 62.953104259660186, 'average_forgetting_pct': 2.2219632116894212, 'backward_transfer_pct': -1.8074215957380853, 'final_pooled_accuracy': 0.6637306713642591, 'final_pooled_accuracy_pct': 66.37306713642592, 'final_pooled_macro_f1': 0.6282154565936443, 'final_pooled_macro_f1_pct': 62.82154565936443, 'final_pooled_weighted_f1': 0.6761319344764697, 'final_pooled_weighted_f1_pct': 67.61319344764696, 'final_pooled_macro_precision': 0.6266867870423803, 'final_pooled_macro_precision_pct': 62.66867870423803, 'final_pooled_macro_recall': 0.6516883123486513, 'final_pooled_macro_recall_pct': 65.16883123486514, 'final_pooled_per_class_f1': {'FAVOR': 0.5321610365571494, 'AGAINST': 0.5862412761714856, 'NONE': 0.7662440570522979}, 'final_pooled_per_class_f1_pct': {'FAVOR': 53.21610365571494, 'AGAINST': 58.62412761714856, 'NONE': 76.62440570522979}} +2026-09-02 03:24:18,627 [INFO] __main__: efficiency={'training_wall_time_seconds': 1786.4002451079432, 'training_wall_time_minutes': 29.77333741846572, 'peak_gpu_memory_allocated_gb': 10.39700698852539, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s5_balanced_replay__x__s1__seed13/environment.json b/outputs/trace_s5_balanced_replay__x__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..3b4fb381bb8f178ebaf660dd533ca23332abcd8d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T03:24:27.031036+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.json b/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..3b5d7e30e2d5a04ba8be97344e3eca9676846e96 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.json @@ -0,0 +1,2276 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6692929292929293, + 0.7163806742119995, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6595153659641549, + 0.7323392778977986, + 0.5420039273166345, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7072334484614694, + 0.6983174603174603, + 0.5494876737118478, + 0.6858835539495209, + null, + null, + null, + null, + null, + null + ], + [ + 0.6371741168796755, + 0.6913604832553354, + 0.5839812053853799, + 0.6529193029193029, + 0.6818639571247734, + null, + null, + null, + null, + null + ], + [ + 0.6664531059086309, + 0.6943405934375072, + 0.5886216163866157, + 0.666953501864226, + 0.6683373549575906, + 0.6060714821401787, + null, + null, + null, + null + ], + [ + 0.6216416095705823, + 0.686111111111111, + 0.5773527482618391, + 0.6051726706593963, + 0.6391296483689461, + 0.5815382213016744, + 0.5955931055979272, + null, + null, + null + ], + [ + 0.6138869358574061, + 0.6576357145240399, + 0.5796522113650435, + 0.6011786516108651, + 0.6109273368658285, + 0.5654158040027605, + 0.6225000972959447, + 0.7036246612466125, + null, + null + ], + [ + 0.6073496899292056, + 0.6831452935896246, + 0.5369842212979468, + 0.7312969656174251, + 0.6105596735635553, + 0.5196620583717358, + 0.5874919355900515, + 0.6997858138460721, + 0.7688943139049944, + null + ], + [ + 0.6527639074953653, + 0.6499918975854805, + 0.5669446686395839, + 0.6830477589306886, + 0.5967464184574379, + 0.47640287769784173, + 0.6324044133382172, + 0.6789680683334199, + 0.7644517402513696, + 0.6203025634948848 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.720836269673479, + "losses": { + "current_ce": 0.6864762581530072, + "replay_ce": 0.15086150431029854, + "total": 0.8373377624012175, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6692929292929293, + "weighted_f1": 0.6958115400581153, + "macro_precision": 0.6687620683563887, + "macro_recall": 0.7478360819590205, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.72, + "NONE": 0.7878787878787878 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.92929292929293, + "weighted_f1_pct": 69.58115400581153, + "macro_precision_pct": 66.87620683563887, + "macro_recall_pct": 74.78360819590205, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 72.0, + "NONE": 78.78787878787878 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8672566371681416, + "macro_f1": 0.7163806742119995, + "weighted_f1": 0.869613611323815, + "macro_precision": 0.7073260073260074, + "macro_recall": 0.7276786797718723, + "per_class_f1": { + "FAVOR": 0.8939393939393939, + "AGAINST": 0.36363636363636365, + "NONE": 0.891566265060241 + }, + "accuracy_pct": 86.72566371681415, + "macro_f1_pct": 71.63806742119995, + "weighted_f1_pct": 86.9613611323815, + "macro_precision_pct": 70.73260073260073, + "macro_recall_pct": 72.76786797718722, + "per_class_f1_pct": { + "FAVOR": 89.39393939393939, + "AGAINST": 36.36363636363637, + "NONE": 89.1566265060241 + } + } + ], + "accuracy": 0.7606177606177607, + "macro_f1": 0.7637632526018869, + "weighted_f1": 0.7554937905098396, + "macro_precision": 0.781627382511276, + "macro_recall": 0.7750485581951292, + "per_class_f1": { + "FAVOR": 0.75, + "AGAINST": 0.6956521739130435, + "NONE": 0.8456375838926175 + }, + "accuracy_pct": 76.06177606177607, + "macro_f1_pct": 76.3763252601887, + "weighted_f1_pct": 75.54937905098396, + "macro_precision_pct": 78.16273825112759, + "macro_recall_pct": 77.50485581951293, + "per_class_f1_pct": { + "FAVOR": 75.0, + "AGAINST": 69.56521739130434, + "NONE": 84.56375838926175 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.5946193643754619, + "losses": { + "current_ce": 0.6479600154992305, + "replay_ce": 0.04948276419644103, + "total": 0.6974427792158994, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6595153659641549, + "weighted_f1": 0.6859080461036724, + "macro_precision": 0.655859887814775, + "macro_recall": 0.7442128935532234, + "per_class_f1": { + "FAVOR": 0.5135135135135135, + "AGAINST": 0.7114093959731543, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 65.9515365964155, + "weighted_f1_pct": 68.59080461036724, + "macro_precision_pct": 65.5859887814775, + "macro_recall_pct": 74.42128935532234, + "per_class_f1_pct": { + "FAVOR": 51.35135135135135, + "AGAINST": 71.14093959731544, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7323392778977986, + "weighted_f1": 0.8554680983704023, + "macro_precision": 0.7052577052577053, + "macro_recall": 0.839291348137362, + "per_class_f1": { + "FAVOR": 0.8688524590163934, + "AGAINST": 0.4444444444444444, + "NONE": 0.8837209302325582 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.23392778977987, + "weighted_f1_pct": 85.54680983704023, + "macro_precision_pct": 70.52577052577053, + "macro_recall_pct": 83.9291348137362, + "per_class_f1_pct": { + "FAVOR": 86.88524590163934, + "AGAINST": 44.44444444444444, + "NONE": 88.37209302325581 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5420039273166345, + "weighted_f1": 0.5670803141001608, + "macro_precision": 0.533840190816935, + "macro_recall": 0.5697167755991285, + "per_class_f1": { + "FAVOR": 0.4528301886792453, + "AGAINST": 0.6382978723404256, + "NONE": 0.5348837209302325 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 54.20039273166345, + "weighted_f1_pct": 56.708031410016076, + "macro_precision_pct": 53.3840190816935, + "macro_recall_pct": 56.97167755991285, + "per_class_f1_pct": { + "FAVOR": 45.28301886792453, + "AGAINST": 63.829787234042556, + "NONE": 53.48837209302325 + } + } + ], + "accuracy": 0.6681514476614699, + "macro_f1": 0.6723027400754172, + "weighted_f1": 0.666102806165604, + "macro_precision": 0.6668070372449936, + "macro_recall": 0.6940041355354221, + "per_class_f1": { + "FAVOR": 0.6357615894039735, + "AGAINST": 0.6591549295774648, + "NONE": 0.7219917012448133 + }, + "accuracy_pct": 66.815144766147, + "macro_f1_pct": 67.23027400754172, + "weighted_f1_pct": 66.6102806165604, + "macro_precision_pct": 66.68070372449935, + "macro_recall_pct": 69.40041355354221, + "per_class_f1_pct": { + "FAVOR": 63.576158940397356, + "AGAINST": 65.91549295774648, + "NONE": 72.19917012448133 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.6646853146853147, + "losses": { + "current_ce": 0.6184006963715408, + "replay_ce": 0.06061562701890415, + "total": 0.6790163220781269, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7072334484614694, + "weighted_f1": 0.7413096089186723, + "macro_precision": 0.695843287147635, + "macro_recall": 0.7750374812593703, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.7770700636942676, + "NONE": 0.78125 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.72334484614694, + "weighted_f1_pct": 74.13096089186723, + "macro_precision_pct": 69.5843287147635, + "macro_recall_pct": 77.50374812593704, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 77.70700636942675, + "NONE": 78.125 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6983174603174603, + "weighted_f1": 0.8375465655288663, + "macro_precision": 0.6842764515178308, + "macro_recall": 0.8117461473122193, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.38095238095238093, + "NONE": 0.85 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 69.83174603174604, + "weighted_f1_pct": 83.75465655288663, + "macro_precision_pct": 68.42764515178308, + "macro_recall_pct": 81.17461473122192, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 38.095238095238095, + "NONE": 85.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5494876737118478, + "weighted_f1": 0.5726232105203484, + "macro_precision": 0.5483303104662328, + "macro_recall": 0.5508351488743646, + "per_class_f1": { + "FAVOR": 0.41509433962264153, + "AGAINST": 0.6536585365853659, + "NONE": 0.5797101449275363 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 54.94876737118478, + "weighted_f1_pct": 57.262321052034835, + "macro_precision_pct": 54.83303104662328, + "macro_recall_pct": 55.08351488743646, + "per_class_f1_pct": { + "FAVOR": 41.509433962264154, + "AGAINST": 65.3658536585366, + "NONE": 57.971014492753625 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6858835539495209, + "weighted_f1": 0.7182667971225305, + "macro_precision": 0.6767676767676768, + "macro_recall": 0.7068462401795735, + "per_class_f1": { + "FAVOR": 0.6410256410256411, + "AGAINST": 0.7729468599033816, + "NONE": 0.6436781609195402 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 68.58835539495209, + "weighted_f1_pct": 71.82667971225305, + "macro_precision_pct": 67.67676767676768, + "macro_recall_pct": 70.68462401795735, + "per_class_f1_pct": { + "FAVOR": 64.1025641025641, + "AGAINST": 77.29468599033817, + "NONE": 64.36781609195403 + } + } + ], + "accuracy": 0.6929133858267716, + "macro_f1": 0.6896065021310337, + "weighted_f1": 0.6943387910587894, + "macro_precision": 0.6865719763494247, + "macro_recall": 0.6948840683095838, + "per_class_f1": { + "FAVOR": 0.6368421052631579, + "AGAINST": 0.7186440677966102, + "NONE": 0.7133333333333334 + }, + "accuracy_pct": 69.29133858267717, + "macro_f1_pct": 68.96065021310336, + "weighted_f1_pct": 69.43387910587894, + "macro_precision_pct": 68.65719763494246, + "macro_recall_pct": 69.48840683095838, + "per_class_f1_pct": { + "FAVOR": 63.68421052631579, + "AGAINST": 71.86440677966101, + "NONE": 71.33333333333334 + } + }, + { + "after_task": 5, + "best_epoch": 8, + "best_dev_macro_f1": 0.7042809042809043, + "losses": { + "current_ce": 0.5784059358365607, + "replay_ce": 0.0685858503492041, + "total": 0.6469917859091904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6438356164383562, + "macro_f1": 0.6371741168796755, + "weighted_f1": 0.6594782994029833, + "macro_precision": 0.6458926875593543, + "macro_recall": 0.7315592203898053, + "per_class_f1": { + "FAVOR": 0.5063291139240507, + "AGAINST": 0.6805555555555556, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 64.38356164383562, + "macro_f1_pct": 63.71741168796755, + "weighted_f1_pct": 65.94782994029833, + "macro_precision_pct": 64.58926875593542, + "macro_recall_pct": 73.15592203898052, + "per_class_f1_pct": { + "FAVOR": 50.63291139240506, + "AGAINST": 68.05555555555556, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6913604832553354, + "weighted_f1": 0.8306529709858229, + "macro_precision": 0.6789604730781201, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8429752066115702, + "AGAINST": 0.36363636363636365, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.13604832553354, + "weighted_f1_pct": 83.06529709858229, + "macro_precision_pct": 67.89604730781201, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 84.29752066115702, + "AGAINST": 36.36363636363637, + "NONE": 86.74698795180723 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5839812053853799, + "weighted_f1": 0.5936310605736351, + "macro_precision": 0.5749650078918371, + "macro_recall": 0.6034858387799564, + "per_class_f1": { + "FAVOR": 0.5042016806722689, + "AGAINST": 0.6344086021505376, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 58.39812053853799, + "weighted_f1_pct": 59.36310605736351, + "macro_precision_pct": 57.49650078918371, + "macro_recall_pct": 60.348583877995644, + "per_class_f1_pct": { + "FAVOR": 50.42016806722689, + "AGAINST": 63.44086021505376, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6529193029193029, + "weighted_f1": 0.6770809835325964, + "macro_precision": 0.6531552851025467, + "macro_recall": 0.7179573512906847, + "per_class_f1": { + "FAVOR": 0.5494505494505495, + "AGAINST": 0.7045454545454546, + "NONE": 0.7047619047619048 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.29193029193029, + "weighted_f1_pct": 67.70809835325964, + "macro_precision_pct": 65.31552851025467, + "macro_recall_pct": 71.79573512906848, + "per_class_f1_pct": { + "FAVOR": 54.94505494505495, + "AGAINST": 70.45454545454545, + "NONE": 70.47619047619048 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6818639571247734, + "weighted_f1": 0.7024238544334525, + "macro_precision": 0.6702858544963809, + "macro_recall": 0.6970696028478433, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.7373271889400922, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.18639571247735, + "weighted_f1_pct": 70.24238544334524, + "macro_precision_pct": 67.02858544963809, + "macro_recall_pct": 69.70696028478433, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 73.73271889400922, + "NONE": 66.0377358490566 + } + } + ], + "accuracy": 0.671875, + "macro_f1": 0.6712501937655931, + "weighted_f1": 0.6733149645584847, + "macro_precision": 0.6636346640457726, + "macro_recall": 0.69586403114705, + "per_class_f1": { + "FAVOR": 0.6195426195426196, + "AGAINST": 0.6818791946308724, + "NONE": 0.7123287671232876 + }, + "accuracy_pct": 67.1875, + "macro_f1_pct": 67.12501937655931, + "weighted_f1_pct": 67.33149645584847, + "macro_precision_pct": 66.36346640457727, + "macro_recall_pct": 69.58640311470499, + "per_class_f1_pct": { + "FAVOR": 61.95426195426196, + "AGAINST": 68.18791946308724, + "NONE": 71.23287671232876 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.6674136321195144, + "losses": { + "current_ce": 0.6722684621810913, + "replay_ce": 0.11681815333043535, + "total": 0.7890866133073966, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6664531059086309, + "weighted_f1": 0.6888903420420369, + "macro_precision": 0.6581346984482482, + "macro_recall": 0.7575462268865567, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.7114093959731543, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.64531059086309, + "weighted_f1_pct": 68.8890342042037, + "macro_precision_pct": 65.81346984482482, + "macro_recall_pct": 75.75462268865567, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 71.14093959731544, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6943405934375072, + "weighted_f1": 0.812736447742959, + "macro_precision": 0.6724439775910365, + "macro_recall": 0.8112607693241111, + "per_class_f1": { + "FAVOR": 0.8305084745762712, + "AGAINST": 0.42105263157894735, + "NONE": 0.8314606741573034 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 69.43405934375072, + "weighted_f1_pct": 81.27364477429589, + "macro_precision_pct": 67.24439775910365, + "macro_recall_pct": 81.12607693241111, + "per_class_f1_pct": { + "FAVOR": 83.05084745762711, + "AGAINST": 42.10526315789473, + "NONE": 83.14606741573034 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5886216163866157, + "weighted_f1": 0.6062542530118332, + "macro_precision": 0.5764873627876089, + "macro_recall": 0.6118373275236021, + "per_class_f1": { + "FAVOR": 0.4672897196261682, + "AGAINST": 0.6735751295336787, + "NONE": 0.625 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.86216163866157, + "weighted_f1_pct": 60.62542530118332, + "macro_precision_pct": 57.64873627876089, + "macro_recall_pct": 61.18373275236021, + "per_class_f1_pct": { + "FAVOR": 46.728971962616825, + "AGAINST": 67.35751295336787, + "NONE": 62.5 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.666953501864226, + "weighted_f1": 0.6830318010546563, + "macro_precision": 0.6582085103211864, + "macro_recall": 0.731144781144781, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.7039106145251397, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 66.6953501864226, + "weighted_f1_pct": 68.30318010546563, + "macro_precision_pct": 65.82085103211864, + "macro_recall_pct": 73.1144781144781, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 70.39106145251397, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6683373549575906, + "weighted_f1": 0.6840740845160486, + "macro_precision": 0.6512781469176818, + "macro_recall": 0.7139168888257904, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7035175879396985, + "NONE": 0.6842105263157895 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.83373549575906, + "weighted_f1_pct": 68.40740845160485, + "macro_precision_pct": 65.12781469176818, + "macro_recall_pct": 71.39168888257905, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 70.35175879396985, + "NONE": 68.42105263157895 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.6060714821401787, + "weighted_f1": 0.593959587734899, + "macro_precision": 0.6056025446396299, + "macro_recall": 0.6102564102564103, + "per_class_f1": { + "FAVOR": 0.6440677966101694, + "AGAINST": 0.49557522123893805, + "NONE": 0.6785714285714286 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 60.60714821401787, + "weighted_f1_pct": 59.3959587734899, + "macro_precision_pct": 60.56025446396299, + "macro_recall_pct": 61.02564102564103, + "per_class_f1_pct": { + "FAVOR": 64.40677966101694, + "AGAINST": 49.557522123893804, + "NONE": 67.85714285714286 + } + } + ], + "accuracy": 0.6663244353182751, + "macro_f1": 0.6655014972869958, + "weighted_f1": 0.6656061558286948, + "macro_precision": 0.6579320554776376, + "macro_recall": 0.6896188617779527, + "per_class_f1": { + "FAVOR": 0.6283524904214559, + "AGAINST": 0.6643192488262911, + "NONE": 0.7038327526132404 + }, + "accuracy_pct": 66.6324435318275, + "macro_f1_pct": 66.55014972869958, + "weighted_f1_pct": 66.56061558286947, + "macro_precision_pct": 65.79320554776376, + "macro_recall_pct": 68.96188617779528, + "per_class_f1_pct": { + "FAVOR": 62.83524904214559, + "AGAINST": 66.43192488262912, + "NONE": 70.38327526132404 + } + }, + { + "after_task": 7, + "best_epoch": 4, + "best_dev_macro_f1": 0.6576176144909298, + "losses": { + "current_ce": 0.6353092351130076, + "replay_ce": 0.05898610468185137, + "total": 0.6942953383165692, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6232876712328768, + "macro_f1": 0.6216416095705823, + "weighted_f1": 0.6386163253367736, + "macro_precision": 0.6382175677288459, + "macro_recall": 0.7206896551724138, + "per_class_f1": { + "FAVOR": 0.4878048780487805, + "AGAINST": 0.6524822695035462, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 62.328767123287676, + "macro_f1_pct": 62.164160957058236, + "weighted_f1_pct": 63.86163253367736, + "macro_precision_pct": 63.821756772884584, + "macro_recall_pct": 72.06896551724138, + "per_class_f1_pct": { + "FAVOR": 48.78048780487805, + "AGAINST": 65.24822695035462, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.686111111111111, + "weighted_f1": 0.8419616519174041, + "macro_precision": 0.6735322964831161, + "macro_recall": 0.7550297294017715, + "per_class_f1": { + "FAVOR": 0.875, + "AGAINST": 0.3333333333333333, + "NONE": 0.85 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 68.6111111111111, + "weighted_f1_pct": 84.19616519174042, + "macro_precision_pct": 67.35322964831161, + "macro_recall_pct": 75.50297294017714, + "per_class_f1_pct": { + "FAVOR": 87.5, + "AGAINST": 33.33333333333333, + "NONE": 85.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5773527482618391, + "weighted_f1": 0.5830160525280143, + "macro_precision": 0.5691624071423703, + "macro_recall": 0.6165577342047931, + "per_class_f1": { + "FAVOR": 0.49586776859504134, + "AGAINST": 0.6171428571428571, + "NONE": 0.6190476190476191 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.73527482618391, + "weighted_f1_pct": 58.30160525280142, + "macro_precision_pct": 56.916240714237034, + "macro_recall_pct": 61.655773420479306, + "per_class_f1_pct": { + "FAVOR": 49.586776859504134, + "AGAINST": 61.71428571428571, + "NONE": 61.904761904761905 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6075268817204301, + "macro_f1": 0.6051726706593963, + "weighted_f1": 0.6085767754623022, + "macro_precision": 0.6301819095936743, + "macro_recall": 0.6996632996632997, + "per_class_f1": { + "FAVOR": 0.5252525252525253, + "AGAINST": 0.6, + "NONE": 0.6902654867256637 + }, + "accuracy_pct": 60.752688172043015, + "macro_f1_pct": 60.51726706593963, + "weighted_f1_pct": 60.857677546230214, + "macro_precision_pct": 63.01819095936742, + "macro_recall_pct": 69.96632996632997, + "per_class_f1_pct": { + "FAVOR": 52.52525252525253, + "AGAINST": 60.0, + "NONE": 69.02654867256636 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6446700507614214, + "macro_f1": 0.6391296483689461, + "weighted_f1": 0.6474340234406144, + "macro_precision": 0.6254399195575666, + "macro_recall": 0.6968329836627598, + "per_class_f1": { + "FAVOR": 0.6190476190476191, + "AGAINST": 0.6596858638743456, + "NONE": 0.6386554621848739 + }, + "accuracy_pct": 64.46700507614213, + "macro_f1_pct": 63.912964836894616, + "weighted_f1_pct": 64.74340234406144, + "macro_precision_pct": 62.54399195575666, + "macro_recall_pct": 69.68329836627598, + "per_class_f1_pct": { + "FAVOR": 61.904761904761905, + "AGAINST": 65.96858638743456, + "NONE": 63.86554621848739 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5815382213016744, + "weighted_f1": 0.5629851398881675, + "macro_precision": 0.5778809441166262, + "macro_recall": 0.6008547008547008, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.42718446601941745, + "NONE": 0.6722689075630253 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 58.15382213016744, + "weighted_f1_pct": 56.298513988816744, + "macro_precision_pct": 57.78809441166261, + "macro_recall_pct": 60.085470085470085, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 42.71844660194174, + "NONE": 67.22689075630252 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.59, + "macro_f1": 0.5955931055979272, + "weighted_f1": 0.5716623381841559, + "macro_precision": 0.6002719167516406, + "macro_recall": 0.6385781236729747, + "per_class_f1": { + "FAVOR": 0.6022727272727273, + "AGAINST": 0.45901639344262296, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 59.0, + "macro_f1_pct": 59.55931055979272, + "weighted_f1_pct": 57.16623381841559, + "macro_precision_pct": 60.02719167516406, + "macro_recall_pct": 63.85781236729747, + "per_class_f1_pct": { + "FAVOR": 60.22727272727273, + "AGAINST": 45.90163934426229, + "NONE": 72.54901960784314 + } + } + ], + "accuracy": 0.6252129471890971, + "macro_f1": 0.6304581617542973, + "weighted_f1": 0.6187068287939513, + "macro_precision": 0.6322013440334051, + "macro_recall": 0.6679960723035784, + "per_class_f1": { + "FAVOR": 0.6143617021276596, + "AGAINST": 0.5802197802197803, + "NONE": 0.6967930029154519 + }, + "accuracy_pct": 62.521294718909715, + "macro_f1_pct": 63.04581617542973, + "weighted_f1_pct": 61.87068287939513, + "macro_precision_pct": 63.22013440334051, + "macro_recall_pct": 66.79960723035784, + "per_class_f1_pct": { + "FAVOR": 61.43617021276596, + "AGAINST": 58.02197802197803, + "NONE": 69.67930029154519 + } + }, + { + "after_task": 8, + "best_epoch": 10, + "best_dev_macro_f1": 0.7020624768261031, + "losses": { + "current_ce": 0.5890163506286732, + "replay_ce": 0.08828706923818243, + "total": 0.6773034204607424, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6164383561643836, + "macro_f1": 0.6138869358574061, + "weighted_f1": 0.632524120672387, + "macro_precision": 0.6280952380952382, + "macro_recall": 0.7073563218390805, + "per_class_f1": { + "FAVOR": 0.4691358024691358, + "AGAINST": 0.647887323943662, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 61.64383561643836, + "macro_f1_pct": 61.388693585740604, + "weighted_f1_pct": 63.252412067238694, + "macro_precision_pct": 62.80952380952382, + "macro_recall_pct": 70.73563218390805, + "per_class_f1_pct": { + "FAVOR": 46.913580246913575, + "AGAINST": 64.7887323943662, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6576357145240399, + "weighted_f1": 0.7941972225068303, + "macro_precision": 0.6387566137566137, + "macro_recall": 0.7301541075112242, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.35294117647058826, + "NONE": 0.7906976744186046 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 65.763571452404, + "weighted_f1_pct": 79.41972225068304, + "macro_precision_pct": 63.87566137566137, + "macro_recall_pct": 73.01541075112242, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 35.294117647058826, + "NONE": 79.06976744186046 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5796522113650435, + "weighted_f1": 0.5833664149066845, + "macro_precision": 0.572156862745098, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.47540983606557374, + "AGAINST": 0.6206896551724138, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.965221136504354, + "weighted_f1_pct": 58.33664149066845, + "macro_precision_pct": 57.2156862745098, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 47.540983606557376, + "AGAINST": 62.06896551724138, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6021505376344086, + "macro_f1": 0.6011786516108651, + "weighted_f1": 0.6058251473359845, + "macro_precision": 0.6209528666751668, + "macro_recall": 0.6906285072951741, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.6012269938650306, + "NONE": 0.6728971962616822 + }, + "accuracy_pct": 60.215053763440864, + "macro_f1_pct": 60.11786516108651, + "weighted_f1_pct": 60.582514733598444, + "macro_precision_pct": 62.09528666751668, + "macro_recall_pct": 69.06285072951741, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 60.122699386503065, + "NONE": 67.28971962616822 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.6109273368658285, + "weighted_f1": 0.6055367246485063, + "macro_precision": 0.6091643949719825, + "macro_recall": 0.6976585217084956, + "per_class_f1": { + "FAVOR": 0.5894736842105263, + "AGAINST": 0.5875706214689266, + "NONE": 0.6557377049180327 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 61.09273368658285, + "weighted_f1_pct": 60.55367246485063, + "macro_precision_pct": 60.91643949719825, + "macro_recall_pct": 69.76585217084957, + "per_class_f1_pct": { + "FAVOR": 58.94736842105262, + "AGAINST": 58.75706214689266, + "NONE": 65.57377049180327 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5654158040027605, + "weighted_f1": 0.5586945586562855, + "macro_precision": 0.5584811702458761, + "macro_recall": 0.5841880341880342, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.4380952380952381, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 56.54158040027605, + "weighted_f1_pct": 55.869455865628545, + "macro_precision_pct": 55.84811702458761, + "macro_recall_pct": 58.41880341880342, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 43.80952380952381, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6225000972959447, + "weighted_f1": 0.6248832931123155, + "macro_precision": 0.6229835903305133, + "macro_recall": 0.6228468410040225, + "per_class_f1": { + "FAVOR": 0.6242038216560509, + "AGAINST": 0.6335403726708074, + "NONE": 0.6097560975609756 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 62.25000972959447, + "weighted_f1_pct": 62.48832931123155, + "macro_precision_pct": 62.298359033051334, + "macro_recall_pct": 62.28468410040225, + "per_class_f1_pct": { + "FAVOR": 62.42038216560509, + "AGAINST": 63.35403726708074, + "NONE": 60.97560975609756 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.7036246612466125, + "weighted_f1": 0.702598069105691, + "macro_precision": 0.7080208523870496, + "macro_recall": 0.717377201112141, + "per_class_f1": { + "FAVOR": 0.7317073170731707, + "AGAINST": 0.6666666666666666, + "NONE": 0.7125 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 70.36246612466125, + "weighted_f1_pct": 70.2598069105691, + "macro_precision_pct": 70.80208523870496, + "macro_recall_pct": 71.7377201112141, + "per_class_f1_pct": { + "FAVOR": 73.17073170731707, + "AGAINST": 66.66666666666666, + "NONE": 71.25 + } + } + ], + "accuracy": 0.631004366812227, + "macro_f1": 0.6341579287686554, + "weighted_f1": 0.6281005051726066, + "macro_precision": 0.6350895756394773, + "macro_recall": 0.6578269108902625, + "per_class_f1": { + "FAVOR": 0.6136101499423299, + "AGAINST": 0.6003787878787878, + "NONE": 0.6884848484848485 + }, + "accuracy_pct": 63.1004366812227, + "macro_f1_pct": 63.41579287686554, + "weighted_f1_pct": 62.81005051726066, + "macro_precision_pct": 63.50895756394773, + "macro_recall_pct": 65.78269108902624, + "per_class_f1_pct": { + "FAVOR": 61.36101499423299, + "AGAINST": 60.03787878787878, + "NONE": 68.84848484848484 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.647694505079904, + "losses": { + "current_ce": 0.44322704955314596, + "replay_ce": 0.06658820668429447, + "total": 0.5098152552110453, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6301369863013698, + "macro_f1": 0.6073496899292056, + "weighted_f1": 0.6453959545724778, + "macro_precision": 0.6044745702640439, + "macro_recall": 0.6836331834082959, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.6845637583892618, + "NONE": 0.7027027027027027 + }, + "accuracy_pct": 63.013698630136986, + "macro_f1_pct": 60.73496899292056, + "weighted_f1_pct": 64.53959545724778, + "macro_precision_pct": 60.44745702640439, + "macro_recall_pct": 68.3633183408296, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 68.45637583892618, + "NONE": 70.27027027027027 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6831452935896246, + "weighted_f1": 0.8180904998533154, + "macro_precision": 0.6626683501683502, + "macro_recall": 0.7545443514136633, + "per_class_f1": { + "FAVOR": 0.8429752066115702, + "AGAINST": 0.375, + "NONE": 0.8314606741573034 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 68.31452935896246, + "weighted_f1_pct": 81.80904998533154, + "macro_precision_pct": 66.26683501683502, + "macro_recall_pct": 75.45443514136633, + "per_class_f1_pct": { + "FAVOR": 84.29752066115702, + "AGAINST": 37.5, + "NONE": 83.14606741573034 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5369842212979468, + "weighted_f1": 0.5551196172248805, + "macro_precision": 0.5262890341321714, + "macro_recall": 0.5668119099491649, + "per_class_f1": { + "FAVOR": 0.42592592592592593, + "AGAINST": 0.6203208556149733, + "NONE": 0.5647058823529412 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 53.69842212979468, + "weighted_f1_pct": 55.511961722488046, + "macro_precision_pct": 52.628903413217145, + "macro_recall_pct": 56.681190994916484, + "per_class_f1_pct": { + "FAVOR": 42.592592592592595, + "AGAINST": 62.03208556149733, + "NONE": 56.470588235294116 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7312969656174251, + "weighted_f1": 0.7456295280210778, + "macro_precision": 0.716617262423714, + "macro_recall": 0.7968013468013467, + "per_class_f1": { + "FAVOR": 0.6506024096385542, + "AGAINST": 0.7582417582417582, + "NONE": 0.7850467289719626 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 73.1296965617425, + "weighted_f1_pct": 74.56295280210779, + "macro_precision_pct": 71.6617262423714, + "macro_recall_pct": 79.68013468013467, + "per_class_f1_pct": { + "FAVOR": 65.06024096385542, + "AGAINST": 75.82417582417582, + "NONE": 78.50467289719626 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6345177664974619, + "macro_f1": 0.6105596735635553, + "weighted_f1": 0.6414083738931068, + "macro_precision": 0.6002956738250856, + "macro_recall": 0.6480263330861978, + "per_class_f1": { + "FAVOR": 0.5128205128205128, + "AGAINST": 0.6802030456852792, + "NONE": 0.6386554621848739 + }, + "accuracy_pct": 63.45177664974619, + "macro_f1_pct": 61.05596735635553, + "weighted_f1_pct": 64.14083738931068, + "macro_precision_pct": 60.02956738250857, + "macro_recall_pct": 64.80263330861979, + "per_class_f1_pct": { + "FAVOR": 51.28205128205128, + "AGAINST": 68.02030456852792, + "NONE": 63.86554621848739 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5211267605633803, + "macro_f1": 0.5196620583717358, + "weighted_f1": 0.5134679041993899, + "macro_precision": 0.5133333333333333, + "macro_recall": 0.5307692307692308, + "per_class_f1": { + "FAVOR": 0.5161290322580645, + "AGAINST": 0.4, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 52.112676056338024, + "macro_f1_pct": 51.96620583717358, + "weighted_f1_pct": 51.34679041993899, + "macro_precision_pct": 51.33333333333333, + "macro_recall_pct": 53.07692307692308, + "per_class_f1_pct": { + "FAVOR": 51.61290322580645, + "AGAINST": 40.0, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.5874919355900515, + "weighted_f1": 0.5661280167476399, + "macro_precision": 0.6233638282899366, + "macro_recall": 0.6243022259282421, + "per_class_f1": { + "FAVOR": 0.41379310344827586, + "AGAINST": 0.6597938144329897, + "NONE": 0.6888888888888889 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 58.74919355900515, + "weighted_f1_pct": 56.61280167476399, + "macro_precision_pct": 62.33638282899366, + "macro_recall_pct": 62.430222592824215, + "per_class_f1_pct": { + "FAVOR": 41.37931034482759, + "AGAINST": 65.97938144329896, + "NONE": 68.88888888888889 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6997858138460721, + "weighted_f1": 0.7033600350709389, + "macro_precision": 0.7035737740547029, + "macro_recall": 0.7090824837812789, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.6495726495726496, + "NONE": 0.7439024390243902 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.97858138460721, + "weighted_f1_pct": 70.33600350709389, + "macro_precision_pct": 70.35737740547029, + "macro_recall_pct": 70.9082483781279, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 64.95726495726495, + "NONE": 74.39024390243902 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7688943139049944, + "weighted_f1": 0.8013426914861154, + "macro_precision": 0.7584770532490221, + "macro_recall": 0.781797989893942, + "per_class_f1": { + "FAVOR": 0.7241379310344828, + "AGAINST": 0.7241379310344828, + "NONE": 0.8584070796460177 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 76.88943139049944, + "weighted_f1_pct": 80.13426914861154, + "macro_precision_pct": 75.84770532490222, + "macro_recall_pct": 78.1797989893942, + "per_class_f1_pct": { + "FAVOR": 72.41379310344827, + "AGAINST": 72.41379310344827, + "NONE": 85.84070796460178 + } + } + ], + "accuracy": 0.6639135959339263, + "macro_f1": 0.6585600316461284, + "weighted_f1": 0.6619221270538519, + "macro_precision": 0.6557057693676772, + "macro_recall": 0.6669425169001139, + "per_class_f1": { + "FAVOR": 0.5872235872235873, + "AGAINST": 0.6529968454258676, + "NONE": 0.7354596622889306 + }, + "accuracy_pct": 66.39135959339264, + "macro_f1_pct": 65.85600316461284, + "weighted_f1_pct": 66.19221270538519, + "macro_precision_pct": 65.57057693676772, + "macro_recall_pct": 66.69425169001138, + "per_class_f1_pct": { + "FAVOR": 58.72235872235873, + "AGAINST": 65.29968454258676, + "NONE": 73.54596622889305 + } + }, + { + "after_task": 10, + "best_epoch": 5, + "best_dev_macro_f1": 0.6522571110806404, + "losses": { + "current_ce": 0.6404847369744228, + "replay_ce": 0.0657416709436056, + "total": 0.7062264072589385, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6527639074953653, + "weighted_f1": 0.6859021949061538, + "macro_precision": 0.6431390406800243, + "macro_recall": 0.7247926036981509, + "per_class_f1": { + "FAVOR": 0.4857142857142857, + "AGAINST": 0.7189542483660131, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 65.27639074953653, + "weighted_f1_pct": 68.59021949061538, + "macro_precision_pct": 64.31390406800243, + "macro_recall_pct": 72.4792603698151, + "per_class_f1_pct": { + "FAVOR": 48.57142857142857, + "AGAINST": 71.89542483660131, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6499918975854805, + "weighted_f1": 0.8020938646796393, + "macro_precision": 0.6404601571268238, + "macro_recall": 0.7333090644339278, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.3, + "NONE": 0.8235294117647058 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 64.99918975854804, + "weighted_f1_pct": 80.20938646796392, + "macro_precision_pct": 64.04601571268238, + "macro_recall_pct": 73.33090644339279, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 30.0, + "NONE": 82.35294117647058 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5669446686395839, + "weighted_f1": 0.5802472282400918, + "macro_precision": 0.5631352282515073, + "macro_recall": 0.592955700798838, + "per_class_f1": { + "FAVOR": 0.5079365079365079, + "AGAINST": 0.6214689265536724, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 56.69446686395839, + "weighted_f1_pct": 58.02472282400918, + "macro_precision_pct": 56.31352282515073, + "macro_recall_pct": 59.295570079883795, + "per_class_f1_pct": { + "FAVOR": 50.79365079365079, + "AGAINST": 62.14689265536724, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6830477589306886, + "weighted_f1": 0.6953830818877265, + "macro_precision": 0.6797799599291771, + "macro_recall": 0.7524130190796857, + "per_class_f1": { + "FAVOR": 0.6185567010309279, + "AGAINST": 0.7071823204419889, + "NONE": 0.723404255319149 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 68.30477589306886, + "weighted_f1_pct": 69.53830818877265, + "macro_precision_pct": 67.97799599291771, + "macro_recall_pct": 75.24130190796858, + "per_class_f1_pct": { + "FAVOR": 61.855670103092784, + "AGAINST": 70.71823204419888, + "NONE": 72.3404255319149 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.5967464184574379, + "weighted_f1": 0.615130933898533, + "macro_precision": 0.5868592748867011, + "macro_recall": 0.6475583529201437, + "per_class_f1": { + "FAVOR": 0.5287356321839081, + "AGAINST": 0.6354166666666666, + "NONE": 0.6260869565217392 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 59.67464184574379, + "weighted_f1_pct": 61.5130933898533, + "macro_precision_pct": 58.68592748867011, + "macro_recall_pct": 64.75583529201437, + "per_class_f1_pct": { + "FAVOR": 52.87356321839081, + "AGAINST": 63.541666666666664, + "NONE": 62.60869565217392 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.49295774647887325, + "macro_f1": 0.47640287769784173, + "weighted_f1": 0.4889613942648698, + "macro_precision": 0.5309599156118143, + "macro_recall": 0.46196581196581193, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.4892086330935252, + "NONE": 0.54 + }, + "accuracy_pct": 49.29577464788733, + "macro_f1_pct": 47.64028776978417, + "weighted_f1_pct": 48.896139426486975, + "macro_precision_pct": 53.095991561181435, + "macro_recall_pct": 46.19658119658119, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 48.92086330935252, + "NONE": 54.0 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6324044133382172, + "weighted_f1": 0.6138473967790088, + "macro_precision": 0.6478678554150252, + "macro_recall": 0.6445038951814019, + "per_class_f1": { + "FAVOR": 0.5112781954887218, + "AGAINST": 0.6630434782608695, + "NONE": 0.7228915662650602 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 63.24044133382172, + "weighted_f1_pct": 61.38473967790088, + "macro_precision_pct": 64.78678554150252, + "macro_recall_pct": 64.4503895181402, + "per_class_f1_pct": { + "FAVOR": 51.127819548872175, + "AGAINST": 66.30434782608695, + "NONE": 72.28915662650603 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.685, + "macro_f1": 0.6789680683334199, + "weighted_f1": 0.6854529861219397, + "macro_precision": 0.6810408373875866, + "macro_recall": 0.6882298424467099, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.628099173553719, + "NONE": 0.7421383647798742 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 67.896806833342, + "weighted_f1_pct": 68.54529861219397, + "macro_precision_pct": 68.10408373875866, + "macro_recall_pct": 68.82298424467099, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.8099173553719, + "NONE": 74.21383647798741 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.79, + "macro_f1": 0.7644517402513696, + "weighted_f1": 0.7952148122784402, + "macro_precision": 0.7512513770578287, + "macro_recall": 0.7873951912932422, + "per_class_f1": { + "FAVOR": 0.7241379310344828, + "AGAINST": 0.71875, + "NONE": 0.8504672897196262 + }, + "accuracy_pct": 79.0, + "macro_f1_pct": 76.44517402513696, + "weighted_f1_pct": 79.52148122784402, + "macro_precision_pct": 75.12513770578288, + "macro_recall_pct": 78.73951912932422, + "per_class_f1_pct": { + "FAVOR": 72.41379310344827, + "AGAINST": 71.875, + "NONE": 85.04672897196261 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.64, + "macro_f1": 0.6203025634948848, + "weighted_f1": 0.6422410238179879, + "macro_precision": 0.6134367681498829, + "macro_recall": 0.6315681082671375, + "per_class_f1": { + "FAVOR": 0.6701030927835051, + "AGAINST": 0.4666666666666667, + "NONE": 0.7241379310344828 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 62.030256349488475, + "weighted_f1_pct": 64.22410238179879, + "macro_precision_pct": 61.34367681498829, + "macro_recall_pct": 63.15681082671375, + "per_class_f1_pct": { + "FAVOR": 67.0103092783505, + "AGAINST": 46.666666666666664, + "NONE": 72.41379310344827 + } + } + ], + "accuracy": 0.6544532130777903, + "macro_f1": 0.6554531233408406, + "weighted_f1": 0.6541930097206508, + "macro_precision": 0.6532636192249955, + "macro_recall": 0.6592388097502925, + "per_class_f1": { + "FAVOR": 0.6108468125594672, + "AGAINST": 0.6324909747292419, + "NONE": 0.7230215827338129 + }, + "accuracy_pct": 65.44532130777903, + "macro_f1_pct": 65.54531233408406, + "weighted_f1_pct": 65.41930097206507, + "macro_precision_pct": 65.32636192249956, + "macro_recall_pct": 65.92388097502925, + "per_class_f1_pct": { + "FAVOR": 61.08468125594671, + "AGAINST": 63.249097472924184, + "NONE": 72.3021582733813 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.632202431422429, + "average_forgetting": 0.04986387910726754, + "backward_transfer": -0.034875357315967324, + "final_mean_target_macro_f1_pct": 63.22024314224291, + "average_forgetting_pct": 4.986387910726754, + "backward_transfer_pct": -3.4875357315967324, + "final_pooled_accuracy": 0.6544532130777903, + "final_pooled_accuracy_pct": 65.44532130777903, + "final_pooled_macro_f1": 0.6554531233408406, + "final_pooled_macro_f1_pct": 65.54531233408406, + "final_pooled_weighted_f1": 0.6541930097206508, + "final_pooled_weighted_f1_pct": 65.41930097206507, + "final_pooled_macro_precision": 0.6532636192249955, + "final_pooled_macro_precision_pct": 65.32636192249956, + "final_pooled_macro_recall": 0.6592388097502925, + "final_pooled_macro_recall_pct": 65.92388097502925, + "final_pooled_per_class_f1": { + "FAVOR": 0.6108468125594672, + "AGAINST": 0.6324909747292419, + "NONE": 0.7230215827338129 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 61.08468125594671, + "AGAINST": 63.249097472924184, + "NONE": 72.3021582733813 + } + }, + "efficiency": { + "training_wall_time_seconds": 927.0189293648582, + "training_wall_time_minutes": 15.450315489414303, + "peak_gpu_memory_allocated_gb": 10.39700698852539, + "peak_gpu_memory_reserved_gb": 10.580078125 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..57251ddd823b0bd11b92ee94ed251671928940c3 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s1__seed13/metrics.partial.json @@ -0,0 +1,2160 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6692929292929293, + 0.7163806742119995, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6595153659641549, + 0.7323392778977986, + 0.5420039273166345, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7072334484614694, + 0.6983174603174603, + 0.5494876737118478, + 0.6858835539495209, + null, + null, + null, + null, + null, + null + ], + [ + 0.6371741168796755, + 0.6913604832553354, + 0.5839812053853799, + 0.6529193029193029, + 0.6818639571247734, + null, + null, + null, + null, + null + ], + [ + 0.6664531059086309, + 0.6943405934375072, + 0.5886216163866157, + 0.666953501864226, + 0.6683373549575906, + 0.6060714821401787, + null, + null, + null, + null + ], + [ + 0.6216416095705823, + 0.686111111111111, + 0.5773527482618391, + 0.6051726706593963, + 0.6391296483689461, + 0.5815382213016744, + 0.5955931055979272, + null, + null, + null + ], + [ + 0.6138869358574061, + 0.6576357145240399, + 0.5796522113650435, + 0.6011786516108651, + 0.6109273368658285, + 0.5654158040027605, + 0.6225000972959447, + 0.7036246612466125, + null, + null + ], + [ + 0.6073496899292056, + 0.6831452935896246, + 0.5369842212979468, + 0.7312969656174251, + 0.6105596735635553, + 0.5196620583717358, + 0.5874919355900515, + 0.6997858138460721, + 0.7688943139049944, + null + ], + [ + 0.6527639074953653, + 0.6499918975854805, + 0.5669446686395839, + 0.6830477589306886, + 0.5967464184574379, + 0.47640287769784173, + 0.6324044133382172, + 0.6789680683334199, + 0.7644517402513696, + 0.6203025634948848 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.720836269673479, + "losses": { + "current_ce": 0.6864762581530072, + "replay_ce": 0.15086150431029854, + "total": 0.8373377624012175, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6692929292929293, + "weighted_f1": 0.6958115400581153, + "macro_precision": 0.6687620683563887, + "macro_recall": 0.7478360819590205, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.72, + "NONE": 0.7878787878787878 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.92929292929293, + "weighted_f1_pct": 69.58115400581153, + "macro_precision_pct": 66.87620683563887, + "macro_recall_pct": 74.78360819590205, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 72.0, + "NONE": 78.78787878787878 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8672566371681416, + "macro_f1": 0.7163806742119995, + "weighted_f1": 0.869613611323815, + "macro_precision": 0.7073260073260074, + "macro_recall": 0.7276786797718723, + "per_class_f1": { + "FAVOR": 0.8939393939393939, + "AGAINST": 0.36363636363636365, + "NONE": 0.891566265060241 + }, + "accuracy_pct": 86.72566371681415, + "macro_f1_pct": 71.63806742119995, + "weighted_f1_pct": 86.9613611323815, + "macro_precision_pct": 70.73260073260073, + "macro_recall_pct": 72.76786797718722, + "per_class_f1_pct": { + "FAVOR": 89.39393939393939, + "AGAINST": 36.36363636363637, + "NONE": 89.1566265060241 + } + } + ], + "accuracy": 0.7606177606177607, + "macro_f1": 0.7637632526018869, + "weighted_f1": 0.7554937905098396, + "macro_precision": 0.781627382511276, + "macro_recall": 0.7750485581951292, + "per_class_f1": { + "FAVOR": 0.75, + "AGAINST": 0.6956521739130435, + "NONE": 0.8456375838926175 + }, + "accuracy_pct": 76.06177606177607, + "macro_f1_pct": 76.3763252601887, + "weighted_f1_pct": 75.54937905098396, + "macro_precision_pct": 78.16273825112759, + "macro_recall_pct": 77.50485581951293, + "per_class_f1_pct": { + "FAVOR": 75.0, + "AGAINST": 69.56521739130434, + "NONE": 84.56375838926175 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.5946193643754619, + "losses": { + "current_ce": 0.6479600154992305, + "replay_ce": 0.04948276419644103, + "total": 0.6974427792158994, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6595153659641549, + "weighted_f1": 0.6859080461036724, + "macro_precision": 0.655859887814775, + "macro_recall": 0.7442128935532234, + "per_class_f1": { + "FAVOR": 0.5135135135135135, + "AGAINST": 0.7114093959731543, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 65.9515365964155, + "weighted_f1_pct": 68.59080461036724, + "macro_precision_pct": 65.5859887814775, + "macro_recall_pct": 74.42128935532234, + "per_class_f1_pct": { + "FAVOR": 51.35135135135135, + "AGAINST": 71.14093959731544, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7323392778977986, + "weighted_f1": 0.8554680983704023, + "macro_precision": 0.7052577052577053, + "macro_recall": 0.839291348137362, + "per_class_f1": { + "FAVOR": 0.8688524590163934, + "AGAINST": 0.4444444444444444, + "NONE": 0.8837209302325582 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.23392778977987, + "weighted_f1_pct": 85.54680983704023, + "macro_precision_pct": 70.52577052577053, + "macro_recall_pct": 83.9291348137362, + "per_class_f1_pct": { + "FAVOR": 86.88524590163934, + "AGAINST": 44.44444444444444, + "NONE": 88.37209302325581 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5420039273166345, + "weighted_f1": 0.5670803141001608, + "macro_precision": 0.533840190816935, + "macro_recall": 0.5697167755991285, + "per_class_f1": { + "FAVOR": 0.4528301886792453, + "AGAINST": 0.6382978723404256, + "NONE": 0.5348837209302325 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 54.20039273166345, + "weighted_f1_pct": 56.708031410016076, + "macro_precision_pct": 53.3840190816935, + "macro_recall_pct": 56.97167755991285, + "per_class_f1_pct": { + "FAVOR": 45.28301886792453, + "AGAINST": 63.829787234042556, + "NONE": 53.48837209302325 + } + } + ], + "accuracy": 0.6681514476614699, + "macro_f1": 0.6723027400754172, + "weighted_f1": 0.666102806165604, + "macro_precision": 0.6668070372449936, + "macro_recall": 0.6940041355354221, + "per_class_f1": { + "FAVOR": 0.6357615894039735, + "AGAINST": 0.6591549295774648, + "NONE": 0.7219917012448133 + }, + "accuracy_pct": 66.815144766147, + "macro_f1_pct": 67.23027400754172, + "weighted_f1_pct": 66.6102806165604, + "macro_precision_pct": 66.68070372449935, + "macro_recall_pct": 69.40041355354221, + "per_class_f1_pct": { + "FAVOR": 63.576158940397356, + "AGAINST": 65.91549295774648, + "NONE": 72.19917012448133 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.6646853146853147, + "losses": { + "current_ce": 0.6184006963715408, + "replay_ce": 0.06061562701890415, + "total": 0.6790163220781269, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7072334484614694, + "weighted_f1": 0.7413096089186723, + "macro_precision": 0.695843287147635, + "macro_recall": 0.7750374812593703, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.7770700636942676, + "NONE": 0.78125 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.72334484614694, + "weighted_f1_pct": 74.13096089186723, + "macro_precision_pct": 69.5843287147635, + "macro_recall_pct": 77.50374812593704, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 77.70700636942675, + "NONE": 78.125 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6983174603174603, + "weighted_f1": 0.8375465655288663, + "macro_precision": 0.6842764515178308, + "macro_recall": 0.8117461473122193, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.38095238095238093, + "NONE": 0.85 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 69.83174603174604, + "weighted_f1_pct": 83.75465655288663, + "macro_precision_pct": 68.42764515178308, + "macro_recall_pct": 81.17461473122192, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 38.095238095238095, + "NONE": 85.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5494876737118478, + "weighted_f1": 0.5726232105203484, + "macro_precision": 0.5483303104662328, + "macro_recall": 0.5508351488743646, + "per_class_f1": { + "FAVOR": 0.41509433962264153, + "AGAINST": 0.6536585365853659, + "NONE": 0.5797101449275363 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 54.94876737118478, + "weighted_f1_pct": 57.262321052034835, + "macro_precision_pct": 54.83303104662328, + "macro_recall_pct": 55.08351488743646, + "per_class_f1_pct": { + "FAVOR": 41.509433962264154, + "AGAINST": 65.3658536585366, + "NONE": 57.971014492753625 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6858835539495209, + "weighted_f1": 0.7182667971225305, + "macro_precision": 0.6767676767676768, + "macro_recall": 0.7068462401795735, + "per_class_f1": { + "FAVOR": 0.6410256410256411, + "AGAINST": 0.7729468599033816, + "NONE": 0.6436781609195402 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 68.58835539495209, + "weighted_f1_pct": 71.82667971225305, + "macro_precision_pct": 67.67676767676768, + "macro_recall_pct": 70.68462401795735, + "per_class_f1_pct": { + "FAVOR": 64.1025641025641, + "AGAINST": 77.29468599033817, + "NONE": 64.36781609195403 + } + } + ], + "accuracy": 0.6929133858267716, + "macro_f1": 0.6896065021310337, + "weighted_f1": 0.6943387910587894, + "macro_precision": 0.6865719763494247, + "macro_recall": 0.6948840683095838, + "per_class_f1": { + "FAVOR": 0.6368421052631579, + "AGAINST": 0.7186440677966102, + "NONE": 0.7133333333333334 + }, + "accuracy_pct": 69.29133858267717, + "macro_f1_pct": 68.96065021310336, + "weighted_f1_pct": 69.43387910587894, + "macro_precision_pct": 68.65719763494246, + "macro_recall_pct": 69.48840683095838, + "per_class_f1_pct": { + "FAVOR": 63.68421052631579, + "AGAINST": 71.86440677966101, + "NONE": 71.33333333333334 + } + }, + { + "after_task": 5, + "best_epoch": 8, + "best_dev_macro_f1": 0.7042809042809043, + "losses": { + "current_ce": 0.5784059358365607, + "replay_ce": 0.0685858503492041, + "total": 0.6469917859091904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6438356164383562, + "macro_f1": 0.6371741168796755, + "weighted_f1": 0.6594782994029833, + "macro_precision": 0.6458926875593543, + "macro_recall": 0.7315592203898053, + "per_class_f1": { + "FAVOR": 0.5063291139240507, + "AGAINST": 0.6805555555555556, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 64.38356164383562, + "macro_f1_pct": 63.71741168796755, + "weighted_f1_pct": 65.94782994029833, + "macro_precision_pct": 64.58926875593542, + "macro_recall_pct": 73.15592203898052, + "per_class_f1_pct": { + "FAVOR": 50.63291139240506, + "AGAINST": 68.05555555555556, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6913604832553354, + "weighted_f1": 0.8306529709858229, + "macro_precision": 0.6789604730781201, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8429752066115702, + "AGAINST": 0.36363636363636365, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.13604832553354, + "weighted_f1_pct": 83.06529709858229, + "macro_precision_pct": 67.89604730781201, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 84.29752066115702, + "AGAINST": 36.36363636363637, + "NONE": 86.74698795180723 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5839812053853799, + "weighted_f1": 0.5936310605736351, + "macro_precision": 0.5749650078918371, + "macro_recall": 0.6034858387799564, + "per_class_f1": { + "FAVOR": 0.5042016806722689, + "AGAINST": 0.6344086021505376, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 58.39812053853799, + "weighted_f1_pct": 59.36310605736351, + "macro_precision_pct": 57.49650078918371, + "macro_recall_pct": 60.348583877995644, + "per_class_f1_pct": { + "FAVOR": 50.42016806722689, + "AGAINST": 63.44086021505376, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6529193029193029, + "weighted_f1": 0.6770809835325964, + "macro_precision": 0.6531552851025467, + "macro_recall": 0.7179573512906847, + "per_class_f1": { + "FAVOR": 0.5494505494505495, + "AGAINST": 0.7045454545454546, + "NONE": 0.7047619047619048 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.29193029193029, + "weighted_f1_pct": 67.70809835325964, + "macro_precision_pct": 65.31552851025467, + "macro_recall_pct": 71.79573512906848, + "per_class_f1_pct": { + "FAVOR": 54.94505494505495, + "AGAINST": 70.45454545454545, + "NONE": 70.47619047619048 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6818639571247734, + "weighted_f1": 0.7024238544334525, + "macro_precision": 0.6702858544963809, + "macro_recall": 0.6970696028478433, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.7373271889400922, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.18639571247735, + "weighted_f1_pct": 70.24238544334524, + "macro_precision_pct": 67.02858544963809, + "macro_recall_pct": 69.70696028478433, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 73.73271889400922, + "NONE": 66.0377358490566 + } + } + ], + "accuracy": 0.671875, + "macro_f1": 0.6712501937655931, + "weighted_f1": 0.6733149645584847, + "macro_precision": 0.6636346640457726, + "macro_recall": 0.69586403114705, + "per_class_f1": { + "FAVOR": 0.6195426195426196, + "AGAINST": 0.6818791946308724, + "NONE": 0.7123287671232876 + }, + "accuracy_pct": 67.1875, + "macro_f1_pct": 67.12501937655931, + "weighted_f1_pct": 67.33149645584847, + "macro_precision_pct": 66.36346640457727, + "macro_recall_pct": 69.58640311470499, + "per_class_f1_pct": { + "FAVOR": 61.95426195426196, + "AGAINST": 68.18791946308724, + "NONE": 71.23287671232876 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.6674136321195144, + "losses": { + "current_ce": 0.6722684621810913, + "replay_ce": 0.11681815333043535, + "total": 0.7890866133073966, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6664531059086309, + "weighted_f1": 0.6888903420420369, + "macro_precision": 0.6581346984482482, + "macro_recall": 0.7575462268865567, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.7114093959731543, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.64531059086309, + "weighted_f1_pct": 68.8890342042037, + "macro_precision_pct": 65.81346984482482, + "macro_recall_pct": 75.75462268865567, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 71.14093959731544, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6943405934375072, + "weighted_f1": 0.812736447742959, + "macro_precision": 0.6724439775910365, + "macro_recall": 0.8112607693241111, + "per_class_f1": { + "FAVOR": 0.8305084745762712, + "AGAINST": 0.42105263157894735, + "NONE": 0.8314606741573034 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 69.43405934375072, + "weighted_f1_pct": 81.27364477429589, + "macro_precision_pct": 67.24439775910365, + "macro_recall_pct": 81.12607693241111, + "per_class_f1_pct": { + "FAVOR": 83.05084745762711, + "AGAINST": 42.10526315789473, + "NONE": 83.14606741573034 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5886216163866157, + "weighted_f1": 0.6062542530118332, + "macro_precision": 0.5764873627876089, + "macro_recall": 0.6118373275236021, + "per_class_f1": { + "FAVOR": 0.4672897196261682, + "AGAINST": 0.6735751295336787, + "NONE": 0.625 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.86216163866157, + "weighted_f1_pct": 60.62542530118332, + "macro_precision_pct": 57.64873627876089, + "macro_recall_pct": 61.18373275236021, + "per_class_f1_pct": { + "FAVOR": 46.728971962616825, + "AGAINST": 67.35751295336787, + "NONE": 62.5 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.666953501864226, + "weighted_f1": 0.6830318010546563, + "macro_precision": 0.6582085103211864, + "macro_recall": 0.731144781144781, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.7039106145251397, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 66.6953501864226, + "weighted_f1_pct": 68.30318010546563, + "macro_precision_pct": 65.82085103211864, + "macro_recall_pct": 73.1144781144781, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 70.39106145251397, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6683373549575906, + "weighted_f1": 0.6840740845160486, + "macro_precision": 0.6512781469176818, + "macro_recall": 0.7139168888257904, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7035175879396985, + "NONE": 0.6842105263157895 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.83373549575906, + "weighted_f1_pct": 68.40740845160485, + "macro_precision_pct": 65.12781469176818, + "macro_recall_pct": 71.39168888257905, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 70.35175879396985, + "NONE": 68.42105263157895 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.6060714821401787, + "weighted_f1": 0.593959587734899, + "macro_precision": 0.6056025446396299, + "macro_recall": 0.6102564102564103, + "per_class_f1": { + "FAVOR": 0.6440677966101694, + "AGAINST": 0.49557522123893805, + "NONE": 0.6785714285714286 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 60.60714821401787, + "weighted_f1_pct": 59.3959587734899, + "macro_precision_pct": 60.56025446396299, + "macro_recall_pct": 61.02564102564103, + "per_class_f1_pct": { + "FAVOR": 64.40677966101694, + "AGAINST": 49.557522123893804, + "NONE": 67.85714285714286 + } + } + ], + "accuracy": 0.6663244353182751, + "macro_f1": 0.6655014972869958, + "weighted_f1": 0.6656061558286948, + "macro_precision": 0.6579320554776376, + "macro_recall": 0.6896188617779527, + "per_class_f1": { + "FAVOR": 0.6283524904214559, + "AGAINST": 0.6643192488262911, + "NONE": 0.7038327526132404 + }, + "accuracy_pct": 66.6324435318275, + "macro_f1_pct": 66.55014972869958, + "weighted_f1_pct": 66.56061558286947, + "macro_precision_pct": 65.79320554776376, + "macro_recall_pct": 68.96188617779528, + "per_class_f1_pct": { + "FAVOR": 62.83524904214559, + "AGAINST": 66.43192488262912, + "NONE": 70.38327526132404 + } + }, + { + "after_task": 7, + "best_epoch": 4, + "best_dev_macro_f1": 0.6576176144909298, + "losses": { + "current_ce": 0.6353092351130076, + "replay_ce": 0.05898610468185137, + "total": 0.6942953383165692, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6232876712328768, + "macro_f1": 0.6216416095705823, + "weighted_f1": 0.6386163253367736, + "macro_precision": 0.6382175677288459, + "macro_recall": 0.7206896551724138, + "per_class_f1": { + "FAVOR": 0.4878048780487805, + "AGAINST": 0.6524822695035462, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 62.328767123287676, + "macro_f1_pct": 62.164160957058236, + "weighted_f1_pct": 63.86163253367736, + "macro_precision_pct": 63.821756772884584, + "macro_recall_pct": 72.06896551724138, + "per_class_f1_pct": { + "FAVOR": 48.78048780487805, + "AGAINST": 65.24822695035462, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.686111111111111, + "weighted_f1": 0.8419616519174041, + "macro_precision": 0.6735322964831161, + "macro_recall": 0.7550297294017715, + "per_class_f1": { + "FAVOR": 0.875, + "AGAINST": 0.3333333333333333, + "NONE": 0.85 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 68.6111111111111, + "weighted_f1_pct": 84.19616519174042, + "macro_precision_pct": 67.35322964831161, + "macro_recall_pct": 75.50297294017714, + "per_class_f1_pct": { + "FAVOR": 87.5, + "AGAINST": 33.33333333333333, + "NONE": 85.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5773527482618391, + "weighted_f1": 0.5830160525280143, + "macro_precision": 0.5691624071423703, + "macro_recall": 0.6165577342047931, + "per_class_f1": { + "FAVOR": 0.49586776859504134, + "AGAINST": 0.6171428571428571, + "NONE": 0.6190476190476191 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.73527482618391, + "weighted_f1_pct": 58.30160525280142, + "macro_precision_pct": 56.916240714237034, + "macro_recall_pct": 61.655773420479306, + "per_class_f1_pct": { + "FAVOR": 49.586776859504134, + "AGAINST": 61.71428571428571, + "NONE": 61.904761904761905 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6075268817204301, + "macro_f1": 0.6051726706593963, + "weighted_f1": 0.6085767754623022, + "macro_precision": 0.6301819095936743, + "macro_recall": 0.6996632996632997, + "per_class_f1": { + "FAVOR": 0.5252525252525253, + "AGAINST": 0.6, + "NONE": 0.6902654867256637 + }, + "accuracy_pct": 60.752688172043015, + "macro_f1_pct": 60.51726706593963, + "weighted_f1_pct": 60.857677546230214, + "macro_precision_pct": 63.01819095936742, + "macro_recall_pct": 69.96632996632997, + "per_class_f1_pct": { + "FAVOR": 52.52525252525253, + "AGAINST": 60.0, + "NONE": 69.02654867256636 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6446700507614214, + "macro_f1": 0.6391296483689461, + "weighted_f1": 0.6474340234406144, + "macro_precision": 0.6254399195575666, + "macro_recall": 0.6968329836627598, + "per_class_f1": { + "FAVOR": 0.6190476190476191, + "AGAINST": 0.6596858638743456, + "NONE": 0.6386554621848739 + }, + "accuracy_pct": 64.46700507614213, + "macro_f1_pct": 63.912964836894616, + "weighted_f1_pct": 64.74340234406144, + "macro_precision_pct": 62.54399195575666, + "macro_recall_pct": 69.68329836627598, + "per_class_f1_pct": { + "FAVOR": 61.904761904761905, + "AGAINST": 65.96858638743456, + "NONE": 63.86554621848739 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5815382213016744, + "weighted_f1": 0.5629851398881675, + "macro_precision": 0.5778809441166262, + "macro_recall": 0.6008547008547008, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.42718446601941745, + "NONE": 0.6722689075630253 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 58.15382213016744, + "weighted_f1_pct": 56.298513988816744, + "macro_precision_pct": 57.78809441166261, + "macro_recall_pct": 60.085470085470085, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 42.71844660194174, + "NONE": 67.22689075630252 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.59, + "macro_f1": 0.5955931055979272, + "weighted_f1": 0.5716623381841559, + "macro_precision": 0.6002719167516406, + "macro_recall": 0.6385781236729747, + "per_class_f1": { + "FAVOR": 0.6022727272727273, + "AGAINST": 0.45901639344262296, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 59.0, + "macro_f1_pct": 59.55931055979272, + "weighted_f1_pct": 57.16623381841559, + "macro_precision_pct": 60.02719167516406, + "macro_recall_pct": 63.85781236729747, + "per_class_f1_pct": { + "FAVOR": 60.22727272727273, + "AGAINST": 45.90163934426229, + "NONE": 72.54901960784314 + } + } + ], + "accuracy": 0.6252129471890971, + "macro_f1": 0.6304581617542973, + "weighted_f1": 0.6187068287939513, + "macro_precision": 0.6322013440334051, + "macro_recall": 0.6679960723035784, + "per_class_f1": { + "FAVOR": 0.6143617021276596, + "AGAINST": 0.5802197802197803, + "NONE": 0.6967930029154519 + }, + "accuracy_pct": 62.521294718909715, + "macro_f1_pct": 63.04581617542973, + "weighted_f1_pct": 61.87068287939513, + "macro_precision_pct": 63.22013440334051, + "macro_recall_pct": 66.79960723035784, + "per_class_f1_pct": { + "FAVOR": 61.43617021276596, + "AGAINST": 58.02197802197803, + "NONE": 69.67930029154519 + } + }, + { + "after_task": 8, + "best_epoch": 10, + "best_dev_macro_f1": 0.7020624768261031, + "losses": { + "current_ce": 0.5890163506286732, + "replay_ce": 0.08828706923818243, + "total": 0.6773034204607424, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6164383561643836, + "macro_f1": 0.6138869358574061, + "weighted_f1": 0.632524120672387, + "macro_precision": 0.6280952380952382, + "macro_recall": 0.7073563218390805, + "per_class_f1": { + "FAVOR": 0.4691358024691358, + "AGAINST": 0.647887323943662, + "NONE": 0.7246376811594203 + }, + "accuracy_pct": 61.64383561643836, + "macro_f1_pct": 61.388693585740604, + "weighted_f1_pct": 63.252412067238694, + "macro_precision_pct": 62.80952380952382, + "macro_recall_pct": 70.73563218390805, + "per_class_f1_pct": { + "FAVOR": 46.913580246913575, + "AGAINST": 64.7887323943662, + "NONE": 72.46376811594203 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6576357145240399, + "weighted_f1": 0.7941972225068303, + "macro_precision": 0.6387566137566137, + "macro_recall": 0.7301541075112242, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.35294117647058826, + "NONE": 0.7906976744186046 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 65.763571452404, + "weighted_f1_pct": 79.41972225068304, + "macro_precision_pct": 63.87566137566137, + "macro_recall_pct": 73.01541075112242, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 35.294117647058826, + "NONE": 79.06976744186046 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5796522113650435, + "weighted_f1": 0.5833664149066845, + "macro_precision": 0.572156862745098, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.47540983606557374, + "AGAINST": 0.6206896551724138, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.965221136504354, + "weighted_f1_pct": 58.33664149066845, + "macro_precision_pct": 57.2156862745098, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 47.540983606557376, + "AGAINST": 62.06896551724138, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6021505376344086, + "macro_f1": 0.6011786516108651, + "weighted_f1": 0.6058251473359845, + "macro_precision": 0.6209528666751668, + "macro_recall": 0.6906285072951741, + "per_class_f1": { + "FAVOR": 0.5294117647058824, + "AGAINST": 0.6012269938650306, + "NONE": 0.6728971962616822 + }, + "accuracy_pct": 60.215053763440864, + "macro_f1_pct": 60.11786516108651, + "weighted_f1_pct": 60.582514733598444, + "macro_precision_pct": 62.09528666751668, + "macro_recall_pct": 69.06285072951741, + "per_class_f1_pct": { + "FAVOR": 52.94117647058824, + "AGAINST": 60.122699386503065, + "NONE": 67.28971962616822 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.6109273368658285, + "weighted_f1": 0.6055367246485063, + "macro_precision": 0.6091643949719825, + "macro_recall": 0.6976585217084956, + "per_class_f1": { + "FAVOR": 0.5894736842105263, + "AGAINST": 0.5875706214689266, + "NONE": 0.6557377049180327 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 61.09273368658285, + "weighted_f1_pct": 60.55367246485063, + "macro_precision_pct": 60.91643949719825, + "macro_recall_pct": 69.76585217084957, + "per_class_f1_pct": { + "FAVOR": 58.94736842105262, + "AGAINST": 58.75706214689266, + "NONE": 65.57377049180327 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5654158040027605, + "weighted_f1": 0.5586945586562855, + "macro_precision": 0.5584811702458761, + "macro_recall": 0.5841880341880342, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.4380952380952381, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 56.54158040027605, + "weighted_f1_pct": 55.869455865628545, + "macro_precision_pct": 55.84811702458761, + "macro_recall_pct": 58.41880341880342, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 43.80952380952381, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6225000972959447, + "weighted_f1": 0.6248832931123155, + "macro_precision": 0.6229835903305133, + "macro_recall": 0.6228468410040225, + "per_class_f1": { + "FAVOR": 0.6242038216560509, + "AGAINST": 0.6335403726708074, + "NONE": 0.6097560975609756 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 62.25000972959447, + "weighted_f1_pct": 62.48832931123155, + "macro_precision_pct": 62.298359033051334, + "macro_recall_pct": 62.28468410040225, + "per_class_f1_pct": { + "FAVOR": 62.42038216560509, + "AGAINST": 63.35403726708074, + "NONE": 60.97560975609756 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.7036246612466125, + "weighted_f1": 0.702598069105691, + "macro_precision": 0.7080208523870496, + "macro_recall": 0.717377201112141, + "per_class_f1": { + "FAVOR": 0.7317073170731707, + "AGAINST": 0.6666666666666666, + "NONE": 0.7125 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 70.36246612466125, + "weighted_f1_pct": 70.2598069105691, + "macro_precision_pct": 70.80208523870496, + "macro_recall_pct": 71.7377201112141, + "per_class_f1_pct": { + "FAVOR": 73.17073170731707, + "AGAINST": 66.66666666666666, + "NONE": 71.25 + } + } + ], + "accuracy": 0.631004366812227, + "macro_f1": 0.6341579287686554, + "weighted_f1": 0.6281005051726066, + "macro_precision": 0.6350895756394773, + "macro_recall": 0.6578269108902625, + "per_class_f1": { + "FAVOR": 0.6136101499423299, + "AGAINST": 0.6003787878787878, + "NONE": 0.6884848484848485 + }, + "accuracy_pct": 63.1004366812227, + "macro_f1_pct": 63.41579287686554, + "weighted_f1_pct": 62.81005051726066, + "macro_precision_pct": 63.50895756394773, + "macro_recall_pct": 65.78269108902624, + "per_class_f1_pct": { + "FAVOR": 61.36101499423299, + "AGAINST": 60.03787878787878, + "NONE": 68.84848484848484 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.647694505079904, + "losses": { + "current_ce": 0.44322704955314596, + "replay_ce": 0.06658820668429447, + "total": 0.5098152552110453, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6301369863013698, + "macro_f1": 0.6073496899292056, + "weighted_f1": 0.6453959545724778, + "macro_precision": 0.6044745702640439, + "macro_recall": 0.6836331834082959, + "per_class_f1": { + "FAVOR": 0.43478260869565216, + "AGAINST": 0.6845637583892618, + "NONE": 0.7027027027027027 + }, + "accuracy_pct": 63.013698630136986, + "macro_f1_pct": 60.73496899292056, + "weighted_f1_pct": 64.53959545724778, + "macro_precision_pct": 60.44745702640439, + "macro_recall_pct": 68.3633183408296, + "per_class_f1_pct": { + "FAVOR": 43.47826086956522, + "AGAINST": 68.45637583892618, + "NONE": 70.27027027027027 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6831452935896246, + "weighted_f1": 0.8180904998533154, + "macro_precision": 0.6626683501683502, + "macro_recall": 0.7545443514136633, + "per_class_f1": { + "FAVOR": 0.8429752066115702, + "AGAINST": 0.375, + "NONE": 0.8314606741573034 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 68.31452935896246, + "weighted_f1_pct": 81.80904998533154, + "macro_precision_pct": 66.26683501683502, + "macro_recall_pct": 75.45443514136633, + "per_class_f1_pct": { + "FAVOR": 84.29752066115702, + "AGAINST": 37.5, + "NONE": 83.14606741573034 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5369842212979468, + "weighted_f1": 0.5551196172248805, + "macro_precision": 0.5262890341321714, + "macro_recall": 0.5668119099491649, + "per_class_f1": { + "FAVOR": 0.42592592592592593, + "AGAINST": 0.6203208556149733, + "NONE": 0.5647058823529412 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 53.69842212979468, + "weighted_f1_pct": 55.511961722488046, + "macro_precision_pct": 52.628903413217145, + "macro_recall_pct": 56.681190994916484, + "per_class_f1_pct": { + "FAVOR": 42.592592592592595, + "AGAINST": 62.03208556149733, + "NONE": 56.470588235294116 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7312969656174251, + "weighted_f1": 0.7456295280210778, + "macro_precision": 0.716617262423714, + "macro_recall": 0.7968013468013467, + "per_class_f1": { + "FAVOR": 0.6506024096385542, + "AGAINST": 0.7582417582417582, + "NONE": 0.7850467289719626 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 73.1296965617425, + "weighted_f1_pct": 74.56295280210779, + "macro_precision_pct": 71.6617262423714, + "macro_recall_pct": 79.68013468013467, + "per_class_f1_pct": { + "FAVOR": 65.06024096385542, + "AGAINST": 75.82417582417582, + "NONE": 78.50467289719626 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6345177664974619, + "macro_f1": 0.6105596735635553, + "weighted_f1": 0.6414083738931068, + "macro_precision": 0.6002956738250856, + "macro_recall": 0.6480263330861978, + "per_class_f1": { + "FAVOR": 0.5128205128205128, + "AGAINST": 0.6802030456852792, + "NONE": 0.6386554621848739 + }, + "accuracy_pct": 63.45177664974619, + "macro_f1_pct": 61.05596735635553, + "weighted_f1_pct": 64.14083738931068, + "macro_precision_pct": 60.02956738250857, + "macro_recall_pct": 64.80263330861979, + "per_class_f1_pct": { + "FAVOR": 51.28205128205128, + "AGAINST": 68.02030456852792, + "NONE": 63.86554621848739 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5211267605633803, + "macro_f1": 0.5196620583717358, + "weighted_f1": 0.5134679041993899, + "macro_precision": 0.5133333333333333, + "macro_recall": 0.5307692307692308, + "per_class_f1": { + "FAVOR": 0.5161290322580645, + "AGAINST": 0.4, + "NONE": 0.6428571428571429 + }, + "accuracy_pct": 52.112676056338024, + "macro_f1_pct": 51.96620583717358, + "weighted_f1_pct": 51.34679041993899, + "macro_precision_pct": 51.33333333333333, + "macro_recall_pct": 53.07692307692308, + "per_class_f1_pct": { + "FAVOR": 51.61290322580645, + "AGAINST": 40.0, + "NONE": 64.28571428571429 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.5874919355900515, + "weighted_f1": 0.5661280167476399, + "macro_precision": 0.6233638282899366, + "macro_recall": 0.6243022259282421, + "per_class_f1": { + "FAVOR": 0.41379310344827586, + "AGAINST": 0.6597938144329897, + "NONE": 0.6888888888888889 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 58.74919355900515, + "weighted_f1_pct": 56.61280167476399, + "macro_precision_pct": 62.33638282899366, + "macro_recall_pct": 62.430222592824215, + "per_class_f1_pct": { + "FAVOR": 41.37931034482759, + "AGAINST": 65.97938144329896, + "NONE": 68.88888888888889 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6997858138460721, + "weighted_f1": 0.7033600350709389, + "macro_precision": 0.7035737740547029, + "macro_recall": 0.7090824837812789, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.6495726495726496, + "NONE": 0.7439024390243902 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.97858138460721, + "weighted_f1_pct": 70.33600350709389, + "macro_precision_pct": 70.35737740547029, + "macro_recall_pct": 70.9082483781279, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 64.95726495726495, + "NONE": 74.39024390243902 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7688943139049944, + "weighted_f1": 0.8013426914861154, + "macro_precision": 0.7584770532490221, + "macro_recall": 0.781797989893942, + "per_class_f1": { + "FAVOR": 0.7241379310344828, + "AGAINST": 0.7241379310344828, + "NONE": 0.8584070796460177 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 76.88943139049944, + "weighted_f1_pct": 80.13426914861154, + "macro_precision_pct": 75.84770532490222, + "macro_recall_pct": 78.1797989893942, + "per_class_f1_pct": { + "FAVOR": 72.41379310344827, + "AGAINST": 72.41379310344827, + "NONE": 85.84070796460178 + } + } + ], + "accuracy": 0.6639135959339263, + "macro_f1": 0.6585600316461284, + "weighted_f1": 0.6619221270538519, + "macro_precision": 0.6557057693676772, + "macro_recall": 0.6669425169001139, + "per_class_f1": { + "FAVOR": 0.5872235872235873, + "AGAINST": 0.6529968454258676, + "NONE": 0.7354596622889306 + }, + "accuracy_pct": 66.39135959339264, + "macro_f1_pct": 65.85600316461284, + "weighted_f1_pct": 66.19221270538519, + "macro_precision_pct": 65.57057693676772, + "macro_recall_pct": 66.69425169001138, + "per_class_f1_pct": { + "FAVOR": 58.72235872235873, + "AGAINST": 65.29968454258676, + "NONE": 73.54596622889305 + } + }, + { + "after_task": 10, + "best_epoch": 5, + "best_dev_macro_f1": 0.6522571110806404, + "losses": { + "current_ce": 0.6404847369744228, + "replay_ce": 0.0657416709436056, + "total": 0.7062264072589385, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6527639074953653, + "weighted_f1": 0.6859021949061538, + "macro_precision": 0.6431390406800243, + "macro_recall": 0.7247926036981509, + "per_class_f1": { + "FAVOR": 0.4857142857142857, + "AGAINST": 0.7189542483660131, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 65.27639074953653, + "weighted_f1_pct": 68.59021949061538, + "macro_precision_pct": 64.31390406800243, + "macro_recall_pct": 72.4792603698151, + "per_class_f1_pct": { + "FAVOR": 48.57142857142857, + "AGAINST": 71.89542483660131, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6499918975854805, + "weighted_f1": 0.8020938646796393, + "macro_precision": 0.6404601571268238, + "macro_recall": 0.7333090644339278, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.3, + "NONE": 0.8235294117647058 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 64.99918975854804, + "weighted_f1_pct": 80.20938646796392, + "macro_precision_pct": 64.04601571268238, + "macro_recall_pct": 73.33090644339279, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 30.0, + "NONE": 82.35294117647058 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5669446686395839, + "weighted_f1": 0.5802472282400918, + "macro_precision": 0.5631352282515073, + "macro_recall": 0.592955700798838, + "per_class_f1": { + "FAVOR": 0.5079365079365079, + "AGAINST": 0.6214689265536724, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 56.69446686395839, + "weighted_f1_pct": 58.02472282400918, + "macro_precision_pct": 56.31352282515073, + "macro_recall_pct": 59.295570079883795, + "per_class_f1_pct": { + "FAVOR": 50.79365079365079, + "AGAINST": 62.14689265536724, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6830477589306886, + "weighted_f1": 0.6953830818877265, + "macro_precision": 0.6797799599291771, + "macro_recall": 0.7524130190796857, + "per_class_f1": { + "FAVOR": 0.6185567010309279, + "AGAINST": 0.7071823204419889, + "NONE": 0.723404255319149 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 68.30477589306886, + "weighted_f1_pct": 69.53830818877265, + "macro_precision_pct": 67.97799599291771, + "macro_recall_pct": 75.24130190796858, + "per_class_f1_pct": { + "FAVOR": 61.855670103092784, + "AGAINST": 70.71823204419888, + "NONE": 72.3404255319149 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.5967464184574379, + "weighted_f1": 0.615130933898533, + "macro_precision": 0.5868592748867011, + "macro_recall": 0.6475583529201437, + "per_class_f1": { + "FAVOR": 0.5287356321839081, + "AGAINST": 0.6354166666666666, + "NONE": 0.6260869565217392 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 59.67464184574379, + "weighted_f1_pct": 61.5130933898533, + "macro_precision_pct": 58.68592748867011, + "macro_recall_pct": 64.75583529201437, + "per_class_f1_pct": { + "FAVOR": 52.87356321839081, + "AGAINST": 63.541666666666664, + "NONE": 62.60869565217392 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.49295774647887325, + "macro_f1": 0.47640287769784173, + "weighted_f1": 0.4889613942648698, + "macro_precision": 0.5309599156118143, + "macro_recall": 0.46196581196581193, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.4892086330935252, + "NONE": 0.54 + }, + "accuracy_pct": 49.29577464788733, + "macro_f1_pct": 47.64028776978417, + "weighted_f1_pct": 48.896139426486975, + "macro_precision_pct": 53.095991561181435, + "macro_recall_pct": 46.19658119658119, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 48.92086330935252, + "NONE": 54.0 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6324044133382172, + "weighted_f1": 0.6138473967790088, + "macro_precision": 0.6478678554150252, + "macro_recall": 0.6445038951814019, + "per_class_f1": { + "FAVOR": 0.5112781954887218, + "AGAINST": 0.6630434782608695, + "NONE": 0.7228915662650602 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 63.24044133382172, + "weighted_f1_pct": 61.38473967790088, + "macro_precision_pct": 64.78678554150252, + "macro_recall_pct": 64.4503895181402, + "per_class_f1_pct": { + "FAVOR": 51.127819548872175, + "AGAINST": 66.30434782608695, + "NONE": 72.28915662650603 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.685, + "macro_f1": 0.6789680683334199, + "weighted_f1": 0.6854529861219397, + "macro_precision": 0.6810408373875866, + "macro_recall": 0.6882298424467099, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.628099173553719, + "NONE": 0.7421383647798742 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 67.896806833342, + "weighted_f1_pct": 68.54529861219397, + "macro_precision_pct": 68.10408373875866, + "macro_recall_pct": 68.82298424467099, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.8099173553719, + "NONE": 74.21383647798741 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.79, + "macro_f1": 0.7644517402513696, + "weighted_f1": 0.7952148122784402, + "macro_precision": 0.7512513770578287, + "macro_recall": 0.7873951912932422, + "per_class_f1": { + "FAVOR": 0.7241379310344828, + "AGAINST": 0.71875, + "NONE": 0.8504672897196262 + }, + "accuracy_pct": 79.0, + "macro_f1_pct": 76.44517402513696, + "weighted_f1_pct": 79.52148122784402, + "macro_precision_pct": 75.12513770578288, + "macro_recall_pct": 78.73951912932422, + "per_class_f1_pct": { + "FAVOR": 72.41379310344827, + "AGAINST": 71.875, + "NONE": 85.04672897196261 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.64, + "macro_f1": 0.6203025634948848, + "weighted_f1": 0.6422410238179879, + "macro_precision": 0.6134367681498829, + "macro_recall": 0.6315681082671375, + "per_class_f1": { + "FAVOR": 0.6701030927835051, + "AGAINST": 0.4666666666666667, + "NONE": 0.7241379310344828 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 62.030256349488475, + "weighted_f1_pct": 64.22410238179879, + "macro_precision_pct": 61.34367681498829, + "macro_recall_pct": 63.15681082671375, + "per_class_f1_pct": { + "FAVOR": 67.0103092783505, + "AGAINST": 46.666666666666664, + "NONE": 72.41379310344827 + } + } + ], + "accuracy": 0.6544532130777903, + "macro_f1": 0.6554531233408406, + "weighted_f1": 0.6541930097206508, + "macro_precision": 0.6532636192249955, + "macro_recall": 0.6592388097502925, + "per_class_f1": { + "FAVOR": 0.6108468125594672, + "AGAINST": 0.6324909747292419, + "NONE": 0.7230215827338129 + }, + "accuracy_pct": 65.44532130777903, + "macro_f1_pct": 65.54531233408406, + "weighted_f1_pct": 65.41930097206507, + "macro_precision_pct": 65.32636192249956, + "macro_recall_pct": 65.92388097502925, + "per_class_f1_pct": { + "FAVOR": 61.08468125594671, + "AGAINST": 63.249097472924184, + "NONE": 72.3021582733813 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s1__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__x__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..25dcab4544bae6c6655daf863851f816797fe8ba --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__x__s1__seed13/run.log b/outputs/trace_s5_balanced_replay__x__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..8ea74a9a1809babab9729e3bbef7fb8726a09e05 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 03:24:27,082 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__x__s1__seed13 device=cuda +2026-09-02 03:39:55,440 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.632202431422429, 'average_forgetting': 0.04986387910726754, 'backward_transfer': -0.034875357315967324, 'final_mean_target_macro_f1_pct': 63.22024314224291, 'average_forgetting_pct': 4.986387910726754, 'backward_transfer_pct': -3.4875357315967324, 'final_pooled_accuracy': 0.6544532130777903, 'final_pooled_accuracy_pct': 65.44532130777903, 'final_pooled_macro_f1': 0.6554531233408406, 'final_pooled_macro_f1_pct': 65.54531233408406, 'final_pooled_weighted_f1': 0.6541930097206508, 'final_pooled_weighted_f1_pct': 65.41930097206507, 'final_pooled_macro_precision': 0.6532636192249955, 'final_pooled_macro_precision_pct': 65.32636192249956, 'final_pooled_macro_recall': 0.6592388097502925, 'final_pooled_macro_recall_pct': 65.92388097502925, 'final_pooled_per_class_f1': {'FAVOR': 0.6108468125594672, 'AGAINST': 0.6324909747292419, 'NONE': 0.7230215827338129}, 'final_pooled_per_class_f1_pct': {'FAVOR': 61.08468125594671, 'AGAINST': 63.249097472924184, 'NONE': 72.3021582733813}} +2026-09-02 03:39:55,440 [INFO] __main__: efficiency={'training_wall_time_seconds': 927.0189293648582, 'training_wall_time_minutes': 15.450315489414303, 'peak_gpu_memory_allocated_gb': 10.39700698852539, 'peak_gpu_memory_reserved_gb': 10.580078125} diff --git a/outputs/trace_s5_balanced_replay__x__s2__seed13/environment.json b/outputs/trace_s5_balanced_replay__x__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..b152c749bf20c1660cc61fd77df59bd759d6785d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T03:40:03.863547+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.json b/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..0a38787091c8dea9ef87f6f84859390ae8603140 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.json @@ -0,0 +1,2276 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6835748792270532, + 0.7275443510737629, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6309871698533299, + 0.6747236981691619, + 0.5953655162370759, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6362433862433862, + 0.607804309842749, + 0.6022097727980081, + 0.7336015762237098, + null, + null, + null, + null, + null, + null + ], + [ + 0.5896931604065522, + 0.6227161545026481, + 0.6024216524216525, + 0.7257798450929585, + 0.5980716253443527, + null, + null, + null, + null, + null + ], + [ + 0.6055861526357199, + 0.6544615024523103, + 0.6313757672120497, + 0.7027618594883743, + 0.5968065700920923, + 0.716434819660626, + null, + null, + null, + null + ], + [ + 0.6202986279257465, + 0.6397075299987921, + 0.6044879324959876, + 0.7098464592448549, + 0.5460723223825078, + 0.6324786324786325, + 0.5390796693275042, + null, + null, + null + ], + [ + 0.6483353704673768, + 0.6396166606271483, + 0.6176777921199827, + 0.6956594603653428, + 0.5450956787570798, + 0.6570826231272987, + 0.5910622710622712, + 0.6950415945968206, + null, + null + ], + [ + 0.6737522819394165, + 0.6358240346679653, + 0.6225693035835024, + 0.6926146010186757, + 0.5699683033016366, + 0.6358024691358025, + 0.5945609945609945, + 0.6708072624173669, + 0.6102701608182971, + null + ], + [ + 0.6741477609162733, + 0.6526305034308139, + 0.5704807811148189, + 0.6442254732178467, + 0.5218626947440507, + 0.6230395626868019, + 0.5634840385505098, + 0.6294462912109972, + 0.5833630520422973, + 0.7681733457595527 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6374794038024278, + "losses": { + "current_ce": 0.6368974308172862, + "replay_ce": 0.0639723111113364, + "total": 0.7008697423971061, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6835748792270532, + "weighted_f1": 0.8047069385661152, + "macro_precision": 0.6688131313131312, + "macro_recall": 0.7590340978036646, + "per_class_f1": { + "FAVOR": 0.8173913043478261, + "AGAINST": 0.4, + "NONE": 0.8333333333333334 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.35748792270532, + "weighted_f1_pct": 80.47069385661152, + "macro_precision_pct": 66.88131313131312, + "macro_recall_pct": 75.90340978036646, + "per_class_f1_pct": { + "FAVOR": 81.73913043478261, + "AGAINST": 40.0, + "NONE": 83.33333333333334 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7526881720430108, + "macro_f1": 0.7275443510737629, + "weighted_f1": 0.7566278124152888, + "macro_precision": 0.7134365486768233, + "macro_recall": 0.7533108866442199, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.8, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 75.26881720430107, + "macro_f1_pct": 72.7544351073763, + "weighted_f1_pct": 75.66278124152888, + "macro_precision_pct": 71.34365486768233, + "macro_recall_pct": 75.33108866442198, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 80.0, + "NONE": 72.54901960784314 + } + } + ], + "accuracy": 0.7692307692307693, + "macro_f1": 0.7688758525967828, + "weighted_f1": 0.7685989170316774, + "macro_precision": 0.7749183006535948, + "macro_recall": 0.7766207038485285, + "per_class_f1": { + "FAVOR": 0.7567567567567568, + "AGAINST": 0.772093023255814, + "NONE": 0.7777777777777778 + }, + "accuracy_pct": 76.92307692307693, + "macro_f1_pct": 76.88758525967827, + "weighted_f1_pct": 76.85989170316773, + "macro_precision_pct": 77.49183006535948, + "macro_recall_pct": 77.66207038485285, + "per_class_f1_pct": { + "FAVOR": 75.67567567567568, + "AGAINST": 77.20930232558139, + "NONE": 77.77777777777779 + } + }, + { + "after_task": 3, + "best_epoch": 14, + "best_dev_macro_f1": 0.6110201429499674, + "losses": { + "current_ce": 0.7396190399924915, + "replay_ce": 0.12601877210351328, + "total": 0.8656378117700417, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7522123893805309, + "macro_f1": 0.6309871698533299, + "weighted_f1": 0.7751937526276396, + "macro_precision": 0.6240231599886331, + "macro_recall": 0.7215386482223031, + "per_class_f1": { + "FAVOR": 0.7796610169491526, + "AGAINST": 0.2857142857142857, + "NONE": 0.8275862068965517 + }, + "accuracy_pct": 75.22123893805309, + "macro_f1_pct": 63.09871698533299, + "weighted_f1_pct": 77.51937526276396, + "macro_precision_pct": 62.40231599886331, + "macro_recall_pct": 72.1538648222303, + "per_class_f1_pct": { + "FAVOR": 77.96610169491525, + "AGAINST": 28.57142857142857, + "NONE": 82.75862068965517 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6747236981691619, + "weighted_f1": 0.6865144508431477, + "macro_precision": 0.6615521771771772, + "macro_recall": 0.7342312008978675, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7032967032967034, + "NONE": 0.6788990825688074 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 67.47236981691618, + "weighted_f1_pct": 68.65144508431477, + "macro_precision_pct": 66.15521771771772, + "macro_recall_pct": 73.42312008978675, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 70.32967032967034, + "NONE": 67.88990825688074 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5953655162370759, + "weighted_f1": 0.5766202720912632, + "macro_precision": 0.5877447948876521, + "macro_recall": 0.6094017094017095, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.45871559633027525, + "NONE": 0.6607142857142857 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 59.53655162370759, + "weighted_f1_pct": 57.66202720912632, + "macro_precision_pct": 58.77447948876521, + "macro_recall_pct": 60.940170940170944, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 45.87155963302752, + "NONE": 66.07142857142857 + } + } + ], + "accuracy": 0.6689342403628118, + "macro_f1": 0.6713176560504804, + "weighted_f1": 0.6641431880565701, + "macro_precision": 0.6711582605573295, + "macro_recall": 0.6814259907248096, + "per_class_f1": { + "FAVOR": 0.7099236641221374, + "AGAINST": 0.5897435897435898, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 66.89342403628117, + "macro_f1_pct": 67.13176560504805, + "weighted_f1_pct": 66.41431880565702, + "macro_precision_pct": 67.11582605573295, + "macro_recall_pct": 68.14259907248096, + "per_class_f1_pct": { + "FAVOR": 70.99236641221374, + "AGAINST": 58.97435897435898, + "NONE": 71.42857142857143 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.7357608178159442, + "losses": { + "current_ce": 0.6458759644757146, + "replay_ce": 0.04429892726458501, + "total": 0.6901748897372812, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6362433862433862, + "weighted_f1": 0.8028164067987077, + "macro_precision": 0.6257092007978882, + "macro_recall": 0.6765926465234801, + "per_class_f1": { + "FAVOR": 0.8253968253968254, + "AGAINST": 0.25, + "NONE": 0.8333333333333334 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 63.62433862433862, + "weighted_f1_pct": 80.28164067987078, + "macro_precision_pct": 62.57092007978882, + "macro_recall_pct": 67.659264652348, + "per_class_f1_pct": { + "FAVOR": 82.53968253968253, + "AGAINST": 25.0, + "NONE": 83.33333333333334 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6129032258064516, + "macro_f1": 0.607804309842749, + "weighted_f1": 0.6205572359892536, + "macro_precision": 0.6180536730821361, + "macro_recall": 0.6865319865319867, + "per_class_f1": { + "FAVOR": 0.5544554455445545, + "AGAINST": 0.6352941176470588, + "NONE": 0.6336633663366337 + }, + "accuracy_pct": 61.29032258064516, + "macro_f1_pct": 60.7804309842749, + "weighted_f1_pct": 62.05572359892536, + "macro_precision_pct": 61.80536730821361, + "macro_recall_pct": 68.65319865319867, + "per_class_f1_pct": { + "FAVOR": 55.44554455445545, + "AGAINST": 63.52941176470588, + "NONE": 63.366336633663366 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.6022097727980081, + "weighted_f1": 0.5793124761641746, + "macro_precision": 0.6089615931721194, + "macro_recall": 0.6504273504273504, + "per_class_f1": { + "FAVOR": 0.6764705882352942, + "AGAINST": 0.4, + "NONE": 0.7301587301587301 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 60.22097727980081, + "weighted_f1_pct": 57.93124761641746, + "macro_precision_pct": 60.89615931721194, + "macro_recall_pct": 65.04273504273505, + "per_class_f1_pct": { + "FAVOR": 67.64705882352942, + "AGAINST": 40.0, + "NONE": 73.01587301587301 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.74, + "macro_f1": 0.7336015762237098, + "weighted_f1": 0.7366693134881072, + "macro_precision": 0.7445993315399037, + "macro_recall": 0.7419833178869323, + "per_class_f1": { + "FAVOR": 0.7457627118644068, + "AGAINST": 0.6785714285714286, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 74.0, + "macro_f1_pct": 73.36015762237098, + "weighted_f1_pct": 73.66693134881072, + "macro_precision_pct": 74.45993315399036, + "macro_recall_pct": 74.19833178869322, + "per_class_f1_pct": { + "FAVOR": 74.57627118644068, + "AGAINST": 67.85714285714286, + "NONE": 77.64705882352942 + } + } + ], + "accuracy": 0.6833073322932918, + "macro_f1": 0.6778222462304239, + "weighted_f1": 0.673085601210694, + "macro_precision": 0.6904972804972805, + "macro_recall": 0.6960784313725489, + "per_class_f1": { + "FAVOR": 0.711864406779661, + "AGAINST": 0.5773195876288659, + "NONE": 0.7442827442827443 + }, + "accuracy_pct": 68.33073322932918, + "macro_f1_pct": 67.78222462304238, + "weighted_f1_pct": 67.3085601210694, + "macro_precision_pct": 69.04972804972806, + "macro_recall_pct": 69.60784313725489, + "per_class_f1_pct": { + "FAVOR": 71.1864406779661, + "AGAINST": 57.73195876288659, + "NONE": 74.42827442827443 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6250533987663583, + "losses": { + "current_ce": 0.6485345646356925, + "replay_ce": 0.05797687631864578, + "total": 0.7065114399561515, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.5896931604065522, + "weighted_f1": 0.7784830717478828, + "macro_precision": 0.5883526850507983, + "macro_recall": 0.6144157262468147, + "per_class_f1": { + "FAVOR": 0.8067226890756303, + "AGAINST": 0.15384615384615385, + "NONE": 0.8085106382978723 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 58.969316040655215, + "weighted_f1_pct": 77.84830717478827, + "macro_precision_pct": 58.83526850507983, + "macro_recall_pct": 61.44157262468147, + "per_class_f1_pct": { + "FAVOR": 80.67226890756302, + "AGAINST": 15.384615384615385, + "NONE": 80.85106382978722 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.6227161545026481, + "weighted_f1": 0.6323454433402401, + "macro_precision": 0.624428471598283, + "macro_recall": 0.6963524130190798, + "per_class_f1": { + "FAVOR": 0.5813953488372093, + "AGAINST": 0.6432748538011696, + "NONE": 0.6434782608695652 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.271615450264804, + "weighted_f1_pct": 63.234544334024015, + "macro_precision_pct": 62.4428471598283, + "macro_recall_pct": 69.63524130190798, + "per_class_f1_pct": { + "FAVOR": 58.139534883720934, + "AGAINST": 64.32748538011695, + "NONE": 64.34782608695652 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6024216524216525, + "weighted_f1": 0.5867461177320332, + "macro_precision": 0.5988185017596782, + "macro_recall": 0.6311965811965812, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.4444444444444444, + "NONE": 0.7166666666666667 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.24216524216525, + "weighted_f1_pct": 58.674611773203324, + "macro_precision_pct": 59.88185017596782, + "macro_recall_pct": 63.11965811965812, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 44.44444444444444, + "NONE": 71.66666666666667 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7257798450929585, + "weighted_f1": 0.7338070281636294, + "macro_precision": 0.7293711885241011, + "macro_recall": 0.730614766759345, + "per_class_f1": { + "FAVOR": 0.7017543859649122, + "AGAINST": 0.6779661016949152, + "NONE": 0.7976190476190477 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.57798450929585, + "weighted_f1_pct": 73.38070281636294, + "macro_precision_pct": 72.93711885241011, + "macro_recall_pct": 73.0614766759345, + "per_class_f1_pct": { + "FAVOR": 70.17543859649122, + "AGAINST": 67.79661016949152, + "NONE": 79.76190476190477 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.615, + "macro_f1": 0.5980716253443527, + "weighted_f1": 0.6218535353535354, + "macro_precision": 0.5961874382927014, + "macro_recall": 0.6200871404754901, + "per_class_f1": { + "FAVOR": 0.6555555555555556, + "AGAINST": 0.4444444444444444, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 61.5, + "macro_f1_pct": 59.807162534435264, + "weighted_f1_pct": 62.185353535353535, + "macro_precision_pct": 59.61874382927014, + "macro_recall_pct": 62.008714047549006, + "per_class_f1_pct": { + "FAVOR": 65.55555555555556, + "AGAINST": 44.44444444444444, + "NONE": 69.42148760330579 + } + } + ], + "accuracy": 0.6658739595719382, + "macro_f1": 0.6596749982785927, + "weighted_f1": 0.6594654851875861, + "macro_precision": 0.6639541996078498, + "macro_recall": 0.6665522501906941, + "per_class_f1": { + "FAVOR": 0.6843971631205674, + "AGAINST": 0.56, + "NONE": 0.7346278317152104 + }, + "accuracy_pct": 66.58739595719382, + "macro_f1_pct": 65.96749982785927, + "weighted_f1_pct": 65.9465485187586, + "macro_precision_pct": 66.39541996078499, + "macro_recall_pct": 66.6552250190694, + "per_class_f1_pct": { + "FAVOR": 68.43971631205675, + "AGAINST": 56.00000000000001, + "NONE": 73.46278317152104 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.7789046653144016, + "losses": { + "current_ce": 0.5580742065001417, + "replay_ce": 0.10421581771914605, + "total": 0.6622900280687544, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6055861526357199, + "weighted_f1": 0.8202968883813874, + "macro_precision": 0.6189987163029524, + "macro_recall": 0.6180560611576266, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.1, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 60.558615263571994, + "weighted_f1_pct": 82.02968883813874, + "macro_precision_pct": 61.899871630295245, + "macro_recall_pct": 61.80560611576266, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 10.0, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6544615024523103, + "weighted_f1": 0.6723758083563534, + "macro_precision": 0.6400875547217011, + "macro_recall": 0.7006734006734007, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.6947368421052632, + "NONE": 0.6804123711340206 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.44615024523102, + "weighted_f1_pct": 67.23758083563534, + "macro_precision_pct": 64.00875547217011, + "macro_recall_pct": 70.06734006734007, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 69.47368421052632, + "NONE": 68.04123711340206 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6313757672120497, + "weighted_f1": 0.6139715246910502, + "macro_precision": 0.6301006889242183, + "macro_recall": 0.6653846153846154, + "per_class_f1": { + "FAVOR": 0.6865671641791045, + "AGAINST": 0.4742268041237113, + "NONE": 0.7333333333333333 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 63.137576721204965, + "weighted_f1_pct": 61.39715246910502, + "macro_precision_pct": 63.010068892421835, + "macro_recall_pct": 66.53846153846153, + "per_class_f1_pct": { + "FAVOR": 68.65671641791045, + "AGAINST": 47.42268041237113, + "NONE": 73.33333333333333 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7027618594883743, + "weighted_f1": 0.7101193392850614, + "macro_precision": 0.7105063766207097, + "macro_recall": 0.7130985480383071, + "per_class_f1": { + "FAVOR": 0.6829268292682927, + "AGAINST": 0.6551724137931034, + "NONE": 0.7701863354037267 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.27618594883744, + "weighted_f1_pct": 71.01193392850614, + "macro_precision_pct": 71.05063766207097, + "macro_recall_pct": 71.3098548038307, + "per_class_f1_pct": { + "FAVOR": 68.29268292682927, + "AGAINST": 65.51724137931035, + "NONE": 77.01863354037268 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.5968065700920923, + "weighted_f1": 0.6235730786470365, + "macro_precision": 0.5914617128902843, + "macro_recall": 0.606398240378823, + "per_class_f1": { + "FAVOR": 0.6597938144329897, + "AGAINST": 0.4175824175824176, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 59.68065700920923, + "weighted_f1_pct": 62.35730786470365, + "macro_precision_pct": 59.14617128902842, + "macro_recall_pct": 60.63982403788229, + "per_class_f1_pct": { + "FAVOR": 65.97938144329896, + "AGAINST": 41.75824175824176, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.716434819660626, + "weighted_f1": 0.7657230381932105, + "macro_precision": 0.6993911719939118, + "macro_recall": 0.7659170414792603, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.8242424242424242, + "NONE": 0.7096774193548387 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 71.64348196606261, + "weighted_f1_pct": 76.57230381932105, + "macro_precision_pct": 69.93911719939118, + "macro_recall_pct": 76.59170414792604, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 82.42424242424242, + "NONE": 70.96774193548387 + } + } + ], + "accuracy": 0.6879432624113475, + "macro_f1": 0.6891556933232739, + "weighted_f1": 0.6853517264734997, + "macro_precision": 0.6888728300449541, + "macro_recall": 0.6957723720547624, + "per_class_f1": { + "FAVOR": 0.6869300911854104, + "AGAINST": 0.6332842415316642, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 68.79432624113475, + "macro_f1_pct": 68.91556933232738, + "weighted_f1_pct": 68.53517264734997, + "macro_precision_pct": 68.88728300449542, + "macro_recall_pct": 69.57723720547624, + "per_class_f1_pct": { + "FAVOR": 68.69300911854104, + "AGAINST": 63.32842415316642, + "NONE": 74.72527472527473 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.5543190249072603, + "losses": { + "current_ce": 0.6003178588368676, + "replay_ce": 0.0892754226358551, + "total": 0.6895932826128873, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6202986279257465, + "weighted_f1": 0.7902151321005378, + "macro_precision": 0.6412660939515159, + "macro_recall": 0.7102536099987865, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.21428571428571427, + "NONE": 0.85 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 62.029862792574654, + "weighted_f1_pct": 79.02151321005378, + "macro_precision_pct": 64.12660939515159, + "macro_recall_pct": 71.02536099987866, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 21.428571428571427, + "NONE": 85.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6397075299987921, + "weighted_f1": 0.6633240707935353, + "macro_precision": 0.6285324997292543, + "macro_recall": 0.6891133557800225, + "per_class_f1": { + "FAVOR": 0.5476190476190477, + "AGAINST": 0.6918918918918919, + "NONE": 0.6796116504854369 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 63.97075299987921, + "weighted_f1_pct": 66.33240707935353, + "macro_precision_pct": 62.853249972925425, + "macro_recall_pct": 68.91133557800225, + "per_class_f1_pct": { + "FAVOR": 54.761904761904766, + "AGAINST": 69.1891891891892, + "NONE": 67.96116504854369 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6044879324959876, + "weighted_f1": 0.5896358958480641, + "macro_precision": 0.5991130820399113, + "macro_recall": 0.6303418803418803, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.45544554455445546, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.448793249598765, + "weighted_f1_pct": 58.96358958480641, + "macro_precision_pct": 59.91130820399113, + "macro_recall_pct": 63.034188034188034, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 45.54455445544555, + "NONE": 71.1864406779661 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7098464592448549, + "weighted_f1": 0.7148429448225571, + "macro_precision": 0.7111197260450992, + "macro_recall": 0.7193388940376892, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.6611570247933884, + "NONE": 0.7625 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.98464592448549, + "weighted_f1_pct": 71.4842944822557, + "macro_precision_pct": 71.11197260450992, + "macro_recall_pct": 71.93388940376893, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 66.11570247933885, + "NONE": 76.25 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.54, + "macro_f1": 0.5460723223825078, + "weighted_f1": 0.5475513032983659, + "macro_precision": 0.5652777777777778, + "macro_recall": 0.5828198769946343, + "per_class_f1": { + "FAVOR": 0.5153374233128835, + "AGAINST": 0.4098360655737705, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 54.0, + "macro_f1_pct": 54.60723223825078, + "weighted_f1_pct": 54.75513032983659, + "macro_precision_pct": 56.52777777777778, + "macro_recall_pct": 58.281987699463436, + "per_class_f1_pct": { + "FAVOR": 51.533742331288344, + "AGAINST": 40.98360655737705, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6643835616438356, + "macro_f1": 0.6324786324786325, + "weighted_f1": 0.6785212504390586, + "macro_precision": 0.6196736483800437, + "macro_recall": 0.6957171414292853, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7307692307692307, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.43835616438356, + "macro_f1_pct": 63.24786324786324, + "weighted_f1_pct": 67.85212504390586, + "macro_precision_pct": 61.96736483800437, + "macro_recall_pct": 69.57171414292853, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 73.07692307692307, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5390796693275042, + "weighted_f1": 0.5740354553566935, + "macro_precision": 0.5531156582880721, + "macro_recall": 0.5421205519244735, + "per_class_f1": { + "FAVOR": 0.42696629213483145, + "AGAINST": 0.6697247706422018, + "NONE": 0.5205479452054794 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 53.90796693275042, + "weighted_f1_pct": 57.40354553566935, + "macro_precision_pct": 55.311565828807204, + "macro_recall_pct": 54.21205519244735, + "per_class_f1_pct": { + "FAVOR": 42.69662921348314, + "AGAINST": 66.97247706422019, + "NONE": 52.054794520547944 + } + } + ], + "accuracy": 0.6380628717077316, + "macro_f1": 0.6402201076264876, + "weighted_f1": 0.6361207949344883, + "macro_precision": 0.6383637427770354, + "macro_recall": 0.6444503805319308, + "per_class_f1": { + "FAVOR": 0.5982905982905983, + "AGAINST": 0.6122448979591837, + "NONE": 0.710124826629681 + }, + "accuracy_pct": 63.80628717077316, + "macro_f1_pct": 64.02201076264876, + "weighted_f1_pct": 63.612079493448825, + "macro_precision_pct": 63.83637427770354, + "macro_recall_pct": 64.44503805319309, + "per_class_f1_pct": { + "FAVOR": 59.82905982905983, + "AGAINST": 61.224489795918366, + "NONE": 71.0124826629681 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.723030707610147, + "losses": { + "current_ce": 0.5816113097198081, + "replay_ce": 0.08163069700540015, + "total": 0.6632420072952906, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6483353704673768, + "weighted_f1": 0.8105068986990671, + "macro_precision": 0.649445941898772, + "macro_recall": 0.7364640213566315, + "per_class_f1": { + "FAVOR": 0.8166666666666667, + "AGAINST": 0.2608695652173913, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 64.83353704673767, + "weighted_f1_pct": 81.0506898699067, + "macro_precision_pct": 64.94459418987721, + "macro_recall_pct": 73.64640213566315, + "per_class_f1_pct": { + "FAVOR": 81.66666666666667, + "AGAINST": 26.08695652173913, + "NONE": 86.74698795180723 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6451612903225806, + "macro_f1": 0.6396166606271483, + "weighted_f1": 0.6499728930980277, + "macro_precision": 0.6390290037831021, + "macro_recall": 0.7153198653198652, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6589595375722543, + "NONE": 0.6792452830188679 + }, + "accuracy_pct": 64.51612903225806, + "macro_f1_pct": 63.961666062714826, + "weighted_f1_pct": 64.99728930980278, + "macro_precision_pct": 63.902900378310214, + "macro_recall_pct": 71.53198653198652, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 65.89595375722543, + "NONE": 67.9245283018868 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6176777921199827, + "weighted_f1": 0.5924115595286638, + "macro_precision": 0.6106150793650794, + "macro_recall": 0.6435897435897436, + "per_class_f1": { + "FAVOR": 0.7272727272727273, + "AGAINST": 0.47058823529411764, + "NONE": 0.6551724137931034 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 61.767779211998274, + "weighted_f1_pct": 59.24115595286638, + "macro_precision_pct": 61.061507936507944, + "macro_recall_pct": 64.35897435897436, + "per_class_f1_pct": { + "FAVOR": 72.72727272727273, + "AGAINST": 47.05882352941176, + "NONE": 65.51724137931035 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6956594603653428, + "weighted_f1": 0.700326797385621, + "macro_precision": 0.7073260073260071, + "macro_recall": 0.7122489959839359, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.6495726495726496, + "NONE": 0.7450980392156863 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 69.56594603653427, + "weighted_f1_pct": 70.0326797385621, + "macro_precision_pct": 70.73260073260072, + "macro_recall_pct": 71.22489959839359, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 64.95726495726495, + "NONE": 74.50980392156863 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.555, + "macro_f1": 0.5450956787570798, + "weighted_f1": 0.5667382028873635, + "macro_precision": 0.5503921365990331, + "macro_recall": 0.5605542246318945, + "per_class_f1": { + "FAVOR": 0.5888888888888889, + "AGAINST": 0.37383177570093457, + "NONE": 0.672566371681416 + }, + "accuracy_pct": 55.50000000000001, + "macro_f1_pct": 54.50956787570797, + "weighted_f1_pct": 56.67382028873635, + "macro_precision_pct": 55.03921365990331, + "macro_recall_pct": 56.05542246318945, + "per_class_f1_pct": { + "FAVOR": 58.88888888888889, + "AGAINST": 37.38317757009346, + "NONE": 67.2566371681416 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6570826231272987, + "weighted_f1": 0.6966709514569311, + "macro_precision": 0.6405372405372406, + "macro_recall": 0.7223288355822088, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7388535031847133, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 65.70826231272987, + "weighted_f1_pct": 69.66709514569311, + "macro_precision_pct": 64.05372405372405, + "macro_recall_pct": 72.23288355822089, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 73.88535031847134, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.5910622710622712, + "weighted_f1": 0.6038079814921921, + "macro_precision": 0.5885399264951553, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.6285714285714286, + "NONE": 0.56 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 59.10622710622712, + "weighted_f1_pct": 60.3807981492192, + "macro_precision_pct": 58.85399264951553, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 62.857142857142854, + "NONE": 56.00000000000001 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6950415945968206, + "weighted_f1": 0.719056793338175, + "macro_precision": 0.6798478323284526, + "macro_recall": 0.7202214755572381, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7605633802816901, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.50415945968206, + "weighted_f1_pct": 71.9056793338175, + "macro_precision_pct": 67.98478323284526, + "macro_recall_pct": 72.02214755572382, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 76.05633802816901, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6550218340611353, + "macro_f1": 0.6576786651825738, + "weighted_f1": 0.6536170435293784, + "macro_precision": 0.6546396504729838, + "macro_recall": 0.6686774925571864, + "per_class_f1": { + "FAVOR": 0.6449359720605355, + "AGAINST": 0.6298031865042174, + "NONE": 0.6982968369829684 + }, + "accuracy_pct": 65.50218340611353, + "macro_f1_pct": 65.76786651825738, + "weighted_f1_pct": 65.36170435293785, + "macro_precision_pct": 65.46396504729837, + "macro_recall_pct": 66.86774925571865, + "per_class_f1_pct": { + "FAVOR": 64.49359720605355, + "AGAINST": 62.980318650421744, + "NONE": 69.82968369829685 + } + }, + { + "after_task": 9, + "best_epoch": 4, + "best_dev_macro_f1": 0.6380892976053002, + "losses": { + "current_ce": 0.5619961916927307, + "replay_ce": 0.06968050430573168, + "total": 0.6316766947980911, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6737522819394165, + "weighted_f1": 0.8263373937958629, + "macro_precision": 0.6625302663438256, + "macro_recall": 0.7482344375682562, + "per_class_f1": { + "FAVOR": 0.8412698412698413, + "AGAINST": 0.3157894736842105, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.37522819394165, + "weighted_f1_pct": 82.63373937958629, + "macro_precision_pct": 66.25302663438256, + "macro_recall_pct": 74.82344375682563, + "per_class_f1_pct": { + "FAVOR": 84.12698412698413, + "AGAINST": 31.57894736842105, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6397849462365591, + "macro_f1": 0.6358240346679653, + "weighted_f1": 0.6424811436931519, + "macro_precision": 0.632828091448781, + "macro_recall": 0.7122334455667789, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.6473988439306358, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 63.97849462365591, + "macro_f1_pct": 63.582403466796535, + "weighted_f1_pct": 64.24811436931519, + "macro_precision_pct": 63.28280914487811, + "macro_recall_pct": 71.22334455667789, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 64.73988439306359, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.6225693035835024, + "weighted_f1": 0.5959460617661344, + "macro_precision": 0.6140899122807018, + "macro_recall": 0.6555555555555556, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.46, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 62.25693035835024, + "weighted_f1_pct": 59.59460617661344, + "macro_precision_pct": 61.40899122807018, + "macro_recall_pct": 65.55555555555556, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 46.0, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.695, + "macro_f1": 0.6926146010186757, + "weighted_f1": 0.6963465316517099, + "macro_precision": 0.7139479905437351, + "macro_recall": 0.7082329317269077, + "per_class_f1": { + "FAVOR": 0.6766917293233082, + "AGAINST": 0.6785714285714286, + "NONE": 0.7225806451612903 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 69.26146010186757, + "weighted_f1_pct": 69.63465316517099, + "macro_precision_pct": 71.39479905437352, + "macro_recall_pct": 70.82329317269077, + "per_class_f1_pct": { + "FAVOR": 67.66917293233082, + "AGAINST": 67.85714285714286, + "NONE": 72.25806451612902 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.575, + "macro_f1": 0.5699683033016366, + "weighted_f1": 0.5861599511599512, + "macro_precision": 0.5753228487405703, + "macro_recall": 0.5857240925202091, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.4074074074074074, + "NONE": 0.7090909090909091 + }, + "accuracy_pct": 57.49999999999999, + "macro_f1_pct": 56.99683033016366, + "weighted_f1_pct": 58.615995115995126, + "macro_precision_pct": 57.53228487405703, + "macro_recall_pct": 58.572409252020904, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 40.74074074074074, + "NONE": 70.9090909090909 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6358024691358025, + "weighted_f1": 0.6848046676813799, + "macro_precision": 0.6206646206646207, + "macro_recall": 0.6835982008995503, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7407407407407407, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 63.580246913580254, + "weighted_f1_pct": 68.48046676813799, + "macro_precision_pct": 62.066462066462066, + "macro_recall_pct": 68.35982008995502, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 74.07407407407408, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.5945609945609945, + "weighted_f1": 0.6132842596000491, + "macro_precision": 0.5838985088985089, + "macro_recall": 0.6136528685548294, + "per_class_f1": { + "FAVOR": 0.5272727272727272, + "AGAINST": 0.6666666666666666, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 59.45609945609945, + "weighted_f1_pct": 61.32842596000491, + "macro_precision_pct": 58.38985088985089, + "macro_recall_pct": 61.365286855482935, + "per_class_f1_pct": { + "FAVOR": 52.72727272727272, + "AGAINST": 66.66666666666666, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6852791878172588, + "macro_f1": 0.6708072624173669, + "weighted_f1": 0.6879735318933035, + "macro_precision": 0.6564478114478115, + "macro_recall": 0.7025013276965385, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7169811320754716, + "NONE": 0.6534653465346535 + }, + "accuracy_pct": 68.52791878172589, + "macro_f1_pct": 67.08072624173668, + "weighted_f1_pct": 68.79735318933035, + "macro_precision_pct": 65.64478114478115, + "macro_recall_pct": 70.25013276965385, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 71.69811320754717, + "NONE": 65.34653465346535 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.6102701608182971, + "weighted_f1": 0.5910209383334757, + "macro_precision": 0.6083450210378681, + "macro_recall": 0.6200287216547379, + "per_class_f1": { + "FAVOR": 0.5895953757225434, + "AGAINST": 0.5285714285714286, + "NONE": 0.7126436781609196 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 61.02701608182971, + "weighted_f1_pct": 59.10209383334757, + "macro_precision_pct": 60.83450210378681, + "macro_recall_pct": 62.00287216547379, + "per_class_f1_pct": { + "FAVOR": 58.95953757225434, + "AGAINST": 52.85714285714286, + "NONE": 71.26436781609196 + } + } + ], + "accuracy": 0.6473951715374842, + "macro_f1": 0.6509345117010061, + "weighted_f1": 0.6461837852060663, + "macro_precision": 0.648248853818474, + "macro_recall": 0.6601571544311856, + "per_class_f1": { + "FAVOR": 0.6342412451361867, + "AGAINST": 0.6223316912972086, + "NONE": 0.6962305986696231 + }, + "accuracy_pct": 64.73951715374841, + "macro_f1_pct": 65.09345117010061, + "weighted_f1_pct": 64.61837852060664, + "macro_precision_pct": 64.8248853818474, + "macro_recall_pct": 66.01571544311857, + "per_class_f1_pct": { + "FAVOR": 63.42412451361867, + "AGAINST": 62.23316912972086, + "NONE": 69.62305986696231 + } + }, + { + "after_task": 10, + "best_epoch": 14, + "best_dev_macro_f1": 0.7130811162421837, + "losses": { + "current_ce": 0.4171718764739732, + "replay_ce": 0.08801929204103848, + "total": 0.5051911684994896, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6741477609162733, + "weighted_f1": 0.8255704288327685, + "macro_precision": 0.6608225108225109, + "macro_recall": 0.7545443514136633, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.3157894736842105, + "NONE": 0.8705882352941177 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.41477609162733, + "weighted_f1_pct": 82.55704288327685, + "macro_precision_pct": 66.08225108225109, + "macro_recall_pct": 75.45443514136633, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 31.57894736842105, + "NONE": 87.05882352941177 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6526305034308139, + "weighted_f1": 0.6589506187144791, + "macro_precision": 0.6438228339209605, + "macro_recall": 0.7241863075196409, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6629213483146067, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 65.26305034308139, + "weighted_f1_pct": 65.89506187144791, + "macro_precision_pct": 64.38228339209606, + "macro_recall_pct": 72.41863075196409, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 66.29213483146067, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5704807811148189, + "weighted_f1": 0.5304297419668176, + "macro_precision": 0.556523569023569, + "macro_recall": 0.6307692307692307, + "per_class_f1": { + "FAVOR": 0.7466666666666667, + "AGAINST": 0.34408602150537637, + "NONE": 0.6206896551724138 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 57.04807811148189, + "weighted_f1_pct": 53.042974196681755, + "macro_precision_pct": 55.6523569023569, + "macro_recall_pct": 63.07692307692307, + "per_class_f1_pct": { + "FAVOR": 74.66666666666667, + "AGAINST": 34.40860215053764, + "NONE": 62.06896551724138 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.65, + "macro_f1": 0.6442254732178467, + "weighted_f1": 0.6508110530014348, + "macro_precision": 0.644696809128886, + "macro_recall": 0.6531047265987024, + "per_class_f1": { + "FAVOR": 0.6386554621848739, + "AGAINST": 0.5806451612903226, + "NONE": 0.7133757961783439 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 64.42254732178468, + "weighted_f1_pct": 65.08110530014348, + "macro_precision_pct": 64.4696809128886, + "macro_recall_pct": 65.31047265987024, + "per_class_f1_pct": { + "FAVOR": 63.86554621848739, + "AGAINST": 58.06451612903226, + "NONE": 71.3375796178344 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.52, + "macro_f1": 0.5218626947440507, + "weighted_f1": 0.5273795582948125, + "macro_precision": 0.536361063950198, + "macro_recall": 0.5548256490004063, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.3898305084745763, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 52.0, + "macro_f1_pct": 52.186269474405066, + "weighted_f1_pct": 52.73795582948125, + "macro_precision_pct": 53.6361063950198, + "macro_recall_pct": 55.48256490004063, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 38.983050847457626, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6575342465753424, + "macro_f1": 0.6230395626868019, + "weighted_f1": 0.6716561912119313, + "macro_precision": 0.6096163877854018, + "macro_recall": 0.6648025987006497, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.7239263803680982, + "NONE": 0.7076923076923077 + }, + "accuracy_pct": 65.75342465753424, + "macro_f1_pct": 62.303956268680196, + "weighted_f1_pct": 67.16561912119313, + "macro_precision_pct": 60.961638778540184, + "macro_recall_pct": 66.48025987006497, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 72.39263803680981, + "NONE": 70.76923076923077 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5684210526315789, + "macro_f1": 0.5634840385505098, + "weighted_f1": 0.5728725680400185, + "macro_precision": 0.5578569455951395, + "macro_recall": 0.6027596223674655, + "per_class_f1": { + "FAVOR": 0.5245901639344263, + "AGAINST": 0.6011560693641619, + "NONE": 0.5647058823529412 + }, + "accuracy_pct": 56.84210526315789, + "macro_f1_pct": 56.348403855050975, + "weighted_f1_pct": 57.28725680400185, + "macro_precision_pct": 55.785694559513956, + "macro_recall_pct": 60.27596223674655, + "per_class_f1_pct": { + "FAVOR": 52.459016393442624, + "AGAINST": 60.115606936416185, + "NONE": 56.470588235294116 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6294416243654822, + "macro_f1": 0.6294462912109972, + "weighted_f1": 0.6307876934183294, + "macro_precision": 0.6205205205205205, + "macro_recall": 0.7022857413279069, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6203208556149733, + "NONE": 0.6846846846846847 + }, + "accuracy_pct": 62.944162436548226, + "macro_f1_pct": 62.944629121099716, + "weighted_f1_pct": 63.07876934183294, + "macro_precision_pct": 62.052052052052055, + "macro_recall_pct": 70.22857413279068, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 62.03208556149733, + "NONE": 68.46846846846847 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.565, + "macro_f1": 0.5833630520422973, + "weighted_f1": 0.5598248807795978, + "macro_precision": 0.573147253533727, + "macro_recall": 0.6004601641458011, + "per_class_f1": { + "FAVOR": 0.49333333333333335, + "AGAINST": 0.5534591194968553, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 58.33630520422973, + "weighted_f1_pct": 55.98248807795978, + "macro_precision_pct": 57.314725353372694, + "macro_recall_pct": 60.04601641458011, + "per_class_f1_pct": { + "FAVOR": 49.333333333333336, + "AGAINST": 55.34591194968554, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7681733457595527, + "weighted_f1": 0.8079007455731593, + "macro_precision": 0.7560546678725757, + "macro_recall": 0.7837951394673034, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.75, + "NONE": 0.8648648648648649 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 76.81733457595527, + "weighted_f1_pct": 80.79007455731593, + "macro_precision_pct": 75.60546678725757, + "macro_recall_pct": 78.37951394673034, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 75.0, + "NONE": 86.48648648648648 + } + } + ], + "accuracy": 0.6364148816234498, + "macro_f1": 0.6377867577257841, + "weighted_f1": 0.6341511980057811, + "macro_precision": 0.6346589857664919, + "macro_recall": 0.6459808162203298, + "per_class_f1": { + "FAVOR": 0.5983066792097836, + "AGAINST": 0.5922038980509745, + "NONE": 0.7228496959165943 + }, + "accuracy_pct": 63.641488162344984, + "macro_f1_pct": 63.77867577257841, + "weighted_f1_pct": 63.41511980057811, + "macro_precision_pct": 63.465898576649195, + "macro_recall_pct": 64.59808162203298, + "per_class_f1_pct": { + "FAVOR": 59.830667920978364, + "AGAINST": 59.220389805097454, + "NONE": 72.28496959165943 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6230853503673963, + "average_forgetting": 0.06809766323097238, + "backward_transfer": -0.05793193254114304, + "final_mean_target_macro_f1_pct": 62.308535036739634, + "average_forgetting_pct": 6.8097663230972385, + "backward_transfer_pct": -5.793193254114303, + "final_pooled_accuracy": 0.6364148816234498, + "final_pooled_accuracy_pct": 63.641488162344984, + "final_pooled_macro_f1": 0.6377867577257841, + "final_pooled_macro_f1_pct": 63.77867577257841, + "final_pooled_weighted_f1": 0.6341511980057811, + "final_pooled_weighted_f1_pct": 63.41511980057811, + "final_pooled_macro_precision": 0.6346589857664919, + "final_pooled_macro_precision_pct": 63.465898576649195, + "final_pooled_macro_recall": 0.6459808162203298, + "final_pooled_macro_recall_pct": 64.59808162203298, + "final_pooled_per_class_f1": { + "FAVOR": 0.5983066792097836, + "AGAINST": 0.5922038980509745, + "NONE": 0.7228496959165943 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 59.830667920978364, + "AGAINST": 59.220389805097454, + "NONE": 72.28496959165943 + } + }, + "efficiency": { + "training_wall_time_seconds": 941.6750660918187, + "training_wall_time_minutes": 15.694584434863646, + "peak_gpu_memory_allocated_gb": 10.39810562133789, + "peak_gpu_memory_reserved_gb": 10.62109375 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..519dfcafdb2975cd531765e92ef356a54d88272d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s2__seed13/metrics.partial.json @@ -0,0 +1,2160 @@ +{ + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6835748792270532, + 0.7275443510737629, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6309871698533299, + 0.6747236981691619, + 0.5953655162370759, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6362433862433862, + 0.607804309842749, + 0.6022097727980081, + 0.7336015762237098, + null, + null, + null, + null, + null, + null + ], + [ + 0.5896931604065522, + 0.6227161545026481, + 0.6024216524216525, + 0.7257798450929585, + 0.5980716253443527, + null, + null, + null, + null, + null + ], + [ + 0.6055861526357199, + 0.6544615024523103, + 0.6313757672120497, + 0.7027618594883743, + 0.5968065700920923, + 0.716434819660626, + null, + null, + null, + null + ], + [ + 0.6202986279257465, + 0.6397075299987921, + 0.6044879324959876, + 0.7098464592448549, + 0.5460723223825078, + 0.6324786324786325, + 0.5390796693275042, + null, + null, + null + ], + [ + 0.6483353704673768, + 0.6396166606271483, + 0.6176777921199827, + 0.6956594603653428, + 0.5450956787570798, + 0.6570826231272987, + 0.5910622710622712, + 0.6950415945968206, + null, + null + ], + [ + 0.6737522819394165, + 0.6358240346679653, + 0.6225693035835024, + 0.6926146010186757, + 0.5699683033016366, + 0.6358024691358025, + 0.5945609945609945, + 0.6708072624173669, + 0.6102701608182971, + null + ], + [ + 0.6741477609162733, + 0.6526305034308139, + 0.5704807811148189, + 0.6442254732178467, + 0.5218626947440507, + 0.6230395626868019, + 0.5634840385505098, + 0.6294462912109972, + 0.5833630520422973, + 0.7681733457595527 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6374794038024278, + "losses": { + "current_ce": 0.6368974308172862, + "replay_ce": 0.0639723111113364, + "total": 0.7008697423971061, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6835748792270532, + "weighted_f1": 0.8047069385661152, + "macro_precision": 0.6688131313131312, + "macro_recall": 0.7590340978036646, + "per_class_f1": { + "FAVOR": 0.8173913043478261, + "AGAINST": 0.4, + "NONE": 0.8333333333333334 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.35748792270532, + "weighted_f1_pct": 80.47069385661152, + "macro_precision_pct": 66.88131313131312, + "macro_recall_pct": 75.90340978036646, + "per_class_f1_pct": { + "FAVOR": 81.73913043478261, + "AGAINST": 40.0, + "NONE": 83.33333333333334 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7526881720430108, + "macro_f1": 0.7275443510737629, + "weighted_f1": 0.7566278124152888, + "macro_precision": 0.7134365486768233, + "macro_recall": 0.7533108866442199, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.8, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 75.26881720430107, + "macro_f1_pct": 72.7544351073763, + "weighted_f1_pct": 75.66278124152888, + "macro_precision_pct": 71.34365486768233, + "macro_recall_pct": 75.33108866442198, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 80.0, + "NONE": 72.54901960784314 + } + } + ], + "accuracy": 0.7692307692307693, + "macro_f1": 0.7688758525967828, + "weighted_f1": 0.7685989170316774, + "macro_precision": 0.7749183006535948, + "macro_recall": 0.7766207038485285, + "per_class_f1": { + "FAVOR": 0.7567567567567568, + "AGAINST": 0.772093023255814, + "NONE": 0.7777777777777778 + }, + "accuracy_pct": 76.92307692307693, + "macro_f1_pct": 76.88758525967827, + "weighted_f1_pct": 76.85989170316773, + "macro_precision_pct": 77.49183006535948, + "macro_recall_pct": 77.66207038485285, + "per_class_f1_pct": { + "FAVOR": 75.67567567567568, + "AGAINST": 77.20930232558139, + "NONE": 77.77777777777779 + } + }, + { + "after_task": 3, + "best_epoch": 14, + "best_dev_macro_f1": 0.6110201429499674, + "losses": { + "current_ce": 0.7396190399924915, + "replay_ce": 0.12601877210351328, + "total": 0.8656378117700417, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7522123893805309, + "macro_f1": 0.6309871698533299, + "weighted_f1": 0.7751937526276396, + "macro_precision": 0.6240231599886331, + "macro_recall": 0.7215386482223031, + "per_class_f1": { + "FAVOR": 0.7796610169491526, + "AGAINST": 0.2857142857142857, + "NONE": 0.8275862068965517 + }, + "accuracy_pct": 75.22123893805309, + "macro_f1_pct": 63.09871698533299, + "weighted_f1_pct": 77.51937526276396, + "macro_precision_pct": 62.40231599886331, + "macro_recall_pct": 72.1538648222303, + "per_class_f1_pct": { + "FAVOR": 77.96610169491525, + "AGAINST": 28.57142857142857, + "NONE": 82.75862068965517 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6747236981691619, + "weighted_f1": 0.6865144508431477, + "macro_precision": 0.6615521771771772, + "macro_recall": 0.7342312008978675, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7032967032967034, + "NONE": 0.6788990825688074 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 67.47236981691618, + "weighted_f1_pct": 68.65144508431477, + "macro_precision_pct": 66.15521771771772, + "macro_recall_pct": 73.42312008978675, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 70.32967032967034, + "NONE": 67.88990825688074 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5953655162370759, + "weighted_f1": 0.5766202720912632, + "macro_precision": 0.5877447948876521, + "macro_recall": 0.6094017094017095, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.45871559633027525, + "NONE": 0.6607142857142857 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 59.53655162370759, + "weighted_f1_pct": 57.66202720912632, + "macro_precision_pct": 58.77447948876521, + "macro_recall_pct": 60.940170940170944, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 45.87155963302752, + "NONE": 66.07142857142857 + } + } + ], + "accuracy": 0.6689342403628118, + "macro_f1": 0.6713176560504804, + "weighted_f1": 0.6641431880565701, + "macro_precision": 0.6711582605573295, + "macro_recall": 0.6814259907248096, + "per_class_f1": { + "FAVOR": 0.7099236641221374, + "AGAINST": 0.5897435897435898, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 66.89342403628117, + "macro_f1_pct": 67.13176560504805, + "weighted_f1_pct": 66.41431880565702, + "macro_precision_pct": 67.11582605573295, + "macro_recall_pct": 68.14259907248096, + "per_class_f1_pct": { + "FAVOR": 70.99236641221374, + "AGAINST": 58.97435897435898, + "NONE": 71.42857142857143 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.7357608178159442, + "losses": { + "current_ce": 0.6458759644757146, + "replay_ce": 0.04429892726458501, + "total": 0.6901748897372812, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6362433862433862, + "weighted_f1": 0.8028164067987077, + "macro_precision": 0.6257092007978882, + "macro_recall": 0.6765926465234801, + "per_class_f1": { + "FAVOR": 0.8253968253968254, + "AGAINST": 0.25, + "NONE": 0.8333333333333334 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 63.62433862433862, + "weighted_f1_pct": 80.28164067987078, + "macro_precision_pct": 62.57092007978882, + "macro_recall_pct": 67.659264652348, + "per_class_f1_pct": { + "FAVOR": 82.53968253968253, + "AGAINST": 25.0, + "NONE": 83.33333333333334 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6129032258064516, + "macro_f1": 0.607804309842749, + "weighted_f1": 0.6205572359892536, + "macro_precision": 0.6180536730821361, + "macro_recall": 0.6865319865319867, + "per_class_f1": { + "FAVOR": 0.5544554455445545, + "AGAINST": 0.6352941176470588, + "NONE": 0.6336633663366337 + }, + "accuracy_pct": 61.29032258064516, + "macro_f1_pct": 60.7804309842749, + "weighted_f1_pct": 62.05572359892536, + "macro_precision_pct": 61.80536730821361, + "macro_recall_pct": 68.65319865319867, + "per_class_f1_pct": { + "FAVOR": 55.44554455445545, + "AGAINST": 63.52941176470588, + "NONE": 63.366336633663366 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.6022097727980081, + "weighted_f1": 0.5793124761641746, + "macro_precision": 0.6089615931721194, + "macro_recall": 0.6504273504273504, + "per_class_f1": { + "FAVOR": 0.6764705882352942, + "AGAINST": 0.4, + "NONE": 0.7301587301587301 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 60.22097727980081, + "weighted_f1_pct": 57.93124761641746, + "macro_precision_pct": 60.89615931721194, + "macro_recall_pct": 65.04273504273505, + "per_class_f1_pct": { + "FAVOR": 67.64705882352942, + "AGAINST": 40.0, + "NONE": 73.01587301587301 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.74, + "macro_f1": 0.7336015762237098, + "weighted_f1": 0.7366693134881072, + "macro_precision": 0.7445993315399037, + "macro_recall": 0.7419833178869323, + "per_class_f1": { + "FAVOR": 0.7457627118644068, + "AGAINST": 0.6785714285714286, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 74.0, + "macro_f1_pct": 73.36015762237098, + "weighted_f1_pct": 73.66693134881072, + "macro_precision_pct": 74.45993315399036, + "macro_recall_pct": 74.19833178869322, + "per_class_f1_pct": { + "FAVOR": 74.57627118644068, + "AGAINST": 67.85714285714286, + "NONE": 77.64705882352942 + } + } + ], + "accuracy": 0.6833073322932918, + "macro_f1": 0.6778222462304239, + "weighted_f1": 0.673085601210694, + "macro_precision": 0.6904972804972805, + "macro_recall": 0.6960784313725489, + "per_class_f1": { + "FAVOR": 0.711864406779661, + "AGAINST": 0.5773195876288659, + "NONE": 0.7442827442827443 + }, + "accuracy_pct": 68.33073322932918, + "macro_f1_pct": 67.78222462304238, + "weighted_f1_pct": 67.3085601210694, + "macro_precision_pct": 69.04972804972806, + "macro_recall_pct": 69.60784313725489, + "per_class_f1_pct": { + "FAVOR": 71.1864406779661, + "AGAINST": 57.73195876288659, + "NONE": 74.42827442827443 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6250533987663583, + "losses": { + "current_ce": 0.6485345646356925, + "replay_ce": 0.05797687631864578, + "total": 0.7065114399561515, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.5896931604065522, + "weighted_f1": 0.7784830717478828, + "macro_precision": 0.5883526850507983, + "macro_recall": 0.6144157262468147, + "per_class_f1": { + "FAVOR": 0.8067226890756303, + "AGAINST": 0.15384615384615385, + "NONE": 0.8085106382978723 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 58.969316040655215, + "weighted_f1_pct": 77.84830717478827, + "macro_precision_pct": 58.83526850507983, + "macro_recall_pct": 61.44157262468147, + "per_class_f1_pct": { + "FAVOR": 80.67226890756302, + "AGAINST": 15.384615384615385, + "NONE": 80.85106382978722 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.6227161545026481, + "weighted_f1": 0.6323454433402401, + "macro_precision": 0.624428471598283, + "macro_recall": 0.6963524130190798, + "per_class_f1": { + "FAVOR": 0.5813953488372093, + "AGAINST": 0.6432748538011696, + "NONE": 0.6434782608695652 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.271615450264804, + "weighted_f1_pct": 63.234544334024015, + "macro_precision_pct": 62.4428471598283, + "macro_recall_pct": 69.63524130190798, + "per_class_f1_pct": { + "FAVOR": 58.139534883720934, + "AGAINST": 64.32748538011695, + "NONE": 64.34782608695652 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6024216524216525, + "weighted_f1": 0.5867461177320332, + "macro_precision": 0.5988185017596782, + "macro_recall": 0.6311965811965812, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.4444444444444444, + "NONE": 0.7166666666666667 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.24216524216525, + "weighted_f1_pct": 58.674611773203324, + "macro_precision_pct": 59.88185017596782, + "macro_recall_pct": 63.11965811965812, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 44.44444444444444, + "NONE": 71.66666666666667 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7257798450929585, + "weighted_f1": 0.7338070281636294, + "macro_precision": 0.7293711885241011, + "macro_recall": 0.730614766759345, + "per_class_f1": { + "FAVOR": 0.7017543859649122, + "AGAINST": 0.6779661016949152, + "NONE": 0.7976190476190477 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.57798450929585, + "weighted_f1_pct": 73.38070281636294, + "macro_precision_pct": 72.93711885241011, + "macro_recall_pct": 73.0614766759345, + "per_class_f1_pct": { + "FAVOR": 70.17543859649122, + "AGAINST": 67.79661016949152, + "NONE": 79.76190476190477 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.615, + "macro_f1": 0.5980716253443527, + "weighted_f1": 0.6218535353535354, + "macro_precision": 0.5961874382927014, + "macro_recall": 0.6200871404754901, + "per_class_f1": { + "FAVOR": 0.6555555555555556, + "AGAINST": 0.4444444444444444, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 61.5, + "macro_f1_pct": 59.807162534435264, + "weighted_f1_pct": 62.185353535353535, + "macro_precision_pct": 59.61874382927014, + "macro_recall_pct": 62.008714047549006, + "per_class_f1_pct": { + "FAVOR": 65.55555555555556, + "AGAINST": 44.44444444444444, + "NONE": 69.42148760330579 + } + } + ], + "accuracy": 0.6658739595719382, + "macro_f1": 0.6596749982785927, + "weighted_f1": 0.6594654851875861, + "macro_precision": 0.6639541996078498, + "macro_recall": 0.6665522501906941, + "per_class_f1": { + "FAVOR": 0.6843971631205674, + "AGAINST": 0.56, + "NONE": 0.7346278317152104 + }, + "accuracy_pct": 66.58739595719382, + "macro_f1_pct": 65.96749982785927, + "weighted_f1_pct": 65.9465485187586, + "macro_precision_pct": 66.39541996078499, + "macro_recall_pct": 66.6552250190694, + "per_class_f1_pct": { + "FAVOR": 68.43971631205675, + "AGAINST": 56.00000000000001, + "NONE": 73.46278317152104 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.7789046653144016, + "losses": { + "current_ce": 0.5580742065001417, + "replay_ce": 0.10421581771914605, + "total": 0.6622900280687544, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6055861526357199, + "weighted_f1": 0.8202968883813874, + "macro_precision": 0.6189987163029524, + "macro_recall": 0.6180560611576266, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.1, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 60.558615263571994, + "weighted_f1_pct": 82.02968883813874, + "macro_precision_pct": 61.899871630295245, + "macro_recall_pct": 61.80560611576266, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 10.0, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6544615024523103, + "weighted_f1": 0.6723758083563534, + "macro_precision": 0.6400875547217011, + "macro_recall": 0.7006734006734007, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.6947368421052632, + "NONE": 0.6804123711340206 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.44615024523102, + "weighted_f1_pct": 67.23758083563534, + "macro_precision_pct": 64.00875547217011, + "macro_recall_pct": 70.06734006734007, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 69.47368421052632, + "NONE": 68.04123711340206 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6313757672120497, + "weighted_f1": 0.6139715246910502, + "macro_precision": 0.6301006889242183, + "macro_recall": 0.6653846153846154, + "per_class_f1": { + "FAVOR": 0.6865671641791045, + "AGAINST": 0.4742268041237113, + "NONE": 0.7333333333333333 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 63.137576721204965, + "weighted_f1_pct": 61.39715246910502, + "macro_precision_pct": 63.010068892421835, + "macro_recall_pct": 66.53846153846153, + "per_class_f1_pct": { + "FAVOR": 68.65671641791045, + "AGAINST": 47.42268041237113, + "NONE": 73.33333333333333 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7027618594883743, + "weighted_f1": 0.7101193392850614, + "macro_precision": 0.7105063766207097, + "macro_recall": 0.7130985480383071, + "per_class_f1": { + "FAVOR": 0.6829268292682927, + "AGAINST": 0.6551724137931034, + "NONE": 0.7701863354037267 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.27618594883744, + "weighted_f1_pct": 71.01193392850614, + "macro_precision_pct": 71.05063766207097, + "macro_recall_pct": 71.3098548038307, + "per_class_f1_pct": { + "FAVOR": 68.29268292682927, + "AGAINST": 65.51724137931035, + "NONE": 77.01863354037268 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.5968065700920923, + "weighted_f1": 0.6235730786470365, + "macro_precision": 0.5914617128902843, + "macro_recall": 0.606398240378823, + "per_class_f1": { + "FAVOR": 0.6597938144329897, + "AGAINST": 0.4175824175824176, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 59.68065700920923, + "weighted_f1_pct": 62.35730786470365, + "macro_precision_pct": 59.14617128902842, + "macro_recall_pct": 60.63982403788229, + "per_class_f1_pct": { + "FAVOR": 65.97938144329896, + "AGAINST": 41.75824175824176, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.716434819660626, + "weighted_f1": 0.7657230381932105, + "macro_precision": 0.6993911719939118, + "macro_recall": 0.7659170414792603, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.8242424242424242, + "NONE": 0.7096774193548387 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 71.64348196606261, + "weighted_f1_pct": 76.57230381932105, + "macro_precision_pct": 69.93911719939118, + "macro_recall_pct": 76.59170414792604, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 82.42424242424242, + "NONE": 70.96774193548387 + } + } + ], + "accuracy": 0.6879432624113475, + "macro_f1": 0.6891556933232739, + "weighted_f1": 0.6853517264734997, + "macro_precision": 0.6888728300449541, + "macro_recall": 0.6957723720547624, + "per_class_f1": { + "FAVOR": 0.6869300911854104, + "AGAINST": 0.6332842415316642, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 68.79432624113475, + "macro_f1_pct": 68.91556933232738, + "weighted_f1_pct": 68.53517264734997, + "macro_precision_pct": 68.88728300449542, + "macro_recall_pct": 69.57723720547624, + "per_class_f1_pct": { + "FAVOR": 68.69300911854104, + "AGAINST": 63.32842415316642, + "NONE": 74.72527472527473 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.5543190249072603, + "losses": { + "current_ce": 0.6003178588368676, + "replay_ce": 0.0892754226358551, + "total": 0.6895932826128873, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6202986279257465, + "weighted_f1": 0.7902151321005378, + "macro_precision": 0.6412660939515159, + "macro_recall": 0.7102536099987865, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.21428571428571427, + "NONE": 0.85 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 62.029862792574654, + "weighted_f1_pct": 79.02151321005378, + "macro_precision_pct": 64.12660939515159, + "macro_recall_pct": 71.02536099987866, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 21.428571428571427, + "NONE": 85.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6397075299987921, + "weighted_f1": 0.6633240707935353, + "macro_precision": 0.6285324997292543, + "macro_recall": 0.6891133557800225, + "per_class_f1": { + "FAVOR": 0.5476190476190477, + "AGAINST": 0.6918918918918919, + "NONE": 0.6796116504854369 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 63.97075299987921, + "weighted_f1_pct": 66.33240707935353, + "macro_precision_pct": 62.853249972925425, + "macro_recall_pct": 68.91133557800225, + "per_class_f1_pct": { + "FAVOR": 54.761904761904766, + "AGAINST": 69.1891891891892, + "NONE": 67.96116504854369 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6044879324959876, + "weighted_f1": 0.5896358958480641, + "macro_precision": 0.5991130820399113, + "macro_recall": 0.6303418803418803, + "per_class_f1": { + "FAVOR": 0.6461538461538462, + "AGAINST": 0.45544554455445546, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 60.448793249598765, + "weighted_f1_pct": 58.96358958480641, + "macro_precision_pct": 59.91130820399113, + "macro_recall_pct": 63.034188034188034, + "per_class_f1_pct": { + "FAVOR": 64.61538461538461, + "AGAINST": 45.54455445544555, + "NONE": 71.1864406779661 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7098464592448549, + "weighted_f1": 0.7148429448225571, + "macro_precision": 0.7111197260450992, + "macro_recall": 0.7193388940376892, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.6611570247933884, + "NONE": 0.7625 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.98464592448549, + "weighted_f1_pct": 71.4842944822557, + "macro_precision_pct": 71.11197260450992, + "macro_recall_pct": 71.93388940376893, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 66.11570247933885, + "NONE": 76.25 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.54, + "macro_f1": 0.5460723223825078, + "weighted_f1": 0.5475513032983659, + "macro_precision": 0.5652777777777778, + "macro_recall": 0.5828198769946343, + "per_class_f1": { + "FAVOR": 0.5153374233128835, + "AGAINST": 0.4098360655737705, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 54.0, + "macro_f1_pct": 54.60723223825078, + "weighted_f1_pct": 54.75513032983659, + "macro_precision_pct": 56.52777777777778, + "macro_recall_pct": 58.281987699463436, + "per_class_f1_pct": { + "FAVOR": 51.533742331288344, + "AGAINST": 40.98360655737705, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6643835616438356, + "macro_f1": 0.6324786324786325, + "weighted_f1": 0.6785212504390586, + "macro_precision": 0.6196736483800437, + "macro_recall": 0.6957171414292853, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7307692307692307, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.43835616438356, + "macro_f1_pct": 63.24786324786324, + "weighted_f1_pct": 67.85212504390586, + "macro_precision_pct": 61.96736483800437, + "macro_recall_pct": 69.57171414292853, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 73.07692307692307, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5390796693275042, + "weighted_f1": 0.5740354553566935, + "macro_precision": 0.5531156582880721, + "macro_recall": 0.5421205519244735, + "per_class_f1": { + "FAVOR": 0.42696629213483145, + "AGAINST": 0.6697247706422018, + "NONE": 0.5205479452054794 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 53.90796693275042, + "weighted_f1_pct": 57.40354553566935, + "macro_precision_pct": 55.311565828807204, + "macro_recall_pct": 54.21205519244735, + "per_class_f1_pct": { + "FAVOR": 42.69662921348314, + "AGAINST": 66.97247706422019, + "NONE": 52.054794520547944 + } + } + ], + "accuracy": 0.6380628717077316, + "macro_f1": 0.6402201076264876, + "weighted_f1": 0.6361207949344883, + "macro_precision": 0.6383637427770354, + "macro_recall": 0.6444503805319308, + "per_class_f1": { + "FAVOR": 0.5982905982905983, + "AGAINST": 0.6122448979591837, + "NONE": 0.710124826629681 + }, + "accuracy_pct": 63.80628717077316, + "macro_f1_pct": 64.02201076264876, + "weighted_f1_pct": 63.612079493448825, + "macro_precision_pct": 63.83637427770354, + "macro_recall_pct": 64.44503805319309, + "per_class_f1_pct": { + "FAVOR": 59.82905982905983, + "AGAINST": 61.224489795918366, + "NONE": 71.0124826629681 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.723030707610147, + "losses": { + "current_ce": 0.5816113097198081, + "replay_ce": 0.08163069700540015, + "total": 0.6632420072952906, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6483353704673768, + "weighted_f1": 0.8105068986990671, + "macro_precision": 0.649445941898772, + "macro_recall": 0.7364640213566315, + "per_class_f1": { + "FAVOR": 0.8166666666666667, + "AGAINST": 0.2608695652173913, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 64.83353704673767, + "weighted_f1_pct": 81.0506898699067, + "macro_precision_pct": 64.94459418987721, + "macro_recall_pct": 73.64640213566315, + "per_class_f1_pct": { + "FAVOR": 81.66666666666667, + "AGAINST": 26.08695652173913, + "NONE": 86.74698795180723 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6451612903225806, + "macro_f1": 0.6396166606271483, + "weighted_f1": 0.6499728930980277, + "macro_precision": 0.6390290037831021, + "macro_recall": 0.7153198653198652, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6589595375722543, + "NONE": 0.6792452830188679 + }, + "accuracy_pct": 64.51612903225806, + "macro_f1_pct": 63.961666062714826, + "weighted_f1_pct": 64.99728930980278, + "macro_precision_pct": 63.902900378310214, + "macro_recall_pct": 71.53198653198652, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 65.89595375722543, + "NONE": 67.9245283018868 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6176777921199827, + "weighted_f1": 0.5924115595286638, + "macro_precision": 0.6106150793650794, + "macro_recall": 0.6435897435897436, + "per_class_f1": { + "FAVOR": 0.7272727272727273, + "AGAINST": 0.47058823529411764, + "NONE": 0.6551724137931034 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 61.767779211998274, + "weighted_f1_pct": 59.24115595286638, + "macro_precision_pct": 61.061507936507944, + "macro_recall_pct": 64.35897435897436, + "per_class_f1_pct": { + "FAVOR": 72.72727272727273, + "AGAINST": 47.05882352941176, + "NONE": 65.51724137931035 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6956594603653428, + "weighted_f1": 0.700326797385621, + "macro_precision": 0.7073260073260071, + "macro_recall": 0.7122489959839359, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.6495726495726496, + "NONE": 0.7450980392156863 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 69.56594603653427, + "weighted_f1_pct": 70.0326797385621, + "macro_precision_pct": 70.73260073260072, + "macro_recall_pct": 71.22489959839359, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 64.95726495726495, + "NONE": 74.50980392156863 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.555, + "macro_f1": 0.5450956787570798, + "weighted_f1": 0.5667382028873635, + "macro_precision": 0.5503921365990331, + "macro_recall": 0.5605542246318945, + "per_class_f1": { + "FAVOR": 0.5888888888888889, + "AGAINST": 0.37383177570093457, + "NONE": 0.672566371681416 + }, + "accuracy_pct": 55.50000000000001, + "macro_f1_pct": 54.50956787570797, + "weighted_f1_pct": 56.67382028873635, + "macro_precision_pct": 55.03921365990331, + "macro_recall_pct": 56.05542246318945, + "per_class_f1_pct": { + "FAVOR": 58.88888888888889, + "AGAINST": 37.38317757009346, + "NONE": 67.2566371681416 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6570826231272987, + "weighted_f1": 0.6966709514569311, + "macro_precision": 0.6405372405372406, + "macro_recall": 0.7223288355822088, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7388535031847133, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 65.70826231272987, + "weighted_f1_pct": 69.66709514569311, + "macro_precision_pct": 64.05372405372405, + "macro_recall_pct": 72.23288355822089, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 73.88535031847134, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.5910622710622712, + "weighted_f1": 0.6038079814921921, + "macro_precision": 0.5885399264951553, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.6285714285714286, + "NONE": 0.56 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 59.10622710622712, + "weighted_f1_pct": 60.3807981492192, + "macro_precision_pct": 58.85399264951553, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 62.857142857142854, + "NONE": 56.00000000000001 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6950415945968206, + "weighted_f1": 0.719056793338175, + "macro_precision": 0.6798478323284526, + "macro_recall": 0.7202214755572381, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7605633802816901, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.50415945968206, + "weighted_f1_pct": 71.9056793338175, + "macro_precision_pct": 67.98478323284526, + "macro_recall_pct": 72.02214755572382, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 76.05633802816901, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6550218340611353, + "macro_f1": 0.6576786651825738, + "weighted_f1": 0.6536170435293784, + "macro_precision": 0.6546396504729838, + "macro_recall": 0.6686774925571864, + "per_class_f1": { + "FAVOR": 0.6449359720605355, + "AGAINST": 0.6298031865042174, + "NONE": 0.6982968369829684 + }, + "accuracy_pct": 65.50218340611353, + "macro_f1_pct": 65.76786651825738, + "weighted_f1_pct": 65.36170435293785, + "macro_precision_pct": 65.46396504729837, + "macro_recall_pct": 66.86774925571865, + "per_class_f1_pct": { + "FAVOR": 64.49359720605355, + "AGAINST": 62.980318650421744, + "NONE": 69.82968369829685 + } + }, + { + "after_task": 9, + "best_epoch": 4, + "best_dev_macro_f1": 0.6380892976053002, + "losses": { + "current_ce": 0.5619961916927307, + "replay_ce": 0.06968050430573168, + "total": 0.6316766947980911, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6737522819394165, + "weighted_f1": 0.8263373937958629, + "macro_precision": 0.6625302663438256, + "macro_recall": 0.7482344375682562, + "per_class_f1": { + "FAVOR": 0.8412698412698413, + "AGAINST": 0.3157894736842105, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.37522819394165, + "weighted_f1_pct": 82.63373937958629, + "macro_precision_pct": 66.25302663438256, + "macro_recall_pct": 74.82344375682563, + "per_class_f1_pct": { + "FAVOR": 84.12698412698413, + "AGAINST": 31.57894736842105, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6397849462365591, + "macro_f1": 0.6358240346679653, + "weighted_f1": 0.6424811436931519, + "macro_precision": 0.632828091448781, + "macro_recall": 0.7122334455667789, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.6473988439306358, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 63.97849462365591, + "macro_f1_pct": 63.582403466796535, + "weighted_f1_pct": 64.24811436931519, + "macro_precision_pct": 63.28280914487811, + "macro_recall_pct": 71.22334455667789, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 64.73988439306359, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.6225693035835024, + "weighted_f1": 0.5959460617661344, + "macro_precision": 0.6140899122807018, + "macro_recall": 0.6555555555555556, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.46, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 62.25693035835024, + "weighted_f1_pct": 59.59460617661344, + "macro_precision_pct": 61.40899122807018, + "macro_recall_pct": 65.55555555555556, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 46.0, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.695, + "macro_f1": 0.6926146010186757, + "weighted_f1": 0.6963465316517099, + "macro_precision": 0.7139479905437351, + "macro_recall": 0.7082329317269077, + "per_class_f1": { + "FAVOR": 0.6766917293233082, + "AGAINST": 0.6785714285714286, + "NONE": 0.7225806451612903 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 69.26146010186757, + "weighted_f1_pct": 69.63465316517099, + "macro_precision_pct": 71.39479905437352, + "macro_recall_pct": 70.82329317269077, + "per_class_f1_pct": { + "FAVOR": 67.66917293233082, + "AGAINST": 67.85714285714286, + "NONE": 72.25806451612902 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.575, + "macro_f1": 0.5699683033016366, + "weighted_f1": 0.5861599511599512, + "macro_precision": 0.5753228487405703, + "macro_recall": 0.5857240925202091, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.4074074074074074, + "NONE": 0.7090909090909091 + }, + "accuracy_pct": 57.49999999999999, + "macro_f1_pct": 56.99683033016366, + "weighted_f1_pct": 58.615995115995126, + "macro_precision_pct": 57.53228487405703, + "macro_recall_pct": 58.572409252020904, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 40.74074074074074, + "NONE": 70.9090909090909 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6712328767123288, + "macro_f1": 0.6358024691358025, + "weighted_f1": 0.6848046676813799, + "macro_precision": 0.6206646206646207, + "macro_recall": 0.6835982008995503, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7407407407407407, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.12328767123287, + "macro_f1_pct": 63.580246913580254, + "weighted_f1_pct": 68.48046676813799, + "macro_precision_pct": 62.066462066462066, + "macro_recall_pct": 68.35982008995502, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 74.07407407407408, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.5945609945609945, + "weighted_f1": 0.6132842596000491, + "macro_precision": 0.5838985088985089, + "macro_recall": 0.6136528685548294, + "per_class_f1": { + "FAVOR": 0.5272727272727272, + "AGAINST": 0.6666666666666666, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 59.45609945609945, + "weighted_f1_pct": 61.32842596000491, + "macro_precision_pct": 58.38985088985089, + "macro_recall_pct": 61.365286855482935, + "per_class_f1_pct": { + "FAVOR": 52.72727272727272, + "AGAINST": 66.66666666666666, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6852791878172588, + "macro_f1": 0.6708072624173669, + "weighted_f1": 0.6879735318933035, + "macro_precision": 0.6564478114478115, + "macro_recall": 0.7025013276965385, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7169811320754716, + "NONE": 0.6534653465346535 + }, + "accuracy_pct": 68.52791878172589, + "macro_f1_pct": 67.08072624173668, + "weighted_f1_pct": 68.79735318933035, + "macro_precision_pct": 65.64478114478115, + "macro_recall_pct": 70.25013276965385, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 71.69811320754717, + "NONE": 65.34653465346535 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.595, + "macro_f1": 0.6102701608182971, + "weighted_f1": 0.5910209383334757, + "macro_precision": 0.6083450210378681, + "macro_recall": 0.6200287216547379, + "per_class_f1": { + "FAVOR": 0.5895953757225434, + "AGAINST": 0.5285714285714286, + "NONE": 0.7126436781609196 + }, + "accuracy_pct": 59.5, + "macro_f1_pct": 61.02701608182971, + "weighted_f1_pct": 59.10209383334757, + "macro_precision_pct": 60.83450210378681, + "macro_recall_pct": 62.00287216547379, + "per_class_f1_pct": { + "FAVOR": 58.95953757225434, + "AGAINST": 52.85714285714286, + "NONE": 71.26436781609196 + } + } + ], + "accuracy": 0.6473951715374842, + "macro_f1": 0.6509345117010061, + "weighted_f1": 0.6461837852060663, + "macro_precision": 0.648248853818474, + "macro_recall": 0.6601571544311856, + "per_class_f1": { + "FAVOR": 0.6342412451361867, + "AGAINST": 0.6223316912972086, + "NONE": 0.6962305986696231 + }, + "accuracy_pct": 64.73951715374841, + "macro_f1_pct": 65.09345117010061, + "weighted_f1_pct": 64.61837852060664, + "macro_precision_pct": 64.8248853818474, + "macro_recall_pct": 66.01571544311857, + "per_class_f1_pct": { + "FAVOR": 63.42412451361867, + "AGAINST": 62.23316912972086, + "NONE": 69.62305986696231 + } + }, + { + "after_task": 10, + "best_epoch": 14, + "best_dev_macro_f1": 0.7130811162421837, + "losses": { + "current_ce": 0.4171718764739732, + "replay_ce": 0.08801929204103848, + "total": 0.5051911684994896, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6741477609162733, + "weighted_f1": 0.8255704288327685, + "macro_precision": 0.6608225108225109, + "macro_recall": 0.7545443514136633, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.3157894736842105, + "NONE": 0.8705882352941177 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.41477609162733, + "weighted_f1_pct": 82.55704288327685, + "macro_precision_pct": 66.08225108225109, + "macro_recall_pct": 75.45443514136633, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 31.57894736842105, + "NONE": 87.05882352941177 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6526305034308139, + "weighted_f1": 0.6589506187144791, + "macro_precision": 0.6438228339209605, + "macro_recall": 0.7241863075196409, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6629213483146067, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 65.26305034308139, + "weighted_f1_pct": 65.89506187144791, + "macro_precision_pct": 64.38228339209606, + "macro_recall_pct": 72.41863075196409, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 66.29213483146067, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5704807811148189, + "weighted_f1": 0.5304297419668176, + "macro_precision": 0.556523569023569, + "macro_recall": 0.6307692307692307, + "per_class_f1": { + "FAVOR": 0.7466666666666667, + "AGAINST": 0.34408602150537637, + "NONE": 0.6206896551724138 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 57.04807811148189, + "weighted_f1_pct": 53.042974196681755, + "macro_precision_pct": 55.6523569023569, + "macro_recall_pct": 63.07692307692307, + "per_class_f1_pct": { + "FAVOR": 74.66666666666667, + "AGAINST": 34.40860215053764, + "NONE": 62.06896551724138 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.65, + "macro_f1": 0.6442254732178467, + "weighted_f1": 0.6508110530014348, + "macro_precision": 0.644696809128886, + "macro_recall": 0.6531047265987024, + "per_class_f1": { + "FAVOR": 0.6386554621848739, + "AGAINST": 0.5806451612903226, + "NONE": 0.7133757961783439 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 64.42254732178468, + "weighted_f1_pct": 65.08110530014348, + "macro_precision_pct": 64.4696809128886, + "macro_recall_pct": 65.31047265987024, + "per_class_f1_pct": { + "FAVOR": 63.86554621848739, + "AGAINST": 58.06451612903226, + "NONE": 71.3375796178344 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.52, + "macro_f1": 0.5218626947440507, + "weighted_f1": 0.5273795582948125, + "macro_precision": 0.536361063950198, + "macro_recall": 0.5548256490004063, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.3898305084745763, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 52.0, + "macro_f1_pct": 52.186269474405066, + "weighted_f1_pct": 52.73795582948125, + "macro_precision_pct": 53.6361063950198, + "macro_recall_pct": 55.48256490004063, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 38.983050847457626, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6575342465753424, + "macro_f1": 0.6230395626868019, + "weighted_f1": 0.6716561912119313, + "macro_precision": 0.6096163877854018, + "macro_recall": 0.6648025987006497, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.7239263803680982, + "NONE": 0.7076923076923077 + }, + "accuracy_pct": 65.75342465753424, + "macro_f1_pct": 62.303956268680196, + "weighted_f1_pct": 67.16561912119313, + "macro_precision_pct": 60.961638778540184, + "macro_recall_pct": 66.48025987006497, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 72.39263803680981, + "NONE": 70.76923076923077 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5684210526315789, + "macro_f1": 0.5634840385505098, + "weighted_f1": 0.5728725680400185, + "macro_precision": 0.5578569455951395, + "macro_recall": 0.6027596223674655, + "per_class_f1": { + "FAVOR": 0.5245901639344263, + "AGAINST": 0.6011560693641619, + "NONE": 0.5647058823529412 + }, + "accuracy_pct": 56.84210526315789, + "macro_f1_pct": 56.348403855050975, + "weighted_f1_pct": 57.28725680400185, + "macro_precision_pct": 55.785694559513956, + "macro_recall_pct": 60.27596223674655, + "per_class_f1_pct": { + "FAVOR": 52.459016393442624, + "AGAINST": 60.115606936416185, + "NONE": 56.470588235294116 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6294416243654822, + "macro_f1": 0.6294462912109972, + "weighted_f1": 0.6307876934183294, + "macro_precision": 0.6205205205205205, + "macro_recall": 0.7022857413279069, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6203208556149733, + "NONE": 0.6846846846846847 + }, + "accuracy_pct": 62.944162436548226, + "macro_f1_pct": 62.944629121099716, + "weighted_f1_pct": 63.07876934183294, + "macro_precision_pct": 62.052052052052055, + "macro_recall_pct": 70.22857413279068, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 62.03208556149733, + "NONE": 68.46846846846847 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.565, + "macro_f1": 0.5833630520422973, + "weighted_f1": 0.5598248807795978, + "macro_precision": 0.573147253533727, + "macro_recall": 0.6004601641458011, + "per_class_f1": { + "FAVOR": 0.49333333333333335, + "AGAINST": 0.5534591194968553, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 56.49999999999999, + "macro_f1_pct": 58.33630520422973, + "weighted_f1_pct": 55.98248807795978, + "macro_precision_pct": 57.314725353372694, + "macro_recall_pct": 60.04601641458011, + "per_class_f1_pct": { + "FAVOR": 49.333333333333336, + "AGAINST": 55.34591194968554, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7681733457595527, + "weighted_f1": 0.8079007455731593, + "macro_precision": 0.7560546678725757, + "macro_recall": 0.7837951394673034, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.75, + "NONE": 0.8648648648648649 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 76.81733457595527, + "weighted_f1_pct": 80.79007455731593, + "macro_precision_pct": 75.60546678725757, + "macro_recall_pct": 78.37951394673034, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 75.0, + "NONE": 86.48648648648648 + } + } + ], + "accuracy": 0.6364148816234498, + "macro_f1": 0.6377867577257841, + "weighted_f1": 0.6341511980057811, + "macro_precision": 0.6346589857664919, + "macro_recall": 0.6459808162203298, + "per_class_f1": { + "FAVOR": 0.5983066792097836, + "AGAINST": 0.5922038980509745, + "NONE": 0.7228496959165943 + }, + "accuracy_pct": 63.641488162344984, + "macro_f1_pct": 63.77867577257841, + "weighted_f1_pct": 63.41511980057811, + "macro_precision_pct": 63.465898576649195, + "macro_recall_pct": 64.59808162203298, + "per_class_f1_pct": { + "FAVOR": 59.830667920978364, + "AGAINST": 59.220389805097454, + "NONE": 72.28496959165943 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s2__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__x__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1acc09101c3e7210118edf3faf952eb14ce8ac27 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__x__s2__seed13/run.log b/outputs/trace_s5_balanced_replay__x__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..b46865363ec0ca77fe05ce2891a0a586b7b5b295 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 03:40:03,913 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__x__s2__seed13 device=cuda +2026-09-02 03:55:46,902 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6230853503673963, 'average_forgetting': 0.06809766323097238, 'backward_transfer': -0.05793193254114304, 'final_mean_target_macro_f1_pct': 62.308535036739634, 'average_forgetting_pct': 6.8097663230972385, 'backward_transfer_pct': -5.793193254114303, 'final_pooled_accuracy': 0.6364148816234498, 'final_pooled_accuracy_pct': 63.641488162344984, 'final_pooled_macro_f1': 0.6377867577257841, 'final_pooled_macro_f1_pct': 63.77867577257841, 'final_pooled_weighted_f1': 0.6341511980057811, 'final_pooled_weighted_f1_pct': 63.41511980057811, 'final_pooled_macro_precision': 0.6346589857664919, 'final_pooled_macro_precision_pct': 63.465898576649195, 'final_pooled_macro_recall': 0.6459808162203298, 'final_pooled_macro_recall_pct': 64.59808162203298, 'final_pooled_per_class_f1': {'FAVOR': 0.5983066792097836, 'AGAINST': 0.5922038980509745, 'NONE': 0.7228496959165943}, 'final_pooled_per_class_f1_pct': {'FAVOR': 59.830667920978364, 'AGAINST': 59.220389805097454, 'NONE': 72.28496959165943}} +2026-09-02 03:55:46,902 [INFO] __main__: efficiency={'training_wall_time_seconds': 941.6750660918187, 'training_wall_time_minutes': 15.694584434863646, 'peak_gpu_memory_allocated_gb': 10.39810562133789, 'peak_gpu_memory_reserved_gb': 10.62109375} diff --git a/outputs/trace_s5_balanced_replay__x__s3__seed13/environment.json b/outputs/trace_s5_balanced_replay__x__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..86236d581532f5aa94360dcdd19646b56c16a59c --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T03:55:54.911278+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.json b/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..1636e8a2fb8c7d7f54a4c0b2b54baeec38ad39c1 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.json @@ -0,0 +1,2276 @@ +{ + "experiment": "trace_s5_balanced_replay", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 5, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": false, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6257435384871667, + 0.5801455928038206, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6148132977195213, + 0.5843972330769116, + 0.689773423147581, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6575077512577513, + 0.5612618577370799, + 0.6355831948161411, + 0.6473686832456559, + null, + null, + null, + null, + null, + null + ], + [ + 0.6240188739167913, + 0.5790715807302451, + 0.6419220830848738, + 0.6668790147051017, + 0.7678039287234689, + null, + null, + null, + null, + null + ], + [ + 0.5935381342019203, + 0.6017646342947548, + 0.5867230816531613, + 0.6716079503097759, + 0.6880941202772188, + 0.6675550817341862, + null, + null, + null, + null + ], + [ + 0.6980533093740641, + 0.5305169409804351, + 0.588327414422413, + 0.6573469634465147, + 0.7542734407611359, + 0.6641549648474715, + 0.6482415398078049, + null, + null, + null + ], + [ + 0.6333001658374793, + 0.5644083840432724, + 0.5833529089343042, + 0.6588995354239257, + 0.7177545352149332, + 0.7190476190476192, + 0.6610989010989011, + 0.6298371525246491, + null, + null + ], + [ + 0.642644192179796, + 0.5843412557129862, + 0.6124567537055245, + 0.6517855187281563, + 0.6896673769222789, + 0.6156214551936477, + 0.6469386545490466, + 0.6154266637317485, + 0.7161246311043882, + null + ], + [ + 0.7032826981017516, + 0.5600520561789933, + 0.5941655259125205, + 0.6083229036295369, + 0.7317886032261228, + 0.5899363794100636, + 0.66781238811314, + 0.5890674997241532, + 0.704869741971979, + 0.6301013890839536 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6440947940947941, + "losses": { + "current_ce": 0.6911601904666785, + "replay_ce": 0.09865059299450932, + "total": 0.7898107823097344, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6301369863013698, + "macro_f1": 0.6257435384871667, + "weighted_f1": 0.64591697828641, + "macro_precision": 0.6317610062893082, + "macro_recall": 0.7127636181909045, + "per_class_f1": { + "FAVOR": 0.46153846153846156, + "AGAINST": 0.6620689655172414, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 63.013698630136986, + "macro_f1_pct": 62.57435384871667, + "weighted_f1_pct": 64.591697828641, + "macro_precision_pct": 63.17610062893082, + "macro_recall_pct": 71.27636181909045, + "per_class_f1_pct": { + "FAVOR": 46.15384615384615, + "AGAINST": 66.20689655172414, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5801455928038206, + "weighted_f1": 0.5997056907116867, + "macro_precision": 0.5715811965811967, + "macro_recall": 0.6053013798111837, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6373626373626373, + "NONE": 0.5316455696202531 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.01455928038206, + "weighted_f1_pct": 59.97056907116867, + "macro_precision_pct": 57.158119658119666, + "macro_recall_pct": 60.53013798111837, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 63.73626373626373, + "NONE": 53.16455696202531 + } + } + ], + "accuracy": 0.6101190476190477, + "macro_f1": 0.6037906532248264, + "weighted_f1": 0.6175380880228246, + "macro_precision": 0.5968705412584302, + "macro_recall": 0.6502171155701104, + "per_class_f1": { + "FAVOR": 0.5279187817258884, + "AGAINST": 0.6483180428134556, + "NONE": 0.6351351351351351 + }, + "accuracy_pct": 61.011904761904766, + "macro_f1_pct": 60.37906532248264, + "weighted_f1_pct": 61.753808802282464, + "macro_precision_pct": 59.68705412584302, + "macro_recall_pct": 65.02171155701105, + "per_class_f1_pct": { + "FAVOR": 52.79187817258884, + "AGAINST": 64.83180428134557, + "NONE": 63.51351351351351 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.7430085580262827, + "losses": { + "current_ce": 0.5712620150862318, + "replay_ce": 0.06368760656791203, + "total": 0.6349496190295075, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6506849315068494, + "macro_f1": 0.6148132977195213, + "weighted_f1": 0.6697217757636779, + "macro_precision": 0.6068755322486665, + "macro_recall": 0.6611794102948526, + "per_class_f1": { + "FAVOR": 0.417910447761194, + "AGAINST": 0.7295597484276729, + "NONE": 0.696969696969697 + }, + "accuracy_pct": 65.06849315068493, + "macro_f1_pct": 61.48132977195213, + "weighted_f1_pct": 66.97217757636778, + "macro_precision_pct": 60.687553224866654, + "macro_recall_pct": 66.11794102948527, + "per_class_f1_pct": { + "FAVOR": 41.7910447761194, + "AGAINST": 72.95597484276729, + "NONE": 69.6969696969697 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5843972330769116, + "weighted_f1": 0.5878669158504798, + "macro_precision": 0.5861925835936597, + "macro_recall": 0.620551924473493, + "per_class_f1": { + "FAVOR": 0.5522388059701493, + "AGAINST": 0.6035502958579881, + "NONE": 0.5974025974025974 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.43972330769116, + "weighted_f1_pct": 58.786691585047976, + "macro_precision_pct": 58.619258359365965, + "macro_recall_pct": 62.055192447349306, + "per_class_f1_pct": { + "FAVOR": 55.223880597014926, + "AGAINST": 60.35502958579882, + "NONE": 59.74025974025974 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.689773423147581, + "weighted_f1": 0.7082535231990782, + "macro_precision": 0.6834722222222221, + "macro_recall": 0.7036265833767135, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7417840375586855, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 68.9773423147581, + "weighted_f1_pct": 70.82535231990782, + "macro_precision_pct": 68.34722222222221, + "macro_recall_pct": 70.36265833767135, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 74.17840375586854, + "NONE": 66.08695652173913 + } + } + ], + "accuracy": 0.6472795497185742, + "macro_f1": 0.6309961707475337, + "weighted_f1": 0.6544912592320474, + "macro_precision": 0.6192399595625403, + "macro_recall": 0.6670018899039128, + "per_class_f1": { + "FAVOR": 0.5468164794007491, + "AGAINST": 0.6950092421441775, + "NONE": 0.6511627906976745 + }, + "accuracy_pct": 64.72795497185741, + "macro_f1_pct": 63.09961707475337, + "weighted_f1_pct": 65.44912592320475, + "macro_precision_pct": 61.92399595625403, + "macro_recall_pct": 66.70018899039128, + "per_class_f1_pct": { + "FAVOR": 54.68164794007491, + "AGAINST": 69.50092421441775, + "NONE": 65.11627906976744 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.7040973474410007, + "losses": { + "current_ce": 0.6197271093489631, + "replay_ce": 0.0379117925873115, + "total": 0.6576389004313756, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6575077512577513, + "weighted_f1": 0.697160848915986, + "macro_precision": 0.6547509326311168, + "macro_recall": 0.7242228885557221, + "per_class_f1": { + "FAVOR": 0.5135135135135135, + "AGAINST": 0.7402597402597403, + "NONE": 0.71875 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 65.75077512577514, + "weighted_f1_pct": 69.7160848915986, + "macro_precision_pct": 65.47509326311169, + "macro_recall_pct": 72.42228885557222, + "per_class_f1_pct": { + "FAVOR": 51.35135135135135, + "AGAINST": 74.02597402597402, + "NONE": 71.875 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5578947368421052, + "macro_f1": 0.5612618577370799, + "weighted_f1": 0.5593623754376514, + "macro_precision": 0.5745849297573435, + "macro_recall": 0.6071169208424111, + "per_class_f1": { + "FAVOR": 0.5390070921985816, + "AGAINST": 0.5625, + "NONE": 0.5822784810126582 + }, + "accuracy_pct": 55.78947368421052, + "macro_f1_pct": 56.126185773707995, + "weighted_f1_pct": 55.936237543765145, + "macro_precision_pct": 57.45849297573435, + "macro_recall_pct": 60.71169208424111, + "per_class_f1_pct": { + "FAVOR": 53.90070921985816, + "AGAINST": 56.25, + "NONE": 58.22784810126582 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.649746192893401, + "macro_f1": 0.6355831948161411, + "weighted_f1": 0.6551613699905482, + "macro_precision": 0.6212430300557595, + "macro_recall": 0.6854174225335079, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6934673366834171, + "NONE": 0.5925925925925926 + }, + "accuracy_pct": 64.9746192893401, + "macro_f1_pct": 63.55831948161411, + "weighted_f1_pct": 65.51613699905482, + "macro_precision_pct": 62.12430300557595, + "macro_recall_pct": 68.54174225335079, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 69.34673366834171, + "NONE": 59.25925925925925 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6473686832456559, + "weighted_f1": 0.646007756456241, + "macro_precision": 0.6799890799890801, + "macro_recall": 0.6572085948779716, + "per_class_f1": { + "FAVOR": 0.6979166666666666, + "AGAINST": 0.5853658536585366, + "NONE": 0.6588235294117647 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 64.73686832456559, + "weighted_f1_pct": 64.6007756456241, + "macro_precision_pct": 67.99890799890801, + "macro_recall_pct": 65.72085948779716, + "per_class_f1_pct": { + "FAVOR": 69.79166666666666, + "AGAINST": 58.536585365853654, + "NONE": 65.88235294117646 + } + } + ], + "accuracy": 0.6330150068212824, + "macro_f1": 0.6310772698558199, + "weighted_f1": 0.6362874227091418, + "macro_precision": 0.637332561301084, + "macro_recall": 0.6679722416977975, + "per_class_f1": { + "FAVOR": 0.611336032388664, + "AGAINST": 0.6509433962264151, + "NONE": 0.6309523809523809 + }, + "accuracy_pct": 63.30150068212824, + "macro_f1_pct": 63.10772698558199, + "weighted_f1_pct": 63.62874227091419, + "macro_precision_pct": 63.733256130108394, + "macro_recall_pct": 66.79722416977974, + "per_class_f1_pct": { + "FAVOR": 61.1336032388664, + "AGAINST": 65.09433962264151, + "NONE": 63.095238095238095 + } + }, + { + "after_task": 5, + "best_epoch": 4, + "best_dev_macro_f1": 0.6714847733985473, + "losses": { + "current_ce": 0.44290814095487197, + "replay_ce": 0.04538771553391901, + "total": 0.48829585754623017, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6575342465753424, + "macro_f1": 0.6240188739167913, + "weighted_f1": 0.6699305093309914, + "macro_precision": 0.60906043464183, + "macro_recall": 0.682383808095952, + "per_class_f1": { + "FAVOR": 0.4838709677419355, + "AGAINST": 0.7215189873417721, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 65.75342465753424, + "macro_f1_pct": 62.40188739167913, + "weighted_f1_pct": 66.99305093309914, + "macro_precision_pct": 60.906043464183, + "macro_recall_pct": 68.2383808095952, + "per_class_f1_pct": { + "FAVOR": 48.38709677419355, + "AGAINST": 72.15189873417721, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5790715807302451, + "weighted_f1": 0.5722874759002398, + "macro_precision": 0.5995658662325329, + "macro_recall": 0.632171387073348, + "per_class_f1": { + "FAVOR": 0.5655172413793104, + "AGAINST": 0.5641025641025641, + "NONE": 0.6075949367088608 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 57.90715807302451, + "weighted_f1_pct": 57.22874759002398, + "macro_precision_pct": 59.956586623253294, + "macro_recall_pct": 63.2171387073348, + "per_class_f1_pct": { + "FAVOR": 56.55172413793104, + "AGAINST": 56.41025641025641, + "NONE": 60.75949367088608 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6649746192893401, + "macro_f1": 0.6419220830848738, + "weighted_f1": 0.6679359443531324, + "macro_precision": 0.6291422229269066, + "macro_recall": 0.6620762544760462, + "per_class_f1": { + "FAVOR": 0.6133333333333333, + "AGAINST": 0.7162790697674418, + "NONE": 0.5961538461538461 + }, + "accuracy_pct": 66.49746192893402, + "macro_f1_pct": 64.19220830848738, + "weighted_f1_pct": 66.79359443531324, + "macro_precision_pct": 62.91422229269066, + "macro_recall_pct": 66.20762544760463, + "per_class_f1_pct": { + "FAVOR": 61.33333333333333, + "AGAINST": 71.62790697674419, + "NONE": 59.61538461538461 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6668790147051017, + "weighted_f1": 0.6552030578117535, + "macro_precision": 0.6699151103565365, + "macro_recall": 0.6847219326894124, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6666666666666666, + "NONE": 0.7252747252747253 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 66.68790147051017, + "weighted_f1_pct": 65.52030578117535, + "macro_precision_pct": 66.99151103565364, + "macro_recall_pct": 68.47219326894124, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 66.66666666666666, + "NONE": 72.52747252747253 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7678039287234689, + "weighted_f1": 0.8038706440430579, + "macro_precision": 0.7539782441242227, + "macro_recall": 0.7894423158790976, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.7619047619047619, + "NONE": 0.8518518518518519 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 76.7803928723469, + "weighted_f1_pct": 80.38706440430579, + "macro_precision_pct": 75.39782441242227, + "macro_recall_pct": 78.94423158790977, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 76.19047619047619, + "NONE": 85.18518518518519 + } + } + ], + "accuracy": 0.6730975348338692, + "macro_f1": 0.6670556102985746, + "weighted_f1": 0.674557267503662, + "macro_precision": 0.661348386383051, + "macro_recall": 0.6784302771887422, + "per_class_f1": { + "FAVOR": 0.5899581589958159, + "AGAINST": 0.6852300242130751, + "NONE": 0.7259786476868327 + }, + "accuracy_pct": 67.30975348338693, + "macro_f1_pct": 66.70556102985746, + "weighted_f1_pct": 67.4557267503662, + "macro_precision_pct": 66.1348386383051, + "macro_recall_pct": 67.84302771887421, + "per_class_f1_pct": { + "FAVOR": 58.995815899581594, + "AGAINST": 68.52300242130751, + "NONE": 72.59786476868328 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.7633699633699633, + "losses": { + "current_ce": 0.3926368802431084, + "replay_ce": 0.1155484979440059, + "total": 0.508185380129587, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6095890410958904, + "macro_f1": 0.5935381342019203, + "weighted_f1": 0.6247331889566794, + "macro_precision": 0.591597133076594, + "macro_recall": 0.6685707146426787, + "per_class_f1": { + "FAVOR": 0.4657534246575342, + "AGAINST": 0.6577181208053692, + "NONE": 0.6571428571428571 + }, + "accuracy_pct": 60.95890410958904, + "macro_f1_pct": 59.353813420192026, + "weighted_f1_pct": 62.47331889566794, + "macro_precision_pct": 59.1597133076594, + "macro_recall_pct": 66.85707146426788, + "per_class_f1_pct": { + "FAVOR": 46.57534246575342, + "AGAINST": 65.77181208053692, + "NONE": 65.71428571428571 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.6017646342947548, + "weighted_f1": 0.6033077765607887, + "macro_precision": 0.6048316762602477, + "macro_recall": 0.6499636891793754, + "per_class_f1": { + "FAVOR": 0.5606060606060606, + "AGAINST": 0.6181818181818182, + "NONE": 0.6265060240963856 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.17646342947548, + "weighted_f1_pct": 60.33077765607887, + "macro_precision_pct": 60.483167626024766, + "macro_recall_pct": 64.99636891793755, + "per_class_f1_pct": { + "FAVOR": 56.060606060606055, + "AGAINST": 61.81818181818181, + "NONE": 62.65060240963856 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5989847715736041, + "macro_f1": 0.5867230816531613, + "weighted_f1": 0.6048285049791856, + "macro_precision": 0.5779239766081872, + "macro_recall": 0.6451606118446306, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6424870466321243, + "NONE": 0.5370370370370371 + }, + "accuracy_pct": 59.89847715736041, + "macro_f1_pct": 58.67230816531613, + "weighted_f1_pct": 60.48285049791856, + "macro_precision_pct": 57.792397660818715, + "macro_recall_pct": 64.51606118446306, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 64.24870466321244, + "NONE": 53.70370370370371 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6716079503097759, + "weighted_f1": 0.663824574343844, + "macro_precision": 0.6702729447722119, + "macro_recall": 0.6790663918035191, + "per_class_f1": { + "FAVOR": 0.6588235294117647, + "AGAINST": 0.6433566433566433, + "NONE": 0.7126436781609196 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.1607950309776, + "weighted_f1_pct": 66.3824574343844, + "macro_precision_pct": 67.02729447722119, + "macro_recall_pct": 67.90663918035192, + "per_class_f1_pct": { + "FAVOR": 65.88235294117646, + "AGAINST": 64.33566433566433, + "NONE": 71.26436781609196 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.73, + "macro_f1": 0.6880941202772188, + "weighted_f1": 0.7405254527162978, + "macro_precision": 0.6770912068531117, + "macro_recall": 0.7364225294760027, + "per_class_f1": { + "FAVOR": 0.5866666666666667, + "AGAINST": 0.6607142857142857, + "NONE": 0.8169014084507042 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 68.80941202772189, + "weighted_f1_pct": 74.05254527162978, + "macro_precision_pct": 67.70912068531118, + "macro_recall_pct": 73.64225294760027, + "per_class_f1_pct": { + "FAVOR": 58.666666666666664, + "AGAINST": 66.07142857142857, + "NONE": 81.69014084507043 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.6675550817341862, + "weighted_f1": 0.8530341340075853, + "macro_precision": 0.6886883412395388, + "macro_recall": 0.6578570561825021, + "per_class_f1": { + "FAVOR": 0.8955223880597015, + "AGAINST": 0.25, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 66.75550817341862, + "weighted_f1_pct": 85.30341340075853, + "macro_precision_pct": 68.86883412395389, + "macro_recall_pct": 65.7857056182502, + "per_class_f1_pct": { + "FAVOR": 89.55223880597015, + "AGAINST": 25.0, + "NONE": 85.71428571428571 + } + } + ], + "accuracy": 0.6663479923518164, + "macro_f1": 0.6684376063098881, + "weighted_f1": 0.6650694288571806, + "macro_precision": 0.6733747984203595, + "macro_recall": 0.6860513507810779, + "per_class_f1": { + "FAVOR": 0.6469719350073855, + "AGAINST": 0.638961038961039, + "NONE": 0.7193798449612403 + }, + "accuracy_pct": 66.63479923518165, + "macro_f1_pct": 66.84376063098881, + "weighted_f1_pct": 66.50694288571806, + "macro_precision_pct": 67.33747984203595, + "macro_recall_pct": 68.60513507810779, + "per_class_f1_pct": { + "FAVOR": 64.69719350073855, + "AGAINST": 63.896103896103895, + "NONE": 71.93798449612403 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6140893470790378, + "losses": { + "current_ce": 0.6150477621591453, + "replay_ce": 0.06797022510207061, + "total": 0.683017986832243, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6980533093740641, + "weighted_f1": 0.7368287609693658, + "macro_precision": 0.6777345621060294, + "macro_recall": 0.763488255872064, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.779874213836478, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.80533093740641, + "weighted_f1_pct": 73.68287609693658, + "macro_precision_pct": 67.77345621060294, + "macro_recall_pct": 76.3488255872064, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 77.9874213836478, + "NONE": 74.28571428571429 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5210526315789473, + "macro_f1": 0.5305169409804351, + "weighted_f1": 0.5097290020476164, + "macro_precision": 0.5580531030379055, + "macro_recall": 0.5929557007988381, + "per_class_f1": { + "FAVOR": 0.5394736842105263, + "AGAINST": 0.4697986577181208, + "NONE": 0.5822784810126582 + }, + "accuracy_pct": 52.10526315789473, + "macro_f1_pct": 53.05169409804351, + "weighted_f1_pct": 50.97290020476164, + "macro_precision_pct": 55.80531030379055, + "macro_recall_pct": 59.29557007988381, + "per_class_f1_pct": { + "FAVOR": 53.94736842105263, + "AGAINST": 46.97986577181208, + "NONE": 58.22784810126582 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5989847715736041, + "macro_f1": 0.588327414422413, + "weighted_f1": 0.6064658468324062, + "macro_precision": 0.5819001891459333, + "macro_recall": 0.6416376150889426, + "per_class_f1": { + "FAVOR": 0.5161290322580645, + "AGAINST": 0.625, + "NONE": 0.6238532110091743 + }, + "accuracy_pct": 59.89847715736041, + "macro_f1_pct": 58.8327414422413, + "weighted_f1_pct": 60.646584683240626, + "macro_precision_pct": 58.19001891459333, + "macro_recall_pct": 64.16376150889425, + "per_class_f1_pct": { + "FAVOR": 51.61290322580645, + "AGAINST": 62.5, + "NONE": 62.38532110091744 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6573469634465147, + "weighted_f1": 0.6494706242350061, + "macro_precision": 0.6528194861528195, + "macro_recall": 0.6631806916359763, + "per_class_f1": { + "FAVOR": 0.6296296296296297, + "AGAINST": 0.6447368421052632, + "NONE": 0.6976744186046512 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 65.73469634465147, + "weighted_f1_pct": 64.94706242350061, + "macro_precision_pct": 65.28194861528195, + "macro_recall_pct": 66.31806916359763, + "per_class_f1_pct": { + "FAVOR": 62.96296296296296, + "AGAINST": 64.47368421052632, + "NONE": 69.76744186046511 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.79, + "macro_f1": 0.7542734407611359, + "weighted_f1": 0.790148024466555, + "macro_precision": 0.7444395301980441, + "macro_recall": 0.7740537138837987, + "per_class_f1": { + "FAVOR": 0.7096774193548387, + "AGAINST": 0.6972477064220184, + "NONE": 0.8558951965065502 + }, + "accuracy_pct": 79.0, + "macro_f1_pct": 75.42734407611358, + "weighted_f1_pct": 79.0148024466555, + "macro_precision_pct": 74.44395301980441, + "macro_recall_pct": 77.40537138837988, + "per_class_f1_pct": { + "FAVOR": 70.96774193548387, + "AGAINST": 69.72477064220183, + "NONE": 85.58951965065502 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6641549648474715, + "weighted_f1": 0.8025306701316162, + "macro_precision": 0.646070460704607, + "macro_recall": 0.7319742749666304, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.35294117647058826, + "NONE": 0.8048780487804879 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 66.41549648474715, + "weighted_f1_pct": 80.25306701316161, + "macro_precision_pct": 64.60704607046071, + "macro_recall_pct": 73.19742749666304, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 35.294117647058826, + "NONE": 80.48780487804879 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6482415398078049, + "weighted_f1": 0.6726790504481914, + "macro_precision": 0.6366505636070853, + "macro_recall": 0.680695847362514, + "per_class_f1": { + "FAVOR": 0.5783132530120482, + "AGAINST": 0.7070707070707071, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 64.82415398078048, + "weighted_f1_pct": 67.26790504481914, + "macro_precision_pct": 63.665056360708526, + "macro_recall_pct": 68.0695847362514, + "per_class_f1_pct": { + "FAVOR": 57.831325301204814, + "AGAINST": 70.70707070707071, + "NONE": 65.93406593406593 + } + } + ], + "accuracy": 0.6688311688311688, + "macro_f1": 0.6707356531166827, + "weighted_f1": 0.6685507677719296, + "macro_precision": 0.6698195695826028, + "macro_recall": 0.6893106020801337, + "per_class_f1": { + "FAVOR": 0.628032345013477, + "AGAINST": 0.6495901639344263, + "NONE": 0.7345844504021448 + }, + "accuracy_pct": 66.88311688311688, + "macro_f1_pct": 67.07356531166828, + "weighted_f1_pct": 66.85507677719296, + "macro_precision_pct": 66.98195695826028, + "macro_recall_pct": 68.93106020801338, + "per_class_f1_pct": { + "FAVOR": 62.8032345013477, + "AGAINST": 64.95901639344262, + "NONE": 73.45844504021449 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.6789210789210789, + "losses": { + "current_ce": 0.7102673407644033, + "replay_ce": 0.09995939262832204, + "total": 0.8102267354726791, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6506849315068494, + "macro_f1": 0.6333001658374793, + "weighted_f1": 0.66842772439174, + "macro_precision": 0.6330187537810043, + "macro_recall": 0.7097301349325337, + "per_class_f1": { + "FAVOR": 0.5066666666666667, + "AGAINST": 0.7066666666666667, + "NONE": 0.6865671641791045 + }, + "accuracy_pct": 65.06849315068493, + "macro_f1_pct": 63.33001658374793, + "weighted_f1_pct": 66.842772439174, + "macro_precision_pct": 63.301875378100426, + "macro_recall_pct": 70.97301349325336, + "per_class_f1_pct": { + "FAVOR": 50.66666666666667, + "AGAINST": 70.66666666666667, + "NONE": 68.65671641791045 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5644083840432724, + "weighted_f1": 0.5457277676950998, + "macro_precision": 0.5901610644257703, + "macro_recall": 0.6328976034858388, + "per_class_f1": { + "FAVOR": 0.5379310344827586, + "AGAINST": 0.52, + "NONE": 0.6352941176470588 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 56.440838404327245, + "weighted_f1_pct": 54.57277676950998, + "macro_precision_pct": 59.01610644257703, + "macro_recall_pct": 63.289760348583876, + "per_class_f1_pct": { + "FAVOR": 53.79310344827586, + "AGAINST": 52.0, + "NONE": 63.52941176470588 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.583756345177665, + "macro_f1": 0.5833529089343042, + "weighted_f1": 0.5779575950748185, + "macro_precision": 0.5902247772601624, + "macro_recall": 0.6686279768008033, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.5581395348837209, + "NONE": 0.6363636363636364 + }, + "accuracy_pct": 58.37563451776649, + "macro_f1_pct": 58.335290893430425, + "weighted_f1_pct": 57.79575950748185, + "macro_precision_pct": 59.02247772601624, + "macro_recall_pct": 66.86279768008033, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 55.81395348837209, + "NONE": 63.63636363636363 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6588995354239257, + "weighted_f1": 0.6534200783972125, + "macro_precision": 0.6547868894934576, + "macro_recall": 0.6787073711192951, + "per_class_f1": { + "FAVOR": 0.6463414634146342, + "AGAINST": 0.6428571428571429, + "NONE": 0.6875 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 65.88995354239256, + "weighted_f1_pct": 65.34200783972125, + "macro_precision_pct": 65.47868894934577, + "macro_recall_pct": 67.87073711192951, + "per_class_f1_pct": { + "FAVOR": 64.63414634146342, + "AGAINST": 64.28571428571429, + "NONE": 68.75 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.7177545352149332, + "weighted_f1": 0.7579640520473425, + "macro_precision": 0.7041669796909824, + "macro_recall": 0.7377181779480629, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.6724137931034483, + "NONE": 0.8251121076233184 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 71.77545352149332, + "weighted_f1_pct": 75.79640520473426, + "macro_precision_pct": 70.41669796909824, + "macro_recall_pct": 73.7718177948063, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 67.24137931034483, + "NONE": 82.51121076233184 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7190476190476192, + "weighted_f1": 0.8206068268015171, + "macro_precision": 0.7676767676767676, + "macro_recall": 0.6933381871132144, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.5, + "NONE": 0.8 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.90476190476191, + "weighted_f1_pct": 82.06068268015171, + "macro_precision_pct": 76.76767676767676, + "macro_recall_pct": 69.33381871132144, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 50.0, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6610989010989011, + "weighted_f1": 0.6793335696561502, + "macro_precision": 0.6522436311909996, + "macro_recall": 0.7221099887766554, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7032967032967034, + "NONE": 0.68 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.10989010989012, + "weighted_f1_pct": 67.93335696561502, + "macro_precision_pct": 65.22436311909996, + "macro_recall_pct": 72.21099887766555, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 70.32967032967034, + "NONE": 68.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6298371525246491, + "weighted_f1": 0.6204217067479211, + "macro_precision": 0.6287891935670484, + "macro_recall": 0.6487179487179487, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.5048543689320388, + "NONE": 0.7394957983193278 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 62.98371525246491, + "weighted_f1_pct": 62.04217067479211, + "macro_precision_pct": 62.87891935670484, + "macro_recall_pct": 64.87179487179488, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 50.48543689320388, + "NONE": 73.94957983193278 + } + } + ], + "accuracy": 0.6586608442503639, + "macro_f1": 0.6595386328747067, + "weighted_f1": 0.6544738009866116, + "macro_precision": 0.6642523467388327, + "macro_recall": 0.6874055682879213, + "per_class_f1": { + "FAVOR": 0.6341463414634146, + "AGAINST": 0.6190009794319294, + "NONE": 0.7254685777287762 + }, + "accuracy_pct": 65.86608442503639, + "macro_f1_pct": 65.95386328747067, + "weighted_f1_pct": 65.44738009866116, + "macro_precision_pct": 66.42523467388327, + "macro_recall_pct": 68.74055682879214, + "per_class_f1_pct": { + "FAVOR": 63.41463414634146, + "AGAINST": 61.90009794319295, + "NONE": 72.54685777287762 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.6914158975692187, + "losses": { + "current_ce": 0.5853335958028185, + "replay_ce": 0.05354579712493696, + "total": 0.6388793919397437, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6643835616438356, + "macro_f1": 0.642644192179796, + "weighted_f1": 0.6807605402170481, + "macro_precision": 0.6368885251863975, + "macro_recall": 0.7169765117441279, + "per_class_f1": { + "FAVOR": 0.5277777777777778, + "AGAINST": 0.7236842105263158, + "NONE": 0.6764705882352942 + }, + "accuracy_pct": 66.43835616438356, + "macro_f1_pct": 64.2644192179796, + "weighted_f1_pct": 68.0760540217048, + "macro_precision_pct": 63.68885251863975, + "macro_recall_pct": 71.6976511744128, + "per_class_f1_pct": { + "FAVOR": 52.77777777777778, + "AGAINST": 72.36842105263158, + "NONE": 67.64705882352942 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5843412557129862, + "weighted_f1": 0.5948759480524669, + "macro_precision": 0.58288657928925, + "macro_recall": 0.6245461147421932, + "per_class_f1": { + "FAVOR": 0.5511811023622047, + "AGAINST": 0.6235294117647059, + "NONE": 0.5783132530120482 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 58.43412557129862, + "weighted_f1_pct": 59.487594805246694, + "macro_precision_pct": 58.288657928925005, + "macro_recall_pct": 62.45461147421933, + "per_class_f1_pct": { + "FAVOR": 55.118110236220474, + "AGAINST": 62.35294117647059, + "NONE": 57.831325301204814 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6243654822335025, + "macro_f1": 0.6124567537055245, + "weighted_f1": 0.6283434573336195, + "macro_precision": 0.5969364108268446, + "macro_recall": 0.6528218152372239, + "per_class_f1": { + "FAVOR": 0.5679012345679012, + "AGAINST": 0.65, + "NONE": 0.6194690265486725 + }, + "accuracy_pct": 62.43654822335025, + "macro_f1_pct": 61.245675370552455, + "weighted_f1_pct": 62.83434573336195, + "macro_precision_pct": 59.69364108268446, + "macro_recall_pct": 65.2821815237224, + "per_class_f1_pct": { + "FAVOR": 56.79012345679012, + "AGAINST": 65.0, + "NONE": 61.94690265486725 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6517855187281563, + "weighted_f1": 0.6298121708766298, + "macro_precision": 0.652063969382176, + "macro_recall": 0.6675275442213112, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6198830409356725, + "NONE": 0.7640449438202247 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 65.17855187281563, + "weighted_f1_pct": 62.981217087662976, + "macro_precision_pct": 65.20639693821761, + "macro_recall_pct": 66.75275442213112, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 61.98830409356725, + "NONE": 76.40449438202246 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.715, + "macro_f1": 0.6896673769222789, + "weighted_f1": 0.7193930965548613, + "macro_precision": 0.6702318204210412, + "macro_recall": 0.7248783015899457, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.6050420168067226, + "NONE": 0.7870370370370371 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 68.96673769222788, + "weighted_f1_pct": 71.93930965548613, + "macro_precision_pct": 67.02318204210412, + "macro_recall_pct": 72.48783015899457, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 60.50420168067227, + "NONE": 78.70370370370371 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6156214551936477, + "weighted_f1": 0.7776898223635246, + "macro_precision": 0.6071710007880221, + "macro_recall": 0.6648222303118554, + "per_class_f1": { + "FAVOR": 0.7933884297520661, + "AGAINST": 0.23529411764705882, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 61.562145519364776, + "weighted_f1_pct": 77.76898223635246, + "macro_precision_pct": 60.71710007880221, + "macro_recall_pct": 66.48222303118554, + "per_class_f1_pct": { + "FAVOR": 79.33884297520662, + "AGAINST": 23.52941176470588, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6469386545490466, + "weighted_f1": 0.6710519959338842, + "macro_precision": 0.63246428336296, + "macro_recall": 0.6855780022446689, + "per_class_f1": { + "FAVOR": 0.56, + "AGAINST": 0.7015706806282722, + "NONE": 0.6792452830188679 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 64.69386545490467, + "weighted_f1_pct": 67.10519959338842, + "macro_precision_pct": 63.246428336296, + "macro_recall_pct": 68.55780022446689, + "per_class_f1_pct": { + "FAVOR": 56.00000000000001, + "AGAINST": 70.15706806282722, + "NONE": 67.9245283018868 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6154266637317485, + "weighted_f1": 0.6117436514309276, + "macro_precision": 0.6234032634032634, + "macro_recall": 0.6192307692307693, + "per_class_f1": { + "FAVOR": 0.6071428571428571, + "AGAINST": 0.5272727272727272, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 61.542666373174846, + "weighted_f1_pct": 61.17436514309276, + "macro_precision_pct": 62.340326340326335, + "macro_recall_pct": 61.92307692307693, + "per_class_f1_pct": { + "FAVOR": 60.71428571428571, + "AGAINST": 52.72727272727272, + "NONE": 71.1864406779661 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7161246311043882, + "weighted_f1": 0.7199586243715799, + "macro_precision": 0.7222734392037907, + "macro_recall": 0.7121563175780045, + "per_class_f1": { + "FAVOR": 0.7142857142857143, + "AGAINST": 0.6766917293233082, + "NONE": 0.757396449704142 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 71.61246311043882, + "weighted_f1_pct": 71.99586243715798, + "macro_precision_pct": 72.22734392037907, + "macro_recall_pct": 71.21563175780045, + "per_class_f1_pct": { + "FAVOR": 71.42857142857143, + "AGAINST": 67.66917293233082, + "NONE": 75.7396449704142 + } + } + ], + "accuracy": 0.6632782719186785, + "macro_f1": 0.6619964730645033, + "weighted_f1": 0.6617207031225307, + "macro_precision": 0.6581115348688988, + "macro_recall": 0.6711246397427494, + "per_class_f1": { + "FAVOR": 0.6227544910179641, + "AGAINST": 0.6413301662707839, + "NONE": 0.7219047619047619 + }, + "accuracy_pct": 66.32782719186785, + "macro_f1_pct": 66.19964730645033, + "weighted_f1_pct": 66.17207031225307, + "macro_precision_pct": 65.81115348688988, + "macro_recall_pct": 67.11246397427495, + "per_class_f1_pct": { + "FAVOR": 62.27544910179641, + "AGAINST": 64.13301662707839, + "NONE": 72.19047619047619 + } + }, + { + "after_task": 10, + "best_epoch": 11, + "best_dev_macro_f1": 0.6462567564262479, + "losses": { + "current_ce": 0.6599301584256001, + "replay_ce": 0.06963025346780434, + "total": 0.7295604129632314, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7032826981017516, + "weighted_f1": 0.7439610942332237, + "macro_precision": 0.6831572113592247, + "macro_recall": 0.768950524737631, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.7924528301886793, + "NONE": 0.6944444444444444 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 70.32826981017516, + "weighted_f1_pct": 74.39610942332237, + "macro_precision_pct": 68.31572113592247, + "macro_recall_pct": 76.8950524737631, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 79.24528301886792, + "NONE": 69.44444444444444 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5600520561789933, + "weighted_f1": 0.5479510731898503, + "macro_precision": 0.5871044069801834, + "macro_recall": 0.6227305737109659, + "per_class_f1": { + "FAVOR": 0.547945205479452, + "AGAINST": 0.5298013245033113, + "NONE": 0.6024096385542169 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 56.005205617899335, + "weighted_f1_pct": 54.795107318985025, + "macro_precision_pct": 58.71044069801834, + "macro_recall_pct": 62.273057371096584, + "per_class_f1_pct": { + "FAVOR": 54.794520547945204, + "AGAINST": 52.980132450331126, + "NONE": 60.24096385542169 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.5941655259125205, + "weighted_f1": 0.6133009760209528, + "macro_precision": 0.5797833406529059, + "macro_recall": 0.629669942527829, + "per_class_f1": { + "FAVOR": 0.5316455696202531, + "AGAINST": 0.6368159203980099, + "NONE": 0.6140350877192983 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 59.41655259125205, + "weighted_f1_pct": 61.33009760209528, + "macro_precision_pct": 57.978334065290596, + "macro_recall_pct": 62.966994252782904, + "per_class_f1_pct": { + "FAVOR": 53.16455696202531, + "AGAINST": 63.681592039801, + "NONE": 61.40350877192983 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.6, + "macro_f1": 0.6083229036295369, + "weighted_f1": 0.5939421151439299, + "macro_precision": 0.6102828945732965, + "macro_recall": 0.6303746940603309, + "per_class_f1": { + "FAVOR": 0.5441176470588235, + "AGAINST": 0.6, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.83229036295369, + "weighted_f1_pct": 59.39421151439299, + "macro_precision_pct": 61.028289457329656, + "macro_recall_pct": 63.03746940603309, + "per_class_f1_pct": { + "FAVOR": 54.41176470588235, + "AGAINST": 60.0, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.7317886032261228, + "weighted_f1": 0.7613162789814841, + "macro_precision": 0.7157410157410157, + "macro_recall": 0.7699539119329225, + "per_class_f1": { + "FAVOR": 0.6774193548387096, + "AGAINST": 0.7121212121212122, + "NONE": 0.8058252427184466 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 73.17886032261228, + "weighted_f1_pct": 76.13162789814841, + "macro_precision_pct": 71.57410157410156, + "macro_recall_pct": 76.99539119329225, + "per_class_f1_pct": { + "FAVOR": 67.74193548387096, + "AGAINST": 71.21212121212122, + "NONE": 80.58252427184466 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6991150442477876, + "macro_f1": 0.5899363794100636, + "weighted_f1": 0.720658009898811, + "macro_precision": 0.5870390070921986, + "macro_recall": 0.6853779881082392, + "per_class_f1": { + "FAVOR": 0.7368421052631579, + "AGAINST": 0.2857142857142857, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 69.91150442477876, + "macro_f1_pct": 58.99363794100636, + "weighted_f1_pct": 72.0658009898811, + "macro_precision_pct": 58.70390070921986, + "macro_recall_pct": 68.53779881082393, + "per_class_f1_pct": { + "FAVOR": 73.68421052631578, + "AGAINST": 28.57142857142857, + "NONE": 74.72527472527473 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.66781238811314, + "weighted_f1": 0.6927428017047227, + "macro_precision": 0.6516032417195207, + "macro_recall": 0.7006172839506172, + "per_class_f1": { + "FAVOR": 0.5789473684210527, + "AGAINST": 0.7244897959183674, + "NONE": 0.7 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.781238811314, + "weighted_f1_pct": 69.27428017047227, + "macro_precision_pct": 65.16032417195207, + "macro_recall_pct": 70.06172839506172, + "per_class_f1_pct": { + "FAVOR": 57.89473684210527, + "AGAINST": 72.44897959183673, + "NONE": 70.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5890674997241532, + "weighted_f1": 0.581729682131185, + "macro_precision": 0.5824876935346368, + "macro_recall": 0.6064102564102565, + "per_class_f1": { + "FAVOR": 0.59375, + "AGAINST": 0.4716981132075472, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 58.90674997241532, + "weighted_f1_pct": 58.1729682131185, + "macro_precision_pct": 58.24876935346368, + "macro_recall_pct": 60.64102564102565, + "per_class_f1_pct": { + "FAVOR": 59.375, + "AGAINST": 47.16981132075472, + "NONE": 70.17543859649122 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.704869741971979, + "weighted_f1": 0.7096493278784337, + "macro_precision": 0.7089016443131015, + "macro_recall": 0.7105344454742045, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.6779661016949152, + "NONE": 0.7469879518072289 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.48697419719791, + "weighted_f1_pct": 70.96493278784337, + "macro_precision_pct": 70.89016443131015, + "macro_recall_pct": 71.05344454742045, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 67.79661016949152, + "NONE": 74.69879518072288 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.66, + "macro_f1": 0.6301013890839536, + "weighted_f1": 0.6623238398105977, + "macro_precision": 0.6247969986637886, + "macro_recall": 0.6379369282281905, + "per_class_f1": { + "FAVOR": 0.7106598984771574, + "AGAINST": 0.4318181818181818, + "NONE": 0.7478260869565218 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 63.01013890839536, + "weighted_f1_pct": 66.23238398105977, + "macro_precision_pct": 62.47969986637886, + "macro_recall_pct": 63.79369282281905, + "per_class_f1_pct": { + "FAVOR": 71.06598984771574, + "AGAINST": 43.18181818181818, + "NONE": 74.78260869565217 + } + } + ], + "accuracy": 0.6583990980834273, + "macro_f1": 0.6590539365538975, + "weighted_f1": 0.6568695837049182, + "macro_precision": 0.6564634401245174, + "macro_recall": 0.666249847820038, + "per_class_f1": { + "FAVOR": 0.6298763082778306, + "AGAINST": 0.6304023845007451, + "NONE": 0.7168831168831169 + }, + "accuracy_pct": 65.83990980834274, + "macro_f1_pct": 65.90539365538976, + "weighted_f1_pct": 65.68695837049182, + "macro_precision_pct": 65.64634401245173, + "macro_recall_pct": 66.6249847820038, + "per_class_f1_pct": { + "FAVOR": 62.98763082778306, + "AGAINST": 63.04023845007451, + "NONE": 71.68831168831169 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6379399185352214, + "average_forgetting": 0.04700497056259408, + "backward_transfer": -0.034759614211529254, + "final_mean_target_macro_f1_pct": 63.79399185352214, + "average_forgetting_pct": 4.700497056259408, + "backward_transfer_pct": -3.4759614211529253, + "final_pooled_accuracy": 0.6583990980834273, + "final_pooled_accuracy_pct": 65.83990980834274, + "final_pooled_macro_f1": 0.6590539365538975, + "final_pooled_macro_f1_pct": 65.90539365538976, + "final_pooled_weighted_f1": 0.6568695837049182, + "final_pooled_weighted_f1_pct": 65.68695837049182, + "final_pooled_macro_precision": 0.6564634401245174, + "final_pooled_macro_precision_pct": 65.64634401245173, + "final_pooled_macro_recall": 0.666249847820038, + "final_pooled_macro_recall_pct": 66.6249847820038, + "final_pooled_per_class_f1": { + "FAVOR": 0.6298763082778306, + "AGAINST": 0.6304023845007451, + "NONE": 0.7168831168831169 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 62.98763082778306, + "AGAINST": 63.04023845007451, + "NONE": 71.68831168831169 + } + }, + "efficiency": { + "training_wall_time_seconds": 929.187706858851, + "training_wall_time_minutes": 15.48646178098085, + "peak_gpu_memory_allocated_gb": 10.39700698852539, + "peak_gpu_memory_reserved_gb": 10.580078125 + } +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.partial.json b/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..59a83a68a3ad58ea4a82e013c673cfbb8fc446d1 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s3__seed13/metrics.partial.json @@ -0,0 +1,2160 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6257435384871667, + 0.5801455928038206, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6148132977195213, + 0.5843972330769116, + 0.689773423147581, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6575077512577513, + 0.5612618577370799, + 0.6355831948161411, + 0.6473686832456559, + null, + null, + null, + null, + null, + null + ], + [ + 0.6240188739167913, + 0.5790715807302451, + 0.6419220830848738, + 0.6668790147051017, + 0.7678039287234689, + null, + null, + null, + null, + null + ], + [ + 0.5935381342019203, + 0.6017646342947548, + 0.5867230816531613, + 0.6716079503097759, + 0.6880941202772188, + 0.6675550817341862, + null, + null, + null, + null + ], + [ + 0.6980533093740641, + 0.5305169409804351, + 0.588327414422413, + 0.6573469634465147, + 0.7542734407611359, + 0.6641549648474715, + 0.6482415398078049, + null, + null, + null + ], + [ + 0.6333001658374793, + 0.5644083840432724, + 0.5833529089343042, + 0.6588995354239257, + 0.7177545352149332, + 0.7190476190476192, + 0.6610989010989011, + 0.6298371525246491, + null, + null + ], + [ + 0.642644192179796, + 0.5843412557129862, + 0.6124567537055245, + 0.6517855187281563, + 0.6896673769222789, + 0.6156214551936477, + 0.6469386545490466, + 0.6154266637317485, + 0.7161246311043882, + null + ], + [ + 0.7032826981017516, + 0.5600520561789933, + 0.5941655259125205, + 0.6083229036295369, + 0.7317886032261228, + 0.5899363794100636, + 0.66781238811314, + 0.5890674997241532, + 0.704869741971979, + 0.6301013890839536 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6440947940947941, + "losses": { + "current_ce": 0.6911601904666785, + "replay_ce": 0.09865059299450932, + "total": 0.7898107823097344, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6301369863013698, + "macro_f1": 0.6257435384871667, + "weighted_f1": 0.64591697828641, + "macro_precision": 0.6317610062893082, + "macro_recall": 0.7127636181909045, + "per_class_f1": { + "FAVOR": 0.46153846153846156, + "AGAINST": 0.6620689655172414, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 63.013698630136986, + "macro_f1_pct": 62.57435384871667, + "weighted_f1_pct": 64.591697828641, + "macro_precision_pct": 63.17610062893082, + "macro_recall_pct": 71.27636181909045, + "per_class_f1_pct": { + "FAVOR": 46.15384615384615, + "AGAINST": 66.20689655172414, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5801455928038206, + "weighted_f1": 0.5997056907116867, + "macro_precision": 0.5715811965811967, + "macro_recall": 0.6053013798111837, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6373626373626373, + "NONE": 0.5316455696202531 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.01455928038206, + "weighted_f1_pct": 59.97056907116867, + "macro_precision_pct": 57.158119658119666, + "macro_recall_pct": 60.53013798111837, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 63.73626373626373, + "NONE": 53.16455696202531 + } + } + ], + "accuracy": 0.6101190476190477, + "macro_f1": 0.6037906532248264, + "weighted_f1": 0.6175380880228246, + "macro_precision": 0.5968705412584302, + "macro_recall": 0.6502171155701104, + "per_class_f1": { + "FAVOR": 0.5279187817258884, + "AGAINST": 0.6483180428134556, + "NONE": 0.6351351351351351 + }, + "accuracy_pct": 61.011904761904766, + "macro_f1_pct": 60.37906532248264, + "weighted_f1_pct": 61.753808802282464, + "macro_precision_pct": 59.68705412584302, + "macro_recall_pct": 65.02171155701105, + "per_class_f1_pct": { + "FAVOR": 52.79187817258884, + "AGAINST": 64.83180428134557, + "NONE": 63.51351351351351 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.7430085580262827, + "losses": { + "current_ce": 0.5712620150862318, + "replay_ce": 0.06368760656791203, + "total": 0.6349496190295075, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6506849315068494, + "macro_f1": 0.6148132977195213, + "weighted_f1": 0.6697217757636779, + "macro_precision": 0.6068755322486665, + "macro_recall": 0.6611794102948526, + "per_class_f1": { + "FAVOR": 0.417910447761194, + "AGAINST": 0.7295597484276729, + "NONE": 0.696969696969697 + }, + "accuracy_pct": 65.06849315068493, + "macro_f1_pct": 61.48132977195213, + "weighted_f1_pct": 66.97217757636778, + "macro_precision_pct": 60.687553224866654, + "macro_recall_pct": 66.11794102948527, + "per_class_f1_pct": { + "FAVOR": 41.7910447761194, + "AGAINST": 72.95597484276729, + "NONE": 69.6969696969697 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5843972330769116, + "weighted_f1": 0.5878669158504798, + "macro_precision": 0.5861925835936597, + "macro_recall": 0.620551924473493, + "per_class_f1": { + "FAVOR": 0.5522388059701493, + "AGAINST": 0.6035502958579881, + "NONE": 0.5974025974025974 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.43972330769116, + "weighted_f1_pct": 58.786691585047976, + "macro_precision_pct": 58.619258359365965, + "macro_recall_pct": 62.055192447349306, + "per_class_f1_pct": { + "FAVOR": 55.223880597014926, + "AGAINST": 60.35502958579882, + "NONE": 59.74025974025974 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.689773423147581, + "weighted_f1": 0.7082535231990782, + "macro_precision": 0.6834722222222221, + "macro_recall": 0.7036265833767135, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7417840375586855, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 68.9773423147581, + "weighted_f1_pct": 70.82535231990782, + "macro_precision_pct": 68.34722222222221, + "macro_recall_pct": 70.36265833767135, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 74.17840375586854, + "NONE": 66.08695652173913 + } + } + ], + "accuracy": 0.6472795497185742, + "macro_f1": 0.6309961707475337, + "weighted_f1": 0.6544912592320474, + "macro_precision": 0.6192399595625403, + "macro_recall": 0.6670018899039128, + "per_class_f1": { + "FAVOR": 0.5468164794007491, + "AGAINST": 0.6950092421441775, + "NONE": 0.6511627906976745 + }, + "accuracy_pct": 64.72795497185741, + "macro_f1_pct": 63.09961707475337, + "weighted_f1_pct": 65.44912592320475, + "macro_precision_pct": 61.92399595625403, + "macro_recall_pct": 66.70018899039128, + "per_class_f1_pct": { + "FAVOR": 54.68164794007491, + "AGAINST": 69.50092421441775, + "NONE": 65.11627906976744 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.7040973474410007, + "losses": { + "current_ce": 0.6197271093489631, + "replay_ce": 0.0379117925873115, + "total": 0.6576389004313756, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.6575077512577513, + "weighted_f1": 0.697160848915986, + "macro_precision": 0.6547509326311168, + "macro_recall": 0.7242228885557221, + "per_class_f1": { + "FAVOR": 0.5135135135135135, + "AGAINST": 0.7402597402597403, + "NONE": 0.71875 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 65.75077512577514, + "weighted_f1_pct": 69.7160848915986, + "macro_precision_pct": 65.47509326311169, + "macro_recall_pct": 72.42228885557222, + "per_class_f1_pct": { + "FAVOR": 51.35135135135135, + "AGAINST": 74.02597402597402, + "NONE": 71.875 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5578947368421052, + "macro_f1": 0.5612618577370799, + "weighted_f1": 0.5593623754376514, + "macro_precision": 0.5745849297573435, + "macro_recall": 0.6071169208424111, + "per_class_f1": { + "FAVOR": 0.5390070921985816, + "AGAINST": 0.5625, + "NONE": 0.5822784810126582 + }, + "accuracy_pct": 55.78947368421052, + "macro_f1_pct": 56.126185773707995, + "weighted_f1_pct": 55.936237543765145, + "macro_precision_pct": 57.45849297573435, + "macro_recall_pct": 60.71169208424111, + "per_class_f1_pct": { + "FAVOR": 53.90070921985816, + "AGAINST": 56.25, + "NONE": 58.22784810126582 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.649746192893401, + "macro_f1": 0.6355831948161411, + "weighted_f1": 0.6551613699905482, + "macro_precision": 0.6212430300557595, + "macro_recall": 0.6854174225335079, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6934673366834171, + "NONE": 0.5925925925925926 + }, + "accuracy_pct": 64.9746192893401, + "macro_f1_pct": 63.55831948161411, + "weighted_f1_pct": 65.51613699905482, + "macro_precision_pct": 62.12430300557595, + "macro_recall_pct": 68.54174225335079, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 69.34673366834171, + "NONE": 59.25925925925925 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6473686832456559, + "weighted_f1": 0.646007756456241, + "macro_precision": 0.6799890799890801, + "macro_recall": 0.6572085948779716, + "per_class_f1": { + "FAVOR": 0.6979166666666666, + "AGAINST": 0.5853658536585366, + "NONE": 0.6588235294117647 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 64.73686832456559, + "weighted_f1_pct": 64.6007756456241, + "macro_precision_pct": 67.99890799890801, + "macro_recall_pct": 65.72085948779716, + "per_class_f1_pct": { + "FAVOR": 69.79166666666666, + "AGAINST": 58.536585365853654, + "NONE": 65.88235294117646 + } + } + ], + "accuracy": 0.6330150068212824, + "macro_f1": 0.6310772698558199, + "weighted_f1": 0.6362874227091418, + "macro_precision": 0.637332561301084, + "macro_recall": 0.6679722416977975, + "per_class_f1": { + "FAVOR": 0.611336032388664, + "AGAINST": 0.6509433962264151, + "NONE": 0.6309523809523809 + }, + "accuracy_pct": 63.30150068212824, + "macro_f1_pct": 63.10772698558199, + "weighted_f1_pct": 63.62874227091419, + "macro_precision_pct": 63.733256130108394, + "macro_recall_pct": 66.79722416977974, + "per_class_f1_pct": { + "FAVOR": 61.1336032388664, + "AGAINST": 65.09433962264151, + "NONE": 63.095238095238095 + } + }, + { + "after_task": 5, + "best_epoch": 4, + "best_dev_macro_f1": 0.6714847733985473, + "losses": { + "current_ce": 0.44290814095487197, + "replay_ce": 0.04538771553391901, + "total": 0.48829585754623017, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6575342465753424, + "macro_f1": 0.6240188739167913, + "weighted_f1": 0.6699305093309914, + "macro_precision": 0.60906043464183, + "macro_recall": 0.682383808095952, + "per_class_f1": { + "FAVOR": 0.4838709677419355, + "AGAINST": 0.7215189873417721, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 65.75342465753424, + "macro_f1_pct": 62.40188739167913, + "weighted_f1_pct": 66.99305093309914, + "macro_precision_pct": 60.906043464183, + "macro_recall_pct": 68.2383808095952, + "per_class_f1_pct": { + "FAVOR": 48.38709677419355, + "AGAINST": 72.15189873417721, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5790715807302451, + "weighted_f1": 0.5722874759002398, + "macro_precision": 0.5995658662325329, + "macro_recall": 0.632171387073348, + "per_class_f1": { + "FAVOR": 0.5655172413793104, + "AGAINST": 0.5641025641025641, + "NONE": 0.6075949367088608 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 57.90715807302451, + "weighted_f1_pct": 57.22874759002398, + "macro_precision_pct": 59.956586623253294, + "macro_recall_pct": 63.2171387073348, + "per_class_f1_pct": { + "FAVOR": 56.55172413793104, + "AGAINST": 56.41025641025641, + "NONE": 60.75949367088608 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6649746192893401, + "macro_f1": 0.6419220830848738, + "weighted_f1": 0.6679359443531324, + "macro_precision": 0.6291422229269066, + "macro_recall": 0.6620762544760462, + "per_class_f1": { + "FAVOR": 0.6133333333333333, + "AGAINST": 0.7162790697674418, + "NONE": 0.5961538461538461 + }, + "accuracy_pct": 66.49746192893402, + "macro_f1_pct": 64.19220830848738, + "weighted_f1_pct": 66.79359443531324, + "macro_precision_pct": 62.91422229269066, + "macro_recall_pct": 66.20762544760463, + "per_class_f1_pct": { + "FAVOR": 61.33333333333333, + "AGAINST": 71.62790697674419, + "NONE": 59.61538461538461 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6668790147051017, + "weighted_f1": 0.6552030578117535, + "macro_precision": 0.6699151103565365, + "macro_recall": 0.6847219326894124, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6666666666666666, + "NONE": 0.7252747252747253 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 66.68790147051017, + "weighted_f1_pct": 65.52030578117535, + "macro_precision_pct": 66.99151103565364, + "macro_recall_pct": 68.47219326894124, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 66.66666666666666, + "NONE": 72.52747252747253 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7678039287234689, + "weighted_f1": 0.8038706440430579, + "macro_precision": 0.7539782441242227, + "macro_recall": 0.7894423158790976, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.7619047619047619, + "NONE": 0.8518518518518519 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 76.7803928723469, + "weighted_f1_pct": 80.38706440430579, + "macro_precision_pct": 75.39782441242227, + "macro_recall_pct": 78.94423158790977, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 76.19047619047619, + "NONE": 85.18518518518519 + } + } + ], + "accuracy": 0.6730975348338692, + "macro_f1": 0.6670556102985746, + "weighted_f1": 0.674557267503662, + "macro_precision": 0.661348386383051, + "macro_recall": 0.6784302771887422, + "per_class_f1": { + "FAVOR": 0.5899581589958159, + "AGAINST": 0.6852300242130751, + "NONE": 0.7259786476868327 + }, + "accuracy_pct": 67.30975348338693, + "macro_f1_pct": 66.70556102985746, + "weighted_f1_pct": 67.4557267503662, + "macro_precision_pct": 66.1348386383051, + "macro_recall_pct": 67.84302771887421, + "per_class_f1_pct": { + "FAVOR": 58.995815899581594, + "AGAINST": 68.52300242130751, + "NONE": 72.59786476868328 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.7633699633699633, + "losses": { + "current_ce": 0.3926368802431084, + "replay_ce": 0.1155484979440059, + "total": 0.508185380129587, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6095890410958904, + "macro_f1": 0.5935381342019203, + "weighted_f1": 0.6247331889566794, + "macro_precision": 0.591597133076594, + "macro_recall": 0.6685707146426787, + "per_class_f1": { + "FAVOR": 0.4657534246575342, + "AGAINST": 0.6577181208053692, + "NONE": 0.6571428571428571 + }, + "accuracy_pct": 60.95890410958904, + "macro_f1_pct": 59.353813420192026, + "weighted_f1_pct": 62.47331889566794, + "macro_precision_pct": 59.1597133076594, + "macro_recall_pct": 66.85707146426788, + "per_class_f1_pct": { + "FAVOR": 46.57534246575342, + "AGAINST": 65.77181208053692, + "NONE": 65.71428571428571 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.6017646342947548, + "weighted_f1": 0.6033077765607887, + "macro_precision": 0.6048316762602477, + "macro_recall": 0.6499636891793754, + "per_class_f1": { + "FAVOR": 0.5606060606060606, + "AGAINST": 0.6181818181818182, + "NONE": 0.6265060240963856 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.17646342947548, + "weighted_f1_pct": 60.33077765607887, + "macro_precision_pct": 60.483167626024766, + "macro_recall_pct": 64.99636891793755, + "per_class_f1_pct": { + "FAVOR": 56.060606060606055, + "AGAINST": 61.81818181818181, + "NONE": 62.65060240963856 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5989847715736041, + "macro_f1": 0.5867230816531613, + "weighted_f1": 0.6048285049791856, + "macro_precision": 0.5779239766081872, + "macro_recall": 0.6451606118446306, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6424870466321243, + "NONE": 0.5370370370370371 + }, + "accuracy_pct": 59.89847715736041, + "macro_f1_pct": 58.67230816531613, + "weighted_f1_pct": 60.48285049791856, + "macro_precision_pct": 57.792397660818715, + "macro_recall_pct": 64.51606118446306, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 64.24870466321244, + "NONE": 53.70370370370371 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6716079503097759, + "weighted_f1": 0.663824574343844, + "macro_precision": 0.6702729447722119, + "macro_recall": 0.6790663918035191, + "per_class_f1": { + "FAVOR": 0.6588235294117647, + "AGAINST": 0.6433566433566433, + "NONE": 0.7126436781609196 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.1607950309776, + "weighted_f1_pct": 66.3824574343844, + "macro_precision_pct": 67.02729447722119, + "macro_recall_pct": 67.90663918035192, + "per_class_f1_pct": { + "FAVOR": 65.88235294117646, + "AGAINST": 64.33566433566433, + "NONE": 71.26436781609196 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.73, + "macro_f1": 0.6880941202772188, + "weighted_f1": 0.7405254527162978, + "macro_precision": 0.6770912068531117, + "macro_recall": 0.7364225294760027, + "per_class_f1": { + "FAVOR": 0.5866666666666667, + "AGAINST": 0.6607142857142857, + "NONE": 0.8169014084507042 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 68.80941202772189, + "weighted_f1_pct": 74.05254527162978, + "macro_precision_pct": 67.70912068531118, + "macro_recall_pct": 73.64225294760027, + "per_class_f1_pct": { + "FAVOR": 58.666666666666664, + "AGAINST": 66.07142857142857, + "NONE": 81.69014084507043 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.6675550817341862, + "weighted_f1": 0.8530341340075853, + "macro_precision": 0.6886883412395388, + "macro_recall": 0.6578570561825021, + "per_class_f1": { + "FAVOR": 0.8955223880597015, + "AGAINST": 0.25, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 66.75550817341862, + "weighted_f1_pct": 85.30341340075853, + "macro_precision_pct": 68.86883412395389, + "macro_recall_pct": 65.7857056182502, + "per_class_f1_pct": { + "FAVOR": 89.55223880597015, + "AGAINST": 25.0, + "NONE": 85.71428571428571 + } + } + ], + "accuracy": 0.6663479923518164, + "macro_f1": 0.6684376063098881, + "weighted_f1": 0.6650694288571806, + "macro_precision": 0.6733747984203595, + "macro_recall": 0.6860513507810779, + "per_class_f1": { + "FAVOR": 0.6469719350073855, + "AGAINST": 0.638961038961039, + "NONE": 0.7193798449612403 + }, + "accuracy_pct": 66.63479923518165, + "macro_f1_pct": 66.84376063098881, + "weighted_f1_pct": 66.50694288571806, + "macro_precision_pct": 67.33747984203595, + "macro_recall_pct": 68.60513507810779, + "per_class_f1_pct": { + "FAVOR": 64.69719350073855, + "AGAINST": 63.896103896103895, + "NONE": 71.93798449612403 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6140893470790378, + "losses": { + "current_ce": 0.6150477621591453, + "replay_ce": 0.06797022510207061, + "total": 0.683017986832243, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6980533093740641, + "weighted_f1": 0.7368287609693658, + "macro_precision": 0.6777345621060294, + "macro_recall": 0.763488255872064, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.779874213836478, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.80533093740641, + "weighted_f1_pct": 73.68287609693658, + "macro_precision_pct": 67.77345621060294, + "macro_recall_pct": 76.3488255872064, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 77.9874213836478, + "NONE": 74.28571428571429 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5210526315789473, + "macro_f1": 0.5305169409804351, + "weighted_f1": 0.5097290020476164, + "macro_precision": 0.5580531030379055, + "macro_recall": 0.5929557007988381, + "per_class_f1": { + "FAVOR": 0.5394736842105263, + "AGAINST": 0.4697986577181208, + "NONE": 0.5822784810126582 + }, + "accuracy_pct": 52.10526315789473, + "macro_f1_pct": 53.05169409804351, + "weighted_f1_pct": 50.97290020476164, + "macro_precision_pct": 55.80531030379055, + "macro_recall_pct": 59.29557007988381, + "per_class_f1_pct": { + "FAVOR": 53.94736842105263, + "AGAINST": 46.97986577181208, + "NONE": 58.22784810126582 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.5989847715736041, + "macro_f1": 0.588327414422413, + "weighted_f1": 0.6064658468324062, + "macro_precision": 0.5819001891459333, + "macro_recall": 0.6416376150889426, + "per_class_f1": { + "FAVOR": 0.5161290322580645, + "AGAINST": 0.625, + "NONE": 0.6238532110091743 + }, + "accuracy_pct": 59.89847715736041, + "macro_f1_pct": 58.8327414422413, + "weighted_f1_pct": 60.646584683240626, + "macro_precision_pct": 58.19001891459333, + "macro_recall_pct": 64.16376150889425, + "per_class_f1_pct": { + "FAVOR": 51.61290322580645, + "AGAINST": 62.5, + "NONE": 62.38532110091744 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6573469634465147, + "weighted_f1": 0.6494706242350061, + "macro_precision": 0.6528194861528195, + "macro_recall": 0.6631806916359763, + "per_class_f1": { + "FAVOR": 0.6296296296296297, + "AGAINST": 0.6447368421052632, + "NONE": 0.6976744186046512 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 65.73469634465147, + "weighted_f1_pct": 64.94706242350061, + "macro_precision_pct": 65.28194861528195, + "macro_recall_pct": 66.31806916359763, + "per_class_f1_pct": { + "FAVOR": 62.96296296296296, + "AGAINST": 64.47368421052632, + "NONE": 69.76744186046511 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.79, + "macro_f1": 0.7542734407611359, + "weighted_f1": 0.790148024466555, + "macro_precision": 0.7444395301980441, + "macro_recall": 0.7740537138837987, + "per_class_f1": { + "FAVOR": 0.7096774193548387, + "AGAINST": 0.6972477064220184, + "NONE": 0.8558951965065502 + }, + "accuracy_pct": 79.0, + "macro_f1_pct": 75.42734407611358, + "weighted_f1_pct": 79.0148024466555, + "macro_precision_pct": 74.44395301980441, + "macro_recall_pct": 77.40537138837988, + "per_class_f1_pct": { + "FAVOR": 70.96774193548387, + "AGAINST": 69.72477064220183, + "NONE": 85.58951965065502 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6641549648474715, + "weighted_f1": 0.8025306701316162, + "macro_precision": 0.646070460704607, + "macro_recall": 0.7319742749666304, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.35294117647058826, + "NONE": 0.8048780487804879 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 66.41549648474715, + "weighted_f1_pct": 80.25306701316161, + "macro_precision_pct": 64.60704607046071, + "macro_recall_pct": 73.19742749666304, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 35.294117647058826, + "NONE": 80.48780487804879 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6482415398078049, + "weighted_f1": 0.6726790504481914, + "macro_precision": 0.6366505636070853, + "macro_recall": 0.680695847362514, + "per_class_f1": { + "FAVOR": 0.5783132530120482, + "AGAINST": 0.7070707070707071, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 64.82415398078048, + "weighted_f1_pct": 67.26790504481914, + "macro_precision_pct": 63.665056360708526, + "macro_recall_pct": 68.0695847362514, + "per_class_f1_pct": { + "FAVOR": 57.831325301204814, + "AGAINST": 70.70707070707071, + "NONE": 65.93406593406593 + } + } + ], + "accuracy": 0.6688311688311688, + "macro_f1": 0.6707356531166827, + "weighted_f1": 0.6685507677719296, + "macro_precision": 0.6698195695826028, + "macro_recall": 0.6893106020801337, + "per_class_f1": { + "FAVOR": 0.628032345013477, + "AGAINST": 0.6495901639344263, + "NONE": 0.7345844504021448 + }, + "accuracy_pct": 66.88311688311688, + "macro_f1_pct": 67.07356531166828, + "weighted_f1_pct": 66.85507677719296, + "macro_precision_pct": 66.98195695826028, + "macro_recall_pct": 68.93106020801338, + "per_class_f1_pct": { + "FAVOR": 62.8032345013477, + "AGAINST": 64.95901639344262, + "NONE": 73.45844504021449 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.6789210789210789, + "losses": { + "current_ce": 0.7102673407644033, + "replay_ce": 0.09995939262832204, + "total": 0.8102267354726791, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6506849315068494, + "macro_f1": 0.6333001658374793, + "weighted_f1": 0.66842772439174, + "macro_precision": 0.6330187537810043, + "macro_recall": 0.7097301349325337, + "per_class_f1": { + "FAVOR": 0.5066666666666667, + "AGAINST": 0.7066666666666667, + "NONE": 0.6865671641791045 + }, + "accuracy_pct": 65.06849315068493, + "macro_f1_pct": 63.33001658374793, + "weighted_f1_pct": 66.842772439174, + "macro_precision_pct": 63.301875378100426, + "macro_recall_pct": 70.97301349325336, + "per_class_f1_pct": { + "FAVOR": 50.66666666666667, + "AGAINST": 70.66666666666667, + "NONE": 68.65671641791045 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5644083840432724, + "weighted_f1": 0.5457277676950998, + "macro_precision": 0.5901610644257703, + "macro_recall": 0.6328976034858388, + "per_class_f1": { + "FAVOR": 0.5379310344827586, + "AGAINST": 0.52, + "NONE": 0.6352941176470588 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 56.440838404327245, + "weighted_f1_pct": 54.57277676950998, + "macro_precision_pct": 59.01610644257703, + "macro_recall_pct": 63.289760348583876, + "per_class_f1_pct": { + "FAVOR": 53.79310344827586, + "AGAINST": 52.0, + "NONE": 63.52941176470588 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.583756345177665, + "macro_f1": 0.5833529089343042, + "weighted_f1": 0.5779575950748185, + "macro_precision": 0.5902247772601624, + "macro_recall": 0.6686279768008033, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.5581395348837209, + "NONE": 0.6363636363636364 + }, + "accuracy_pct": 58.37563451776649, + "macro_f1_pct": 58.335290893430425, + "weighted_f1_pct": 57.79575950748185, + "macro_precision_pct": 59.02247772601624, + "macro_recall_pct": 66.86279768008033, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 55.81395348837209, + "NONE": 63.63636363636363 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6588995354239257, + "weighted_f1": 0.6534200783972125, + "macro_precision": 0.6547868894934576, + "macro_recall": 0.6787073711192951, + "per_class_f1": { + "FAVOR": 0.6463414634146342, + "AGAINST": 0.6428571428571429, + "NONE": 0.6875 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 65.88995354239256, + "weighted_f1_pct": 65.34200783972125, + "macro_precision_pct": 65.47868894934577, + "macro_recall_pct": 67.87073711192951, + "per_class_f1_pct": { + "FAVOR": 64.63414634146342, + "AGAINST": 64.28571428571429, + "NONE": 68.75 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.7177545352149332, + "weighted_f1": 0.7579640520473425, + "macro_precision": 0.7041669796909824, + "macro_recall": 0.7377181779480629, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.6724137931034483, + "NONE": 0.8251121076233184 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 71.77545352149332, + "weighted_f1_pct": 75.79640520473426, + "macro_precision_pct": 70.41669796909824, + "macro_recall_pct": 73.7718177948063, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 67.24137931034483, + "NONE": 82.51121076233184 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7190476190476192, + "weighted_f1": 0.8206068268015171, + "macro_precision": 0.7676767676767676, + "macro_recall": 0.6933381871132144, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.5, + "NONE": 0.8 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.90476190476191, + "weighted_f1_pct": 82.06068268015171, + "macro_precision_pct": 76.76767676767676, + "macro_recall_pct": 69.33381871132144, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 50.0, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6610989010989011, + "weighted_f1": 0.6793335696561502, + "macro_precision": 0.6522436311909996, + "macro_recall": 0.7221099887766554, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7032967032967034, + "NONE": 0.68 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.10989010989012, + "weighted_f1_pct": 67.93335696561502, + "macro_precision_pct": 65.22436311909996, + "macro_recall_pct": 72.21099887766555, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 70.32967032967034, + "NONE": 68.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6298371525246491, + "weighted_f1": 0.6204217067479211, + "macro_precision": 0.6287891935670484, + "macro_recall": 0.6487179487179487, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.5048543689320388, + "NONE": 0.7394957983193278 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 62.98371525246491, + "weighted_f1_pct": 62.04217067479211, + "macro_precision_pct": 62.87891935670484, + "macro_recall_pct": 64.87179487179488, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 50.48543689320388, + "NONE": 73.94957983193278 + } + } + ], + "accuracy": 0.6586608442503639, + "macro_f1": 0.6595386328747067, + "weighted_f1": 0.6544738009866116, + "macro_precision": 0.6642523467388327, + "macro_recall": 0.6874055682879213, + "per_class_f1": { + "FAVOR": 0.6341463414634146, + "AGAINST": 0.6190009794319294, + "NONE": 0.7254685777287762 + }, + "accuracy_pct": 65.86608442503639, + "macro_f1_pct": 65.95386328747067, + "weighted_f1_pct": 65.44738009866116, + "macro_precision_pct": 66.42523467388327, + "macro_recall_pct": 68.74055682879214, + "per_class_f1_pct": { + "FAVOR": 63.41463414634146, + "AGAINST": 61.90009794319295, + "NONE": 72.54685777287762 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.6914158975692187, + "losses": { + "current_ce": 0.5853335958028185, + "replay_ce": 0.05354579712493696, + "total": 0.6388793919397437, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6643835616438356, + "macro_f1": 0.642644192179796, + "weighted_f1": 0.6807605402170481, + "macro_precision": 0.6368885251863975, + "macro_recall": 0.7169765117441279, + "per_class_f1": { + "FAVOR": 0.5277777777777778, + "AGAINST": 0.7236842105263158, + "NONE": 0.6764705882352942 + }, + "accuracy_pct": 66.43835616438356, + "macro_f1_pct": 64.2644192179796, + "weighted_f1_pct": 68.0760540217048, + "macro_precision_pct": 63.68885251863975, + "macro_recall_pct": 71.6976511744128, + "per_class_f1_pct": { + "FAVOR": 52.77777777777778, + "AGAINST": 72.36842105263158, + "NONE": 67.64705882352942 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5843412557129862, + "weighted_f1": 0.5948759480524669, + "macro_precision": 0.58288657928925, + "macro_recall": 0.6245461147421932, + "per_class_f1": { + "FAVOR": 0.5511811023622047, + "AGAINST": 0.6235294117647059, + "NONE": 0.5783132530120482 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 58.43412557129862, + "weighted_f1_pct": 59.487594805246694, + "macro_precision_pct": 58.288657928925005, + "macro_recall_pct": 62.45461147421933, + "per_class_f1_pct": { + "FAVOR": 55.118110236220474, + "AGAINST": 62.35294117647059, + "NONE": 57.831325301204814 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6243654822335025, + "macro_f1": 0.6124567537055245, + "weighted_f1": 0.6283434573336195, + "macro_precision": 0.5969364108268446, + "macro_recall": 0.6528218152372239, + "per_class_f1": { + "FAVOR": 0.5679012345679012, + "AGAINST": 0.65, + "NONE": 0.6194690265486725 + }, + "accuracy_pct": 62.43654822335025, + "macro_f1_pct": 61.245675370552455, + "weighted_f1_pct": 62.83434573336195, + "macro_precision_pct": 59.69364108268446, + "macro_recall_pct": 65.2821815237224, + "per_class_f1_pct": { + "FAVOR": 56.79012345679012, + "AGAINST": 65.0, + "NONE": 61.94690265486725 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6517855187281563, + "weighted_f1": 0.6298121708766298, + "macro_precision": 0.652063969382176, + "macro_recall": 0.6675275442213112, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6198830409356725, + "NONE": 0.7640449438202247 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 65.17855187281563, + "weighted_f1_pct": 62.981217087662976, + "macro_precision_pct": 65.20639693821761, + "macro_recall_pct": 66.75275442213112, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 61.98830409356725, + "NONE": 76.40449438202246 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.715, + "macro_f1": 0.6896673769222789, + "weighted_f1": 0.7193930965548613, + "macro_precision": 0.6702318204210412, + "macro_recall": 0.7248783015899457, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.6050420168067226, + "NONE": 0.7870370370370371 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 68.96673769222788, + "weighted_f1_pct": 71.93930965548613, + "macro_precision_pct": 67.02318204210412, + "macro_recall_pct": 72.48783015899457, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 60.50420168067227, + "NONE": 78.70370370370371 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6156214551936477, + "weighted_f1": 0.7776898223635246, + "macro_precision": 0.6071710007880221, + "macro_recall": 0.6648222303118554, + "per_class_f1": { + "FAVOR": 0.7933884297520661, + "AGAINST": 0.23529411764705882, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 61.562145519364776, + "weighted_f1_pct": 77.76898223635246, + "macro_precision_pct": 60.71710007880221, + "macro_recall_pct": 66.48222303118554, + "per_class_f1_pct": { + "FAVOR": 79.33884297520662, + "AGAINST": 23.52941176470588, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6469386545490466, + "weighted_f1": 0.6710519959338842, + "macro_precision": 0.63246428336296, + "macro_recall": 0.6855780022446689, + "per_class_f1": { + "FAVOR": 0.56, + "AGAINST": 0.7015706806282722, + "NONE": 0.6792452830188679 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 64.69386545490467, + "weighted_f1_pct": 67.10519959338842, + "macro_precision_pct": 63.246428336296, + "macro_recall_pct": 68.55780022446689, + "per_class_f1_pct": { + "FAVOR": 56.00000000000001, + "AGAINST": 70.15706806282722, + "NONE": 67.9245283018868 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6154266637317485, + "weighted_f1": 0.6117436514309276, + "macro_precision": 0.6234032634032634, + "macro_recall": 0.6192307692307693, + "per_class_f1": { + "FAVOR": 0.6071428571428571, + "AGAINST": 0.5272727272727272, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 61.542666373174846, + "weighted_f1_pct": 61.17436514309276, + "macro_precision_pct": 62.340326340326335, + "macro_recall_pct": 61.92307692307693, + "per_class_f1_pct": { + "FAVOR": 60.71428571428571, + "AGAINST": 52.72727272727272, + "NONE": 71.1864406779661 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7161246311043882, + "weighted_f1": 0.7199586243715799, + "macro_precision": 0.7222734392037907, + "macro_recall": 0.7121563175780045, + "per_class_f1": { + "FAVOR": 0.7142857142857143, + "AGAINST": 0.6766917293233082, + "NONE": 0.757396449704142 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 71.61246311043882, + "weighted_f1_pct": 71.99586243715798, + "macro_precision_pct": 72.22734392037907, + "macro_recall_pct": 71.21563175780045, + "per_class_f1_pct": { + "FAVOR": 71.42857142857143, + "AGAINST": 67.66917293233082, + "NONE": 75.7396449704142 + } + } + ], + "accuracy": 0.6632782719186785, + "macro_f1": 0.6619964730645033, + "weighted_f1": 0.6617207031225307, + "macro_precision": 0.6581115348688988, + "macro_recall": 0.6711246397427494, + "per_class_f1": { + "FAVOR": 0.6227544910179641, + "AGAINST": 0.6413301662707839, + "NONE": 0.7219047619047619 + }, + "accuracy_pct": 66.32782719186785, + "macro_f1_pct": 66.19964730645033, + "weighted_f1_pct": 66.17207031225307, + "macro_precision_pct": 65.81115348688988, + "macro_recall_pct": 67.11246397427495, + "per_class_f1_pct": { + "FAVOR": 62.27544910179641, + "AGAINST": 64.13301662707839, + "NONE": 72.19047619047619 + } + }, + { + "after_task": 10, + "best_epoch": 11, + "best_dev_macro_f1": 0.6462567564262479, + "losses": { + "current_ce": 0.6599301584256001, + "replay_ce": 0.06963025346780434, + "total": 0.7295604129632314, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7032826981017516, + "weighted_f1": 0.7439610942332237, + "macro_precision": 0.6831572113592247, + "macro_recall": 0.768950524737631, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.7924528301886793, + "NONE": 0.6944444444444444 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 70.32826981017516, + "weighted_f1_pct": 74.39610942332237, + "macro_precision_pct": 68.31572113592247, + "macro_recall_pct": 76.8950524737631, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 79.24528301886792, + "NONE": 69.44444444444444 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5526315789473685, + "macro_f1": 0.5600520561789933, + "weighted_f1": 0.5479510731898503, + "macro_precision": 0.5871044069801834, + "macro_recall": 0.6227305737109659, + "per_class_f1": { + "FAVOR": 0.547945205479452, + "AGAINST": 0.5298013245033113, + "NONE": 0.6024096385542169 + }, + "accuracy_pct": 55.26315789473685, + "macro_f1_pct": 56.005205617899335, + "weighted_f1_pct": 54.795107318985025, + "macro_precision_pct": 58.71044069801834, + "macro_recall_pct": 62.273057371096584, + "per_class_f1_pct": { + "FAVOR": 54.794520547945204, + "AGAINST": 52.980132450331126, + "NONE": 60.24096385542169 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6091370558375635, + "macro_f1": 0.5941655259125205, + "weighted_f1": 0.6133009760209528, + "macro_precision": 0.5797833406529059, + "macro_recall": 0.629669942527829, + "per_class_f1": { + "FAVOR": 0.5316455696202531, + "AGAINST": 0.6368159203980099, + "NONE": 0.6140350877192983 + }, + "accuracy_pct": 60.913705583756354, + "macro_f1_pct": 59.41655259125205, + "weighted_f1_pct": 61.33009760209528, + "macro_precision_pct": 57.978334065290596, + "macro_recall_pct": 62.966994252782904, + "per_class_f1_pct": { + "FAVOR": 53.16455696202531, + "AGAINST": 63.681592039801, + "NONE": 61.40350877192983 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.6, + "macro_f1": 0.6083229036295369, + "weighted_f1": 0.5939421151439299, + "macro_precision": 0.6102828945732965, + "macro_recall": 0.6303746940603309, + "per_class_f1": { + "FAVOR": 0.5441176470588235, + "AGAINST": 0.6, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 60.83229036295369, + "weighted_f1_pct": 59.39421151439299, + "macro_precision_pct": 61.028289457329656, + "macro_recall_pct": 63.03746940603309, + "per_class_f1_pct": { + "FAVOR": 54.41176470588235, + "AGAINST": 60.0, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.755, + "macro_f1": 0.7317886032261228, + "weighted_f1": 0.7613162789814841, + "macro_precision": 0.7157410157410157, + "macro_recall": 0.7699539119329225, + "per_class_f1": { + "FAVOR": 0.6774193548387096, + "AGAINST": 0.7121212121212122, + "NONE": 0.8058252427184466 + }, + "accuracy_pct": 75.5, + "macro_f1_pct": 73.17886032261228, + "weighted_f1_pct": 76.13162789814841, + "macro_precision_pct": 71.57410157410156, + "macro_recall_pct": 76.99539119329225, + "per_class_f1_pct": { + "FAVOR": 67.74193548387096, + "AGAINST": 71.21212121212122, + "NONE": 80.58252427184466 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6991150442477876, + "macro_f1": 0.5899363794100636, + "weighted_f1": 0.720658009898811, + "macro_precision": 0.5870390070921986, + "macro_recall": 0.6853779881082392, + "per_class_f1": { + "FAVOR": 0.7368421052631579, + "AGAINST": 0.2857142857142857, + "NONE": 0.7472527472527473 + }, + "accuracy_pct": 69.91150442477876, + "macro_f1_pct": 58.99363794100636, + "weighted_f1_pct": 72.0658009898811, + "macro_precision_pct": 58.70390070921986, + "macro_recall_pct": 68.53779881082393, + "per_class_f1_pct": { + "FAVOR": 73.68421052631578, + "AGAINST": 28.57142857142857, + "NONE": 74.72527472527473 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.66781238811314, + "weighted_f1": 0.6927428017047227, + "macro_precision": 0.6516032417195207, + "macro_recall": 0.7006172839506172, + "per_class_f1": { + "FAVOR": 0.5789473684210527, + "AGAINST": 0.7244897959183674, + "NONE": 0.7 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.781238811314, + "weighted_f1_pct": 69.27428017047227, + "macro_precision_pct": 65.16032417195207, + "macro_recall_pct": 70.06172839506172, + "per_class_f1_pct": { + "FAVOR": 57.89473684210527, + "AGAINST": 72.44897959183673, + "NONE": 70.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5890674997241532, + "weighted_f1": 0.581729682131185, + "macro_precision": 0.5824876935346368, + "macro_recall": 0.6064102564102565, + "per_class_f1": { + "FAVOR": 0.59375, + "AGAINST": 0.4716981132075472, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 58.90674997241532, + "weighted_f1_pct": 58.1729682131185, + "macro_precision_pct": 58.24876935346368, + "macro_recall_pct": 60.64102564102565, + "per_class_f1_pct": { + "FAVOR": 59.375, + "AGAINST": 47.16981132075472, + "NONE": 70.17543859649122 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.704869741971979, + "weighted_f1": 0.7096493278784337, + "macro_precision": 0.7089016443131015, + "macro_recall": 0.7105344454742045, + "per_class_f1": { + "FAVOR": 0.6896551724137931, + "AGAINST": 0.6779661016949152, + "NONE": 0.7469879518072289 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.48697419719791, + "weighted_f1_pct": 70.96493278784337, + "macro_precision_pct": 70.89016443131015, + "macro_recall_pct": 71.05344454742045, + "per_class_f1_pct": { + "FAVOR": 68.96551724137932, + "AGAINST": 67.79661016949152, + "NONE": 74.69879518072288 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.66, + "macro_f1": 0.6301013890839536, + "weighted_f1": 0.6623238398105977, + "macro_precision": 0.6247969986637886, + "macro_recall": 0.6379369282281905, + "per_class_f1": { + "FAVOR": 0.7106598984771574, + "AGAINST": 0.4318181818181818, + "NONE": 0.7478260869565218 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 63.01013890839536, + "weighted_f1_pct": 66.23238398105977, + "macro_precision_pct": 62.47969986637886, + "macro_recall_pct": 63.79369282281905, + "per_class_f1_pct": { + "FAVOR": 71.06598984771574, + "AGAINST": 43.18181818181818, + "NONE": 74.78260869565217 + } + } + ], + "accuracy": 0.6583990980834273, + "macro_f1": 0.6590539365538975, + "weighted_f1": 0.6568695837049182, + "macro_precision": 0.6564634401245174, + "macro_recall": 0.666249847820038, + "per_class_f1": { + "FAVOR": 0.6298763082778306, + "AGAINST": 0.6304023845007451, + "NONE": 0.7168831168831169 + }, + "accuracy_pct": 65.83990980834274, + "macro_f1_pct": 65.90539365538976, + "weighted_f1_pct": 65.68695837049182, + "macro_precision_pct": 65.64634401245173, + "macro_recall_pct": 66.6249847820038, + "per_class_f1_pct": { + "FAVOR": 62.98763082778306, + "AGAINST": 63.04023845007451, + "NONE": 71.68831168831169 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s5_balanced_replay__x__s3__seed13/resolved_config.yaml b/outputs/trace_s5_balanced_replay__x__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3e40a8a90fa076c82bf8a1e31b36d95ff7c63b74 --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s5_balanced_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 5 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s5_balanced_replay__x__s3__seed13/run.log b/outputs/trace_s5_balanced_replay__x__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..702ec86fa9adba3544063a2eddf6eeed3415fb1d --- /dev/null +++ b/outputs/trace_s5_balanced_replay__x__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 03:55:54,965 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s5_balanced_replay__x__s3__seed13 device=cuda +2026-09-02 04:11:25,509 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6379399185352214, 'average_forgetting': 0.04700497056259408, 'backward_transfer': -0.034759614211529254, 'final_mean_target_macro_f1_pct': 63.79399185352214, 'average_forgetting_pct': 4.700497056259408, 'backward_transfer_pct': -3.4759614211529253, 'final_pooled_accuracy': 0.6583990980834273, 'final_pooled_accuracy_pct': 65.83990980834274, 'final_pooled_macro_f1': 0.6590539365538975, 'final_pooled_macro_f1_pct': 65.90539365538976, 'final_pooled_weighted_f1': 0.6568695837049182, 'final_pooled_weighted_f1_pct': 65.68695837049182, 'final_pooled_macro_precision': 0.6564634401245174, 'final_pooled_macro_precision_pct': 65.64634401245173, 'final_pooled_macro_recall': 0.666249847820038, 'final_pooled_macro_recall_pct': 66.6249847820038, 'final_pooled_per_class_f1': {'FAVOR': 0.6298763082778306, 'AGAINST': 0.6304023845007451, 'NONE': 0.7168831168831169}, 'final_pooled_per_class_f1_pct': {'FAVOR': 62.98763082778306, 'AGAINST': 63.04023845007451, 'NONE': 71.68831168831169}} +2026-09-02 04:11:25,510 [INFO] __main__: efficiency={'training_wall_time_seconds': 929.187706858851, 'training_wall_time_minutes': 15.48646178098085, 'peak_gpu_memory_allocated_gb': 10.39700698852539, 'peak_gpu_memory_reserved_gb': 10.580078125} diff --git a/outputs/trace_s6_derpp__doc__s1__seed13/environment.json b/outputs/trace_s6_derpp__doc__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..367bf15586893a8dfcc7cb111dc8d3baac3a162b --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T04:11:33.628513+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s1__seed13/metrics.json b/outputs/trace_s6_derpp__doc__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..f2d167ee2b2c3a817d7cc090262d358b71474a07 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s1__seed13/metrics.json @@ -0,0 +1,1594 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5548005117825808, + 0.7151502562758631, + null, + null, + null, + null, + null, + null + ], + [ + 0.537714535567911, + 0.694451117189438, + 0.6022354228859256, + null, + null, + null, + null, + null + ], + [ + 0.526945225335231, + 0.6679191539915773, + 0.5404608059607088, + 0.5845353133646238, + null, + null, + null, + null + ], + [ + 0.5387149361534006, + 0.6824045711117471, + 0.5750429603013839, + 0.5758109343725782, + 0.6837506623323457, + null, + null, + null + ], + [ + 0.5484171869934005, + 0.689211664551773, + 0.5432439503868075, + 0.5472498875764585, + 0.6891546707270789, + 0.7241294163364437, + null, + null + ], + [ + 0.5158625875667896, + 0.7033105934827252, + 0.5560252292175211, + 0.5500779972995543, + 0.6788728115636937, + 0.6726493491846224, + 0.687297833928024, + null + ], + [ + 0.49153344822436346, + 0.6674158859225438, + 0.5217063257700835, + 0.5440816434804114, + 0.65348068863568, + 0.6233952427500814, + 0.6322903795794608, + 0.7083922773437125 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7294129145980998, + "losses": { + "current_ce": 0.5736874097585678, + "replay_ce": 0.3014473999114264, + "dark_distillation": 0.4032048189640045, + "total": 1.0767372196061271, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6073697585768743, + "macro_f1": 0.5548005117825808, + "weighted_f1": 0.6344356154695512, + "macro_precision": 0.5593971122083318, + "macro_recall": 0.5957745574847971, + "per_class_f1": { + "FAVOR": 0.4598337950138504, + "AGAINST": 0.4631043256997455, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.736975857687426, + "macro_f1_pct": 55.48005117825808, + "weighted_f1_pct": 63.44356154695512, + "macro_precision_pct": 55.939711220833175, + "macro_recall_pct": 59.577455748479714, + "per_class_f1_pct": { + "FAVOR": 45.98337950138504, + "AGAINST": 46.31043256997455, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7323481116584565, + "macro_f1": 0.7151502562758631, + "weighted_f1": 0.7391506707898283, + "macro_precision": 0.7134794872485873, + "macro_recall": 0.7284857416516939, + "per_class_f1": { + "FAVOR": 0.6280487804878049, + "AGAINST": 0.7079646017699115, + "NONE": 0.809437386569873 + }, + "accuracy_pct": 73.23481116584564, + "macro_f1_pct": 71.51502562758631, + "weighted_f1_pct": 73.91506707898283, + "macro_precision_pct": 71.34794872485874, + "macro_recall_pct": 72.8485741651694, + "per_class_f1_pct": { + "FAVOR": 62.80487804878049, + "AGAINST": 70.79646017699115, + "NONE": 80.94373865698729 + } + } + ], + "accuracy": 0.66189111747851, + "macro_f1": 0.6283991868419012, + "weighted_f1": 0.677338825919338, + "macro_precision": 0.6268980686022455, + "macro_recall": 0.658011987150969, + "per_class_f1": { + "FAVOR": 0.5399129172714079, + "AGAINST": 0.5765027322404371, + "NONE": 0.7687819110138585 + }, + "accuracy_pct": 66.189111747851, + "macro_f1_pct": 62.839918684190124, + "weighted_f1_pct": 67.7338825919338, + "macro_precision_pct": 62.68980686022455, + "macro_recall_pct": 65.8011987150969, + "per_class_f1_pct": { + "FAVOR": 53.99129172714079, + "AGAINST": 57.650273224043715, + "NONE": 76.87819110138585 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7155293011717131, + "losses": { + "current_ce": 0.557130595840632, + "replay_ce": 0.20427104764071013, + "dark_distillation": 0.32252677798507706, + "total": 0.9226650311833337, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.537714535567911, + "weighted_f1": 0.6273151938262095, + "macro_precision": 0.5621579666434363, + "macro_recall": 0.5748123940934399, + "per_class_f1": { + "FAVOR": 0.38461538461538464, + "AGAINST": 0.4849699398797595, + "NONE": 0.743558282208589 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.7714535567911, + "weighted_f1_pct": 62.731519382620945, + "macro_precision_pct": 56.21579666434363, + "macro_recall_pct": 57.481239409343985, + "per_class_f1_pct": { + "FAVOR": 38.46153846153847, + "AGAINST": 48.49699398797595, + "NONE": 74.3558282208589 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.694451117189438, + "weighted_f1": 0.7105875630938698, + "macro_precision": 0.7009102009102008, + "macro_recall": 0.7190181221105899, + "per_class_f1": { + "FAVOR": 0.6384364820846905, + "AGAINST": 0.6865671641791045, + "NONE": 0.7583497053045186 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.44511171894379, + "weighted_f1_pct": 71.05875630938698, + "macro_precision_pct": 70.09102009102008, + "macro_recall_pct": 71.901812211059, + "per_class_f1_pct": { + "FAVOR": 63.84364820846905, + "AGAINST": 68.65671641791045, + "NONE": 75.83497053045186 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6470588235294118, + "macro_f1": 0.6022354228859256, + "weighted_f1": 0.6565044064160089, + "macro_precision": 0.6021184956710991, + "macro_recall": 0.6196519091145807, + "per_class_f1": { + "FAVOR": 0.4622222222222222, + "AGAINST": 0.6104868913857678, + "NONE": 0.7339971550497866 + }, + "accuracy_pct": 64.70588235294117, + "macro_f1_pct": 60.223542288592554, + "weighted_f1_pct": 65.6504406416009, + "macro_precision_pct": 60.21184956710991, + "macro_recall_pct": 61.965190911458066, + "per_class_f1_pct": { + "FAVOR": 46.22222222222222, + "AGAINST": 61.04868913857678, + "NONE": 73.39971550497866 + } + } + ], + "accuracy": 0.6469205453690644, + "macro_f1": 0.6123868004142831, + "weighted_f1": 0.6597760182430703, + "macro_precision": 0.6209761437377633, + "macro_recall": 0.6367596787919587, + "per_class_f1": { + "FAVOR": 0.5050505050505051, + "AGAINST": 0.5881533101045296, + "NONE": 0.7439565860878145 + }, + "accuracy_pct": 64.69205453690644, + "macro_f1_pct": 61.2386800414283, + "weighted_f1_pct": 65.97760182430703, + "macro_precision_pct": 62.09761437377633, + "macro_recall_pct": 63.67596787919587, + "per_class_f1_pct": { + "FAVOR": 50.505050505050505, + "AGAINST": 58.815331010452965, + "NONE": 74.39565860878145 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.694872219411752, + "losses": { + "current_ce": 0.5143649857818035, + "replay_ce": 0.1622161221216645, + "dark_distillation": 0.29448629079157845, + "total": 0.823824252475772, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5781448538754765, + "macro_f1": 0.526945225335231, + "weighted_f1": 0.6088219989614784, + "macro_precision": 0.5448579904736014, + "macro_recall": 0.5691460721961811, + "per_class_f1": { + "FAVOR": 0.39263803680981596, + "AGAINST": 0.4726477024070022, + "NONE": 0.7155499367888748 + }, + "accuracy_pct": 57.814485387547656, + "macro_f1_pct": 52.694522533523106, + "weighted_f1_pct": 60.88219989614784, + "macro_precision_pct": 54.48579904736014, + "macro_recall_pct": 56.914607219618105, + "per_class_f1_pct": { + "FAVOR": 39.263803680981596, + "AGAINST": 47.26477024070022, + "NONE": 71.55499367888748 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6765188834154351, + "macro_f1": 0.6679191539915773, + "weighted_f1": 0.6879786813785195, + "macro_precision": 0.6842231963772726, + "macro_recall": 0.6962385610579119, + "per_class_f1": { + "FAVOR": 0.5989010989010989, + "AGAINST": 0.6573816155988857, + "NONE": 0.7474747474747475 + }, + "accuracy_pct": 67.65188834154351, + "macro_f1_pct": 66.79191539915773, + "weighted_f1_pct": 68.79786813785195, + "macro_precision_pct": 68.42231963772726, + "macro_recall_pct": 69.62385610579119, + "per_class_f1_pct": { + "FAVOR": 59.89010989010989, + "AGAINST": 65.73816155988858, + "NONE": 74.74747474747475 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5759233926128591, + "macro_f1": 0.5404608059607088, + "weighted_f1": 0.5944701587514194, + "macro_precision": 0.5505112413750695, + "macro_recall": 0.5741532612138571, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5506692160611855, + "NONE": 0.6707132018209409 + }, + "accuracy_pct": 57.59233926128591, + "macro_f1_pct": 54.04608059607088, + "weighted_f1_pct": 59.44701587514194, + "macro_precision_pct": 55.051124137506946, + "macro_recall_pct": 57.41532612138571, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 55.06692160611855, + "NONE": 67.07132018209408 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6203288490284006, + "macro_f1": 0.5845353133646238, + "weighted_f1": 0.6408424005023907, + "macro_precision": 0.5985396979803594, + "macro_recall": 0.6155684570074976, + "per_class_f1": { + "FAVOR": 0.5397590361445783, + "AGAINST": 0.4709897610921502, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 62.03288490284006, + "macro_f1_pct": 58.45353133646238, + "weighted_f1_pct": 64.08424005023907, + "macro_precision_pct": 59.85396979803595, + "macro_recall_pct": 61.55684570074976, + "per_class_f1_pct": { + "FAVOR": 53.975903614457835, + "AGAINST": 47.098976109215016, + "NONE": 74.28571428571429 + } + } + ], + "accuracy": 0.6090844062947067, + "macro_f1": 0.5824340575083679, + "weighted_f1": 0.6289582904608965, + "macro_precision": 0.5938530887524581, + "macro_recall": 0.6178796909185497, + "per_class_f1": { + "FAVOR": 0.4924187725631769, + "AGAINST": 0.5379901960784313, + "NONE": 0.7168932038834952 + }, + "accuracy_pct": 60.908440629470675, + "macro_f1_pct": 58.243405750836786, + "weighted_f1_pct": 62.895829046089645, + "macro_precision_pct": 59.38530887524581, + "macro_recall_pct": 61.78796909185497, + "per_class_f1_pct": { + "FAVOR": 49.24187725631769, + "AGAINST": 53.799019607843135, + "NONE": 71.68932038834951 + } + }, + { + "after_task": 5, + "best_epoch": 9, + "best_dev_macro_f1": 0.6948108657019548, + "losses": { + "current_ce": 0.4845923680102541, + "replay_ce": 0.15272170562474502, + "dark_distillation": 0.2854618174689157, + "total": 0.780044980772904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.5387149361534006, + "weighted_f1": 0.6221089072615559, + "macro_precision": 0.5482597809076682, + "macro_recall": 0.5774262230144583, + "per_class_f1": { + "FAVOR": 0.4297994269340974, + "AGAINST": 0.4530120481927711, + "NONE": 0.7333333333333333 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 53.87149361534006, + "weighted_f1_pct": 62.21089072615559, + "macro_precision_pct": 54.825978090766824, + "macro_recall_pct": 57.742622301445834, + "per_class_f1_pct": { + "FAVOR": 42.97994269340974, + "AGAINST": 45.30120481927711, + "NONE": 73.33333333333333 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6945812807881774, + "macro_f1": 0.6824045711117471, + "weighted_f1": 0.7027430370289041, + "macro_precision": 0.6847164884795879, + "macro_recall": 0.7038567338628576, + "per_class_f1": { + "FAVOR": 0.6121212121212121, + "AGAINST": 0.6720867208672087, + "NONE": 0.7630057803468208 + }, + "accuracy_pct": 69.45812807881774, + "macro_f1_pct": 68.2404571111747, + "weighted_f1_pct": 70.2743037028904, + "macro_precision_pct": 68.47164884795879, + "macro_recall_pct": 70.38567338628576, + "per_class_f1_pct": { + "FAVOR": 61.212121212121204, + "AGAINST": 67.20867208672087, + "NONE": 76.30057803468208 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6142270861833106, + "macro_f1": 0.5750429603013839, + "weighted_f1": 0.6292428161128569, + "macro_precision": 0.5785754369087702, + "macro_recall": 0.6036725924573263, + "per_class_f1": { + "FAVOR": 0.4357976653696498, + "AGAINST": 0.5822306238185255, + "NONE": 0.7071005917159763 + }, + "accuracy_pct": 61.42270861833106, + "macro_f1_pct": 57.50429603013839, + "weighted_f1_pct": 62.92428161128569, + "macro_precision_pct": 57.85754369087702, + "macro_recall_pct": 60.367259245732626, + "per_class_f1_pct": { + "FAVOR": 43.57976653696498, + "AGAINST": 58.22306238185255, + "NONE": 70.71005917159763 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6128550074738416, + "macro_f1": 0.5758109343725782, + "weighted_f1": 0.6340686679865991, + "macro_precision": 0.5907811371830679, + "macro_recall": 0.6054546821813088, + "per_class_f1": { + "FAVOR": 0.5288461538461539, + "AGAINST": 0.4589041095890411, + "NONE": 0.7396825396825397 + }, + "accuracy_pct": 61.285500747384155, + "macro_f1_pct": 57.58109343725783, + "weighted_f1_pct": 63.40686679865991, + "macro_precision_pct": 59.07811371830679, + "macro_recall_pct": 60.545468218130885, + "per_class_f1_pct": { + "FAVOR": 52.88461538461539, + "AGAINST": 45.89041095890411, + "NONE": 73.96825396825398 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6837506623323457, + "weighted_f1": 0.719306436259082, + "macro_precision": 0.6825023073299521, + "macro_recall": 0.689223533365022, + "per_class_f1": { + "FAVOR": 0.6177606177606177, + "AGAINST": 0.6147540983606558, + "NONE": 0.8187372708757638 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.37506623323458, + "weighted_f1_pct": 71.9306436259082, + "macro_precision_pct": 68.2502307329952, + "macro_recall_pct": 68.9223533365022, + "per_class_f1_pct": { + "FAVOR": 61.77606177606177, + "AGAINST": 61.47540983606557, + "NONE": 81.87372708757637 + } + } + ], + "accuracy": 0.638931065897358, + "macro_f1": 0.6090517577835262, + "weighted_f1": 0.6548431356033134, + "macro_precision": 0.6109001039695932, + "macro_recall": 0.6368971885563848, + "per_class_f1": { + "FAVOR": 0.5238981998758535, + "AGAINST": 0.5559762033531639, + "NONE": 0.7472808701215611 + }, + "accuracy_pct": 63.8931065897358, + "macro_f1_pct": 60.905175778352614, + "weighted_f1_pct": 65.48431356033134, + "macro_precision_pct": 61.090010396959315, + "macro_recall_pct": 63.689718855638475, + "per_class_f1_pct": { + "FAVOR": 52.38981998758535, + "AGAINST": 55.59762033531639, + "NONE": 74.72808701215611 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7138047138047138, + "losses": { + "current_ce": 0.3950907472343672, + "replay_ce": 0.1425374300884349, + "dark_distillation": 0.26292309764595256, + "total": 0.6690897266779627, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5484171869934005, + "weighted_f1": 0.6303660943255325, + "macro_precision": 0.5589420581900147, + "macro_recall": 0.5865693976805089, + "per_class_f1": { + "FAVOR": 0.43312101910828027, + "AGAINST": 0.4732142857142857, + "NONE": 0.7389162561576355 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 54.84171869934005, + "weighted_f1_pct": 63.03660943255325, + "macro_precision_pct": 55.89420581900148, + "macro_recall_pct": 58.65693976805089, + "per_class_f1_pct": { + "FAVOR": 43.31210191082803, + "AGAINST": 47.32142857142857, + "NONE": 73.89162561576354 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7011494252873564, + "macro_f1": 0.689211664551773, + "weighted_f1": 0.7099694629002806, + "macro_precision": 0.6926432932691959, + "macro_recall": 0.7110017967335788, + "per_class_f1": { + "FAVOR": 0.6180758017492711, + "AGAINST": 0.6779661016949152, + "NONE": 0.7715930902111324 + }, + "accuracy_pct": 70.11494252873564, + "macro_f1_pct": 68.9211664551773, + "weighted_f1_pct": 70.99694629002806, + "macro_precision_pct": 69.2643293269196, + "macro_recall_pct": 71.10017967335787, + "per_class_f1_pct": { + "FAVOR": 61.80758017492711, + "AGAINST": 67.79661016949152, + "NONE": 77.15930902111324 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.5432439503868075, + "weighted_f1": 0.6022032494281859, + "macro_precision": 0.5510138915187716, + "macro_recall": 0.5775633807103508, + "per_class_f1": { + "FAVOR": 0.39864864864864863, + "AGAINST": 0.5387755102040817, + "NONE": 0.6923076923076923 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.32439503868075, + "weighted_f1_pct": 60.22032494281859, + "macro_precision_pct": 55.10138915187716, + "macro_recall_pct": 57.75633807103509, + "per_class_f1_pct": { + "FAVOR": 39.86486486486486, + "AGAINST": 53.87755102040816, + "NONE": 69.23076923076923 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5799701046337817, + "macro_f1": 0.5472498875764585, + "weighted_f1": 0.6057905515271336, + "macro_precision": 0.5734714277210937, + "macro_recall": 0.577062675441534, + "per_class_f1": { + "FAVOR": 0.48372093023255813, + "AGAINST": 0.4444444444444444, + "NONE": 0.7135842880523732 + }, + "accuracy_pct": 57.997010463378174, + "macro_f1_pct": 54.72498875764585, + "weighted_f1_pct": 60.57905515271336, + "macro_precision_pct": 57.34714277210937, + "macro_recall_pct": 57.7062675441534, + "per_class_f1_pct": { + "FAVOR": 48.372093023255815, + "AGAINST": 44.44444444444444, + "NONE": 71.35842880523732 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6891546707270789, + "weighted_f1": 0.7209111165413011, + "macro_precision": 0.6894244334606011, + "macro_recall": 0.6997748832678825, + "per_class_f1": { + "FAVOR": 0.635036496350365, + "AGAINST": 0.6224066390041494, + "NONE": 0.8100208768267223 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.9154670727079, + "weighted_f1_pct": 72.09111165413012, + "macro_precision_pct": 68.94244334606012, + "macro_recall_pct": 69.97748832678825, + "per_class_f1_pct": { + "FAVOR": 63.503649635036496, + "AGAINST": 62.24066390041494, + "NONE": 81.00208768267223 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7525150905432596, + "macro_f1": 0.7241294163364437, + "weighted_f1": 0.7627778199802902, + "macro_precision": 0.7235969084162983, + "macro_recall": 0.7421294858076467, + "per_class_f1": { + "FAVOR": 0.6308243727598566, + "AGAINST": 0.6936936936936937, + "NONE": 0.847870182555781 + }, + "accuracy_pct": 75.25150905432596, + "macro_f1_pct": 72.41294163364437, + "weighted_f1_pct": 76.27778199802901, + "macro_precision_pct": 72.35969084162983, + "macro_recall_pct": 74.21294858076467, + "per_class_f1_pct": { + "FAVOR": 63.08243727598566, + "AGAINST": 69.36936936936937, + "NONE": 84.78701825557809 + } + } + ], + "accuracy": 0.6448548812664908, + "macro_f1": 0.6151134214880338, + "weighted_f1": 0.6611358759074459, + "macro_precision": 0.6179736555939875, + "macro_recall": 0.6432313579501682, + "per_class_f1": { + "FAVOR": 0.5289256198347108, + "AGAINST": 0.5614035087719298, + "NONE": 0.755011135857461 + }, + "accuracy_pct": 64.48548812664909, + "macro_f1_pct": 61.51134214880338, + "weighted_f1_pct": 66.1135875907446, + "macro_precision_pct": 61.797365559398756, + "macro_recall_pct": 64.32313579501681, + "per_class_f1_pct": { + "FAVOR": 52.892561983471076, + "AGAINST": 56.14035087719298, + "NONE": 75.50111358574611 + } + }, + { + "after_task": 7, + "best_epoch": 10, + "best_dev_macro_f1": 0.7479179047274543, + "losses": { + "current_ce": 0.34056563904372655, + "replay_ce": 0.13081120861376205, + "dark_distillation": 0.25165544825840774, + "total": 0.5972045729315378, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5756035578144854, + "macro_f1": 0.5158625875667896, + "weighted_f1": 0.6044520173784866, + "macro_precision": 0.5506393876889284, + "macro_recall": 0.5528758903486573, + "per_class_f1": { + "FAVOR": 0.3735408560311284, + "AGAINST": 0.45384615384615384, + "NONE": 0.7202007528230866 + }, + "accuracy_pct": 57.560355781448536, + "macro_f1_pct": 51.58625875667896, + "weighted_f1_pct": 60.44520173784866, + "macro_precision_pct": 55.06393876889284, + "macro_recall_pct": 55.287589034865725, + "per_class_f1_pct": { + "FAVOR": 37.35408560311284, + "AGAINST": 45.38461538461539, + "NONE": 72.02007528230865 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7142857142857143, + "macro_f1": 0.7033105934827252, + "weighted_f1": 0.7202184292391397, + "macro_precision": 0.7055055055055055, + "macro_recall": 0.7243317766128543, + "per_class_f1": { + "FAVOR": 0.6513157894736842, + "AGAINST": 0.6870229007633588, + "NONE": 0.7715930902111324 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 70.33105934827252, + "weighted_f1_pct": 72.02184292391397, + "macro_precision_pct": 70.55055055055055, + "macro_recall_pct": 72.43317766128543, + "per_class_f1_pct": { + "FAVOR": 65.13157894736842, + "AGAINST": 68.70229007633588, + "NONE": 77.15930902111324 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5896032831737346, + "macro_f1": 0.5560252292175211, + "weighted_f1": 0.6048879973992642, + "macro_precision": 0.5701895903423709, + "macro_recall": 0.5891017360300922, + "per_class_f1": { + "FAVOR": 0.4263565891472868, + "AGAINST": 0.5698924731182796, + "NONE": 0.6718266253869969 + }, + "accuracy_pct": 58.96032831737345, + "macro_f1_pct": 55.602522921752104, + "weighted_f1_pct": 60.48879973992642, + "macro_precision_pct": 57.01895903423709, + "macro_recall_pct": 58.91017360300922, + "per_class_f1_pct": { + "FAVOR": 42.63565891472868, + "AGAINST": 56.98924731182796, + "NONE": 67.18266253869969 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5814648729446936, + "macro_f1": 0.5500779972995543, + "weighted_f1": 0.6085357516106584, + "macro_precision": 0.572841929408327, + "macro_recall": 0.5838547376411022, + "per_class_f1": { + "FAVOR": 0.4709141274238227, + "AGAINST": 0.46153846153846156, + "NONE": 0.7177814029363785 + }, + "accuracy_pct": 58.14648729446936, + "macro_f1_pct": 55.00779972995543, + "weighted_f1_pct": 60.85357516106584, + "macro_precision_pct": 57.2841929408327, + "macro_recall_pct": 58.385473764110216, + "per_class_f1_pct": { + "FAVOR": 47.091412742382275, + "AGAINST": 46.15384615384615, + "NONE": 71.77814029363785 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6788728115636937, + "weighted_f1": 0.7141480610224877, + "macro_precision": 0.6782529550337472, + "macro_recall": 0.680990343481132, + "per_class_f1": { + "FAVOR": 0.6147186147186147, + "AGAINST": 0.6090225563909775, + "NONE": 0.8128772635814889 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.88728115636937, + "weighted_f1_pct": 71.41480610224878, + "macro_precision_pct": 67.82529550337472, + "macro_recall_pct": 68.0990343481132, + "per_class_f1_pct": { + "FAVOR": 61.471861471861466, + "AGAINST": 60.902255639097746, + "NONE": 81.28772635814889 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6726493491846224, + "weighted_f1": 0.7183260433705415, + "macro_precision": 0.6688251735574107, + "macro_recall": 0.6943253598426012, + "per_class_f1": { + "FAVOR": 0.5849802371541502, + "AGAINST": 0.6147859922178989, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 67.26493491846225, + "weighted_f1_pct": 71.83260433705415, + "macro_precision_pct": 66.88251735574107, + "macro_recall_pct": 69.43253598426013, + "per_class_f1_pct": { + "FAVOR": 58.498023715415016, + "AGAINST": 61.478599221789885, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7252440725244073, + "macro_f1": 0.687297833928024, + "weighted_f1": 0.7348012320155364, + "macro_precision": 0.6773679936153961, + "macro_recall": 0.7199511585266221, + "per_class_f1": { + "FAVOR": 0.5931558935361216, + "AGAINST": 0.6650485436893204, + "NONE": 0.8036890645586298 + }, + "accuracy_pct": 72.52440725244072, + "macro_f1_pct": 68.7297833928024, + "weighted_f1_pct": 73.48012320155364, + "macro_precision_pct": 67.73679936153961, + "macro_recall_pct": 71.99511585266221, + "per_class_f1_pct": { + "FAVOR": 59.31558935361216, + "AGAINST": 66.50485436893204, + "NONE": 80.36890645586297 + } + } + ], + "accuracy": 0.6507654759263368, + "macro_f1": 0.6190124686243678, + "weighted_f1": 0.6655008809627991, + "macro_precision": 0.6236012548743894, + "macro_recall": 0.6457565134815683, + "per_class_f1": { + "FAVOR": 0.5293201868188895, + "AGAINST": 0.5725631768953069, + "NONE": 0.7551540421589067 + }, + "accuracy_pct": 65.07654759263369, + "macro_f1_pct": 61.90124686243678, + "weighted_f1_pct": 66.55008809627991, + "macro_precision_pct": 62.36012548743894, + "macro_recall_pct": 64.57565134815682, + "per_class_f1_pct": { + "FAVOR": 52.93201868188895, + "AGAINST": 57.25631768953069, + "NONE": 75.51540421589067 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.7884441473983753, + "losses": { + "current_ce": 0.32363019625784134, + "replay_ce": 0.12402305822162067, + "dark_distillation": 0.242403445436674, + "total": 0.5688549762847377, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5412960609911055, + "macro_f1": 0.49153344822436346, + "weighted_f1": 0.5751752658925274, + "macro_precision": 0.5253960679428088, + "macro_recall": 0.5316047182060254, + "per_class_f1": { + "FAVOR": 0.3612040133779264, + "AGAINST": 0.42857142857142855, + "NONE": 0.6848249027237354 + }, + "accuracy_pct": 54.129606099110546, + "macro_f1_pct": 49.153344822436345, + "weighted_f1_pct": 57.517526589252746, + "macro_precision_pct": 52.53960679428088, + "macro_recall_pct": 53.16047182060254, + "per_class_f1_pct": { + "FAVOR": 36.12040133779264, + "AGAINST": 42.857142857142854, + "NONE": 68.48249027237354 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6748768472906403, + "macro_f1": 0.6674158859225438, + "weighted_f1": 0.6834760809478491, + "macro_precision": 0.6791549929854218, + "macro_recall": 0.6970050514906608, + "per_class_f1": { + "FAVOR": 0.6131805157593123, + "AGAINST": 0.6577540106951871, + "NONE": 0.7313131313131314 + }, + "accuracy_pct": 67.48768472906403, + "macro_f1_pct": 66.74158859225437, + "weighted_f1_pct": 68.3476080947849, + "macro_precision_pct": 67.91549929854219, + "macro_recall_pct": 69.70050514906609, + "per_class_f1_pct": { + "FAVOR": 61.31805157593123, + "AGAINST": 65.77540106951871, + "NONE": 73.13131313131314 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5458276333789329, + "macro_f1": 0.5217063257700835, + "weighted_f1": 0.5723896070259619, + "macro_precision": 0.5532984563182359, + "macro_recall": 0.5760526253495789, + "per_class_f1": { + "FAVOR": 0.4011627906976744, + "AGAINST": 0.5111561866125761, + "NONE": 0.6528 + }, + "accuracy_pct": 54.58276333789329, + "macro_f1_pct": 52.17063257700835, + "weighted_f1_pct": 57.23896070259619, + "macro_precision_pct": 55.329845631823595, + "macro_recall_pct": 57.60526253495789, + "per_class_f1_pct": { + "FAVOR": 40.116279069767444, + "AGAINST": 51.115618661257614, + "NONE": 65.28 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5665171898355755, + "macro_f1": 0.5440816434804114, + "weighted_f1": 0.5888137872476565, + "macro_precision": 0.5769497188899014, + "macro_recall": 0.5874678491822617, + "per_class_f1": { + "FAVOR": 0.5174825174825175, + "AGAINST": 0.4458204334365325, + "NONE": 0.6689419795221843 + }, + "accuracy_pct": 56.65171898355755, + "macro_f1_pct": 54.40816434804114, + "weighted_f1_pct": 58.88137872476565, + "macro_precision_pct": 57.694971888990146, + "macro_recall_pct": 58.74678491822617, + "per_class_f1_pct": { + "FAVOR": 51.74825174825175, + "AGAINST": 44.58204334365325, + "NONE": 66.89419795221842 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.65348068863568, + "weighted_f1": 0.6888605100502637, + "macro_precision": 0.6503852924706665, + "macro_recall": 0.6638109700010953, + "per_class_f1": { + "FAVOR": 0.5931558935361216, + "AGAINST": 0.5791505791505791, + "NONE": 0.788135593220339 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 65.348068863568, + "weighted_f1_pct": 68.88605100502637, + "macro_precision_pct": 65.03852924706665, + "macro_recall_pct": 66.38109700010953, + "per_class_f1_pct": { + "FAVOR": 59.31558935361216, + "AGAINST": 57.91505791505791, + "NONE": 78.8135593220339 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6579476861167002, + "macro_f1": 0.6233952427500814, + "weighted_f1": 0.6763584439186437, + "macro_precision": 0.6459392459392459, + "macro_recall": 0.6453286619953286, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.5333333333333333, + "NONE": 0.7913978494623656 + }, + "accuracy_pct": 65.79476861167002, + "macro_f1_pct": 62.33952427500814, + "weighted_f1_pct": 67.63584439186438, + "macro_precision_pct": 64.59392459392458, + "macro_recall_pct": 64.53286619953286, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 53.333333333333336, + "NONE": 79.13978494623656 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6569037656903766, + "macro_f1": 0.6322903795794608, + "weighted_f1": 0.6715322349898079, + "macro_precision": 0.6377831552060228, + "macro_recall": 0.6946489931740717, + "per_class_f1": { + "FAVOR": 0.532051282051282, + "AGAINST": 0.6407322654462243, + "NONE": 0.724087591240876 + }, + "accuracy_pct": 65.69037656903765, + "macro_f1_pct": 63.229037957946076, + "weighted_f1_pct": 67.15322349898078, + "macro_precision_pct": 63.77831552060228, + "macro_recall_pct": 69.46489931740717, + "per_class_f1_pct": { + "FAVOR": 53.205128205128204, + "AGAINST": 64.07322654462243, + "NONE": 72.4087591240876 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7408637873754153, + "macro_f1": 0.7083922773437125, + "weighted_f1": 0.7499359400800085, + "macro_precision": 0.6960853491103766, + "macro_recall": 0.7416147817574217, + "per_class_f1": { + "FAVOR": 0.6367041198501873, + "AGAINST": 0.6688524590163935, + "NONE": 0.819620253164557 + }, + "accuracy_pct": 74.08637873754152, + "macro_f1_pct": 70.83922773437125, + "weighted_f1_pct": 74.99359400800085, + "macro_precision_pct": 69.60853491103765, + "macro_recall_pct": 74.16147817574216, + "per_class_f1_pct": { + "FAVOR": 63.670411985018724, + "AGAINST": 66.88524590163935, + "NONE": 81.9620253164557 + } + } + ], + "accuracy": 0.6259541984732825, + "macro_f1": 0.6007306208930351, + "weighted_f1": 0.6444197561375281, + "macro_precision": 0.6097447703930157, + "macro_recall": 0.6384883349749618, + "per_class_f1": { + "FAVOR": 0.5220759101471727, + "AGAINST": 0.5521514629948365, + "NONE": 0.7279644895370958 + }, + "accuracy_pct": 62.59541984732825, + "macro_f1_pct": 60.073062089303505, + "weighted_f1_pct": 64.44197561375282, + "macro_precision_pct": 60.97447703930157, + "macro_recall_pct": 63.84883349749618, + "per_class_f1_pct": { + "FAVOR": 52.207591014717266, + "AGAINST": 55.21514629948365, + "NONE": 72.79644895370959 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6052869864632922, + "average_forgetting": 0.06561329008898055, + "backward_transfer": -0.06484128888973294, + "final_mean_target_macro_f1_pct": 60.52869864632922, + "average_forgetting_pct": 6.561329008898055, + "backward_transfer_pct": -6.484128888973293, + "final_pooled_accuracy": 0.6259541984732825, + "final_pooled_accuracy_pct": 62.59541984732825, + "final_pooled_macro_f1": 0.6007306208930351, + "final_pooled_macro_f1_pct": 60.073062089303505, + "final_pooled_weighted_f1": 0.6444197561375281, + "final_pooled_weighted_f1_pct": 64.44197561375282, + "final_pooled_macro_precision": 0.6097447703930157, + "final_pooled_macro_precision_pct": 60.97447703930157, + "final_pooled_macro_recall": 0.6384883349749618, + "final_pooled_macro_recall_pct": 63.84883349749618, + "final_pooled_per_class_f1": { + "FAVOR": 0.5220759101471727, + "AGAINST": 0.5521514629948365, + "NONE": 0.7279644895370958 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 52.207591014717266, + "AGAINST": 55.21514629948365, + "NONE": 72.79644895370959 + } + }, + "efficiency": { + "training_wall_time_seconds": 1786.894356549019, + "training_wall_time_minutes": 29.781572609150317, + "peak_gpu_memory_allocated_gb": 10.397007942199707, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s1__seed13/metrics.partial.json b/outputs/trace_s6_derpp__doc__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..837a6e7907c5d74d6fa4f73ee61501e387e7a99f --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s1__seed13/metrics.partial.json @@ -0,0 +1,1488 @@ +{ + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5548005117825808, + 0.7151502562758631, + null, + null, + null, + null, + null, + null + ], + [ + 0.537714535567911, + 0.694451117189438, + 0.6022354228859256, + null, + null, + null, + null, + null + ], + [ + 0.526945225335231, + 0.6679191539915773, + 0.5404608059607088, + 0.5845353133646238, + null, + null, + null, + null + ], + [ + 0.5387149361534006, + 0.6824045711117471, + 0.5750429603013839, + 0.5758109343725782, + 0.6837506623323457, + null, + null, + null + ], + [ + 0.5484171869934005, + 0.689211664551773, + 0.5432439503868075, + 0.5472498875764585, + 0.6891546707270789, + 0.7241294163364437, + null, + null + ], + [ + 0.5158625875667896, + 0.7033105934827252, + 0.5560252292175211, + 0.5500779972995543, + 0.6788728115636937, + 0.6726493491846224, + 0.687297833928024, + null + ], + [ + 0.49153344822436346, + 0.6674158859225438, + 0.5217063257700835, + 0.5440816434804114, + 0.65348068863568, + 0.6233952427500814, + 0.6322903795794608, + 0.7083922773437125 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7294129145980998, + "losses": { + "current_ce": 0.5736874097585678, + "replay_ce": 0.3014473999114264, + "dark_distillation": 0.4032048189640045, + "total": 1.0767372196061271, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6073697585768743, + "macro_f1": 0.5548005117825808, + "weighted_f1": 0.6344356154695512, + "macro_precision": 0.5593971122083318, + "macro_recall": 0.5957745574847971, + "per_class_f1": { + "FAVOR": 0.4598337950138504, + "AGAINST": 0.4631043256997455, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.736975857687426, + "macro_f1_pct": 55.48005117825808, + "weighted_f1_pct": 63.44356154695512, + "macro_precision_pct": 55.939711220833175, + "macro_recall_pct": 59.577455748479714, + "per_class_f1_pct": { + "FAVOR": 45.98337950138504, + "AGAINST": 46.31043256997455, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7323481116584565, + "macro_f1": 0.7151502562758631, + "weighted_f1": 0.7391506707898283, + "macro_precision": 0.7134794872485873, + "macro_recall": 0.7284857416516939, + "per_class_f1": { + "FAVOR": 0.6280487804878049, + "AGAINST": 0.7079646017699115, + "NONE": 0.809437386569873 + }, + "accuracy_pct": 73.23481116584564, + "macro_f1_pct": 71.51502562758631, + "weighted_f1_pct": 73.91506707898283, + "macro_precision_pct": 71.34794872485874, + "macro_recall_pct": 72.8485741651694, + "per_class_f1_pct": { + "FAVOR": 62.80487804878049, + "AGAINST": 70.79646017699115, + "NONE": 80.94373865698729 + } + } + ], + "accuracy": 0.66189111747851, + "macro_f1": 0.6283991868419012, + "weighted_f1": 0.677338825919338, + "macro_precision": 0.6268980686022455, + "macro_recall": 0.658011987150969, + "per_class_f1": { + "FAVOR": 0.5399129172714079, + "AGAINST": 0.5765027322404371, + "NONE": 0.7687819110138585 + }, + "accuracy_pct": 66.189111747851, + "macro_f1_pct": 62.839918684190124, + "weighted_f1_pct": 67.7338825919338, + "macro_precision_pct": 62.68980686022455, + "macro_recall_pct": 65.8011987150969, + "per_class_f1_pct": { + "FAVOR": 53.99129172714079, + "AGAINST": 57.650273224043715, + "NONE": 76.87819110138585 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7155293011717131, + "losses": { + "current_ce": 0.557130595840632, + "replay_ce": 0.20427104764071013, + "dark_distillation": 0.32252677798507706, + "total": 0.9226650311833337, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.537714535567911, + "weighted_f1": 0.6273151938262095, + "macro_precision": 0.5621579666434363, + "macro_recall": 0.5748123940934399, + "per_class_f1": { + "FAVOR": 0.38461538461538464, + "AGAINST": 0.4849699398797595, + "NONE": 0.743558282208589 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.7714535567911, + "weighted_f1_pct": 62.731519382620945, + "macro_precision_pct": 56.21579666434363, + "macro_recall_pct": 57.481239409343985, + "per_class_f1_pct": { + "FAVOR": 38.46153846153847, + "AGAINST": 48.49699398797595, + "NONE": 74.3558282208589 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.694451117189438, + "weighted_f1": 0.7105875630938698, + "macro_precision": 0.7009102009102008, + "macro_recall": 0.7190181221105899, + "per_class_f1": { + "FAVOR": 0.6384364820846905, + "AGAINST": 0.6865671641791045, + "NONE": 0.7583497053045186 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.44511171894379, + "weighted_f1_pct": 71.05875630938698, + "macro_precision_pct": 70.09102009102008, + "macro_recall_pct": 71.901812211059, + "per_class_f1_pct": { + "FAVOR": 63.84364820846905, + "AGAINST": 68.65671641791045, + "NONE": 75.83497053045186 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6470588235294118, + "macro_f1": 0.6022354228859256, + "weighted_f1": 0.6565044064160089, + "macro_precision": 0.6021184956710991, + "macro_recall": 0.6196519091145807, + "per_class_f1": { + "FAVOR": 0.4622222222222222, + "AGAINST": 0.6104868913857678, + "NONE": 0.7339971550497866 + }, + "accuracy_pct": 64.70588235294117, + "macro_f1_pct": 60.223542288592554, + "weighted_f1_pct": 65.6504406416009, + "macro_precision_pct": 60.21184956710991, + "macro_recall_pct": 61.965190911458066, + "per_class_f1_pct": { + "FAVOR": 46.22222222222222, + "AGAINST": 61.04868913857678, + "NONE": 73.39971550497866 + } + } + ], + "accuracy": 0.6469205453690644, + "macro_f1": 0.6123868004142831, + "weighted_f1": 0.6597760182430703, + "macro_precision": 0.6209761437377633, + "macro_recall": 0.6367596787919587, + "per_class_f1": { + "FAVOR": 0.5050505050505051, + "AGAINST": 0.5881533101045296, + "NONE": 0.7439565860878145 + }, + "accuracy_pct": 64.69205453690644, + "macro_f1_pct": 61.2386800414283, + "weighted_f1_pct": 65.97760182430703, + "macro_precision_pct": 62.09761437377633, + "macro_recall_pct": 63.67596787919587, + "per_class_f1_pct": { + "FAVOR": 50.505050505050505, + "AGAINST": 58.815331010452965, + "NONE": 74.39565860878145 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.694872219411752, + "losses": { + "current_ce": 0.5143649857818035, + "replay_ce": 0.1622161221216645, + "dark_distillation": 0.29448629079157845, + "total": 0.823824252475772, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5781448538754765, + "macro_f1": 0.526945225335231, + "weighted_f1": 0.6088219989614784, + "macro_precision": 0.5448579904736014, + "macro_recall": 0.5691460721961811, + "per_class_f1": { + "FAVOR": 0.39263803680981596, + "AGAINST": 0.4726477024070022, + "NONE": 0.7155499367888748 + }, + "accuracy_pct": 57.814485387547656, + "macro_f1_pct": 52.694522533523106, + "weighted_f1_pct": 60.88219989614784, + "macro_precision_pct": 54.48579904736014, + "macro_recall_pct": 56.914607219618105, + "per_class_f1_pct": { + "FAVOR": 39.263803680981596, + "AGAINST": 47.26477024070022, + "NONE": 71.55499367888748 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6765188834154351, + "macro_f1": 0.6679191539915773, + "weighted_f1": 0.6879786813785195, + "macro_precision": 0.6842231963772726, + "macro_recall": 0.6962385610579119, + "per_class_f1": { + "FAVOR": 0.5989010989010989, + "AGAINST": 0.6573816155988857, + "NONE": 0.7474747474747475 + }, + "accuracy_pct": 67.65188834154351, + "macro_f1_pct": 66.79191539915773, + "weighted_f1_pct": 68.79786813785195, + "macro_precision_pct": 68.42231963772726, + "macro_recall_pct": 69.62385610579119, + "per_class_f1_pct": { + "FAVOR": 59.89010989010989, + "AGAINST": 65.73816155988858, + "NONE": 74.74747474747475 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5759233926128591, + "macro_f1": 0.5404608059607088, + "weighted_f1": 0.5944701587514194, + "macro_precision": 0.5505112413750695, + "macro_recall": 0.5741532612138571, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5506692160611855, + "NONE": 0.6707132018209409 + }, + "accuracy_pct": 57.59233926128591, + "macro_f1_pct": 54.04608059607088, + "weighted_f1_pct": 59.44701587514194, + "macro_precision_pct": 55.051124137506946, + "macro_recall_pct": 57.41532612138571, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 55.06692160611855, + "NONE": 67.07132018209408 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6203288490284006, + "macro_f1": 0.5845353133646238, + "weighted_f1": 0.6408424005023907, + "macro_precision": 0.5985396979803594, + "macro_recall": 0.6155684570074976, + "per_class_f1": { + "FAVOR": 0.5397590361445783, + "AGAINST": 0.4709897610921502, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 62.03288490284006, + "macro_f1_pct": 58.45353133646238, + "weighted_f1_pct": 64.08424005023907, + "macro_precision_pct": 59.85396979803595, + "macro_recall_pct": 61.55684570074976, + "per_class_f1_pct": { + "FAVOR": 53.975903614457835, + "AGAINST": 47.098976109215016, + "NONE": 74.28571428571429 + } + } + ], + "accuracy": 0.6090844062947067, + "macro_f1": 0.5824340575083679, + "weighted_f1": 0.6289582904608965, + "macro_precision": 0.5938530887524581, + "macro_recall": 0.6178796909185497, + "per_class_f1": { + "FAVOR": 0.4924187725631769, + "AGAINST": 0.5379901960784313, + "NONE": 0.7168932038834952 + }, + "accuracy_pct": 60.908440629470675, + "macro_f1_pct": 58.243405750836786, + "weighted_f1_pct": 62.895829046089645, + "macro_precision_pct": 59.38530887524581, + "macro_recall_pct": 61.78796909185497, + "per_class_f1_pct": { + "FAVOR": 49.24187725631769, + "AGAINST": 53.799019607843135, + "NONE": 71.68932038834951 + } + }, + { + "after_task": 5, + "best_epoch": 9, + "best_dev_macro_f1": 0.6948108657019548, + "losses": { + "current_ce": 0.4845923680102541, + "replay_ce": 0.15272170562474502, + "dark_distillation": 0.2854618174689157, + "total": 0.780044980772904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.5387149361534006, + "weighted_f1": 0.6221089072615559, + "macro_precision": 0.5482597809076682, + "macro_recall": 0.5774262230144583, + "per_class_f1": { + "FAVOR": 0.4297994269340974, + "AGAINST": 0.4530120481927711, + "NONE": 0.7333333333333333 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 53.87149361534006, + "weighted_f1_pct": 62.21089072615559, + "macro_precision_pct": 54.825978090766824, + "macro_recall_pct": 57.742622301445834, + "per_class_f1_pct": { + "FAVOR": 42.97994269340974, + "AGAINST": 45.30120481927711, + "NONE": 73.33333333333333 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6945812807881774, + "macro_f1": 0.6824045711117471, + "weighted_f1": 0.7027430370289041, + "macro_precision": 0.6847164884795879, + "macro_recall": 0.7038567338628576, + "per_class_f1": { + "FAVOR": 0.6121212121212121, + "AGAINST": 0.6720867208672087, + "NONE": 0.7630057803468208 + }, + "accuracy_pct": 69.45812807881774, + "macro_f1_pct": 68.2404571111747, + "weighted_f1_pct": 70.2743037028904, + "macro_precision_pct": 68.47164884795879, + "macro_recall_pct": 70.38567338628576, + "per_class_f1_pct": { + "FAVOR": 61.212121212121204, + "AGAINST": 67.20867208672087, + "NONE": 76.30057803468208 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6142270861833106, + "macro_f1": 0.5750429603013839, + "weighted_f1": 0.6292428161128569, + "macro_precision": 0.5785754369087702, + "macro_recall": 0.6036725924573263, + "per_class_f1": { + "FAVOR": 0.4357976653696498, + "AGAINST": 0.5822306238185255, + "NONE": 0.7071005917159763 + }, + "accuracy_pct": 61.42270861833106, + "macro_f1_pct": 57.50429603013839, + "weighted_f1_pct": 62.92428161128569, + "macro_precision_pct": 57.85754369087702, + "macro_recall_pct": 60.367259245732626, + "per_class_f1_pct": { + "FAVOR": 43.57976653696498, + "AGAINST": 58.22306238185255, + "NONE": 70.71005917159763 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6128550074738416, + "macro_f1": 0.5758109343725782, + "weighted_f1": 0.6340686679865991, + "macro_precision": 0.5907811371830679, + "macro_recall": 0.6054546821813088, + "per_class_f1": { + "FAVOR": 0.5288461538461539, + "AGAINST": 0.4589041095890411, + "NONE": 0.7396825396825397 + }, + "accuracy_pct": 61.285500747384155, + "macro_f1_pct": 57.58109343725783, + "weighted_f1_pct": 63.40686679865991, + "macro_precision_pct": 59.07811371830679, + "macro_recall_pct": 60.545468218130885, + "per_class_f1_pct": { + "FAVOR": 52.88461538461539, + "AGAINST": 45.89041095890411, + "NONE": 73.96825396825398 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6837506623323457, + "weighted_f1": 0.719306436259082, + "macro_precision": 0.6825023073299521, + "macro_recall": 0.689223533365022, + "per_class_f1": { + "FAVOR": 0.6177606177606177, + "AGAINST": 0.6147540983606558, + "NONE": 0.8187372708757638 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.37506623323458, + "weighted_f1_pct": 71.9306436259082, + "macro_precision_pct": 68.2502307329952, + "macro_recall_pct": 68.9223533365022, + "per_class_f1_pct": { + "FAVOR": 61.77606177606177, + "AGAINST": 61.47540983606557, + "NONE": 81.87372708757637 + } + } + ], + "accuracy": 0.638931065897358, + "macro_f1": 0.6090517577835262, + "weighted_f1": 0.6548431356033134, + "macro_precision": 0.6109001039695932, + "macro_recall": 0.6368971885563848, + "per_class_f1": { + "FAVOR": 0.5238981998758535, + "AGAINST": 0.5559762033531639, + "NONE": 0.7472808701215611 + }, + "accuracy_pct": 63.8931065897358, + "macro_f1_pct": 60.905175778352614, + "weighted_f1_pct": 65.48431356033134, + "macro_precision_pct": 61.090010396959315, + "macro_recall_pct": 63.689718855638475, + "per_class_f1_pct": { + "FAVOR": 52.38981998758535, + "AGAINST": 55.59762033531639, + "NONE": 74.72808701215611 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7138047138047138, + "losses": { + "current_ce": 0.3950907472343672, + "replay_ce": 0.1425374300884349, + "dark_distillation": 0.26292309764595256, + "total": 0.6690897266779627, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5484171869934005, + "weighted_f1": 0.6303660943255325, + "macro_precision": 0.5589420581900147, + "macro_recall": 0.5865693976805089, + "per_class_f1": { + "FAVOR": 0.43312101910828027, + "AGAINST": 0.4732142857142857, + "NONE": 0.7389162561576355 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 54.84171869934005, + "weighted_f1_pct": 63.03660943255325, + "macro_precision_pct": 55.89420581900148, + "macro_recall_pct": 58.65693976805089, + "per_class_f1_pct": { + "FAVOR": 43.31210191082803, + "AGAINST": 47.32142857142857, + "NONE": 73.89162561576354 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7011494252873564, + "macro_f1": 0.689211664551773, + "weighted_f1": 0.7099694629002806, + "macro_precision": 0.6926432932691959, + "macro_recall": 0.7110017967335788, + "per_class_f1": { + "FAVOR": 0.6180758017492711, + "AGAINST": 0.6779661016949152, + "NONE": 0.7715930902111324 + }, + "accuracy_pct": 70.11494252873564, + "macro_f1_pct": 68.9211664551773, + "weighted_f1_pct": 70.99694629002806, + "macro_precision_pct": 69.2643293269196, + "macro_recall_pct": 71.10017967335787, + "per_class_f1_pct": { + "FAVOR": 61.80758017492711, + "AGAINST": 67.79661016949152, + "NONE": 77.15930902111324 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.5432439503868075, + "weighted_f1": 0.6022032494281859, + "macro_precision": 0.5510138915187716, + "macro_recall": 0.5775633807103508, + "per_class_f1": { + "FAVOR": 0.39864864864864863, + "AGAINST": 0.5387755102040817, + "NONE": 0.6923076923076923 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.32439503868075, + "weighted_f1_pct": 60.22032494281859, + "macro_precision_pct": 55.10138915187716, + "macro_recall_pct": 57.75633807103509, + "per_class_f1_pct": { + "FAVOR": 39.86486486486486, + "AGAINST": 53.87755102040816, + "NONE": 69.23076923076923 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5799701046337817, + "macro_f1": 0.5472498875764585, + "weighted_f1": 0.6057905515271336, + "macro_precision": 0.5734714277210937, + "macro_recall": 0.577062675441534, + "per_class_f1": { + "FAVOR": 0.48372093023255813, + "AGAINST": 0.4444444444444444, + "NONE": 0.7135842880523732 + }, + "accuracy_pct": 57.997010463378174, + "macro_f1_pct": 54.72498875764585, + "weighted_f1_pct": 60.57905515271336, + "macro_precision_pct": 57.34714277210937, + "macro_recall_pct": 57.7062675441534, + "per_class_f1_pct": { + "FAVOR": 48.372093023255815, + "AGAINST": 44.44444444444444, + "NONE": 71.35842880523732 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6891546707270789, + "weighted_f1": 0.7209111165413011, + "macro_precision": 0.6894244334606011, + "macro_recall": 0.6997748832678825, + "per_class_f1": { + "FAVOR": 0.635036496350365, + "AGAINST": 0.6224066390041494, + "NONE": 0.8100208768267223 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 68.9154670727079, + "weighted_f1_pct": 72.09111165413012, + "macro_precision_pct": 68.94244334606012, + "macro_recall_pct": 69.97748832678825, + "per_class_f1_pct": { + "FAVOR": 63.503649635036496, + "AGAINST": 62.24066390041494, + "NONE": 81.00208768267223 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7525150905432596, + "macro_f1": 0.7241294163364437, + "weighted_f1": 0.7627778199802902, + "macro_precision": 0.7235969084162983, + "macro_recall": 0.7421294858076467, + "per_class_f1": { + "FAVOR": 0.6308243727598566, + "AGAINST": 0.6936936936936937, + "NONE": 0.847870182555781 + }, + "accuracy_pct": 75.25150905432596, + "macro_f1_pct": 72.41294163364437, + "weighted_f1_pct": 76.27778199802901, + "macro_precision_pct": 72.35969084162983, + "macro_recall_pct": 74.21294858076467, + "per_class_f1_pct": { + "FAVOR": 63.08243727598566, + "AGAINST": 69.36936936936937, + "NONE": 84.78701825557809 + } + } + ], + "accuracy": 0.6448548812664908, + "macro_f1": 0.6151134214880338, + "weighted_f1": 0.6611358759074459, + "macro_precision": 0.6179736555939875, + "macro_recall": 0.6432313579501682, + "per_class_f1": { + "FAVOR": 0.5289256198347108, + "AGAINST": 0.5614035087719298, + "NONE": 0.755011135857461 + }, + "accuracy_pct": 64.48548812664909, + "macro_f1_pct": 61.51134214880338, + "weighted_f1_pct": 66.1135875907446, + "macro_precision_pct": 61.797365559398756, + "macro_recall_pct": 64.32313579501681, + "per_class_f1_pct": { + "FAVOR": 52.892561983471076, + "AGAINST": 56.14035087719298, + "NONE": 75.50111358574611 + } + }, + { + "after_task": 7, + "best_epoch": 10, + "best_dev_macro_f1": 0.7479179047274543, + "losses": { + "current_ce": 0.34056563904372655, + "replay_ce": 0.13081120861376205, + "dark_distillation": 0.25165544825840774, + "total": 0.5972045729315378, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5756035578144854, + "macro_f1": 0.5158625875667896, + "weighted_f1": 0.6044520173784866, + "macro_precision": 0.5506393876889284, + "macro_recall": 0.5528758903486573, + "per_class_f1": { + "FAVOR": 0.3735408560311284, + "AGAINST": 0.45384615384615384, + "NONE": 0.7202007528230866 + }, + "accuracy_pct": 57.560355781448536, + "macro_f1_pct": 51.58625875667896, + "weighted_f1_pct": 60.44520173784866, + "macro_precision_pct": 55.06393876889284, + "macro_recall_pct": 55.287589034865725, + "per_class_f1_pct": { + "FAVOR": 37.35408560311284, + "AGAINST": 45.38461538461539, + "NONE": 72.02007528230865 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7142857142857143, + "macro_f1": 0.7033105934827252, + "weighted_f1": 0.7202184292391397, + "macro_precision": 0.7055055055055055, + "macro_recall": 0.7243317766128543, + "per_class_f1": { + "FAVOR": 0.6513157894736842, + "AGAINST": 0.6870229007633588, + "NONE": 0.7715930902111324 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 70.33105934827252, + "weighted_f1_pct": 72.02184292391397, + "macro_precision_pct": 70.55055055055055, + "macro_recall_pct": 72.43317766128543, + "per_class_f1_pct": { + "FAVOR": 65.13157894736842, + "AGAINST": 68.70229007633588, + "NONE": 77.15930902111324 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5896032831737346, + "macro_f1": 0.5560252292175211, + "weighted_f1": 0.6048879973992642, + "macro_precision": 0.5701895903423709, + "macro_recall": 0.5891017360300922, + "per_class_f1": { + "FAVOR": 0.4263565891472868, + "AGAINST": 0.5698924731182796, + "NONE": 0.6718266253869969 + }, + "accuracy_pct": 58.96032831737345, + "macro_f1_pct": 55.602522921752104, + "weighted_f1_pct": 60.48879973992642, + "macro_precision_pct": 57.01895903423709, + "macro_recall_pct": 58.91017360300922, + "per_class_f1_pct": { + "FAVOR": 42.63565891472868, + "AGAINST": 56.98924731182796, + "NONE": 67.18266253869969 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5814648729446936, + "macro_f1": 0.5500779972995543, + "weighted_f1": 0.6085357516106584, + "macro_precision": 0.572841929408327, + "macro_recall": 0.5838547376411022, + "per_class_f1": { + "FAVOR": 0.4709141274238227, + "AGAINST": 0.46153846153846156, + "NONE": 0.7177814029363785 + }, + "accuracy_pct": 58.14648729446936, + "macro_f1_pct": 55.00779972995543, + "weighted_f1_pct": 60.85357516106584, + "macro_precision_pct": 57.2841929408327, + "macro_recall_pct": 58.385473764110216, + "per_class_f1_pct": { + "FAVOR": 47.091412742382275, + "AGAINST": 46.15384615384615, + "NONE": 71.77814029363785 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6788728115636937, + "weighted_f1": 0.7141480610224877, + "macro_precision": 0.6782529550337472, + "macro_recall": 0.680990343481132, + "per_class_f1": { + "FAVOR": 0.6147186147186147, + "AGAINST": 0.6090225563909775, + "NONE": 0.8128772635814889 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.88728115636937, + "weighted_f1_pct": 71.41480610224878, + "macro_precision_pct": 67.82529550337472, + "macro_recall_pct": 68.0990343481132, + "per_class_f1_pct": { + "FAVOR": 61.471861471861466, + "AGAINST": 60.902255639097746, + "NONE": 81.28772635814889 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6726493491846224, + "weighted_f1": 0.7183260433705415, + "macro_precision": 0.6688251735574107, + "macro_recall": 0.6943253598426012, + "per_class_f1": { + "FAVOR": 0.5849802371541502, + "AGAINST": 0.6147859922178989, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 67.26493491846225, + "weighted_f1_pct": 71.83260433705415, + "macro_precision_pct": 66.88251735574107, + "macro_recall_pct": 69.43253598426013, + "per_class_f1_pct": { + "FAVOR": 58.498023715415016, + "AGAINST": 61.478599221789885, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7252440725244073, + "macro_f1": 0.687297833928024, + "weighted_f1": 0.7348012320155364, + "macro_precision": 0.6773679936153961, + "macro_recall": 0.7199511585266221, + "per_class_f1": { + "FAVOR": 0.5931558935361216, + "AGAINST": 0.6650485436893204, + "NONE": 0.8036890645586298 + }, + "accuracy_pct": 72.52440725244072, + "macro_f1_pct": 68.7297833928024, + "weighted_f1_pct": 73.48012320155364, + "macro_precision_pct": 67.73679936153961, + "macro_recall_pct": 71.99511585266221, + "per_class_f1_pct": { + "FAVOR": 59.31558935361216, + "AGAINST": 66.50485436893204, + "NONE": 80.36890645586297 + } + } + ], + "accuracy": 0.6507654759263368, + "macro_f1": 0.6190124686243678, + "weighted_f1": 0.6655008809627991, + "macro_precision": 0.6236012548743894, + "macro_recall": 0.6457565134815683, + "per_class_f1": { + "FAVOR": 0.5293201868188895, + "AGAINST": 0.5725631768953069, + "NONE": 0.7551540421589067 + }, + "accuracy_pct": 65.07654759263369, + "macro_f1_pct": 61.90124686243678, + "weighted_f1_pct": 66.55008809627991, + "macro_precision_pct": 62.36012548743894, + "macro_recall_pct": 64.57565134815682, + "per_class_f1_pct": { + "FAVOR": 52.93201868188895, + "AGAINST": 57.25631768953069, + "NONE": 75.51540421589067 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.7884441473983753, + "losses": { + "current_ce": 0.32363019625784134, + "replay_ce": 0.12402305822162067, + "dark_distillation": 0.242403445436674, + "total": 0.5688549762847377, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5412960609911055, + "macro_f1": 0.49153344822436346, + "weighted_f1": 0.5751752658925274, + "macro_precision": 0.5253960679428088, + "macro_recall": 0.5316047182060254, + "per_class_f1": { + "FAVOR": 0.3612040133779264, + "AGAINST": 0.42857142857142855, + "NONE": 0.6848249027237354 + }, + "accuracy_pct": 54.129606099110546, + "macro_f1_pct": 49.153344822436345, + "weighted_f1_pct": 57.517526589252746, + "macro_precision_pct": 52.53960679428088, + "macro_recall_pct": 53.16047182060254, + "per_class_f1_pct": { + "FAVOR": 36.12040133779264, + "AGAINST": 42.857142857142854, + "NONE": 68.48249027237354 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6748768472906403, + "macro_f1": 0.6674158859225438, + "weighted_f1": 0.6834760809478491, + "macro_precision": 0.6791549929854218, + "macro_recall": 0.6970050514906608, + "per_class_f1": { + "FAVOR": 0.6131805157593123, + "AGAINST": 0.6577540106951871, + "NONE": 0.7313131313131314 + }, + "accuracy_pct": 67.48768472906403, + "macro_f1_pct": 66.74158859225437, + "weighted_f1_pct": 68.3476080947849, + "macro_precision_pct": 67.91549929854219, + "macro_recall_pct": 69.70050514906609, + "per_class_f1_pct": { + "FAVOR": 61.31805157593123, + "AGAINST": 65.77540106951871, + "NONE": 73.13131313131314 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5458276333789329, + "macro_f1": 0.5217063257700835, + "weighted_f1": 0.5723896070259619, + "macro_precision": 0.5532984563182359, + "macro_recall": 0.5760526253495789, + "per_class_f1": { + "FAVOR": 0.4011627906976744, + "AGAINST": 0.5111561866125761, + "NONE": 0.6528 + }, + "accuracy_pct": 54.58276333789329, + "macro_f1_pct": 52.17063257700835, + "weighted_f1_pct": 57.23896070259619, + "macro_precision_pct": 55.329845631823595, + "macro_recall_pct": 57.60526253495789, + "per_class_f1_pct": { + "FAVOR": 40.116279069767444, + "AGAINST": 51.115618661257614, + "NONE": 65.28 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5665171898355755, + "macro_f1": 0.5440816434804114, + "weighted_f1": 0.5888137872476565, + "macro_precision": 0.5769497188899014, + "macro_recall": 0.5874678491822617, + "per_class_f1": { + "FAVOR": 0.5174825174825175, + "AGAINST": 0.4458204334365325, + "NONE": 0.6689419795221843 + }, + "accuracy_pct": 56.65171898355755, + "macro_f1_pct": 54.40816434804114, + "weighted_f1_pct": 58.88137872476565, + "macro_precision_pct": 57.694971888990146, + "macro_recall_pct": 58.74678491822617, + "per_class_f1_pct": { + "FAVOR": 51.74825174825175, + "AGAINST": 44.58204334365325, + "NONE": 66.89419795221842 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.65348068863568, + "weighted_f1": 0.6888605100502637, + "macro_precision": 0.6503852924706665, + "macro_recall": 0.6638109700010953, + "per_class_f1": { + "FAVOR": 0.5931558935361216, + "AGAINST": 0.5791505791505791, + "NONE": 0.788135593220339 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 65.348068863568, + "weighted_f1_pct": 68.88605100502637, + "macro_precision_pct": 65.03852924706665, + "macro_recall_pct": 66.38109700010953, + "per_class_f1_pct": { + "FAVOR": 59.31558935361216, + "AGAINST": 57.91505791505791, + "NONE": 78.8135593220339 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6579476861167002, + "macro_f1": 0.6233952427500814, + "weighted_f1": 0.6763584439186437, + "macro_precision": 0.6459392459392459, + "macro_recall": 0.6453286619953286, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.5333333333333333, + "NONE": 0.7913978494623656 + }, + "accuracy_pct": 65.79476861167002, + "macro_f1_pct": 62.33952427500814, + "weighted_f1_pct": 67.63584439186438, + "macro_precision_pct": 64.59392459392458, + "macro_recall_pct": 64.53286619953286, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 53.333333333333336, + "NONE": 79.13978494623656 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6569037656903766, + "macro_f1": 0.6322903795794608, + "weighted_f1": 0.6715322349898079, + "macro_precision": 0.6377831552060228, + "macro_recall": 0.6946489931740717, + "per_class_f1": { + "FAVOR": 0.532051282051282, + "AGAINST": 0.6407322654462243, + "NONE": 0.724087591240876 + }, + "accuracy_pct": 65.69037656903765, + "macro_f1_pct": 63.229037957946076, + "weighted_f1_pct": 67.15322349898078, + "macro_precision_pct": 63.77831552060228, + "macro_recall_pct": 69.46489931740717, + "per_class_f1_pct": { + "FAVOR": 53.205128205128204, + "AGAINST": 64.07322654462243, + "NONE": 72.4087591240876 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7408637873754153, + "macro_f1": 0.7083922773437125, + "weighted_f1": 0.7499359400800085, + "macro_precision": 0.6960853491103766, + "macro_recall": 0.7416147817574217, + "per_class_f1": { + "FAVOR": 0.6367041198501873, + "AGAINST": 0.6688524590163935, + "NONE": 0.819620253164557 + }, + "accuracy_pct": 74.08637873754152, + "macro_f1_pct": 70.83922773437125, + "weighted_f1_pct": 74.99359400800085, + "macro_precision_pct": 69.60853491103765, + "macro_recall_pct": 74.16147817574216, + "per_class_f1_pct": { + "FAVOR": 63.670411985018724, + "AGAINST": 66.88524590163935, + "NONE": 81.9620253164557 + } + } + ], + "accuracy": 0.6259541984732825, + "macro_f1": 0.6007306208930351, + "weighted_f1": 0.6444197561375281, + "macro_precision": 0.6097447703930157, + "macro_recall": 0.6384883349749618, + "per_class_f1": { + "FAVOR": 0.5220759101471727, + "AGAINST": 0.5521514629948365, + "NONE": 0.7279644895370958 + }, + "accuracy_pct": 62.59541984732825, + "macro_f1_pct": 60.073062089303505, + "weighted_f1_pct": 64.44197561375282, + "macro_precision_pct": 60.97447703930157, + "macro_recall_pct": 63.84883349749618, + "per_class_f1_pct": { + "FAVOR": 52.207591014717266, + "AGAINST": 55.21514629948365, + "NONE": 72.79644895370959 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s1__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__doc__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0a9b0da17e34a9b27171de06b154ab10b8e70ab7 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__doc__s1__seed13/run.log b/outputs/trace_s6_derpp__doc__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..5e6fcad6f896b76bb35778d3e4be25ad6598b538 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 04:11:33,686 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__doc__s1__seed13 device=cuda +2026-09-02 04:41:21,953 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6052869864632922, 'average_forgetting': 0.06561329008898055, 'backward_transfer': -0.06484128888973294, 'final_mean_target_macro_f1_pct': 60.52869864632922, 'average_forgetting_pct': 6.561329008898055, 'backward_transfer_pct': -6.484128888973293, 'final_pooled_accuracy': 0.6259541984732825, 'final_pooled_accuracy_pct': 62.59541984732825, 'final_pooled_macro_f1': 0.6007306208930351, 'final_pooled_macro_f1_pct': 60.073062089303505, 'final_pooled_weighted_f1': 0.6444197561375281, 'final_pooled_weighted_f1_pct': 64.44197561375282, 'final_pooled_macro_precision': 0.6097447703930157, 'final_pooled_macro_precision_pct': 60.97447703930157, 'final_pooled_macro_recall': 0.6384883349749618, 'final_pooled_macro_recall_pct': 63.84883349749618, 'final_pooled_per_class_f1': {'FAVOR': 0.5220759101471727, 'AGAINST': 0.5521514629948365, 'NONE': 0.7279644895370958}, 'final_pooled_per_class_f1_pct': {'FAVOR': 52.207591014717266, 'AGAINST': 55.21514629948365, 'NONE': 72.79644895370959}} +2026-09-02 04:41:21,953 [INFO] __main__: efficiency={'training_wall_time_seconds': 1786.894356549019, 'training_wall_time_minutes': 29.781572609150317, 'peak_gpu_memory_allocated_gb': 10.397007942199707, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s6_derpp__doc__s2__seed13/environment.json b/outputs/trace_s6_derpp__doc__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..c24025239117541e653d5fbeaabb074fd02bed85 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T04:41:29.927825+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s2__seed13/metrics.json b/outputs/trace_s6_derpp__doc__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..ca90126381a81261ebe9f1e813ca2ff86094eed3 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s2__seed13/metrics.json @@ -0,0 +1,1594 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7176203235269606, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6997250537468859, + 0.5857817613723736, + null, + null, + null, + null, + null, + null + ], + [ + 0.6779960113876987, + 0.5943608903768479, + 0.6847550482587769, + null, + null, + null, + null, + null + ], + [ + 0.6837636239081027, + 0.5637068176489537, + 0.6657500074887452, + 0.7073801658959114, + null, + null, + null, + null + ], + [ + 0.6930014867215958, + 0.5788485199097199, + 0.6764314706569566, + 0.7033861443248277, + 0.6543311300198124, + null, + null, + null + ], + [ + 0.6908739704060062, + 0.5579064984450782, + 0.6355294538811335, + 0.6807779130631797, + 0.5982431302220618, + 0.5915705381258056, + null, + null + ], + [ + 0.6832901548051473, + 0.5548206707083708, + 0.6621212121212122, + 0.6780403217021785, + 0.5987180637877826, + 0.5723530309478079, + 0.6773714211178503, + null + ], + [ + 0.6885807637083147, + 0.560765590811874, + 0.6620625584370708, + 0.6688332240670034, + 0.6025389275724815, + 0.5628969860548808, + 0.6629092304992381, + 0.6815813864386988 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 8, + "best_dev_macro_f1": 0.6924205015674249, + "losses": { + "current_ce": 0.5369736462547665, + "total": 0.5369736462547665, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7438423645320197, + "macro_f1": 0.7176203235269606, + "weighted_f1": 0.7435449773575725, + "macro_precision": 0.7190878766965724, + "macro_recall": 0.7164253637768642, + "per_class_f1": { + "FAVOR": 0.6223776223776224, + "AGAINST": 0.7112462006079028, + "NONE": 0.8192371475953566 + }, + "accuracy_pct": 74.38423645320196, + "macro_f1_pct": 71.76203235269605, + "weighted_f1_pct": 74.35449773575725, + "macro_precision_pct": 71.90878766965724, + "macro_recall_pct": 71.64253637768641, + "per_class_f1_pct": { + "FAVOR": 62.23776223776224, + "AGAINST": 71.12462006079028, + "NONE": 81.92371475953566 + } + } + ], + "accuracy": 0.7438423645320197, + "macro_f1": 0.7176203235269606, + "weighted_f1": 0.7435449773575725, + "macro_precision": 0.7190878766965724, + "macro_recall": 0.7164253637768642, + "per_class_f1": { + "FAVOR": 0.6223776223776224, + "AGAINST": 0.7112462006079028, + "NONE": 0.8192371475953566 + }, + "accuracy_pct": 74.38423645320196, + "macro_f1_pct": 71.76203235269605, + "weighted_f1_pct": 74.35449773575725, + "macro_precision_pct": 71.90878766965724, + "macro_recall_pct": 71.64253637768641, + "per_class_f1_pct": { + "FAVOR": 62.23776223776224, + "AGAINST": 71.12462006079028, + "NONE": 81.92371475953566 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6790623136362042, + "losses": { + "current_ce": 0.5934215384617186, + "replay_ce": 0.13203714067737263, + "dark_distillation": 0.24501096462471444, + "total": 0.8479641610070279, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7208538587848933, + "macro_f1": 0.6997250537468859, + "weighted_f1": 0.7250734015481426, + "macro_precision": 0.7052684903748734, + "macro_recall": 0.706409863865467, + "per_class_f1": { + "FAVOR": 0.6181818181818182, + "AGAINST": 0.6796116504854369, + "NONE": 0.8013816925734024 + }, + "accuracy_pct": 72.08538587848933, + "macro_f1_pct": 69.97250537468858, + "weighted_f1_pct": 72.50734015481426, + "macro_precision_pct": 70.52684903748734, + "macro_recall_pct": 70.6409863865467, + "per_class_f1_pct": { + "FAVOR": 61.81818181818181, + "AGAINST": 67.96116504854369, + "NONE": 80.13816925734024 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6547085201793722, + "macro_f1": 0.5857817613723736, + "weighted_f1": 0.660940171610141, + "macro_precision": 0.5994686289296266, + "macro_recall": 0.6079151230161667, + "per_class_f1": { + "FAVOR": 0.5775656324582339, + "AGAINST": 0.3879310344827586, + "NONE": 0.7918486171761281 + }, + "accuracy_pct": 65.47085201793722, + "macro_f1_pct": 58.57817613723736, + "weighted_f1_pct": 66.0940171610141, + "macro_precision_pct": 59.94686289296266, + "macro_recall_pct": 60.79151230161667, + "per_class_f1_pct": { + "FAVOR": 57.756563245823386, + "AGAINST": 38.793103448275865, + "NONE": 79.1848617176128 + } + } + ], + "accuracy": 0.6862284820031299, + "macro_f1": 0.6487325985349467, + "weighted_f1": 0.6921631325867317, + "macro_precision": 0.659115616291249, + "macro_recall": 0.6620443534953514, + "per_class_f1": { + "FAVOR": 0.595460614152203, + "AGAINST": 0.5545286506469501, + "NONE": 0.7962085308056872 + }, + "accuracy_pct": 68.62284820031299, + "macro_f1_pct": 64.87325985349467, + "weighted_f1_pct": 69.21631325867317, + "macro_precision_pct": 65.9115616291249, + "macro_recall_pct": 66.20443534953515, + "per_class_f1_pct": { + "FAVOR": 59.546061415220294, + "AGAINST": 55.45286506469501, + "NONE": 79.62085308056872 + } + }, + { + "after_task": 3, + "best_epoch": 8, + "best_dev_macro_f1": 0.7318545035275269, + "losses": { + "current_ce": 0.4831791052151294, + "replay_ce": 0.1696243535372473, + "dark_distillation": 0.2909816677371661, + "total": 0.7982942916807674, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6912972085385879, + "macro_f1": 0.6779960113876987, + "weighted_f1": 0.7017810716608873, + "macro_precision": 0.6913085499934661, + "macro_recall": 0.6933752272086626, + "per_class_f1": { + "FAVOR": 0.5863013698630137, + "AGAINST": 0.6773162939297125, + "NONE": 0.7703703703703704 + }, + "accuracy_pct": 69.12972085385879, + "macro_f1_pct": 67.79960113876987, + "weighted_f1_pct": 70.17810716608874, + "macro_precision_pct": 69.13085499934661, + "macro_recall_pct": 69.33752272086626, + "per_class_f1_pct": { + "FAVOR": 58.63013698630137, + "AGAINST": 67.73162939297124, + "NONE": 77.03703703703704 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6442451420029895, + "macro_f1": 0.5943608903768479, + "weighted_f1": 0.655091026621031, + "macro_precision": 0.6386307202762116, + "macro_recall": 0.6254301952725226, + "per_class_f1": { + "FAVOR": 0.5732484076433121, + "AGAINST": 0.4474885844748858, + "NONE": 0.7623456790123457 + }, + "accuracy_pct": 64.42451420029896, + "macro_f1_pct": 59.436089037684795, + "weighted_f1_pct": 65.5091026621031, + "macro_precision_pct": 63.86307202762116, + "macro_recall_pct": 62.54301952725226, + "per_class_f1_pct": { + "FAVOR": 57.324840764331206, + "AGAINST": 44.74885844748858, + "NONE": 76.23456790123457 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6847550482587769, + "weighted_f1": 0.7299989614214312, + "macro_precision": 0.6799814011626156, + "macro_recall": 0.7051401976689333, + "per_class_f1": { + "FAVOR": 0.5769230769230769, + "AGAINST": 0.6477732793522267, + "NONE": 0.8295687885010267 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 68.47550482587769, + "weighted_f1_pct": 72.99989614214311, + "macro_precision_pct": 67.99814011626157, + "macro_recall_pct": 70.51401976689333, + "per_class_f1_pct": { + "FAVOR": 57.692307692307686, + "AGAINST": 64.77732793522267, + "NONE": 82.95687885010267 + } + } + ], + "accuracy": 0.6811267605633803, + "macro_f1": 0.6554274591828205, + "weighted_f1": 0.6941508341133322, + "macro_precision": 0.6696586578458673, + "macro_recall": 0.6753924598834047, + "per_class_f1": { + "FAVOR": 0.5784671532846716, + "AGAINST": 0.6033376123234917, + "NONE": 0.7844776119402985 + }, + "accuracy_pct": 68.11267605633803, + "macro_f1_pct": 65.54274591828205, + "weighted_f1_pct": 69.41508341133323, + "macro_precision_pct": 66.96586578458673, + "macro_recall_pct": 67.53924598834047, + "per_class_f1_pct": { + "FAVOR": 57.846715328467155, + "AGAINST": 60.33376123234917, + "NONE": 78.44776119402985 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7966480918608578, + "losses": { + "current_ce": 0.39983995662016025, + "replay_ce": 0.16771390817913354, + "dark_distillation": 0.28156477116486606, + "total": 0.7083362503963359, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6962233169129721, + "macro_f1": 0.6837636239081027, + "weighted_f1": 0.7053104725429554, + "macro_precision": 0.6902782721361426, + "macro_recall": 0.6978744447360491, + "per_class_f1": { + "FAVOR": 0.5868945868945868, + "AGAINST": 0.6996904024767802, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 69.6223316912972, + "macro_f1_pct": 68.37636239081027, + "weighted_f1_pct": 70.53104725429554, + "macro_precision_pct": 69.02782721361426, + "macro_recall_pct": 69.7874444736049, + "per_class_f1_pct": { + "FAVOR": 58.68945868945868, + "AGAINST": 69.96904024767801, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5637068176489537, + "weighted_f1": 0.6203031301083258, + "macro_precision": 0.5920481788842785, + "macro_recall": 0.598603282702772, + "per_class_f1": { + "FAVOR": 0.5414847161572053, + "AGAINST": 0.42911877394636017, + "NONE": 0.7205169628432956 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.370681764895366, + "weighted_f1_pct": 62.030313010832586, + "macro_precision_pct": 59.204817888427854, + "macro_recall_pct": 59.86032827027719, + "per_class_f1_pct": { + "FAVOR": 54.14847161572053, + "AGAINST": 42.911877394636015, + "NONE": 72.05169628432955 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6657500074887452, + "weighted_f1": 0.6983665873170453, + "macro_precision": 0.6717438763043523, + "macro_recall": 0.6986446216331275, + "per_class_f1": { + "FAVOR": 0.5850340136054422, + "AGAINST": 0.6419753086419753, + "NONE": 0.7702407002188184 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.57500074887453, + "weighted_f1_pct": 69.83665873170453, + "macro_precision_pct": 67.17438763043523, + "macro_recall_pct": 69.86446216331275, + "per_class_f1_pct": { + "FAVOR": 58.50340136054422, + "AGAINST": 64.19753086419753, + "NONE": 77.02407002188184 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7073801658959114, + "weighted_f1": 0.7481791371490906, + "macro_precision": 0.6965198323688889, + "macro_recall": 0.7230161348757737, + "per_class_f1": { + "FAVOR": 0.6610878661087866, + "AGAINST": 0.64, + "NONE": 0.8210526315789474 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.73801658959114, + "weighted_f1_pct": 74.81791371490905, + "macro_precision_pct": 69.6519832368889, + "macro_recall_pct": 72.30161348757737, + "per_class_f1_pct": { + "FAVOR": 66.10878661087865, + "AGAINST": 64.0, + "NONE": 82.10526315789474 + } + } + ], + "accuracy": 0.6794278502313841, + "macro_f1": 0.6543796783186898, + "weighted_f1": 0.691228459546465, + "macro_precision": 0.6568693465678945, + "macro_recall": 0.680474610323541, + "per_class_f1": { + "FAVOR": 0.5842026825633383, + "AGAINST": 0.6086956521739131, + "NONE": 0.7702407002188184 + }, + "accuracy_pct": 67.94278502313841, + "macro_f1_pct": 65.43796783186899, + "weighted_f1_pct": 69.1228459546465, + "macro_precision_pct": 65.68693465678945, + "macro_recall_pct": 68.0474610323541, + "per_class_f1_pct": { + "FAVOR": 58.42026825633383, + "AGAINST": 60.86956521739131, + "NONE": 77.02407002188184 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.6796392243035284, + "losses": { + "current_ce": 0.5303162590883397, + "replay_ce": 0.14063591205411488, + "dark_distillation": 0.2745294854817567, + "total": 0.8082169146449477, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6930014867215958, + "weighted_f1": 0.7177760793026335, + "macro_precision": 0.7033397945944256, + "macro_recall": 0.7056383264282836, + "per_class_f1": { + "FAVOR": 0.6045197740112994, + "AGAINST": 0.6838709677419355, + "NONE": 0.7906137184115524 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.30014867215958, + "weighted_f1_pct": 71.77760793026336, + "macro_precision_pct": 70.33397945944256, + "macro_recall_pct": 70.56383264282836, + "per_class_f1_pct": { + "FAVOR": 60.451977401129945, + "AGAINST": 68.38709677419355, + "NONE": 79.06137184115524 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6278026905829597, + "macro_f1": 0.5788485199097199, + "weighted_f1": 0.6427858279901902, + "macro_precision": 0.5927825261158595, + "macro_recall": 0.6058306807807141, + "per_class_f1": { + "FAVOR": 0.5560747663551402, + "AGAINST": 0.4247104247104247, + "NONE": 0.7557603686635944 + }, + "accuracy_pct": 62.78026905829597, + "macro_f1_pct": 57.88485199097199, + "weighted_f1_pct": 64.27858279901902, + "macro_precision_pct": 59.27825261158595, + "macro_recall_pct": 60.58306807807141, + "per_class_f1_pct": { + "FAVOR": 55.60747663551402, + "AGAINST": 42.471042471042466, + "NONE": 75.57603686635944 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7022132796780685, + "macro_f1": 0.6764314706569566, + "weighted_f1": 0.7160911965774714, + "macro_precision": 0.6771663315824875, + "macro_recall": 0.7014313163738451, + "per_class_f1": { + "FAVOR": 0.5815602836879432, + "AGAINST": 0.6443514644351465, + "NONE": 0.8033826638477801 + }, + "accuracy_pct": 70.22132796780684, + "macro_f1_pct": 67.64314706569566, + "weighted_f1_pct": 71.60911965774714, + "macro_precision_pct": 67.71663315824875, + "macro_recall_pct": 70.14313163738451, + "per_class_f1_pct": { + "FAVOR": 58.156028368794324, + "AGAINST": 64.43514644351464, + "NONE": 80.33826638477801 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7033861443248277, + "weighted_f1": 0.7481900701852747, + "macro_precision": 0.69354395455942, + "macro_recall": 0.7257961470859762, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6126760563380281, + "NONE": 0.8308157099697885 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.33861443248277, + "weighted_f1_pct": 74.81900701852747, + "macro_precision_pct": 69.354395455942, + "macro_recall_pct": 72.57961470859762, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 61.267605633802816, + "NONE": 83.08157099697885 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.7179161372299873, + "macro_f1": 0.6543311300198124, + "weighted_f1": 0.724501712540731, + "macro_precision": 0.6472427238786292, + "macro_recall": 0.6836030970562561, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.532051282051282, + "NONE": 0.822246455834242 + }, + "accuracy_pct": 71.79161372299873, + "macro_f1_pct": 65.43311300198124, + "weighted_f1_pct": 72.4501712540731, + "macro_precision_pct": 64.72427238786292, + "macro_recall_pct": 68.36030970562561, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 53.205128205128204, + "NONE": 82.2246455834242 + } + } + ], + "accuracy": 0.6997471554993678, + "macro_f1": 0.6608186352096336, + "weighted_f1": 0.7093942826604703, + "macro_precision": 0.6600148758733049, + "macro_recall": 0.684500602785573, + "per_class_f1": { + "FAVOR": 0.5986802639472105, + "AGAINST": 0.5811965811965812, + "NONE": 0.802579060485109 + }, + "accuracy_pct": 69.97471554993679, + "macro_f1_pct": 66.08186352096335, + "weighted_f1_pct": 70.93942826604703, + "macro_precision_pct": 66.00148758733049, + "macro_recall_pct": 68.4500602785573, + "per_class_f1_pct": { + "FAVOR": 59.86802639472105, + "AGAINST": 58.119658119658126, + "NONE": 80.25790604851089 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.7102031403401267, + "losses": { + "current_ce": 0.47710517739492747, + "replay_ce": 0.13272886373220927, + "dark_distillation": 0.26405621474934005, + "total": 0.7418621473369145, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7060755336617406, + "macro_f1": 0.6908739704060062, + "weighted_f1": 0.7148688471671196, + "macro_precision": 0.6927430871918562, + "macro_recall": 0.7068080164375328, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6882352941176471, + "NONE": 0.7843866171003717 + }, + "accuracy_pct": 70.60755336617406, + "macro_f1_pct": 69.08739704060062, + "weighted_f1_pct": 71.48688471671196, + "macro_precision_pct": 69.27430871918561, + "macro_recall_pct": 70.68080164375328, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.82352941176471, + "NONE": 78.43866171003717 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5579064984450782, + "weighted_f1": 0.6201200934448771, + "macro_precision": 0.5742529488348886, + "macro_recall": 0.5881176217618589, + "per_class_f1": { + "FAVOR": 0.5401459854014599, + "AGAINST": 0.40397350993377484, + "NONE": 0.7296 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 55.790649844507826, + "weighted_f1_pct": 62.01200934448771, + "macro_precision_pct": 57.42529488348887, + "macro_recall_pct": 58.81176217618589, + "per_class_f1_pct": { + "FAVOR": 54.01459854014598, + "AGAINST": 40.397350993377486, + "NONE": 72.96000000000001 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6639839034205232, + "macro_f1": 0.6355294538811335, + "weighted_f1": 0.6797693545119485, + "macro_precision": 0.6391735048969369, + "macro_recall": 0.6638707673190432, + "per_class_f1": { + "FAVOR": 0.5171102661596958, + "AGAINST": 0.6119402985074627, + "NONE": 0.7775377969762419 + }, + "accuracy_pct": 66.39839034205231, + "macro_f1_pct": 63.55294538811334, + "weighted_f1_pct": 67.97693545119485, + "macro_precision_pct": 63.91735048969369, + "macro_recall_pct": 66.38707673190432, + "per_class_f1_pct": { + "FAVOR": 51.71102661596958, + "AGAINST": 61.19402985074627, + "NONE": 77.7537796976242 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.6807779130631797, + "weighted_f1": 0.7278891493400268, + "macro_precision": 0.6693367820260195, + "macro_recall": 0.7036369356569177, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.6081504702194357, + "NONE": 0.8112324492979719 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 68.07779130631798, + "weighted_f1_pct": 72.78891493400268, + "macro_precision_pct": 66.93367820260195, + "macro_recall_pct": 70.36369356569176, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 60.81504702194357, + "NONE": 81.1232449297972 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6645489199491741, + "macro_f1": 0.5982431302220618, + "weighted_f1": 0.6775638594257173, + "macro_precision": 0.5928028798411122, + "macro_recall": 0.6335312089669388, + "per_class_f1": { + "FAVOR": 0.5439560439560439, + "AGAINST": 0.46296296296296297, + "NONE": 0.7878103837471784 + }, + "accuracy_pct": 66.4548919949174, + "macro_f1_pct": 59.82431302220618, + "weighted_f1_pct": 67.75638594257173, + "macro_precision_pct": 59.280287984111226, + "macro_recall_pct": 63.35312089669388, + "per_class_f1_pct": { + "FAVOR": 54.395604395604394, + "AGAINST": 46.2962962962963, + "NONE": 78.78103837471784 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6347469220246238, + "macro_f1": 0.5915705381258056, + "weighted_f1": 0.6457690219601727, + "macro_precision": 0.5947171424498188, + "macro_recall": 0.6144562746086014, + "per_class_f1": { + "FAVOR": 0.4444444444444444, + "AGAINST": 0.6128440366972477, + "NONE": 0.7174231332357247 + }, + "accuracy_pct": 63.47469220246238, + "macro_f1_pct": 59.157053812580564, + "weighted_f1_pct": 64.57690219601727, + "macro_precision_pct": 59.471714244981875, + "macro_recall_pct": 61.445627460860145, + "per_class_f1_pct": { + "FAVOR": 44.44444444444444, + "AGAINST": 61.284403669724774, + "NONE": 71.74231332357247 + } + } + ], + "accuracy": 0.6623876765083441, + "macro_f1": 0.6288070997116048, + "weighted_f1": 0.6764285052056415, + "macro_precision": 0.6255367851184187, + "macro_recall": 0.6563589146391541, + "per_class_f1": { + "FAVOR": 0.5474137931034483, + "AGAINST": 0.5710200190657769, + "NONE": 0.7679874869655892 + }, + "accuracy_pct": 66.2387676508344, + "macro_f1_pct": 62.88070997116048, + "weighted_f1_pct": 67.64285052056415, + "macro_precision_pct": 62.55367851184187, + "macro_recall_pct": 65.63589146391541, + "per_class_f1_pct": { + "FAVOR": 54.741379310344826, + "AGAINST": 57.10200190657769, + "NONE": 76.79874869655892 + } + }, + { + "after_task": 7, + "best_epoch": 8, + "best_dev_macro_f1": 0.6412283431891275, + "losses": { + "current_ce": 0.4483082293044953, + "replay_ce": 0.12987337100009125, + "dark_distillation": 0.2574882027648744, + "total": 0.7069257031594004, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6978653530377669, + "macro_f1": 0.6832901548051473, + "weighted_f1": 0.7047975373868168, + "macro_precision": 0.6832971718127546, + "macro_recall": 0.6993525627940814, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6646884272997032, + "NONE": 0.7697974217311234 + }, + "accuracy_pct": 69.78653530377669, + "macro_f1_pct": 68.32901548051473, + "weighted_f1_pct": 70.47975373868168, + "macro_precision_pct": 68.32971718127546, + "macro_recall_pct": 69.93525627940814, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 66.46884272997032, + "NONE": 76.97974217311234 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5949177877428998, + "macro_f1": 0.5548206707083708, + "weighted_f1": 0.6175579794295194, + "macro_precision": 0.5696183632486281, + "macro_recall": 0.5838982124791584, + "per_class_f1": { + "FAVOR": 0.543640897755611, + "AGAINST": 0.3935483870967742, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 59.49177877428998, + "macro_f1_pct": 55.48206707083708, + "weighted_f1_pct": 61.75579794295194, + "macro_precision_pct": 56.961836324862816, + "macro_recall_pct": 58.38982124791584, + "per_class_f1_pct": { + "FAVOR": 54.3640897755611, + "AGAINST": 39.35483870967742, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6621212121212122, + "weighted_f1": 0.6985993885390264, + "macro_precision": 0.6645584673931846, + "macro_recall": 0.692476959718339, + "per_class_f1": { + "FAVOR": 0.5642857142857143, + "AGAINST": 0.6428571428571429, + "NONE": 0.7792207792207793 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.21212121212122, + "weighted_f1_pct": 69.85993885390263, + "macro_precision_pct": 66.45584673931846, + "macro_recall_pct": 69.2476959718339, + "per_class_f1_pct": { + "FAVOR": 56.42857142857143, + "AGAINST": 64.28571428571429, + "NONE": 77.92207792207793 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.6780403217021785, + "weighted_f1": 0.7272923095618483, + "macro_precision": 0.666627227025094, + "macro_recall": 0.7012194628742362, + "per_class_f1": { + "FAVOR": 0.616600790513834, + "AGAINST": 0.6032786885245902, + "NONE": 0.8142414860681114 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 67.80403217021785, + "weighted_f1_pct": 72.72923095618484, + "macro_precision_pct": 66.6627227025094, + "macro_recall_pct": 70.12194628742363, + "per_class_f1_pct": { + "FAVOR": 61.6600790513834, + "AGAINST": 60.32786885245902, + "NONE": 81.42414860681114 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6607369758576874, + "macro_f1": 0.5987180637877826, + "weighted_f1": 0.672786084679873, + "macro_precision": 0.6028899994816727, + "macro_recall": 0.6481791407172671, + "per_class_f1": { + "FAVOR": 0.5621890547263682, + "AGAINST": 0.45695364238410596, + "NONE": 0.7770114942528735 + }, + "accuracy_pct": 66.07369758576874, + "macro_f1_pct": 59.87180637877826, + "weighted_f1_pct": 67.2786084679873, + "macro_precision_pct": 60.288999948167266, + "macro_recall_pct": 64.81791407172672, + "per_class_f1_pct": { + "FAVOR": 56.21890547263681, + "AGAINST": 45.6953642384106, + "NONE": 77.70114942528735 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6114911080711354, + "macro_f1": 0.5723530309478079, + "weighted_f1": 0.6268051119802884, + "macro_precision": 0.5819247469423293, + "macro_recall": 0.601365315920053, + "per_class_f1": { + "FAVOR": 0.42857142857142855, + "AGAINST": 0.5865209471766849, + "NONE": 0.7019667170953101 + }, + "accuracy_pct": 61.14911080711354, + "macro_f1_pct": 57.23530309478079, + "weighted_f1_pct": 62.680511198028846, + "macro_precision_pct": 58.192474694232935, + "macro_recall_pct": 60.136531592005305, + "per_class_f1_pct": { + "FAVOR": 42.857142857142854, + "AGAINST": 58.65209471766849, + "NONE": 70.19667170953102 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7022132796780685, + "macro_f1": 0.6773714211178503, + "weighted_f1": 0.7074057162383418, + "macro_precision": 0.674701993170269, + "macro_recall": 0.6886861739477804, + "per_class_f1": { + "FAVOR": 0.6245353159851301, + "AGAINST": 0.616, + "NONE": 0.791578947368421 + }, + "accuracy_pct": 70.22132796780684, + "macro_f1_pct": 67.73714211178503, + "weighted_f1_pct": 70.74057162383419, + "macro_precision_pct": 67.4701993170269, + "macro_recall_pct": 68.86861739477804, + "per_class_f1_pct": { + "FAVOR": 62.45353159851301, + "AGAINST": 61.6, + "NONE": 79.15789473684211 + } + } + ], + "accuracy": 0.663023679417122, + "macro_f1": 0.6323772258398118, + "weighted_f1": 0.6760069255986815, + "macro_precision": 0.6298260936113816, + "macro_recall": 0.6608345076467886, + "per_class_f1": { + "FAVOR": 0.5658314350797267, + "AGAINST": 0.5665943600867679, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 66.3023679417122, + "macro_f1_pct": 63.23772258398118, + "weighted_f1_pct": 67.60069255986815, + "macro_precision_pct": 62.98260936113817, + "macro_recall_pct": 66.08345076467886, + "per_class_f1_pct": { + "FAVOR": 56.583143507972665, + "AGAINST": 56.65943600867679, + "NONE": 76.47058823529412 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7339951640488313, + "losses": { + "current_ce": 0.33631183962996414, + "replay_ce": 0.13349357755567, + "dark_distillation": 0.24566642065358354, + "total": 0.5926386286088121, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7060755336617406, + "macro_f1": 0.6885807637083147, + "weighted_f1": 0.7122082796004581, + "macro_precision": 0.6915634273806406, + "macro_recall": 0.7025393722423728, + "per_class_f1": { + "FAVOR": 0.6239067055393586, + "AGAINST": 0.65625, + "NONE": 0.7855855855855856 + }, + "accuracy_pct": 70.60755336617406, + "macro_f1_pct": 68.85807637083147, + "weighted_f1_pct": 71.22082796004581, + "macro_precision_pct": 69.15634273806405, + "macro_recall_pct": 70.25393722423729, + "per_class_f1_pct": { + "FAVOR": 62.390670553935855, + "AGAINST": 65.625, + "NONE": 78.55855855855856 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.560765590811874, + "weighted_f1": 0.6236961875233046, + "macro_precision": 0.581924601445215, + "macro_recall": 0.5920291821912963, + "per_class_f1": { + "FAVOR": 0.535796766743649, + "AGAINST": 0.41134751773049644, + "NONE": 0.7351524879614767 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.076559081187405, + "weighted_f1_pct": 62.36961875233046, + "macro_precision_pct": 58.1924601445215, + "macro_recall_pct": 59.20291821912963, + "per_class_f1_pct": { + "FAVOR": 53.5796766743649, + "AGAINST": 41.13475177304964, + "NONE": 73.51524879614767 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6800804828973843, + "macro_f1": 0.6620625584370708, + "weighted_f1": 0.696596387834992, + "macro_precision": 0.6770893308983955, + "macro_recall": 0.6938892915904411, + "per_class_f1": { + "FAVOR": 0.5787781350482315, + "AGAINST": 0.6347826086956522, + "NONE": 0.7726269315673289 + }, + "accuracy_pct": 68.00804828973843, + "macro_f1_pct": 66.20625584370708, + "weighted_f1_pct": 69.6596387834992, + "macro_precision_pct": 67.70893308983955, + "macro_recall_pct": 69.38892915904411, + "per_class_f1_pct": { + "FAVOR": 57.87781350482315, + "AGAINST": 63.47826086956522, + "NONE": 77.2626931567329 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.707641196013289, + "macro_f1": 0.6688332240670034, + "weighted_f1": 0.7180799109608949, + "macro_precision": 0.6582557434672385, + "macro_recall": 0.6979151387313921, + "per_class_f1": { + "FAVOR": 0.6007604562737643, + "AGAINST": 0.6019417475728155, + "NONE": 0.8037974683544303 + }, + "accuracy_pct": 70.7641196013289, + "macro_f1_pct": 66.88332240670034, + "weighted_f1_pct": 71.80799109608948, + "macro_precision_pct": 65.82557434672384, + "macro_recall_pct": 69.79151387313921, + "per_class_f1_pct": { + "FAVOR": 60.07604562737643, + "AGAINST": 60.19417475728155, + "NONE": 80.37974683544303 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6709021601016518, + "macro_f1": 0.6025389275724815, + "weighted_f1": 0.6814779959474174, + "macro_precision": 0.6132854156375706, + "macro_recall": 0.6478980069394012, + "per_class_f1": { + "FAVOR": 0.5558312655086849, + "AGAINST": 0.45993031358885017, + "NONE": 0.7918552036199095 + }, + "accuracy_pct": 67.09021601016518, + "macro_f1_pct": 60.25389275724815, + "weighted_f1_pct": 68.14779959474174, + "macro_precision_pct": 61.32854156375706, + "macro_recall_pct": 64.78980069394012, + "per_class_f1_pct": { + "FAVOR": 55.58312655086849, + "AGAINST": 45.99303135888502, + "NONE": 79.18552036199095 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5950752393980848, + "macro_f1": 0.5628969860548808, + "weighted_f1": 0.6101661182314937, + "macro_precision": 0.573739219548047, + "macro_recall": 0.5990571939684762, + "per_class_f1": { + "FAVOR": 0.43609022556390975, + "AGAINST": 0.5787545787545788, + "NONE": 0.6738461538461539 + }, + "accuracy_pct": 59.50752393980848, + "macro_f1_pct": 56.289698605488084, + "weighted_f1_pct": 61.01661182314937, + "macro_precision_pct": 57.37392195480469, + "macro_recall_pct": 59.90571939684762, + "per_class_f1_pct": { + "FAVOR": 43.609022556390975, + "AGAINST": 57.87545787545788, + "NONE": 67.38461538461539 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6629092304992381, + "weighted_f1": 0.6893054865972921, + "macro_precision": 0.6593890151254332, + "macro_recall": 0.6789568138131145, + "per_class_f1": { + "FAVOR": 0.6044776119402985, + "AGAINST": 0.6217228464419475, + "NONE": 0.7625272331154684 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 66.2909230499238, + "weighted_f1_pct": 68.93054865972921, + "macro_precision_pct": 65.93890151254332, + "macro_recall_pct": 67.89568138131145, + "per_class_f1_pct": { + "FAVOR": 60.447761194029844, + "AGAINST": 62.172284644194754, + "NONE": 76.25272331154684 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.7140864714086471, + "macro_f1": 0.6815813864386988, + "weighted_f1": 0.7272952637949974, + "macro_precision": 0.6706977605820508, + "macro_recall": 0.7273293080796296, + "per_class_f1": { + "FAVOR": 0.5796178343949044, + "AGAINST": 0.6737400530503979, + "NONE": 0.7913862718707941 + }, + "accuracy_pct": 71.40864714086472, + "macro_f1_pct": 68.15813864386988, + "weighted_f1_pct": 72.72952637949975, + "macro_precision_pct": 67.06977605820508, + "macro_recall_pct": 72.73293080796296, + "per_class_f1_pct": { + "FAVOR": 57.961783439490446, + "AGAINST": 67.37400530503979, + "NONE": 79.13862718707941 + } + } + ], + "accuracy": 0.6676453317674692, + "macro_f1": 0.6376797198359435, + "weighted_f1": 0.6813711179083157, + "macro_precision": 0.6366896058993522, + "macro_recall": 0.6704577942829801, + "per_class_f1": { + "FAVOR": 0.564398308342945, + "AGAINST": 0.5828877005347594, + "NONE": 0.7657531506301261 + }, + "accuracy_pct": 66.76453317674692, + "macro_f1_pct": 63.76797198359435, + "weighted_f1_pct": 68.13711179083157, + "macro_precision_pct": 63.66896058993522, + "macro_recall_pct": 67.04577942829802, + "per_class_f1_pct": { + "FAVOR": 56.439830834294504, + "AGAINST": 58.288770053475936, + "NONE": 76.57531506301261 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6362710834486953, + "average_forgetting": 0.03125746231015737, + "backward_transfer": -0.030031872452375326, + "final_mean_target_macro_f1_pct": 63.627108344869534, + "average_forgetting_pct": 3.125746231015737, + "backward_transfer_pct": -3.0031872452375326, + "final_pooled_accuracy": 0.6676453317674692, + "final_pooled_accuracy_pct": 66.76453317674692, + "final_pooled_macro_f1": 0.6376797198359435, + "final_pooled_macro_f1_pct": 63.76797198359435, + "final_pooled_weighted_f1": 0.6813711179083157, + "final_pooled_weighted_f1_pct": 68.13711179083157, + "final_pooled_macro_precision": 0.6366896058993522, + "final_pooled_macro_precision_pct": 63.66896058993522, + "final_pooled_macro_recall": 0.6704577942829801, + "final_pooled_macro_recall_pct": 67.04577942829802, + "final_pooled_per_class_f1": { + "FAVOR": 0.564398308342945, + "AGAINST": 0.5828877005347594, + "NONE": 0.7657531506301261 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 56.439830834294504, + "AGAINST": 58.288770053475936, + "NONE": 76.57531506301261 + } + }, + "efficiency": { + "training_wall_time_seconds": 1801.5697203811724, + "training_wall_time_minutes": 30.026162006352873, + "peak_gpu_memory_allocated_gb": 10.397007942199707, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s2__seed13/metrics.partial.json b/outputs/trace_s6_derpp__doc__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..5036ead57e80faa2ae48952feeba81cb0cb110b8 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s2__seed13/metrics.partial.json @@ -0,0 +1,1488 @@ +{ + "matrix": [ + [ + 0.7176203235269606, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6997250537468859, + 0.5857817613723736, + null, + null, + null, + null, + null, + null + ], + [ + 0.6779960113876987, + 0.5943608903768479, + 0.6847550482587769, + null, + null, + null, + null, + null + ], + [ + 0.6837636239081027, + 0.5637068176489537, + 0.6657500074887452, + 0.7073801658959114, + null, + null, + null, + null + ], + [ + 0.6930014867215958, + 0.5788485199097199, + 0.6764314706569566, + 0.7033861443248277, + 0.6543311300198124, + null, + null, + null + ], + [ + 0.6908739704060062, + 0.5579064984450782, + 0.6355294538811335, + 0.6807779130631797, + 0.5982431302220618, + 0.5915705381258056, + null, + null + ], + [ + 0.6832901548051473, + 0.5548206707083708, + 0.6621212121212122, + 0.6780403217021785, + 0.5987180637877826, + 0.5723530309478079, + 0.6773714211178503, + null + ], + [ + 0.6885807637083147, + 0.560765590811874, + 0.6620625584370708, + 0.6688332240670034, + 0.6025389275724815, + 0.5628969860548808, + 0.6629092304992381, + 0.6815813864386988 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 8, + "best_dev_macro_f1": 0.6924205015674249, + "losses": { + "current_ce": 0.5369736462547665, + "total": 0.5369736462547665, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7438423645320197, + "macro_f1": 0.7176203235269606, + "weighted_f1": 0.7435449773575725, + "macro_precision": 0.7190878766965724, + "macro_recall": 0.7164253637768642, + "per_class_f1": { + "FAVOR": 0.6223776223776224, + "AGAINST": 0.7112462006079028, + "NONE": 0.8192371475953566 + }, + "accuracy_pct": 74.38423645320196, + "macro_f1_pct": 71.76203235269605, + "weighted_f1_pct": 74.35449773575725, + "macro_precision_pct": 71.90878766965724, + "macro_recall_pct": 71.64253637768641, + "per_class_f1_pct": { + "FAVOR": 62.23776223776224, + "AGAINST": 71.12462006079028, + "NONE": 81.92371475953566 + } + } + ], + "accuracy": 0.7438423645320197, + "macro_f1": 0.7176203235269606, + "weighted_f1": 0.7435449773575725, + "macro_precision": 0.7190878766965724, + "macro_recall": 0.7164253637768642, + "per_class_f1": { + "FAVOR": 0.6223776223776224, + "AGAINST": 0.7112462006079028, + "NONE": 0.8192371475953566 + }, + "accuracy_pct": 74.38423645320196, + "macro_f1_pct": 71.76203235269605, + "weighted_f1_pct": 74.35449773575725, + "macro_precision_pct": 71.90878766965724, + "macro_recall_pct": 71.64253637768641, + "per_class_f1_pct": { + "FAVOR": 62.23776223776224, + "AGAINST": 71.12462006079028, + "NONE": 81.92371475953566 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6790623136362042, + "losses": { + "current_ce": 0.5934215384617186, + "replay_ce": 0.13203714067737263, + "dark_distillation": 0.24501096462471444, + "total": 0.8479641610070279, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7208538587848933, + "macro_f1": 0.6997250537468859, + "weighted_f1": 0.7250734015481426, + "macro_precision": 0.7052684903748734, + "macro_recall": 0.706409863865467, + "per_class_f1": { + "FAVOR": 0.6181818181818182, + "AGAINST": 0.6796116504854369, + "NONE": 0.8013816925734024 + }, + "accuracy_pct": 72.08538587848933, + "macro_f1_pct": 69.97250537468858, + "weighted_f1_pct": 72.50734015481426, + "macro_precision_pct": 70.52684903748734, + "macro_recall_pct": 70.6409863865467, + "per_class_f1_pct": { + "FAVOR": 61.81818181818181, + "AGAINST": 67.96116504854369, + "NONE": 80.13816925734024 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6547085201793722, + "macro_f1": 0.5857817613723736, + "weighted_f1": 0.660940171610141, + "macro_precision": 0.5994686289296266, + "macro_recall": 0.6079151230161667, + "per_class_f1": { + "FAVOR": 0.5775656324582339, + "AGAINST": 0.3879310344827586, + "NONE": 0.7918486171761281 + }, + "accuracy_pct": 65.47085201793722, + "macro_f1_pct": 58.57817613723736, + "weighted_f1_pct": 66.0940171610141, + "macro_precision_pct": 59.94686289296266, + "macro_recall_pct": 60.79151230161667, + "per_class_f1_pct": { + "FAVOR": 57.756563245823386, + "AGAINST": 38.793103448275865, + "NONE": 79.1848617176128 + } + } + ], + "accuracy": 0.6862284820031299, + "macro_f1": 0.6487325985349467, + "weighted_f1": 0.6921631325867317, + "macro_precision": 0.659115616291249, + "macro_recall": 0.6620443534953514, + "per_class_f1": { + "FAVOR": 0.595460614152203, + "AGAINST": 0.5545286506469501, + "NONE": 0.7962085308056872 + }, + "accuracy_pct": 68.62284820031299, + "macro_f1_pct": 64.87325985349467, + "weighted_f1_pct": 69.21631325867317, + "macro_precision_pct": 65.9115616291249, + "macro_recall_pct": 66.20443534953515, + "per_class_f1_pct": { + "FAVOR": 59.546061415220294, + "AGAINST": 55.45286506469501, + "NONE": 79.62085308056872 + } + }, + { + "after_task": 3, + "best_epoch": 8, + "best_dev_macro_f1": 0.7318545035275269, + "losses": { + "current_ce": 0.4831791052151294, + "replay_ce": 0.1696243535372473, + "dark_distillation": 0.2909816677371661, + "total": 0.7982942916807674, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6912972085385879, + "macro_f1": 0.6779960113876987, + "weighted_f1": 0.7017810716608873, + "macro_precision": 0.6913085499934661, + "macro_recall": 0.6933752272086626, + "per_class_f1": { + "FAVOR": 0.5863013698630137, + "AGAINST": 0.6773162939297125, + "NONE": 0.7703703703703704 + }, + "accuracy_pct": 69.12972085385879, + "macro_f1_pct": 67.79960113876987, + "weighted_f1_pct": 70.17810716608874, + "macro_precision_pct": 69.13085499934661, + "macro_recall_pct": 69.33752272086626, + "per_class_f1_pct": { + "FAVOR": 58.63013698630137, + "AGAINST": 67.73162939297124, + "NONE": 77.03703703703704 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6442451420029895, + "macro_f1": 0.5943608903768479, + "weighted_f1": 0.655091026621031, + "macro_precision": 0.6386307202762116, + "macro_recall": 0.6254301952725226, + "per_class_f1": { + "FAVOR": 0.5732484076433121, + "AGAINST": 0.4474885844748858, + "NONE": 0.7623456790123457 + }, + "accuracy_pct": 64.42451420029896, + "macro_f1_pct": 59.436089037684795, + "weighted_f1_pct": 65.5091026621031, + "macro_precision_pct": 63.86307202762116, + "macro_recall_pct": 62.54301952725226, + "per_class_f1_pct": { + "FAVOR": 57.324840764331206, + "AGAINST": 44.74885844748858, + "NONE": 76.23456790123457 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6847550482587769, + "weighted_f1": 0.7299989614214312, + "macro_precision": 0.6799814011626156, + "macro_recall": 0.7051401976689333, + "per_class_f1": { + "FAVOR": 0.5769230769230769, + "AGAINST": 0.6477732793522267, + "NONE": 0.8295687885010267 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 68.47550482587769, + "weighted_f1_pct": 72.99989614214311, + "macro_precision_pct": 67.99814011626157, + "macro_recall_pct": 70.51401976689333, + "per_class_f1_pct": { + "FAVOR": 57.692307692307686, + "AGAINST": 64.77732793522267, + "NONE": 82.95687885010267 + } + } + ], + "accuracy": 0.6811267605633803, + "macro_f1": 0.6554274591828205, + "weighted_f1": 0.6941508341133322, + "macro_precision": 0.6696586578458673, + "macro_recall": 0.6753924598834047, + "per_class_f1": { + "FAVOR": 0.5784671532846716, + "AGAINST": 0.6033376123234917, + "NONE": 0.7844776119402985 + }, + "accuracy_pct": 68.11267605633803, + "macro_f1_pct": 65.54274591828205, + "weighted_f1_pct": 69.41508341133323, + "macro_precision_pct": 66.96586578458673, + "macro_recall_pct": 67.53924598834047, + "per_class_f1_pct": { + "FAVOR": 57.846715328467155, + "AGAINST": 60.33376123234917, + "NONE": 78.44776119402985 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7966480918608578, + "losses": { + "current_ce": 0.39983995662016025, + "replay_ce": 0.16771390817913354, + "dark_distillation": 0.28156477116486606, + "total": 0.7083362503963359, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6962233169129721, + "macro_f1": 0.6837636239081027, + "weighted_f1": 0.7053104725429554, + "macro_precision": 0.6902782721361426, + "macro_recall": 0.6978744447360491, + "per_class_f1": { + "FAVOR": 0.5868945868945868, + "AGAINST": 0.6996904024767802, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 69.6223316912972, + "macro_f1_pct": 68.37636239081027, + "weighted_f1_pct": 70.53104725429554, + "macro_precision_pct": 69.02782721361426, + "macro_recall_pct": 69.7874444736049, + "per_class_f1_pct": { + "FAVOR": 58.68945868945868, + "AGAINST": 69.96904024767801, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5637068176489537, + "weighted_f1": 0.6203031301083258, + "macro_precision": 0.5920481788842785, + "macro_recall": 0.598603282702772, + "per_class_f1": { + "FAVOR": 0.5414847161572053, + "AGAINST": 0.42911877394636017, + "NONE": 0.7205169628432956 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.370681764895366, + "weighted_f1_pct": 62.030313010832586, + "macro_precision_pct": 59.204817888427854, + "macro_recall_pct": 59.86032827027719, + "per_class_f1_pct": { + "FAVOR": 54.14847161572053, + "AGAINST": 42.911877394636015, + "NONE": 72.05169628432955 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6657500074887452, + "weighted_f1": 0.6983665873170453, + "macro_precision": 0.6717438763043523, + "macro_recall": 0.6986446216331275, + "per_class_f1": { + "FAVOR": 0.5850340136054422, + "AGAINST": 0.6419753086419753, + "NONE": 0.7702407002188184 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.57500074887453, + "weighted_f1_pct": 69.83665873170453, + "macro_precision_pct": 67.17438763043523, + "macro_recall_pct": 69.86446216331275, + "per_class_f1_pct": { + "FAVOR": 58.50340136054422, + "AGAINST": 64.19753086419753, + "NONE": 77.02407002188184 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7073801658959114, + "weighted_f1": 0.7481791371490906, + "macro_precision": 0.6965198323688889, + "macro_recall": 0.7230161348757737, + "per_class_f1": { + "FAVOR": 0.6610878661087866, + "AGAINST": 0.64, + "NONE": 0.8210526315789474 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.73801658959114, + "weighted_f1_pct": 74.81791371490905, + "macro_precision_pct": 69.6519832368889, + "macro_recall_pct": 72.30161348757737, + "per_class_f1_pct": { + "FAVOR": 66.10878661087865, + "AGAINST": 64.0, + "NONE": 82.10526315789474 + } + } + ], + "accuracy": 0.6794278502313841, + "macro_f1": 0.6543796783186898, + "weighted_f1": 0.691228459546465, + "macro_precision": 0.6568693465678945, + "macro_recall": 0.680474610323541, + "per_class_f1": { + "FAVOR": 0.5842026825633383, + "AGAINST": 0.6086956521739131, + "NONE": 0.7702407002188184 + }, + "accuracy_pct": 67.94278502313841, + "macro_f1_pct": 65.43796783186899, + "weighted_f1_pct": 69.1228459546465, + "macro_precision_pct": 65.68693465678945, + "macro_recall_pct": 68.0474610323541, + "per_class_f1_pct": { + "FAVOR": 58.42026825633383, + "AGAINST": 60.86956521739131, + "NONE": 77.02407002188184 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.6796392243035284, + "losses": { + "current_ce": 0.5303162590883397, + "replay_ce": 0.14063591205411488, + "dark_distillation": 0.2745294854817567, + "total": 0.8082169146449477, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6930014867215958, + "weighted_f1": 0.7177760793026335, + "macro_precision": 0.7033397945944256, + "macro_recall": 0.7056383264282836, + "per_class_f1": { + "FAVOR": 0.6045197740112994, + "AGAINST": 0.6838709677419355, + "NONE": 0.7906137184115524 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.30014867215958, + "weighted_f1_pct": 71.77760793026336, + "macro_precision_pct": 70.33397945944256, + "macro_recall_pct": 70.56383264282836, + "per_class_f1_pct": { + "FAVOR": 60.451977401129945, + "AGAINST": 68.38709677419355, + "NONE": 79.06137184115524 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6278026905829597, + "macro_f1": 0.5788485199097199, + "weighted_f1": 0.6427858279901902, + "macro_precision": 0.5927825261158595, + "macro_recall": 0.6058306807807141, + "per_class_f1": { + "FAVOR": 0.5560747663551402, + "AGAINST": 0.4247104247104247, + "NONE": 0.7557603686635944 + }, + "accuracy_pct": 62.78026905829597, + "macro_f1_pct": 57.88485199097199, + "weighted_f1_pct": 64.27858279901902, + "macro_precision_pct": 59.27825261158595, + "macro_recall_pct": 60.58306807807141, + "per_class_f1_pct": { + "FAVOR": 55.60747663551402, + "AGAINST": 42.471042471042466, + "NONE": 75.57603686635944 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7022132796780685, + "macro_f1": 0.6764314706569566, + "weighted_f1": 0.7160911965774714, + "macro_precision": 0.6771663315824875, + "macro_recall": 0.7014313163738451, + "per_class_f1": { + "FAVOR": 0.5815602836879432, + "AGAINST": 0.6443514644351465, + "NONE": 0.8033826638477801 + }, + "accuracy_pct": 70.22132796780684, + "macro_f1_pct": 67.64314706569566, + "weighted_f1_pct": 71.60911965774714, + "macro_precision_pct": 67.71663315824875, + "macro_recall_pct": 70.14313163738451, + "per_class_f1_pct": { + "FAVOR": 58.156028368794324, + "AGAINST": 64.43514644351464, + "NONE": 80.33826638477801 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7033861443248277, + "weighted_f1": 0.7481900701852747, + "macro_precision": 0.69354395455942, + "macro_recall": 0.7257961470859762, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6126760563380281, + "NONE": 0.8308157099697885 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.33861443248277, + "weighted_f1_pct": 74.81900701852747, + "macro_precision_pct": 69.354395455942, + "macro_recall_pct": 72.57961470859762, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 61.267605633802816, + "NONE": 83.08157099697885 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.7179161372299873, + "macro_f1": 0.6543311300198124, + "weighted_f1": 0.724501712540731, + "macro_precision": 0.6472427238786292, + "macro_recall": 0.6836030970562561, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.532051282051282, + "NONE": 0.822246455834242 + }, + "accuracy_pct": 71.79161372299873, + "macro_f1_pct": 65.43311300198124, + "weighted_f1_pct": 72.4501712540731, + "macro_precision_pct": 64.72427238786292, + "macro_recall_pct": 68.36030970562561, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 53.205128205128204, + "NONE": 82.2246455834242 + } + } + ], + "accuracy": 0.6997471554993678, + "macro_f1": 0.6608186352096336, + "weighted_f1": 0.7093942826604703, + "macro_precision": 0.6600148758733049, + "macro_recall": 0.684500602785573, + "per_class_f1": { + "FAVOR": 0.5986802639472105, + "AGAINST": 0.5811965811965812, + "NONE": 0.802579060485109 + }, + "accuracy_pct": 69.97471554993679, + "macro_f1_pct": 66.08186352096335, + "weighted_f1_pct": 70.93942826604703, + "macro_precision_pct": 66.00148758733049, + "macro_recall_pct": 68.4500602785573, + "per_class_f1_pct": { + "FAVOR": 59.86802639472105, + "AGAINST": 58.119658119658126, + "NONE": 80.25790604851089 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.7102031403401267, + "losses": { + "current_ce": 0.47710517739492747, + "replay_ce": 0.13272886373220927, + "dark_distillation": 0.26405621474934005, + "total": 0.7418621473369145, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7060755336617406, + "macro_f1": 0.6908739704060062, + "weighted_f1": 0.7148688471671196, + "macro_precision": 0.6927430871918562, + "macro_recall": 0.7068080164375328, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6882352941176471, + "NONE": 0.7843866171003717 + }, + "accuracy_pct": 70.60755336617406, + "macro_f1_pct": 69.08739704060062, + "weighted_f1_pct": 71.48688471671196, + "macro_precision_pct": 69.27430871918561, + "macro_recall_pct": 70.68080164375328, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.82352941176471, + "NONE": 78.43866171003717 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5579064984450782, + "weighted_f1": 0.6201200934448771, + "macro_precision": 0.5742529488348886, + "macro_recall": 0.5881176217618589, + "per_class_f1": { + "FAVOR": 0.5401459854014599, + "AGAINST": 0.40397350993377484, + "NONE": 0.7296 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 55.790649844507826, + "weighted_f1_pct": 62.01200934448771, + "macro_precision_pct": 57.42529488348887, + "macro_recall_pct": 58.81176217618589, + "per_class_f1_pct": { + "FAVOR": 54.01459854014598, + "AGAINST": 40.397350993377486, + "NONE": 72.96000000000001 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6639839034205232, + "macro_f1": 0.6355294538811335, + "weighted_f1": 0.6797693545119485, + "macro_precision": 0.6391735048969369, + "macro_recall": 0.6638707673190432, + "per_class_f1": { + "FAVOR": 0.5171102661596958, + "AGAINST": 0.6119402985074627, + "NONE": 0.7775377969762419 + }, + "accuracy_pct": 66.39839034205231, + "macro_f1_pct": 63.55294538811334, + "weighted_f1_pct": 67.97693545119485, + "macro_precision_pct": 63.91735048969369, + "macro_recall_pct": 66.38707673190432, + "per_class_f1_pct": { + "FAVOR": 51.71102661596958, + "AGAINST": 61.19402985074627, + "NONE": 77.7537796976242 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.6807779130631797, + "weighted_f1": 0.7278891493400268, + "macro_precision": 0.6693367820260195, + "macro_recall": 0.7036369356569177, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.6081504702194357, + "NONE": 0.8112324492979719 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 68.07779130631798, + "weighted_f1_pct": 72.78891493400268, + "macro_precision_pct": 66.93367820260195, + "macro_recall_pct": 70.36369356569176, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 60.81504702194357, + "NONE": 81.1232449297972 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6645489199491741, + "macro_f1": 0.5982431302220618, + "weighted_f1": 0.6775638594257173, + "macro_precision": 0.5928028798411122, + "macro_recall": 0.6335312089669388, + "per_class_f1": { + "FAVOR": 0.5439560439560439, + "AGAINST": 0.46296296296296297, + "NONE": 0.7878103837471784 + }, + "accuracy_pct": 66.4548919949174, + "macro_f1_pct": 59.82431302220618, + "weighted_f1_pct": 67.75638594257173, + "macro_precision_pct": 59.280287984111226, + "macro_recall_pct": 63.35312089669388, + "per_class_f1_pct": { + "FAVOR": 54.395604395604394, + "AGAINST": 46.2962962962963, + "NONE": 78.78103837471784 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6347469220246238, + "macro_f1": 0.5915705381258056, + "weighted_f1": 0.6457690219601727, + "macro_precision": 0.5947171424498188, + "macro_recall": 0.6144562746086014, + "per_class_f1": { + "FAVOR": 0.4444444444444444, + "AGAINST": 0.6128440366972477, + "NONE": 0.7174231332357247 + }, + "accuracy_pct": 63.47469220246238, + "macro_f1_pct": 59.157053812580564, + "weighted_f1_pct": 64.57690219601727, + "macro_precision_pct": 59.471714244981875, + "macro_recall_pct": 61.445627460860145, + "per_class_f1_pct": { + "FAVOR": 44.44444444444444, + "AGAINST": 61.284403669724774, + "NONE": 71.74231332357247 + } + } + ], + "accuracy": 0.6623876765083441, + "macro_f1": 0.6288070997116048, + "weighted_f1": 0.6764285052056415, + "macro_precision": 0.6255367851184187, + "macro_recall": 0.6563589146391541, + "per_class_f1": { + "FAVOR": 0.5474137931034483, + "AGAINST": 0.5710200190657769, + "NONE": 0.7679874869655892 + }, + "accuracy_pct": 66.2387676508344, + "macro_f1_pct": 62.88070997116048, + "weighted_f1_pct": 67.64285052056415, + "macro_precision_pct": 62.55367851184187, + "macro_recall_pct": 65.63589146391541, + "per_class_f1_pct": { + "FAVOR": 54.741379310344826, + "AGAINST": 57.10200190657769, + "NONE": 76.79874869655892 + } + }, + { + "after_task": 7, + "best_epoch": 8, + "best_dev_macro_f1": 0.6412283431891275, + "losses": { + "current_ce": 0.4483082293044953, + "replay_ce": 0.12987337100009125, + "dark_distillation": 0.2574882027648744, + "total": 0.7069257031594004, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6978653530377669, + "macro_f1": 0.6832901548051473, + "weighted_f1": 0.7047975373868168, + "macro_precision": 0.6832971718127546, + "macro_recall": 0.6993525627940814, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6646884272997032, + "NONE": 0.7697974217311234 + }, + "accuracy_pct": 69.78653530377669, + "macro_f1_pct": 68.32901548051473, + "weighted_f1_pct": 70.47975373868168, + "macro_precision_pct": 68.32971718127546, + "macro_recall_pct": 69.93525627940814, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 66.46884272997032, + "NONE": 76.97974217311234 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5949177877428998, + "macro_f1": 0.5548206707083708, + "weighted_f1": 0.6175579794295194, + "macro_precision": 0.5696183632486281, + "macro_recall": 0.5838982124791584, + "per_class_f1": { + "FAVOR": 0.543640897755611, + "AGAINST": 0.3935483870967742, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 59.49177877428998, + "macro_f1_pct": 55.48206707083708, + "weighted_f1_pct": 61.75579794295194, + "macro_precision_pct": 56.961836324862816, + "macro_recall_pct": 58.38982124791584, + "per_class_f1_pct": { + "FAVOR": 54.3640897755611, + "AGAINST": 39.35483870967742, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6621212121212122, + "weighted_f1": 0.6985993885390264, + "macro_precision": 0.6645584673931846, + "macro_recall": 0.692476959718339, + "per_class_f1": { + "FAVOR": 0.5642857142857143, + "AGAINST": 0.6428571428571429, + "NONE": 0.7792207792207793 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 66.21212121212122, + "weighted_f1_pct": 69.85993885390263, + "macro_precision_pct": 66.45584673931846, + "macro_recall_pct": 69.2476959718339, + "per_class_f1_pct": { + "FAVOR": 56.42857142857143, + "AGAINST": 64.28571428571429, + "NONE": 77.92207792207793 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.6780403217021785, + "weighted_f1": 0.7272923095618483, + "macro_precision": 0.666627227025094, + "macro_recall": 0.7012194628742362, + "per_class_f1": { + "FAVOR": 0.616600790513834, + "AGAINST": 0.6032786885245902, + "NONE": 0.8142414860681114 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 67.80403217021785, + "weighted_f1_pct": 72.72923095618484, + "macro_precision_pct": 66.6627227025094, + "macro_recall_pct": 70.12194628742363, + "per_class_f1_pct": { + "FAVOR": 61.6600790513834, + "AGAINST": 60.32786885245902, + "NONE": 81.42414860681114 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6607369758576874, + "macro_f1": 0.5987180637877826, + "weighted_f1": 0.672786084679873, + "macro_precision": 0.6028899994816727, + "macro_recall": 0.6481791407172671, + "per_class_f1": { + "FAVOR": 0.5621890547263682, + "AGAINST": 0.45695364238410596, + "NONE": 0.7770114942528735 + }, + "accuracy_pct": 66.07369758576874, + "macro_f1_pct": 59.87180637877826, + "weighted_f1_pct": 67.2786084679873, + "macro_precision_pct": 60.288999948167266, + "macro_recall_pct": 64.81791407172672, + "per_class_f1_pct": { + "FAVOR": 56.21890547263681, + "AGAINST": 45.6953642384106, + "NONE": 77.70114942528735 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6114911080711354, + "macro_f1": 0.5723530309478079, + "weighted_f1": 0.6268051119802884, + "macro_precision": 0.5819247469423293, + "macro_recall": 0.601365315920053, + "per_class_f1": { + "FAVOR": 0.42857142857142855, + "AGAINST": 0.5865209471766849, + "NONE": 0.7019667170953101 + }, + "accuracy_pct": 61.14911080711354, + "macro_f1_pct": 57.23530309478079, + "weighted_f1_pct": 62.680511198028846, + "macro_precision_pct": 58.192474694232935, + "macro_recall_pct": 60.136531592005305, + "per_class_f1_pct": { + "FAVOR": 42.857142857142854, + "AGAINST": 58.65209471766849, + "NONE": 70.19667170953102 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7022132796780685, + "macro_f1": 0.6773714211178503, + "weighted_f1": 0.7074057162383418, + "macro_precision": 0.674701993170269, + "macro_recall": 0.6886861739477804, + "per_class_f1": { + "FAVOR": 0.6245353159851301, + "AGAINST": 0.616, + "NONE": 0.791578947368421 + }, + "accuracy_pct": 70.22132796780684, + "macro_f1_pct": 67.73714211178503, + "weighted_f1_pct": 70.74057162383419, + "macro_precision_pct": 67.4701993170269, + "macro_recall_pct": 68.86861739477804, + "per_class_f1_pct": { + "FAVOR": 62.45353159851301, + "AGAINST": 61.6, + "NONE": 79.15789473684211 + } + } + ], + "accuracy": 0.663023679417122, + "macro_f1": 0.6323772258398118, + "weighted_f1": 0.6760069255986815, + "macro_precision": 0.6298260936113816, + "macro_recall": 0.6608345076467886, + "per_class_f1": { + "FAVOR": 0.5658314350797267, + "AGAINST": 0.5665943600867679, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 66.3023679417122, + "macro_f1_pct": 63.23772258398118, + "weighted_f1_pct": 67.60069255986815, + "macro_precision_pct": 62.98260936113817, + "macro_recall_pct": 66.08345076467886, + "per_class_f1_pct": { + "FAVOR": 56.583143507972665, + "AGAINST": 56.65943600867679, + "NONE": 76.47058823529412 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7339951640488313, + "losses": { + "current_ce": 0.33631183962996414, + "replay_ce": 0.13349357755567, + "dark_distillation": 0.24566642065358354, + "total": 0.5926386286088121, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.7060755336617406, + "macro_f1": 0.6885807637083147, + "weighted_f1": 0.7122082796004581, + "macro_precision": 0.6915634273806406, + "macro_recall": 0.7025393722423728, + "per_class_f1": { + "FAVOR": 0.6239067055393586, + "AGAINST": 0.65625, + "NONE": 0.7855855855855856 + }, + "accuracy_pct": 70.60755336617406, + "macro_f1_pct": 68.85807637083147, + "weighted_f1_pct": 71.22082796004581, + "macro_precision_pct": 69.15634273806405, + "macro_recall_pct": 70.25393722423729, + "per_class_f1_pct": { + "FAVOR": 62.390670553935855, + "AGAINST": 65.625, + "NONE": 78.55855855855856 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.560765590811874, + "weighted_f1": 0.6236961875233046, + "macro_precision": 0.581924601445215, + "macro_recall": 0.5920291821912963, + "per_class_f1": { + "FAVOR": 0.535796766743649, + "AGAINST": 0.41134751773049644, + "NONE": 0.7351524879614767 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.076559081187405, + "weighted_f1_pct": 62.36961875233046, + "macro_precision_pct": 58.1924601445215, + "macro_recall_pct": 59.20291821912963, + "per_class_f1_pct": { + "FAVOR": 53.5796766743649, + "AGAINST": 41.13475177304964, + "NONE": 73.51524879614767 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6800804828973843, + "macro_f1": 0.6620625584370708, + "weighted_f1": 0.696596387834992, + "macro_precision": 0.6770893308983955, + "macro_recall": 0.6938892915904411, + "per_class_f1": { + "FAVOR": 0.5787781350482315, + "AGAINST": 0.6347826086956522, + "NONE": 0.7726269315673289 + }, + "accuracy_pct": 68.00804828973843, + "macro_f1_pct": 66.20625584370708, + "weighted_f1_pct": 69.6596387834992, + "macro_precision_pct": 67.70893308983955, + "macro_recall_pct": 69.38892915904411, + "per_class_f1_pct": { + "FAVOR": 57.87781350482315, + "AGAINST": 63.47826086956522, + "NONE": 77.2626931567329 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.707641196013289, + "macro_f1": 0.6688332240670034, + "weighted_f1": 0.7180799109608949, + "macro_precision": 0.6582557434672385, + "macro_recall": 0.6979151387313921, + "per_class_f1": { + "FAVOR": 0.6007604562737643, + "AGAINST": 0.6019417475728155, + "NONE": 0.8037974683544303 + }, + "accuracy_pct": 70.7641196013289, + "macro_f1_pct": 66.88332240670034, + "weighted_f1_pct": 71.80799109608948, + "macro_precision_pct": 65.82557434672384, + "macro_recall_pct": 69.79151387313921, + "per_class_f1_pct": { + "FAVOR": 60.07604562737643, + "AGAINST": 60.19417475728155, + "NONE": 80.37974683544303 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6709021601016518, + "macro_f1": 0.6025389275724815, + "weighted_f1": 0.6814779959474174, + "macro_precision": 0.6132854156375706, + "macro_recall": 0.6478980069394012, + "per_class_f1": { + "FAVOR": 0.5558312655086849, + "AGAINST": 0.45993031358885017, + "NONE": 0.7918552036199095 + }, + "accuracy_pct": 67.09021601016518, + "macro_f1_pct": 60.25389275724815, + "weighted_f1_pct": 68.14779959474174, + "macro_precision_pct": 61.32854156375706, + "macro_recall_pct": 64.78980069394012, + "per_class_f1_pct": { + "FAVOR": 55.58312655086849, + "AGAINST": 45.99303135888502, + "NONE": 79.18552036199095 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5950752393980848, + "macro_f1": 0.5628969860548808, + "weighted_f1": 0.6101661182314937, + "macro_precision": 0.573739219548047, + "macro_recall": 0.5990571939684762, + "per_class_f1": { + "FAVOR": 0.43609022556390975, + "AGAINST": 0.5787545787545788, + "NONE": 0.6738461538461539 + }, + "accuracy_pct": 59.50752393980848, + "macro_f1_pct": 56.289698605488084, + "weighted_f1_pct": 61.01661182314937, + "macro_precision_pct": 57.37392195480469, + "macro_recall_pct": 59.90571939684762, + "per_class_f1_pct": { + "FAVOR": 43.609022556390975, + "AGAINST": 57.87545787545788, + "NONE": 67.38461538461539 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6629092304992381, + "weighted_f1": 0.6893054865972921, + "macro_precision": 0.6593890151254332, + "macro_recall": 0.6789568138131145, + "per_class_f1": { + "FAVOR": 0.6044776119402985, + "AGAINST": 0.6217228464419475, + "NONE": 0.7625272331154684 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 66.2909230499238, + "weighted_f1_pct": 68.93054865972921, + "macro_precision_pct": 65.93890151254332, + "macro_recall_pct": 67.89568138131145, + "per_class_f1_pct": { + "FAVOR": 60.447761194029844, + "AGAINST": 62.172284644194754, + "NONE": 76.25272331154684 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.7140864714086471, + "macro_f1": 0.6815813864386988, + "weighted_f1": 0.7272952637949974, + "macro_precision": 0.6706977605820508, + "macro_recall": 0.7273293080796296, + "per_class_f1": { + "FAVOR": 0.5796178343949044, + "AGAINST": 0.6737400530503979, + "NONE": 0.7913862718707941 + }, + "accuracy_pct": 71.40864714086472, + "macro_f1_pct": 68.15813864386988, + "weighted_f1_pct": 72.72952637949975, + "macro_precision_pct": 67.06977605820508, + "macro_recall_pct": 72.73293080796296, + "per_class_f1_pct": { + "FAVOR": 57.961783439490446, + "AGAINST": 67.37400530503979, + "NONE": 79.13862718707941 + } + } + ], + "accuracy": 0.6676453317674692, + "macro_f1": 0.6376797198359435, + "weighted_f1": 0.6813711179083157, + "macro_precision": 0.6366896058993522, + "macro_recall": 0.6704577942829801, + "per_class_f1": { + "FAVOR": 0.564398308342945, + "AGAINST": 0.5828877005347594, + "NONE": 0.7657531506301261 + }, + "accuracy_pct": 66.76453317674692, + "macro_f1_pct": 63.76797198359435, + "weighted_f1_pct": 68.13711179083157, + "macro_precision_pct": 63.66896058993522, + "macro_recall_pct": 67.04577942829802, + "per_class_f1_pct": { + "FAVOR": 56.439830834294504, + "AGAINST": 58.288770053475936, + "NONE": 76.57531506301261 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s2__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__doc__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..70c3cad76dd59d5187a09393e8975ef6eee230a8 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__doc__s2__seed13/run.log b/outputs/trace_s6_derpp__doc__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..fd7010b8b42390239f60fbdf10f2de9fec238e80 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 04:41:29,981 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__doc__s2__seed13 device=cuda +2026-09-02 05:11:32,934 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6362710834486953, 'average_forgetting': 0.03125746231015737, 'backward_transfer': -0.030031872452375326, 'final_mean_target_macro_f1_pct': 63.627108344869534, 'average_forgetting_pct': 3.125746231015737, 'backward_transfer_pct': -3.0031872452375326, 'final_pooled_accuracy': 0.6676453317674692, 'final_pooled_accuracy_pct': 66.76453317674692, 'final_pooled_macro_f1': 0.6376797198359435, 'final_pooled_macro_f1_pct': 63.76797198359435, 'final_pooled_weighted_f1': 0.6813711179083157, 'final_pooled_weighted_f1_pct': 68.13711179083157, 'final_pooled_macro_precision': 0.6366896058993522, 'final_pooled_macro_precision_pct': 63.66896058993522, 'final_pooled_macro_recall': 0.6704577942829801, 'final_pooled_macro_recall_pct': 67.04577942829802, 'final_pooled_per_class_f1': {'FAVOR': 0.564398308342945, 'AGAINST': 0.5828877005347594, 'NONE': 0.7657531506301261}, 'final_pooled_per_class_f1_pct': {'FAVOR': 56.439830834294504, 'AGAINST': 58.288770053475936, 'NONE': 76.57531506301261}} +2026-09-02 05:11:32,935 [INFO] __main__: efficiency={'training_wall_time_seconds': 1801.5697203811724, 'training_wall_time_minutes': 30.026162006352873, 'peak_gpu_memory_allocated_gb': 10.397007942199707, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s6_derpp__doc__s3__seed13/environment.json b/outputs/trace_s6_derpp__doc__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..d742e104364fc1d693266c1494eff51fa87fbebe --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T05:11:41.059698+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s3__seed13/metrics.json b/outputs/trace_s6_derpp__doc__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..985f0dbc8f5656414d1f4f2c1ddf3ddc278445b5 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s3__seed13/metrics.json @@ -0,0 +1,1594 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5374945250275344, + 0.5463075011612847, + null, + null, + null, + null, + null, + null + ], + [ + 0.5232522834586256, + 0.5335155435230371, + 0.6543247691104762, + null, + null, + null, + null, + null + ], + [ + 0.5087447364444003, + 0.4977928411587664, + 0.6470368747333941, + 0.6473990654259899, + null, + null, + null, + null + ], + [ + 0.52195274359378, + 0.5231365867193868, + 0.6623767078987792, + 0.6434524888983026, + 0.6970303442493787, + null, + null, + null + ], + [ + 0.5504146207992564, + 0.5314584102938406, + 0.6417148799712963, + 0.6463568595418386, + 0.6964409552570192, + 0.5964318804902767, + null, + null + ], + [ + 0.5321040216218281, + 0.5243885092860165, + 0.6716086760318358, + 0.6345308981292561, + 0.694603676067382, + 0.5825529293536768, + 0.7015197925693059, + null + ], + [ + 0.5428242154468658, + 0.5461933697640239, + 0.6711225503385233, + 0.6574413511480081, + 0.6775454464638669, + 0.5922058089403222, + 0.7026402070750647, + 0.7291183105918587 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 8, + "best_dev_macro_f1": 0.6598148411148085, + "losses": { + "current_ce": 0.6017990967997956, + "replay_ce": 0.29593221313719237, + "dark_distillation": 0.39678630508364193, + "total": 1.0961244651249478, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5374945250275344, + "weighted_f1": 0.6260728523998539, + "macro_precision": 0.5582332112961722, + "macro_recall": 0.5761467250574004, + "per_class_f1": { + "FAVOR": 0.37735849056603776, + "AGAINST": 0.4949290060851927, + "NONE": 0.7401960784313726 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.749452502753435, + "weighted_f1_pct": 62.60728523998539, + "macro_precision_pct": 55.82332112961722, + "macro_recall_pct": 57.61467250574004, + "per_class_f1_pct": { + "FAVOR": 37.735849056603776, + "AGAINST": 49.49290060851927, + "NONE": 74.01960784313727 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.5463075011612847, + "weighted_f1": 0.5943410243796432, + "macro_precision": 0.5625785385642903, + "macro_recall": 0.5774484043650431, + "per_class_f1": { + "FAVOR": 0.4112903225806452, + "AGAINST": 0.5734265734265734, + "NONE": 0.6542056074766355 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.63075011612847, + "weighted_f1_pct": 59.43410243796432, + "macro_precision_pct": 56.25785385642903, + "macro_recall_pct": 57.744840436504305, + "per_class_f1_pct": { + "FAVOR": 41.12903225806452, + "AGAINST": 57.34265734265735, + "NONE": 65.42056074766354 + } + } + ], + "accuracy": 0.5922266139657444, + "macro_f1": 0.5443944489016742, + "weighted_f1": 0.6105338373269887, + "macro_precision": 0.5618819133832902, + "macro_recall": 0.5744986770475305, + "per_class_f1": { + "FAVOR": 0.3937621832358674, + "AGAINST": 0.5370892018779343, + "NONE": 0.7023319615912208 + }, + "accuracy_pct": 59.22266139657444, + "macro_f1_pct": 54.43944489016742, + "weighted_f1_pct": 61.053383732698876, + "macro_precision_pct": 56.18819133832902, + "macro_recall_pct": 57.44986770475305, + "per_class_f1_pct": { + "FAVOR": 39.37621832358674, + "AGAINST": 53.70892018779343, + "NONE": 70.23319615912207 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.6720962418822527, + "losses": { + "current_ce": 0.5767768492656095, + "replay_ce": 0.24032061337715105, + "dark_distillation": 0.3301475648014318, + "total": 0.9821712439968473, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5895806861499364, + "macro_f1": 0.5232522834586256, + "weighted_f1": 0.617346102682281, + "macro_precision": 0.5292465014566562, + "macro_recall": 0.5513429501991594, + "per_class_f1": { + "FAVOR": 0.38006230529595014, + "AGAINST": 0.4486873508353222, + "NONE": 0.7410071942446043 + }, + "accuracy_pct": 58.95806861499364, + "macro_f1_pct": 52.32522834586256, + "weighted_f1_pct": 61.734610268228096, + "macro_precision_pct": 52.924650145665616, + "macro_recall_pct": 55.13429501991594, + "per_class_f1_pct": { + "FAVOR": 38.006230529595015, + "AGAINST": 44.868735083532215, + "NONE": 74.10071942446042 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5704514363885089, + "macro_f1": 0.5335155435230371, + "weighted_f1": 0.5868113170448906, + "macro_precision": 0.5576118831905328, + "macro_recall": 0.5641403151028193, + "per_class_f1": { + "FAVOR": 0.38735177865612647, + "AGAINST": 0.5570934256055363, + "NONE": 0.6561014263074485 + }, + "accuracy_pct": 57.045143638850895, + "macro_f1_pct": 53.35155435230371, + "weighted_f1_pct": 58.68113170448906, + "macro_precision_pct": 55.761188319053275, + "macro_recall_pct": 56.41403151028192, + "per_class_f1_pct": { + "FAVOR": 38.73517786561265, + "AGAINST": 55.70934256055363, + "NONE": 65.61014263074485 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6543247691104762, + "weighted_f1": 0.6928119083435389, + "macro_precision": 0.6519363341443634, + "macro_recall": 0.6578122878786107, + "per_class_f1": { + "FAVOR": 0.5630252100840336, + "AGAINST": 0.6007604562737643, + "NONE": 0.7991886409736308 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 65.43247691104762, + "weighted_f1_pct": 69.2811908343539, + "macro_precision_pct": 65.19363341443633, + "macro_recall_pct": 65.78122878786107, + "per_class_f1_pct": { + "FAVOR": 56.30252100840336, + "AGAINST": 60.07604562737643, + "NONE": 79.91886409736308 + } + } + ], + "accuracy": 0.6074441687344914, + "macro_f1": 0.5648044996747051, + "weighted_f1": 0.6250743616048388, + "macro_precision": 0.5695140800996698, + "macro_recall": 0.5875384929622324, + "per_class_f1": { + "FAVOR": 0.43596059113300495, + "AGAINST": 0.5301587301587302, + "NONE": 0.72829417773238 + }, + "accuracy_pct": 60.74441687344914, + "macro_f1_pct": 56.48044996747051, + "weighted_f1_pct": 62.50743616048388, + "macro_precision_pct": 56.951408009966975, + "macro_recall_pct": 58.75384929622324, + "per_class_f1_pct": { + "FAVOR": 43.5960591133005, + "AGAINST": 53.01587301587302, + "NONE": 72.829417773238 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7330493484446611, + "losses": { + "current_ce": 0.42256592037231944, + "replay_ce": 0.24960727357282872, + "dark_distillation": 0.32572553390409886, + "total": 0.8350359624478875, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.5087447364444003, + "weighted_f1": 0.6035631935656818, + "macro_precision": 0.5311162440200322, + "macro_recall": 0.5420255569493042, + "per_class_f1": { + "FAVOR": 0.3445692883895131, + "AGAINST": 0.45528455284552843, + "NONE": 0.7263803680981595 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 50.87447364444003, + "weighted_f1_pct": 60.35631935656818, + "macro_precision_pct": 53.11162440200322, + "macro_recall_pct": 54.202555694930425, + "per_class_f1_pct": { + "FAVOR": 34.45692883895131, + "AGAINST": 45.52845528455284, + "NONE": 72.63803680981596 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5266757865937073, + "macro_f1": 0.4977928411587664, + "weighted_f1": 0.5400641190222629, + "macro_precision": 0.5420096793143089, + "macro_recall": 0.5330103287083535, + "per_class_f1": { + "FAVOR": 0.37751004016064255, + "AGAINST": 0.5242718446601942, + "NONE": 0.5915966386554622 + }, + "accuracy_pct": 52.667578659370726, + "macro_f1_pct": 49.779284115876635, + "weighted_f1_pct": 54.00641190222629, + "macro_precision_pct": 54.20096793143089, + "macro_recall_pct": 53.30103287083535, + "per_class_f1_pct": { + "FAVOR": 37.75100401606426, + "AGAINST": 52.42718446601942, + "NONE": 59.15966386554622 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6470368747333941, + "weighted_f1": 0.6882543398622264, + "macro_precision": 0.6485199308728721, + "macro_recall": 0.6556293538607467, + "per_class_f1": { + "FAVOR": 0.5414847161572053, + "AGAINST": 0.5979381443298969, + "NONE": 0.8016877637130801 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 64.70368747333941, + "weighted_f1_pct": 68.82543398622263, + "macro_precision_pct": 64.85199308728721, + "macro_recall_pct": 65.56293538607467, + "per_class_f1_pct": { + "FAVOR": 54.14847161572053, + "AGAINST": 59.79381443298969, + "NONE": 80.16877637130801 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6959553695955369, + "macro_f1": 0.6473990654259899, + "weighted_f1": 0.7083102424122704, + "macro_precision": 0.6402361320796545, + "macro_recall": 0.6793283385457837, + "per_class_f1": { + "FAVOR": 0.5055762081784386, + "AGAINST": 0.6440677966101694, + "NONE": 0.7925531914893617 + }, + "accuracy_pct": 69.5955369595537, + "macro_f1_pct": 64.73990654259899, + "weighted_f1_pct": 70.83102424122704, + "macro_precision_pct": 64.02361320796544, + "macro_recall_pct": 67.93283385457836, + "per_class_f1_pct": { + "FAVOR": 50.55762081784386, + "AGAINST": 64.40677966101694, + "NONE": 79.25531914893617 + } + } + ], + "accuracy": 0.6138360175695461, + "macro_f1": 0.570957079351705, + "weighted_f1": 0.6311171624377911, + "macro_precision": 0.5830862536616448, + "macro_recall": 0.5977160478674138, + "per_class_f1": { + "FAVOR": 0.43984220907297833, + "AGAINST": 0.5446527012127894, + "NONE": 0.7283763277693475 + }, + "accuracy_pct": 61.38360175695461, + "macro_f1_pct": 57.0957079351705, + "weighted_f1_pct": 63.111716243779114, + "macro_precision_pct": 58.30862536616448, + "macro_recall_pct": 59.771604786741385, + "per_class_f1_pct": { + "FAVOR": 43.98422090729783, + "AGAINST": 54.46527012127894, + "NONE": 72.83763277693474 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.7418596503125036, + "losses": { + "current_ce": 0.4637480529717037, + "replay_ce": 0.22047008640709378, + "dark_distillation": 0.32636636401925767, + "total": 0.8474013226940519, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5844980940279543, + "macro_f1": 0.52195274359378, + "weighted_f1": 0.6162103840273592, + "macro_precision": 0.5359205428660122, + "macro_recall": 0.5526024588660754, + "per_class_f1": { + "FAVOR": 0.38977635782747605, + "AGAINST": 0.4349775784753363, + "NONE": 0.7411042944785277 + }, + "accuracy_pct": 58.449809402795424, + "macro_f1_pct": 52.195274359378, + "weighted_f1_pct": 61.62103840273592, + "macro_precision_pct": 53.59205428660122, + "macro_recall_pct": 55.26024588660754, + "per_class_f1_pct": { + "FAVOR": 38.977635782747605, + "AGAINST": 43.49775784753363, + "NONE": 74.11042944785277 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5231365867193868, + "weighted_f1": 0.5658044656042314, + "macro_precision": 0.5583250285928314, + "macro_recall": 0.5599019713934344, + "per_class_f1": { + "FAVOR": 0.4108527131782946, + "AGAINST": 0.5335570469798657, + "NONE": 0.625 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 52.313658671938676, + "weighted_f1_pct": 56.580446560423134, + "macro_precision_pct": 55.83250285928314, + "macro_recall_pct": 55.990197139343444, + "per_class_f1_pct": { + "FAVOR": 41.08527131782946, + "AGAINST": 53.355704697986575, + "NONE": 62.5 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6623767078987792, + "weighted_f1": 0.7020537913388651, + "macro_precision": 0.6590542350454245, + "macro_recall": 0.6693015355506144, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.5939849624060151, + "NONE": 0.8125 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 66.23767078987792, + "weighted_f1_pct": 70.20537913388651, + "macro_precision_pct": 65.90542350454245, + "macro_recall_pct": 66.93015355506145, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 59.3984962406015, + "NONE": 81.25 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6694560669456067, + "macro_f1": 0.6434524888983026, + "weighted_f1": 0.6836444678613905, + "macro_precision": 0.6422760832661747, + "macro_recall": 0.7009640018353612, + "per_class_f1": { + "FAVOR": 0.5483870967741935, + "AGAINST": 0.6430260047281324, + "NONE": 0.738944365192582 + }, + "accuracy_pct": 66.94560669456067, + "macro_f1_pct": 64.34524888983026, + "weighted_f1_pct": 68.36444678613906, + "macro_precision_pct": 64.22760832661747, + "macro_recall_pct": 70.09640018353612, + "per_class_f1_pct": { + "FAVOR": 54.83870967741935, + "AGAINST": 64.30260047281324, + "NONE": 73.89443651925819 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6970303442493787, + "weighted_f1": 0.715896851540361, + "macro_precision": 0.6961901536362936, + "macro_recall": 0.7175100958781302, + "per_class_f1": { + "FAVOR": 0.6130030959752322, + "AGAINST": 0.7100271002710027, + "NONE": 0.7680608365019012 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.70303442493787, + "weighted_f1_pct": 71.5896851540361, + "macro_precision_pct": 69.61901536362936, + "macro_recall_pct": 71.75100958781302, + "per_class_f1_pct": { + "FAVOR": 61.30030959752322, + "AGAINST": 71.00271002710026, + "NONE": 76.80608365019012 + } + } + ], + "accuracy": 0.634540556719545, + "macro_f1": 0.6055071774967216, + "weighted_f1": 0.6505926399645455, + "macro_precision": 0.6113435294931517, + "macro_recall": 0.6373525835247188, + "per_class_f1": { + "FAVOR": 0.509641873278237, + "AGAINST": 0.5733333333333334, + "NONE": 0.7335463258785943 + }, + "accuracy_pct": 63.4540556719545, + "macro_f1_pct": 60.55071774967216, + "weighted_f1_pct": 65.05926399645455, + "macro_precision_pct": 61.13435294931517, + "macro_recall_pct": 63.73525835247188, + "per_class_f1_pct": { + "FAVOR": 50.964187327823694, + "AGAINST": 57.333333333333336, + "NONE": 73.35463258785943 + } + }, + { + "after_task": 6, + "best_epoch": 9, + "best_dev_macro_f1": 0.6871266882822313, + "losses": { + "current_ce": 0.501748810996089, + "replay_ce": 0.20201711892558818, + "dark_distillation": 0.31889288313033287, + "total": 0.8632123711339215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6124523506988564, + "macro_f1": 0.5504146207992564, + "weighted_f1": 0.6402539684402383, + "macro_precision": 0.5565610859728506, + "macro_recall": 0.5845231545667277, + "per_class_f1": { + "FAVOR": 0.4086687306501548, + "AGAINST": 0.48470588235294115, + "NONE": 0.7578692493946732 + }, + "accuracy_pct": 61.24523506988564, + "macro_f1_pct": 55.04146207992564, + "weighted_f1_pct": 64.02539684402383, + "macro_precision_pct": 55.65610859728506, + "macro_recall_pct": 58.452315456672764, + "per_class_f1_pct": { + "FAVOR": 40.86687306501548, + "AGAINST": 48.470588235294116, + "NONE": 75.78692493946731 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.560875512995896, + "macro_f1": 0.5314584102938406, + "weighted_f1": 0.5815524288238701, + "macro_precision": 0.5522443612861265, + "macro_recall": 0.5684929306455921, + "per_class_f1": { + "FAVOR": 0.4128113879003559, + "AGAINST": 0.5201465201465202, + "NONE": 0.6614173228346457 + }, + "accuracy_pct": 56.0875512995896, + "macro_f1_pct": 53.14584102938406, + "weighted_f1_pct": 58.155242882387014, + "macro_precision_pct": 55.22443612861265, + "macro_recall_pct": 56.84929306455921, + "per_class_f1_pct": { + "FAVOR": 41.281138790035584, + "AGAINST": 52.01465201465202, + "NONE": 66.14173228346458 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6417148799712963, + "weighted_f1": 0.691986525090873, + "macro_precision": 0.6428967402063139, + "macro_recall": 0.6442462665749328, + "per_class_f1": { + "FAVOR": 0.5178571428571429, + "AGAINST": 0.5766423357664233, + "NONE": 0.8306451612903226 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 64.17148799712963, + "weighted_f1_pct": 69.1986525090873, + "macro_precision_pct": 64.2896740206314, + "macro_recall_pct": 64.42462665749328, + "per_class_f1_pct": { + "FAVOR": 51.78571428571429, + "AGAINST": 57.66423357664233, + "NONE": 83.06451612903226 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6778242677824268, + "macro_f1": 0.6463568595418386, + "weighted_f1": 0.693531099516167, + "macro_precision": 0.6415967338836941, + "macro_recall": 0.696375749233206, + "per_class_f1": { + "FAVOR": 0.5396825396825397, + "AGAINST": 0.64, + "NONE": 0.7593880389429764 + }, + "accuracy_pct": 67.78242677824268, + "macro_f1_pct": 64.63568595418387, + "weighted_f1_pct": 69.3531099516167, + "macro_precision_pct": 64.15967338836941, + "macro_recall_pct": 69.6375749233206, + "per_class_f1_pct": { + "FAVOR": 53.96825396825397, + "AGAINST": 64.0, + "NONE": 75.93880389429764 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.6964409552570192, + "weighted_f1": 0.7109185138427025, + "macro_precision": 0.7007600686915977, + "macro_recall": 0.724026339381514, + "per_class_f1": { + "FAVOR": 0.6445783132530121, + "AGAINST": 0.6912928759894459, + "NONE": 0.7534516765285996 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.64409552570191, + "weighted_f1_pct": 71.09185138427026, + "macro_precision_pct": 70.07600686915977, + "macro_recall_pct": 72.4026339381514, + "per_class_f1_pct": { + "FAVOR": 64.45783132530121, + "AGAINST": 69.12928759894459, + "NONE": 75.34516765285996 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6307922272047832, + "macro_f1": 0.5964318804902767, + "weighted_f1": 0.650760169841083, + "macro_precision": 0.6045797635434466, + "macro_recall": 0.6290321443552623, + "per_class_f1": { + "FAVOR": 0.5684754521963824, + "AGAINST": 0.47318611987381703, + "NONE": 0.7476340694006309 + }, + "accuracy_pct": 63.07922272047832, + "macro_f1_pct": 59.64318804902767, + "weighted_f1_pct": 65.07601698410829, + "macro_precision_pct": 60.45797635434466, + "macro_recall_pct": 62.90321443552623, + "per_class_f1_pct": { + "FAVOR": 56.84754521963824, + "AGAINST": 47.3186119873817, + "NONE": 74.76340694006309 + } + } + ], + "accuracy": 0.6413965087281795, + "macro_f1": 0.6104345237807094, + "weighted_f1": 0.6582819955854231, + "macro_precision": 0.6134178127890829, + "macro_recall": 0.6407169291242313, + "per_class_f1": { + "FAVOR": 0.5198711063372717, + "AGAINST": 0.5621529260999573, + "NONE": 0.7492795389048992 + }, + "accuracy_pct": 64.13965087281795, + "macro_f1_pct": 61.04345237807094, + "weighted_f1_pct": 65.82819955854231, + "macro_precision_pct": 61.34178127890829, + "macro_recall_pct": 64.07169291242313, + "per_class_f1_pct": { + "FAVOR": 51.98711063372718, + "AGAINST": 56.21529260999573, + "NONE": 74.92795389048992 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.7145235097573966, + "losses": { + "current_ce": 0.4077385595334428, + "replay_ce": 0.19610688175473895, + "dark_distillation": 0.29602664301083204, + "total": 0.7518587647449403, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6060991105463787, + "macro_f1": 0.5321040216218281, + "weighted_f1": 0.6299955720448114, + "macro_precision": 0.5483200850355855, + "macro_recall": 0.5626651408786486, + "per_class_f1": { + "FAVOR": 0.36363636363636365, + "AGAINST": 0.47578947368421054, + "NONE": 0.7568862275449102 + }, + "accuracy_pct": 60.60991105463786, + "macro_f1_pct": 53.21040216218281, + "weighted_f1_pct": 62.999557204481135, + "macro_precision_pct": 54.83200850355855, + "macro_recall_pct": 56.26651408786486, + "per_class_f1_pct": { + "FAVOR": 36.36363636363637, + "AGAINST": 47.578947368421055, + "NONE": 75.68862275449102 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5554035567715458, + "macro_f1": 0.5243885092860165, + "weighted_f1": 0.5699230461359269, + "macro_precision": 0.5607555614058181, + "macro_recall": 0.5572611789133383, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5436573311367381, + "NONE": 0.6295081967213115 + }, + "accuracy_pct": 55.54035567715459, + "macro_f1_pct": 52.43885092860165, + "weighted_f1_pct": 56.992304613592694, + "macro_precision_pct": 56.0755561405818, + "macro_recall_pct": 55.72611789133383, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 54.36573311367381, + "NONE": 62.95081967213115 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.710261569416499, + "macro_f1": 0.6716086760318358, + "weighted_f1": 0.7127397731485023, + "macro_precision": 0.6789344718370142, + "macro_recall": 0.6765512797200343, + "per_class_f1": { + "FAVOR": 0.5700934579439252, + "AGAINST": 0.6185567010309279, + "NONE": 0.8261758691206544 + }, + "accuracy_pct": 71.0261569416499, + "macro_f1_pct": 67.16086760318359, + "weighted_f1_pct": 71.27397731485023, + "macro_precision_pct": 67.89344718370141, + "macro_recall_pct": 67.65512797200343, + "per_class_f1_pct": { + "FAVOR": 57.009345794392516, + "AGAINST": 61.855670103092784, + "NONE": 82.61758691206545 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6750348675034867, + "macro_f1": 0.6345308981292561, + "weighted_f1": 0.6913888335630604, + "macro_precision": 0.6300536947396256, + "macro_recall": 0.6753426530377219, + "per_class_f1": { + "FAVOR": 0.5067567567567568, + "AGAINST": 0.6259168704156479, + "NONE": 0.7709190672153635 + }, + "accuracy_pct": 67.50348675034867, + "macro_f1_pct": 63.45308981292561, + "weighted_f1_pct": 69.13888335630604, + "macro_precision_pct": 63.00536947396256, + "macro_recall_pct": 67.53426530377219, + "per_class_f1_pct": { + "FAVOR": 50.67567567567568, + "AGAINST": 62.59168704156479, + "NONE": 77.09190672153635 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.694603676067382, + "weighted_f1": 0.7082629692676564, + "macro_precision": 0.6989239042998294, + "macro_recall": 0.7166956589032524, + "per_class_f1": { + "FAVOR": 0.6416382252559727, + "AGAINST": 0.6945812807881774, + "NONE": 0.7475915221579962 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.4603676067382, + "weighted_f1_pct": 70.82629692676564, + "macro_precision_pct": 69.89239042998294, + "macro_recall_pct": 71.66956589032523, + "per_class_f1_pct": { + "FAVOR": 64.16382252559727, + "AGAINST": 69.45812807881774, + "NONE": 74.75915221579962 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6143497757847534, + "macro_f1": 0.5825529293536768, + "weighted_f1": 0.6367423947672474, + "macro_precision": 0.5972526751747531, + "macro_recall": 0.6177075365394264, + "per_class_f1": { + "FAVOR": 0.5655526992287918, + "AGAINST": 0.44984802431610943, + "NONE": 0.7322580645161291 + }, + "accuracy_pct": 61.43497757847533, + "macro_f1_pct": 58.25529293536768, + "weighted_f1_pct": 63.674239476724736, + "macro_precision_pct": 59.72526751747531, + "macro_recall_pct": 61.77075365394265, + "per_class_f1_pct": { + "FAVOR": 56.555269922879184, + "AGAINST": 44.984802431610944, + "NONE": 73.22580645161291 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.7015197925693059, + "weighted_f1": 0.7329989063210653, + "macro_precision": 0.6979025260929879, + "macro_recall": 0.7291751521636579, + "per_class_f1": { + "FAVOR": 0.6187050359712231, + "AGAINST": 0.6833333333333333, + "NONE": 0.8025210084033614 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 70.15197925693059, + "weighted_f1_pct": 73.29989063210654, + "macro_precision_pct": 69.79025260929879, + "macro_recall_pct": 72.91751521636579, + "per_class_f1_pct": { + "FAVOR": 61.87050359712231, + "AGAINST": 68.33333333333333, + "NONE": 80.25210084033614 + } + } + ], + "accuracy": 0.6476591968049701, + "macro_f1": 0.6177427847929325, + "weighted_f1": 0.6628197068125781, + "macro_precision": 0.6224716082855182, + "macro_recall": 0.6470068460168169, + "per_class_f1": { + "FAVOR": 0.5285497726124305, + "AGAINST": 0.5752629669931084, + "NONE": 0.7494156147732586 + }, + "accuracy_pct": 64.765919680497, + "macro_f1_pct": 61.77427847929325, + "weighted_f1_pct": 66.2819706812578, + "macro_precision_pct": 62.24716082855182, + "macro_recall_pct": 64.70068460168169, + "per_class_f1_pct": { + "FAVOR": 52.85497726124305, + "AGAINST": 57.52629669931084, + "NONE": 74.94156147732586 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.7778474641889276, + "losses": { + "current_ce": 0.337591310269108, + "replay_ce": 0.18207696132800158, + "dark_distillation": 0.2913791876505403, + "total": 0.6653578645458409, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6226175349428208, + "macro_f1": 0.5428242154468658, + "weighted_f1": 0.645728462985635, + "macro_precision": 0.5503162699796812, + "macro_recall": 0.5688708068337698, + "per_class_f1": { + "FAVOR": 0.36823104693140796, + "AGAINST": 0.48089887640449436, + "NONE": 0.7793427230046949 + }, + "accuracy_pct": 62.26175349428208, + "macro_f1_pct": 54.28242154468658, + "weighted_f1_pct": 64.5728462985635, + "macro_precision_pct": 55.03162699796812, + "macro_recall_pct": 56.88708068337698, + "per_class_f1_pct": { + "FAVOR": 36.8231046931408, + "AGAINST": 48.08988764044943, + "NONE": 77.93427230046949 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5759233926128591, + "macro_f1": 0.5461933697640239, + "weighted_f1": 0.588421862814057, + "macro_precision": 0.5728881713913531, + "macro_recall": 0.5788560191807598, + "per_class_f1": { + "FAVOR": 0.4315352697095436, + "AGAINST": 0.5628140703517588, + "NONE": 0.6442307692307693 + }, + "accuracy_pct": 57.59233926128591, + "macro_f1_pct": 54.61933697640239, + "weighted_f1_pct": 58.842186281405695, + "macro_precision_pct": 57.288817139135304, + "macro_recall_pct": 57.88560191807598, + "per_class_f1_pct": { + "FAVOR": 43.15352697095436, + "AGAINST": 56.28140703517588, + "NONE": 64.42307692307693 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6711225503385233, + "weighted_f1": 0.7143077591054621, + "macro_precision": 0.6766716781214205, + "macro_recall": 0.6710286356564912, + "per_class_f1": { + "FAVOR": 0.5581395348837209, + "AGAINST": 0.6222222222222222, + "NONE": 0.8330058939096268 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 67.11225503385234, + "weighted_f1_pct": 71.4307759105462, + "macro_precision_pct": 67.66716781214205, + "macro_recall_pct": 67.10286356564912, + "per_class_f1_pct": { + "FAVOR": 55.81395348837209, + "AGAINST": 62.22222222222222, + "NONE": 83.30058939096267 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6959553695955369, + "macro_f1": 0.6574413511480081, + "weighted_f1": 0.7088029794365435, + "macro_precision": 0.6524016760393453, + "macro_recall": 0.7003100337708897, + "per_class_f1": { + "FAVOR": 0.5342960288808665, + "AGAINST": 0.6588785046728972, + "NONE": 0.7791495198902606 + }, + "accuracy_pct": 69.5955369595537, + "macro_f1_pct": 65.74413511480081, + "weighted_f1_pct": 70.88029794365434, + "macro_precision_pct": 65.24016760393452, + "macro_recall_pct": 70.03100337708898, + "per_class_f1_pct": { + "FAVOR": 53.42960288808665, + "AGAINST": 65.88785046728972, + "NONE": 77.91495198902606 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6880131362889984, + "macro_f1": 0.6775454464638669, + "weighted_f1": 0.692112365014755, + "macro_precision": 0.6848455598455598, + "macro_recall": 0.6997477245487044, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6779661016949152, + "NONE": 0.7339805825242719 + }, + "accuracy_pct": 68.80131362889983, + "macro_f1_pct": 67.75454464638669, + "weighted_f1_pct": 69.2112365014755, + "macro_precision_pct": 68.48455598455598, + "macro_recall_pct": 69.97477245487043, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 67.79661016949152, + "NONE": 73.39805825242719 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6218236173393124, + "macro_f1": 0.5922058089403222, + "weighted_f1": 0.6433191365957746, + "macro_precision": 0.604110460369961, + "macro_recall": 0.6281796850306732, + "per_class_f1": { + "FAVOR": 0.5706666666666667, + "AGAINST": 0.471976401179941, + "NONE": 0.7339743589743589 + }, + "accuracy_pct": 62.18236173393124, + "macro_f1_pct": 59.22058089403222, + "weighted_f1_pct": 64.33191365957745, + "macro_precision_pct": 60.4110460369961, + "macro_recall_pct": 62.81796850306732, + "per_class_f1_pct": { + "FAVOR": 57.06666666666666, + "AGAINST": 47.1976401179941, + "NONE": 73.3974358974359 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.7026402070750647, + "weighted_f1": 0.7318102922973562, + "macro_precision": 0.697317230797503, + "macro_recall": 0.7324531427979704, + "per_class_f1": { + "FAVOR": 0.64, + "AGAINST": 0.6721311475409836, + "NONE": 0.7957894736842105 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 70.26402070750648, + "weighted_f1_pct": 73.18102922973561, + "macro_precision_pct": 69.7317230797503, + "macro_recall_pct": 73.24531427979704, + "per_class_f1_pct": { + "FAVOR": 64.0, + "AGAINST": 67.21311475409836, + "NONE": 79.57894736842105 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7591362126245847, + "macro_f1": 0.7291183105918587, + "weighted_f1": 0.7650915804968302, + "macro_precision": 0.7190885724025161, + "macro_recall": 0.747531550769343, + "per_class_f1": { + "FAVOR": 0.6960352422907489, + "AGAINST": 0.6605504587155964, + "NONE": 0.8307692307692308 + }, + "accuracy_pct": 75.91362126245848, + "macro_f1_pct": 72.91183105918587, + "weighted_f1_pct": 76.50915804968302, + "macro_precision_pct": 71.9088572402516, + "macro_recall_pct": 74.75315507693429, + "per_class_f1_pct": { + "FAVOR": 69.60352422907489, + "AGAINST": 66.05504587155964, + "NONE": 83.07692307692308 + } + } + ], + "accuracy": 0.6688197298884322, + "macro_f1": 0.6373646853691143, + "weighted_f1": 0.6817398364271444, + "macro_precision": 0.6372624859385251, + "macro_recall": 0.6645037793355199, + "per_class_f1": { + "FAVOR": 0.5521359669269638, + "AGAINST": 0.5941887038850799, + "NONE": 0.7657693852952994 + }, + "accuracy_pct": 66.88197298884322, + "macro_f1_pct": 63.73646853691143, + "weighted_f1_pct": 68.17398364271445, + "macro_precision_pct": 63.72624859385251, + "macro_recall_pct": 66.45037793355199, + "per_class_f1_pct": { + "FAVOR": 55.21359669269638, + "AGAINST": 59.418870388508, + "NONE": 76.57693852952994 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6398864074710667, + "average_forgetting": 0.008716863174132259, + "backward_transfer": -0.006247733613938029, + "final_mean_target_macro_f1_pct": 63.98864074710667, + "average_forgetting_pct": 0.8716863174132259, + "backward_transfer_pct": -0.6247733613938029, + "final_pooled_accuracy": 0.6688197298884322, + "final_pooled_accuracy_pct": 66.88197298884322, + "final_pooled_macro_f1": 0.6373646853691143, + "final_pooled_macro_f1_pct": 63.73646853691143, + "final_pooled_weighted_f1": 0.6817398364271444, + "final_pooled_weighted_f1_pct": 68.17398364271445, + "final_pooled_macro_precision": 0.6372624859385251, + "final_pooled_macro_precision_pct": 63.72624859385251, + "final_pooled_macro_recall": 0.6645037793355199, + "final_pooled_macro_recall_pct": 66.45037793355199, + "final_pooled_per_class_f1": { + "FAVOR": 0.5521359669269638, + "AGAINST": 0.5941887038850799, + "NONE": 0.7657693852952994 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 55.21359669269638, + "AGAINST": 59.418870388508, + "NONE": 76.57693852952994 + } + }, + "efficiency": { + "training_wall_time_seconds": 1785.9809627318755, + "training_wall_time_minutes": 29.766349378864593, + "peak_gpu_memory_allocated_gb": 10.397007942199707, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s3__seed13/metrics.partial.json b/outputs/trace_s6_derpp__doc__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..526083f1363a6708e01b6c7dbe8bf0c74cd0b142 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s3__seed13/metrics.partial.json @@ -0,0 +1,1488 @@ +{ + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5374945250275344, + 0.5463075011612847, + null, + null, + null, + null, + null, + null + ], + [ + 0.5232522834586256, + 0.5335155435230371, + 0.6543247691104762, + null, + null, + null, + null, + null + ], + [ + 0.5087447364444003, + 0.4977928411587664, + 0.6470368747333941, + 0.6473990654259899, + null, + null, + null, + null + ], + [ + 0.52195274359378, + 0.5231365867193868, + 0.6623767078987792, + 0.6434524888983026, + 0.6970303442493787, + null, + null, + null + ], + [ + 0.5504146207992564, + 0.5314584102938406, + 0.6417148799712963, + 0.6463568595418386, + 0.6964409552570192, + 0.5964318804902767, + null, + null + ], + [ + 0.5321040216218281, + 0.5243885092860165, + 0.6716086760318358, + 0.6345308981292561, + 0.694603676067382, + 0.5825529293536768, + 0.7015197925693059, + null + ], + [ + 0.5428242154468658, + 0.5461933697640239, + 0.6711225503385233, + 0.6574413511480081, + 0.6775454464638669, + 0.5922058089403222, + 0.7026402070750647, + 0.7291183105918587 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 8, + "best_dev_macro_f1": 0.6598148411148085, + "losses": { + "current_ce": 0.6017990967997956, + "replay_ce": 0.29593221313719237, + "dark_distillation": 0.39678630508364193, + "total": 1.0961244651249478, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5374945250275344, + "weighted_f1": 0.6260728523998539, + "macro_precision": 0.5582332112961722, + "macro_recall": 0.5761467250574004, + "per_class_f1": { + "FAVOR": 0.37735849056603776, + "AGAINST": 0.4949290060851927, + "NONE": 0.7401960784313726 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.749452502753435, + "weighted_f1_pct": 62.60728523998539, + "macro_precision_pct": 55.82332112961722, + "macro_recall_pct": 57.61467250574004, + "per_class_f1_pct": { + "FAVOR": 37.735849056603776, + "AGAINST": 49.49290060851927, + "NONE": 74.01960784313727 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.5463075011612847, + "weighted_f1": 0.5943410243796432, + "macro_precision": 0.5625785385642903, + "macro_recall": 0.5774484043650431, + "per_class_f1": { + "FAVOR": 0.4112903225806452, + "AGAINST": 0.5734265734265734, + "NONE": 0.6542056074766355 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.63075011612847, + "weighted_f1_pct": 59.43410243796432, + "macro_precision_pct": 56.25785385642903, + "macro_recall_pct": 57.744840436504305, + "per_class_f1_pct": { + "FAVOR": 41.12903225806452, + "AGAINST": 57.34265734265735, + "NONE": 65.42056074766354 + } + } + ], + "accuracy": 0.5922266139657444, + "macro_f1": 0.5443944489016742, + "weighted_f1": 0.6105338373269887, + "macro_precision": 0.5618819133832902, + "macro_recall": 0.5744986770475305, + "per_class_f1": { + "FAVOR": 0.3937621832358674, + "AGAINST": 0.5370892018779343, + "NONE": 0.7023319615912208 + }, + "accuracy_pct": 59.22266139657444, + "macro_f1_pct": 54.43944489016742, + "weighted_f1_pct": 61.053383732698876, + "macro_precision_pct": 56.18819133832902, + "macro_recall_pct": 57.44986770475305, + "per_class_f1_pct": { + "FAVOR": 39.37621832358674, + "AGAINST": 53.70892018779343, + "NONE": 70.23319615912207 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.6720962418822527, + "losses": { + "current_ce": 0.5767768492656095, + "replay_ce": 0.24032061337715105, + "dark_distillation": 0.3301475648014318, + "total": 0.9821712439968473, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5895806861499364, + "macro_f1": 0.5232522834586256, + "weighted_f1": 0.617346102682281, + "macro_precision": 0.5292465014566562, + "macro_recall": 0.5513429501991594, + "per_class_f1": { + "FAVOR": 0.38006230529595014, + "AGAINST": 0.4486873508353222, + "NONE": 0.7410071942446043 + }, + "accuracy_pct": 58.95806861499364, + "macro_f1_pct": 52.32522834586256, + "weighted_f1_pct": 61.734610268228096, + "macro_precision_pct": 52.924650145665616, + "macro_recall_pct": 55.13429501991594, + "per_class_f1_pct": { + "FAVOR": 38.006230529595015, + "AGAINST": 44.868735083532215, + "NONE": 74.10071942446042 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5704514363885089, + "macro_f1": 0.5335155435230371, + "weighted_f1": 0.5868113170448906, + "macro_precision": 0.5576118831905328, + "macro_recall": 0.5641403151028193, + "per_class_f1": { + "FAVOR": 0.38735177865612647, + "AGAINST": 0.5570934256055363, + "NONE": 0.6561014263074485 + }, + "accuracy_pct": 57.045143638850895, + "macro_f1_pct": 53.35155435230371, + "weighted_f1_pct": 58.68113170448906, + "macro_precision_pct": 55.761188319053275, + "macro_recall_pct": 56.41403151028192, + "per_class_f1_pct": { + "FAVOR": 38.73517786561265, + "AGAINST": 55.70934256055363, + "NONE": 65.61014263074485 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6543247691104762, + "weighted_f1": 0.6928119083435389, + "macro_precision": 0.6519363341443634, + "macro_recall": 0.6578122878786107, + "per_class_f1": { + "FAVOR": 0.5630252100840336, + "AGAINST": 0.6007604562737643, + "NONE": 0.7991886409736308 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 65.43247691104762, + "weighted_f1_pct": 69.2811908343539, + "macro_precision_pct": 65.19363341443633, + "macro_recall_pct": 65.78122878786107, + "per_class_f1_pct": { + "FAVOR": 56.30252100840336, + "AGAINST": 60.07604562737643, + "NONE": 79.91886409736308 + } + } + ], + "accuracy": 0.6074441687344914, + "macro_f1": 0.5648044996747051, + "weighted_f1": 0.6250743616048388, + "macro_precision": 0.5695140800996698, + "macro_recall": 0.5875384929622324, + "per_class_f1": { + "FAVOR": 0.43596059113300495, + "AGAINST": 0.5301587301587302, + "NONE": 0.72829417773238 + }, + "accuracy_pct": 60.74441687344914, + "macro_f1_pct": 56.48044996747051, + "weighted_f1_pct": 62.50743616048388, + "macro_precision_pct": 56.951408009966975, + "macro_recall_pct": 58.75384929622324, + "per_class_f1_pct": { + "FAVOR": 43.5960591133005, + "AGAINST": 53.01587301587302, + "NONE": 72.829417773238 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7330493484446611, + "losses": { + "current_ce": 0.42256592037231944, + "replay_ce": 0.24960727357282872, + "dark_distillation": 0.32572553390409886, + "total": 0.8350359624478875, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.5087447364444003, + "weighted_f1": 0.6035631935656818, + "macro_precision": 0.5311162440200322, + "macro_recall": 0.5420255569493042, + "per_class_f1": { + "FAVOR": 0.3445692883895131, + "AGAINST": 0.45528455284552843, + "NONE": 0.7263803680981595 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 50.87447364444003, + "weighted_f1_pct": 60.35631935656818, + "macro_precision_pct": 53.11162440200322, + "macro_recall_pct": 54.202555694930425, + "per_class_f1_pct": { + "FAVOR": 34.45692883895131, + "AGAINST": 45.52845528455284, + "NONE": 72.63803680981596 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5266757865937073, + "macro_f1": 0.4977928411587664, + "weighted_f1": 0.5400641190222629, + "macro_precision": 0.5420096793143089, + "macro_recall": 0.5330103287083535, + "per_class_f1": { + "FAVOR": 0.37751004016064255, + "AGAINST": 0.5242718446601942, + "NONE": 0.5915966386554622 + }, + "accuracy_pct": 52.667578659370726, + "macro_f1_pct": 49.779284115876635, + "weighted_f1_pct": 54.00641190222629, + "macro_precision_pct": 54.20096793143089, + "macro_recall_pct": 53.30103287083535, + "per_class_f1_pct": { + "FAVOR": 37.75100401606426, + "AGAINST": 52.42718446601942, + "NONE": 59.15966386554622 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6470368747333941, + "weighted_f1": 0.6882543398622264, + "macro_precision": 0.6485199308728721, + "macro_recall": 0.6556293538607467, + "per_class_f1": { + "FAVOR": 0.5414847161572053, + "AGAINST": 0.5979381443298969, + "NONE": 0.8016877637130801 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 64.70368747333941, + "weighted_f1_pct": 68.82543398622263, + "macro_precision_pct": 64.85199308728721, + "macro_recall_pct": 65.56293538607467, + "per_class_f1_pct": { + "FAVOR": 54.14847161572053, + "AGAINST": 59.79381443298969, + "NONE": 80.16877637130801 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6959553695955369, + "macro_f1": 0.6473990654259899, + "weighted_f1": 0.7083102424122704, + "macro_precision": 0.6402361320796545, + "macro_recall": 0.6793283385457837, + "per_class_f1": { + "FAVOR": 0.5055762081784386, + "AGAINST": 0.6440677966101694, + "NONE": 0.7925531914893617 + }, + "accuracy_pct": 69.5955369595537, + "macro_f1_pct": 64.73990654259899, + "weighted_f1_pct": 70.83102424122704, + "macro_precision_pct": 64.02361320796544, + "macro_recall_pct": 67.93283385457836, + "per_class_f1_pct": { + "FAVOR": 50.55762081784386, + "AGAINST": 64.40677966101694, + "NONE": 79.25531914893617 + } + } + ], + "accuracy": 0.6138360175695461, + "macro_f1": 0.570957079351705, + "weighted_f1": 0.6311171624377911, + "macro_precision": 0.5830862536616448, + "macro_recall": 0.5977160478674138, + "per_class_f1": { + "FAVOR": 0.43984220907297833, + "AGAINST": 0.5446527012127894, + "NONE": 0.7283763277693475 + }, + "accuracy_pct": 61.38360175695461, + "macro_f1_pct": 57.0957079351705, + "weighted_f1_pct": 63.111716243779114, + "macro_precision_pct": 58.30862536616448, + "macro_recall_pct": 59.771604786741385, + "per_class_f1_pct": { + "FAVOR": 43.98422090729783, + "AGAINST": 54.46527012127894, + "NONE": 72.83763277693474 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.7418596503125036, + "losses": { + "current_ce": 0.4637480529717037, + "replay_ce": 0.22047008640709378, + "dark_distillation": 0.32636636401925767, + "total": 0.8474013226940519, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5844980940279543, + "macro_f1": 0.52195274359378, + "weighted_f1": 0.6162103840273592, + "macro_precision": 0.5359205428660122, + "macro_recall": 0.5526024588660754, + "per_class_f1": { + "FAVOR": 0.38977635782747605, + "AGAINST": 0.4349775784753363, + "NONE": 0.7411042944785277 + }, + "accuracy_pct": 58.449809402795424, + "macro_f1_pct": 52.195274359378, + "weighted_f1_pct": 61.62103840273592, + "macro_precision_pct": 53.59205428660122, + "macro_recall_pct": 55.26024588660754, + "per_class_f1_pct": { + "FAVOR": 38.977635782747605, + "AGAINST": 43.49775784753363, + "NONE": 74.11042944785277 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5231365867193868, + "weighted_f1": 0.5658044656042314, + "macro_precision": 0.5583250285928314, + "macro_recall": 0.5599019713934344, + "per_class_f1": { + "FAVOR": 0.4108527131782946, + "AGAINST": 0.5335570469798657, + "NONE": 0.625 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 52.313658671938676, + "weighted_f1_pct": 56.580446560423134, + "macro_precision_pct": 55.83250285928314, + "macro_recall_pct": 55.990197139343444, + "per_class_f1_pct": { + "FAVOR": 41.08527131782946, + "AGAINST": 53.355704697986575, + "NONE": 62.5 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6623767078987792, + "weighted_f1": 0.7020537913388651, + "macro_precision": 0.6590542350454245, + "macro_recall": 0.6693015355506144, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.5939849624060151, + "NONE": 0.8125 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 66.23767078987792, + "weighted_f1_pct": 70.20537913388651, + "macro_precision_pct": 65.90542350454245, + "macro_recall_pct": 66.93015355506145, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 59.3984962406015, + "NONE": 81.25 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6694560669456067, + "macro_f1": 0.6434524888983026, + "weighted_f1": 0.6836444678613905, + "macro_precision": 0.6422760832661747, + "macro_recall": 0.7009640018353612, + "per_class_f1": { + "FAVOR": 0.5483870967741935, + "AGAINST": 0.6430260047281324, + "NONE": 0.738944365192582 + }, + "accuracy_pct": 66.94560669456067, + "macro_f1_pct": 64.34524888983026, + "weighted_f1_pct": 68.36444678613906, + "macro_precision_pct": 64.22760832661747, + "macro_recall_pct": 70.09640018353612, + "per_class_f1_pct": { + "FAVOR": 54.83870967741935, + "AGAINST": 64.30260047281324, + "NONE": 73.89443651925819 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6970303442493787, + "weighted_f1": 0.715896851540361, + "macro_precision": 0.6961901536362936, + "macro_recall": 0.7175100958781302, + "per_class_f1": { + "FAVOR": 0.6130030959752322, + "AGAINST": 0.7100271002710027, + "NONE": 0.7680608365019012 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.70303442493787, + "weighted_f1_pct": 71.5896851540361, + "macro_precision_pct": 69.61901536362936, + "macro_recall_pct": 71.75100958781302, + "per_class_f1_pct": { + "FAVOR": 61.30030959752322, + "AGAINST": 71.00271002710026, + "NONE": 76.80608365019012 + } + } + ], + "accuracy": 0.634540556719545, + "macro_f1": 0.6055071774967216, + "weighted_f1": 0.6505926399645455, + "macro_precision": 0.6113435294931517, + "macro_recall": 0.6373525835247188, + "per_class_f1": { + "FAVOR": 0.509641873278237, + "AGAINST": 0.5733333333333334, + "NONE": 0.7335463258785943 + }, + "accuracy_pct": 63.4540556719545, + "macro_f1_pct": 60.55071774967216, + "weighted_f1_pct": 65.05926399645455, + "macro_precision_pct": 61.13435294931517, + "macro_recall_pct": 63.73525835247188, + "per_class_f1_pct": { + "FAVOR": 50.964187327823694, + "AGAINST": 57.333333333333336, + "NONE": 73.35463258785943 + } + }, + { + "after_task": 6, + "best_epoch": 9, + "best_dev_macro_f1": 0.6871266882822313, + "losses": { + "current_ce": 0.501748810996089, + "replay_ce": 0.20201711892558818, + "dark_distillation": 0.31889288313033287, + "total": 0.8632123711339215, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6124523506988564, + "macro_f1": 0.5504146207992564, + "weighted_f1": 0.6402539684402383, + "macro_precision": 0.5565610859728506, + "macro_recall": 0.5845231545667277, + "per_class_f1": { + "FAVOR": 0.4086687306501548, + "AGAINST": 0.48470588235294115, + "NONE": 0.7578692493946732 + }, + "accuracy_pct": 61.24523506988564, + "macro_f1_pct": 55.04146207992564, + "weighted_f1_pct": 64.02539684402383, + "macro_precision_pct": 55.65610859728506, + "macro_recall_pct": 58.452315456672764, + "per_class_f1_pct": { + "FAVOR": 40.86687306501548, + "AGAINST": 48.470588235294116, + "NONE": 75.78692493946731 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.560875512995896, + "macro_f1": 0.5314584102938406, + "weighted_f1": 0.5815524288238701, + "macro_precision": 0.5522443612861265, + "macro_recall": 0.5684929306455921, + "per_class_f1": { + "FAVOR": 0.4128113879003559, + "AGAINST": 0.5201465201465202, + "NONE": 0.6614173228346457 + }, + "accuracy_pct": 56.0875512995896, + "macro_f1_pct": 53.14584102938406, + "weighted_f1_pct": 58.155242882387014, + "macro_precision_pct": 55.22443612861265, + "macro_recall_pct": 56.84929306455921, + "per_class_f1_pct": { + "FAVOR": 41.281138790035584, + "AGAINST": 52.01465201465202, + "NONE": 66.14173228346458 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6417148799712963, + "weighted_f1": 0.691986525090873, + "macro_precision": 0.6428967402063139, + "macro_recall": 0.6442462665749328, + "per_class_f1": { + "FAVOR": 0.5178571428571429, + "AGAINST": 0.5766423357664233, + "NONE": 0.8306451612903226 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 64.17148799712963, + "weighted_f1_pct": 69.1986525090873, + "macro_precision_pct": 64.2896740206314, + "macro_recall_pct": 64.42462665749328, + "per_class_f1_pct": { + "FAVOR": 51.78571428571429, + "AGAINST": 57.66423357664233, + "NONE": 83.06451612903226 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6778242677824268, + "macro_f1": 0.6463568595418386, + "weighted_f1": 0.693531099516167, + "macro_precision": 0.6415967338836941, + "macro_recall": 0.696375749233206, + "per_class_f1": { + "FAVOR": 0.5396825396825397, + "AGAINST": 0.64, + "NONE": 0.7593880389429764 + }, + "accuracy_pct": 67.78242677824268, + "macro_f1_pct": 64.63568595418387, + "weighted_f1_pct": 69.3531099516167, + "macro_precision_pct": 64.15967338836941, + "macro_recall_pct": 69.6375749233206, + "per_class_f1_pct": { + "FAVOR": 53.96825396825397, + "AGAINST": 64.0, + "NONE": 75.93880389429764 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.6964409552570192, + "weighted_f1": 0.7109185138427025, + "macro_precision": 0.7007600686915977, + "macro_recall": 0.724026339381514, + "per_class_f1": { + "FAVOR": 0.6445783132530121, + "AGAINST": 0.6912928759894459, + "NONE": 0.7534516765285996 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.64409552570191, + "weighted_f1_pct": 71.09185138427026, + "macro_precision_pct": 70.07600686915977, + "macro_recall_pct": 72.4026339381514, + "per_class_f1_pct": { + "FAVOR": 64.45783132530121, + "AGAINST": 69.12928759894459, + "NONE": 75.34516765285996 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6307922272047832, + "macro_f1": 0.5964318804902767, + "weighted_f1": 0.650760169841083, + "macro_precision": 0.6045797635434466, + "macro_recall": 0.6290321443552623, + "per_class_f1": { + "FAVOR": 0.5684754521963824, + "AGAINST": 0.47318611987381703, + "NONE": 0.7476340694006309 + }, + "accuracy_pct": 63.07922272047832, + "macro_f1_pct": 59.64318804902767, + "weighted_f1_pct": 65.07601698410829, + "macro_precision_pct": 60.45797635434466, + "macro_recall_pct": 62.90321443552623, + "per_class_f1_pct": { + "FAVOR": 56.84754521963824, + "AGAINST": 47.3186119873817, + "NONE": 74.76340694006309 + } + } + ], + "accuracy": 0.6413965087281795, + "macro_f1": 0.6104345237807094, + "weighted_f1": 0.6582819955854231, + "macro_precision": 0.6134178127890829, + "macro_recall": 0.6407169291242313, + "per_class_f1": { + "FAVOR": 0.5198711063372717, + "AGAINST": 0.5621529260999573, + "NONE": 0.7492795389048992 + }, + "accuracy_pct": 64.13965087281795, + "macro_f1_pct": 61.04345237807094, + "weighted_f1_pct": 65.82819955854231, + "macro_precision_pct": 61.34178127890829, + "macro_recall_pct": 64.07169291242313, + "per_class_f1_pct": { + "FAVOR": 51.98711063372718, + "AGAINST": 56.21529260999573, + "NONE": 74.92795389048992 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.7145235097573966, + "losses": { + "current_ce": 0.4077385595334428, + "replay_ce": 0.19610688175473895, + "dark_distillation": 0.29602664301083204, + "total": 0.7518587647449403, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6060991105463787, + "macro_f1": 0.5321040216218281, + "weighted_f1": 0.6299955720448114, + "macro_precision": 0.5483200850355855, + "macro_recall": 0.5626651408786486, + "per_class_f1": { + "FAVOR": 0.36363636363636365, + "AGAINST": 0.47578947368421054, + "NONE": 0.7568862275449102 + }, + "accuracy_pct": 60.60991105463786, + "macro_f1_pct": 53.21040216218281, + "weighted_f1_pct": 62.999557204481135, + "macro_precision_pct": 54.83200850355855, + "macro_recall_pct": 56.26651408786486, + "per_class_f1_pct": { + "FAVOR": 36.36363636363637, + "AGAINST": 47.578947368421055, + "NONE": 75.68862275449102 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5554035567715458, + "macro_f1": 0.5243885092860165, + "weighted_f1": 0.5699230461359269, + "macro_precision": 0.5607555614058181, + "macro_recall": 0.5572611789133383, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5436573311367381, + "NONE": 0.6295081967213115 + }, + "accuracy_pct": 55.54035567715459, + "macro_f1_pct": 52.43885092860165, + "weighted_f1_pct": 56.992304613592694, + "macro_precision_pct": 56.0755561405818, + "macro_recall_pct": 55.72611789133383, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 54.36573311367381, + "NONE": 62.95081967213115 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.710261569416499, + "macro_f1": 0.6716086760318358, + "weighted_f1": 0.7127397731485023, + "macro_precision": 0.6789344718370142, + "macro_recall": 0.6765512797200343, + "per_class_f1": { + "FAVOR": 0.5700934579439252, + "AGAINST": 0.6185567010309279, + "NONE": 0.8261758691206544 + }, + "accuracy_pct": 71.0261569416499, + "macro_f1_pct": 67.16086760318359, + "weighted_f1_pct": 71.27397731485023, + "macro_precision_pct": 67.89344718370141, + "macro_recall_pct": 67.65512797200343, + "per_class_f1_pct": { + "FAVOR": 57.009345794392516, + "AGAINST": 61.855670103092784, + "NONE": 82.61758691206545 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6750348675034867, + "macro_f1": 0.6345308981292561, + "weighted_f1": 0.6913888335630604, + "macro_precision": 0.6300536947396256, + "macro_recall": 0.6753426530377219, + "per_class_f1": { + "FAVOR": 0.5067567567567568, + "AGAINST": 0.6259168704156479, + "NONE": 0.7709190672153635 + }, + "accuracy_pct": 67.50348675034867, + "macro_f1_pct": 63.45308981292561, + "weighted_f1_pct": 69.13888335630604, + "macro_precision_pct": 63.00536947396256, + "macro_recall_pct": 67.53426530377219, + "per_class_f1_pct": { + "FAVOR": 50.67567567567568, + "AGAINST": 62.59168704156479, + "NONE": 77.09190672153635 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.694603676067382, + "weighted_f1": 0.7082629692676564, + "macro_precision": 0.6989239042998294, + "macro_recall": 0.7166956589032524, + "per_class_f1": { + "FAVOR": 0.6416382252559727, + "AGAINST": 0.6945812807881774, + "NONE": 0.7475915221579962 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.4603676067382, + "weighted_f1_pct": 70.82629692676564, + "macro_precision_pct": 69.89239042998294, + "macro_recall_pct": 71.66956589032523, + "per_class_f1_pct": { + "FAVOR": 64.16382252559727, + "AGAINST": 69.45812807881774, + "NONE": 74.75915221579962 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6143497757847534, + "macro_f1": 0.5825529293536768, + "weighted_f1": 0.6367423947672474, + "macro_precision": 0.5972526751747531, + "macro_recall": 0.6177075365394264, + "per_class_f1": { + "FAVOR": 0.5655526992287918, + "AGAINST": 0.44984802431610943, + "NONE": 0.7322580645161291 + }, + "accuracy_pct": 61.43497757847533, + "macro_f1_pct": 58.25529293536768, + "weighted_f1_pct": 63.674239476724736, + "macro_precision_pct": 59.72526751747531, + "macro_recall_pct": 61.77075365394265, + "per_class_f1_pct": { + "FAVOR": 56.555269922879184, + "AGAINST": 44.984802431610944, + "NONE": 73.22580645161291 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.7015197925693059, + "weighted_f1": 0.7329989063210653, + "macro_precision": 0.6979025260929879, + "macro_recall": 0.7291751521636579, + "per_class_f1": { + "FAVOR": 0.6187050359712231, + "AGAINST": 0.6833333333333333, + "NONE": 0.8025210084033614 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 70.15197925693059, + "weighted_f1_pct": 73.29989063210654, + "macro_precision_pct": 69.79025260929879, + "macro_recall_pct": 72.91751521636579, + "per_class_f1_pct": { + "FAVOR": 61.87050359712231, + "AGAINST": 68.33333333333333, + "NONE": 80.25210084033614 + } + } + ], + "accuracy": 0.6476591968049701, + "macro_f1": 0.6177427847929325, + "weighted_f1": 0.6628197068125781, + "macro_precision": 0.6224716082855182, + "macro_recall": 0.6470068460168169, + "per_class_f1": { + "FAVOR": 0.5285497726124305, + "AGAINST": 0.5752629669931084, + "NONE": 0.7494156147732586 + }, + "accuracy_pct": 64.765919680497, + "macro_f1_pct": 61.77427847929325, + "weighted_f1_pct": 66.2819706812578, + "macro_precision_pct": 62.24716082855182, + "macro_recall_pct": 64.70068460168169, + "per_class_f1_pct": { + "FAVOR": 52.85497726124305, + "AGAINST": 57.52629669931084, + "NONE": 74.94156147732586 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.7778474641889276, + "losses": { + "current_ce": 0.337591310269108, + "replay_ce": 0.18207696132800158, + "dark_distillation": 0.2913791876505403, + "total": 0.6653578645458409, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6226175349428208, + "macro_f1": 0.5428242154468658, + "weighted_f1": 0.645728462985635, + "macro_precision": 0.5503162699796812, + "macro_recall": 0.5688708068337698, + "per_class_f1": { + "FAVOR": 0.36823104693140796, + "AGAINST": 0.48089887640449436, + "NONE": 0.7793427230046949 + }, + "accuracy_pct": 62.26175349428208, + "macro_f1_pct": 54.28242154468658, + "weighted_f1_pct": 64.5728462985635, + "macro_precision_pct": 55.03162699796812, + "macro_recall_pct": 56.88708068337698, + "per_class_f1_pct": { + "FAVOR": 36.8231046931408, + "AGAINST": 48.08988764044943, + "NONE": 77.93427230046949 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5759233926128591, + "macro_f1": 0.5461933697640239, + "weighted_f1": 0.588421862814057, + "macro_precision": 0.5728881713913531, + "macro_recall": 0.5788560191807598, + "per_class_f1": { + "FAVOR": 0.4315352697095436, + "AGAINST": 0.5628140703517588, + "NONE": 0.6442307692307693 + }, + "accuracy_pct": 57.59233926128591, + "macro_f1_pct": 54.61933697640239, + "weighted_f1_pct": 58.842186281405695, + "macro_precision_pct": 57.288817139135304, + "macro_recall_pct": 57.88560191807598, + "per_class_f1_pct": { + "FAVOR": 43.15352697095436, + "AGAINST": 56.28140703517588, + "NONE": 64.42307692307693 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6711225503385233, + "weighted_f1": 0.7143077591054621, + "macro_precision": 0.6766716781214205, + "macro_recall": 0.6710286356564912, + "per_class_f1": { + "FAVOR": 0.5581395348837209, + "AGAINST": 0.6222222222222222, + "NONE": 0.8330058939096268 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 67.11225503385234, + "weighted_f1_pct": 71.4307759105462, + "macro_precision_pct": 67.66716781214205, + "macro_recall_pct": 67.10286356564912, + "per_class_f1_pct": { + "FAVOR": 55.81395348837209, + "AGAINST": 62.22222222222222, + "NONE": 83.30058939096267 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6959553695955369, + "macro_f1": 0.6574413511480081, + "weighted_f1": 0.7088029794365435, + "macro_precision": 0.6524016760393453, + "macro_recall": 0.7003100337708897, + "per_class_f1": { + "FAVOR": 0.5342960288808665, + "AGAINST": 0.6588785046728972, + "NONE": 0.7791495198902606 + }, + "accuracy_pct": 69.5955369595537, + "macro_f1_pct": 65.74413511480081, + "weighted_f1_pct": 70.88029794365434, + "macro_precision_pct": 65.24016760393452, + "macro_recall_pct": 70.03100337708898, + "per_class_f1_pct": { + "FAVOR": 53.42960288808665, + "AGAINST": 65.88785046728972, + "NONE": 77.91495198902606 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6880131362889984, + "macro_f1": 0.6775454464638669, + "weighted_f1": 0.692112365014755, + "macro_precision": 0.6848455598455598, + "macro_recall": 0.6997477245487044, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6779661016949152, + "NONE": 0.7339805825242719 + }, + "accuracy_pct": 68.80131362889983, + "macro_f1_pct": 67.75454464638669, + "weighted_f1_pct": 69.2112365014755, + "macro_precision_pct": 68.48455598455598, + "macro_recall_pct": 69.97477245487043, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 67.79661016949152, + "NONE": 73.39805825242719 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6218236173393124, + "macro_f1": 0.5922058089403222, + "weighted_f1": 0.6433191365957746, + "macro_precision": 0.604110460369961, + "macro_recall": 0.6281796850306732, + "per_class_f1": { + "FAVOR": 0.5706666666666667, + "AGAINST": 0.471976401179941, + "NONE": 0.7339743589743589 + }, + "accuracy_pct": 62.18236173393124, + "macro_f1_pct": 59.22058089403222, + "weighted_f1_pct": 64.33191365957745, + "macro_precision_pct": 60.4110460369961, + "macro_recall_pct": 62.81796850306732, + "per_class_f1_pct": { + "FAVOR": 57.06666666666666, + "AGAINST": 47.1976401179941, + "NONE": 73.3974358974359 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7223340040241448, + "macro_f1": 0.7026402070750647, + "weighted_f1": 0.7318102922973562, + "macro_precision": 0.697317230797503, + "macro_recall": 0.7324531427979704, + "per_class_f1": { + "FAVOR": 0.64, + "AGAINST": 0.6721311475409836, + "NONE": 0.7957894736842105 + }, + "accuracy_pct": 72.23340040241449, + "macro_f1_pct": 70.26402070750648, + "weighted_f1_pct": 73.18102922973561, + "macro_precision_pct": 69.7317230797503, + "macro_recall_pct": 73.24531427979704, + "per_class_f1_pct": { + "FAVOR": 64.0, + "AGAINST": 67.21311475409836, + "NONE": 79.57894736842105 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7591362126245847, + "macro_f1": 0.7291183105918587, + "weighted_f1": 0.7650915804968302, + "macro_precision": 0.7190885724025161, + "macro_recall": 0.747531550769343, + "per_class_f1": { + "FAVOR": 0.6960352422907489, + "AGAINST": 0.6605504587155964, + "NONE": 0.8307692307692308 + }, + "accuracy_pct": 75.91362126245848, + "macro_f1_pct": 72.91183105918587, + "weighted_f1_pct": 76.50915804968302, + "macro_precision_pct": 71.9088572402516, + "macro_recall_pct": 74.75315507693429, + "per_class_f1_pct": { + "FAVOR": 69.60352422907489, + "AGAINST": 66.05504587155964, + "NONE": 83.07692307692308 + } + } + ], + "accuracy": 0.6688197298884322, + "macro_f1": 0.6373646853691143, + "weighted_f1": 0.6817398364271444, + "macro_precision": 0.6372624859385251, + "macro_recall": 0.6645037793355199, + "per_class_f1": { + "FAVOR": 0.5521359669269638, + "AGAINST": 0.5941887038850799, + "NONE": 0.7657693852952994 + }, + "accuracy_pct": 66.88197298884322, + "macro_f1_pct": 63.73646853691143, + "weighted_f1_pct": 68.17398364271445, + "macro_precision_pct": 63.72624859385251, + "macro_recall_pct": 66.45037793355199, + "per_class_f1_pct": { + "FAVOR": 55.21359669269638, + "AGAINST": 59.418870388508, + "NONE": 76.57693852952994 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__doc__s3__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__doc__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b9bbb88387b93733cd6e35576f2b1c1be0f3aba3 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__doc__s3__seed13/run.log b/outputs/trace_s6_derpp__doc__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..ac25c2501046375e3fd16f53fd437dba44ed66c3 --- /dev/null +++ b/outputs/trace_s6_derpp__doc__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 05:11:41,109 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__doc__s3__seed13 device=cuda +2026-09-02 05:41:28,535 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6398864074710667, 'average_forgetting': 0.008716863174132259, 'backward_transfer': -0.006247733613938029, 'final_mean_target_macro_f1_pct': 63.98864074710667, 'average_forgetting_pct': 0.8716863174132259, 'backward_transfer_pct': -0.6247733613938029, 'final_pooled_accuracy': 0.6688197298884322, 'final_pooled_accuracy_pct': 66.88197298884322, 'final_pooled_macro_f1': 0.6373646853691143, 'final_pooled_macro_f1_pct': 63.73646853691143, 'final_pooled_weighted_f1': 0.6817398364271444, 'final_pooled_weighted_f1_pct': 68.17398364271445, 'final_pooled_macro_precision': 0.6372624859385251, 'final_pooled_macro_precision_pct': 63.72624859385251, 'final_pooled_macro_recall': 0.6645037793355199, 'final_pooled_macro_recall_pct': 66.45037793355199, 'final_pooled_per_class_f1': {'FAVOR': 0.5521359669269638, 'AGAINST': 0.5941887038850799, 'NONE': 0.7657693852952994}, 'final_pooled_per_class_f1_pct': {'FAVOR': 55.21359669269638, 'AGAINST': 59.418870388508, 'NONE': 76.57693852952994}} +2026-09-02 05:41:28,536 [INFO] __main__: efficiency={'training_wall_time_seconds': 1785.9809627318755, 'training_wall_time_minutes': 29.766349378864593, 'peak_gpu_memory_allocated_gb': 10.397007942199707, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s6_derpp__x__s1__seed13/environment.json b/outputs/trace_s6_derpp__x__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..cd5d0e037c0c40e94931e7a9c2afb40f33b4a48c --- /dev/null +++ b/outputs/trace_s6_derpp__x__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T05:41:36.618444+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s1__seed13/metrics.json b/outputs/trace_s6_derpp__x__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..354ec7e6195aa63f5add1731cd170bdfa8067129 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s1__seed13/metrics.json @@ -0,0 +1,2285 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7086325439266616, + 0.5945165945165946, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7116641528406235, + 0.641371628745366, + 0.5594731164480781, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7037323701410388, + 0.7241850145075951, + 0.5746323527807488, + 0.7182016512153498, + null, + null, + null, + null, + null, + null + ], + [ + 0.67046883366121, + 0.7021442860715178, + 0.5939931528166822, + 0.6659498207885304, + 0.6349317422488153, + null, + null, + null, + null, + null + ], + [ + 0.6766321115636185, + 0.7526873906998599, + 0.5896460114605678, + 0.6525647805393971, + 0.6510334139823205, + 0.6472431951444642, + null, + null, + null, + null + ], + [ + 0.7033333333333333, + 0.6857007290611626, + 0.6135822214360529, + 0.6700329703135232, + 0.7313881929908913, + 0.5713784521880473, + 0.6795023188656119, + null, + null, + null + ], + [ + 0.7074314574314574, + 0.7203548525457949, + 0.6013859309350852, + 0.6630824372759857, + 0.6541619058790477, + 0.6379917787175137, + 0.673738353080374, + 0.7290427437266881, + null, + null + ], + [ + 0.7096626856420892, + 0.6924486803519061, + 0.5986009498138878, + 0.7088543640267778, + 0.6672700269191498, + 0.5751948370960899, + 0.6754494932984635, + 0.7135238095238096, + 0.8115569192038392, + null + ], + [ + 0.7042724542724542, + 0.6958754456546625, + 0.5980952380952381, + 0.6813132250261148, + 0.6622393075825238, + 0.574048555564355, + 0.6861699501308447, + 0.7051904438915098, + 0.7991414177492491, + 0.6421550102590051 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6840048840048839, + "losses": { + "current_ce": 0.742766341141292, + "replay_ce": 0.25674224751336233, + "dark_distillation": 0.361050540634564, + "total": 1.18003385918481, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7086325439266616, + "weighted_f1": 0.7367864206703852, + "macro_precision": 0.6917552558871426, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.7662337662337663, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.86325439266616, + "weighted_f1_pct": 73.67864206703852, + "macro_precision_pct": 69.17552558871427, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 76.62337662337663, + "NONE": 77.14285714285715 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5945165945165946, + "weighted_f1": 0.8493276634869555, + "macro_precision": 0.5920095408467502, + "macro_recall": 0.5975003033612426, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.0, + "NONE": 0.9047619047619048 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 59.45165945165945, + "weighted_f1_pct": 84.93276634869555, + "macro_precision_pct": 59.20095408467502, + "macro_recall_pct": 59.75003033612426, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 0.0, + "NONE": 90.47619047619048 + } + } + ], + "accuracy": 0.7799227799227799, + "macro_f1": 0.7812226797592651, + "weighted_f1": 0.7746856834661713, + "macro_precision": 0.7922095869857064, + "macro_recall": 0.7948816460694541, + "per_class_f1": { + "FAVOR": 0.78, + "AGAINST": 0.7195121951219512, + "NONE": 0.8441558441558441 + }, + "accuracy_pct": 77.99227799227799, + "macro_f1_pct": 78.1222679759265, + "weighted_f1_pct": 77.46856834661713, + "macro_precision_pct": 79.22095869857064, + "macro_recall_pct": 79.48816460694542, + "per_class_f1_pct": { + "FAVOR": 78.0, + "AGAINST": 71.95121951219512, + "NONE": 84.4155844155844 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6090658702505393, + "losses": { + "current_ce": 0.6782336643247894, + "replay_ce": 0.22404358504396496, + "dark_distillation": 0.3302716556823615, + "total": 1.0674130804610975, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7116641528406235, + "weighted_f1": 0.7454182937664, + "macro_precision": 0.694969315245478, + "macro_recall": 0.7865317341329335, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.782051282051282, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.16641528406234, + "weighted_f1_pct": 74.54182937664, + "macro_precision_pct": 69.4969315245478, + "macro_recall_pct": 78.65317341329336, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 78.2051282051282, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.641371628745366, + "weighted_f1": 0.8048368281004376, + "macro_precision": 0.6467673814165043, + "macro_recall": 0.7283339400558185, + "per_class_f1": { + "FAVOR": 0.8099173553719008, + "AGAINST": 0.25, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 64.1371628745366, + "weighted_f1_pct": 80.48368281004376, + "macro_precision_pct": 64.67673814165042, + "macro_recall_pct": 72.83339400558185, + "per_class_f1_pct": { + "FAVOR": 80.99173553719008, + "AGAINST": 25.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5594731164480781, + "weighted_f1": 0.5900438372374472, + "macro_precision": 0.5543874115302687, + "macro_recall": 0.5693536673928831, + "per_class_f1": { + "FAVOR": 0.4854368932038835, + "AGAINST": 0.6666666666666666, + "NONE": 0.5263157894736842 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 55.947311644807804, + "weighted_f1_pct": 59.00438372374472, + "macro_precision_pct": 55.43874115302687, + "macro_recall_pct": 56.935366739288305, + "per_class_f1_pct": { + "FAVOR": 48.54368932038835, + "AGAINST": 66.66666666666666, + "NONE": 52.63157894736842 + } + } + ], + "accuracy": 0.6815144766146993, + "macro_f1": 0.6838332194777023, + "weighted_f1": 0.6809023259620387, + "macro_precision": 0.6776791079406271, + "macro_recall": 0.6936577425236474, + "per_class_f1": { + "FAVOR": 0.6438356164383562, + "AGAINST": 0.6876640419947506, + "NONE": 0.72 + }, + "accuracy_pct": 68.15144766146993, + "macro_f1_pct": 68.38332194777023, + "weighted_f1_pct": 68.09023259620388, + "macro_precision_pct": 67.76791079406271, + "macro_recall_pct": 69.36577425236474, + "per_class_f1_pct": { + "FAVOR": 64.38356164383562, + "AGAINST": 68.76640419947506, + "NONE": 72.0 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.6498013902681231, + "losses": { + "current_ce": 0.6360149974172765, + "replay_ce": 0.20808983025225727, + "dark_distillation": 0.33606521649794147, + "total": 1.0121374368667602, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7037323701410388, + "weighted_f1": 0.7326734947764253, + "macro_precision": 0.6933806146572103, + "macro_recall": 0.7889955022488756, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.7631578947368421, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.37323701410388, + "weighted_f1_pct": 73.26734947764253, + "macro_precision_pct": 69.33806146572104, + "macro_recall_pct": 78.89955022488756, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 76.31578947368422, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7241850145075951, + "weighted_f1": 0.8470789210155467, + "macro_precision": 0.6974233436901737, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.8709677419354839, + "AGAINST": 0.4444444444444444, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.4185014507595, + "weighted_f1_pct": 84.70789210155468, + "macro_precision_pct": 69.74233436901737, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 87.09677419354838, + "AGAINST": 44.44444444444444, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.5746323527807488, + "weighted_f1": 0.5998972157087302, + "macro_precision": 0.5712692890910713, + "macro_recall": 0.5787944807552651, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.6699507389162561, + "NONE": 0.5633802816901409 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 57.463235278074876, + "weighted_f1_pct": 59.98972157087302, + "macro_precision_pct": 57.126928909107136, + "macro_recall_pct": 57.879448075526504, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 66.99507389162561, + "NONE": 56.33802816901409 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7182016512153498, + "weighted_f1": 0.746434472877248, + "macro_precision": 0.7022812667740205, + "macro_recall": 0.7455106621773289, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.79, + "NONE": 0.7070707070707071 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 71.82016512153498, + "weighted_f1_pct": 74.6434472877248, + "macro_precision_pct": 70.22812667740205, + "macro_recall_pct": 74.55106621773288, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 79.0, + "NONE": 70.70707070707071 + } + } + ], + "accuracy": 0.710236220472441, + "macro_f1": 0.7076228776661283, + "weighted_f1": 0.7111298601901942, + "macro_precision": 0.6999233219299281, + "macro_recall": 0.721446893544294, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7294938917975567, + "NONE": 0.7267080745341615 + }, + "accuracy_pct": 71.02362204724409, + "macro_f1_pct": 70.76228776661283, + "weighted_f1_pct": 71.11298601901942, + "macro_precision_pct": 69.99233219299282, + "macro_recall_pct": 72.1446893544294, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 72.94938917975567, + "NONE": 72.67080745341616 + } + }, + { + "after_task": 5, + "best_epoch": 4, + "best_dev_macro_f1": 0.687115687115687, + "losses": { + "current_ce": 0.5763537430402005, + "replay_ce": 0.1827620306701371, + "dark_distillation": 0.3391505776932745, + "total": 0.9286910642277111, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.67046883366121, + "weighted_f1": 0.691956730653933, + "macro_precision": 0.672269705603039, + "macro_recall": 0.767256371814093, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.7123287671232876, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 67.04688336612101, + "weighted_f1_pct": 69.19567306539331, + "macro_precision_pct": 67.2269705603039, + "macro_recall_pct": 76.7256371814093, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 71.23287671232876, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7021442860715178, + "weighted_f1": 0.8452115794435334, + "macro_precision": 0.6807726386673755, + "macro_recall": 0.7694697245479918, + "per_class_f1": { + "FAVOR": 0.8709677419354839, + "AGAINST": 0.375, + "NONE": 0.8604651162790697 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 70.21442860715177, + "weighted_f1_pct": 84.52115794435335, + "macro_precision_pct": 68.07726386673755, + "macro_recall_pct": 76.94697245479918, + "per_class_f1_pct": { + "FAVOR": 87.09677419354838, + "AGAINST": 37.5, + "NONE": 86.04651162790698 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5939931528166822, + "weighted_f1": 0.6105508870214753, + "macro_precision": 0.5882079203313323, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5042016806722689, + "AGAINST": 0.6666666666666666, + "NONE": 0.6111111111111112 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 59.39931528166822, + "weighted_f1_pct": 61.05508870214753, + "macro_precision_pct": 58.82079203313323, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 50.42016806722689, + "AGAINST": 66.66666666666666, + "NONE": 61.111111111111114 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6659498207885304, + "weighted_f1": 0.692299687825182, + "macro_precision": 0.6576433542377814, + "macro_recall": 0.7196408529741863, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7311827956989247, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 66.59498207885304, + "weighted_f1_pct": 69.2299687825182, + "macro_precision_pct": 65.76433542377814, + "macro_recall_pct": 71.96408529741863, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 73.11827956989248, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6548223350253807, + "macro_f1": 0.6349317422488153, + "weighted_f1": 0.6603065556389811, + "macro_precision": 0.6201288244766505, + "macro_recall": 0.6669348350764279, + "per_class_f1": { + "FAVOR": 0.5897435897435898, + "AGAINST": 0.7024390243902439, + "NONE": 0.6126126126126126 + }, + "accuracy_pct": 65.48223350253807, + "macro_f1_pct": 63.493174224881535, + "weighted_f1_pct": 66.03065556389811, + "macro_precision_pct": 62.01288244766505, + "macro_recall_pct": 66.6934835076428, + "per_class_f1_pct": { + "FAVOR": 58.97435897435898, + "AGAINST": 70.24390243902438, + "NONE": 61.261261261261254 + } + } + ], + "accuracy": 0.6778846153846154, + "macro_f1": 0.6755054086529896, + "weighted_f1": 0.680190035766943, + "macro_precision": 0.6693785280741803, + "macro_recall": 0.7001392632524709, + "per_class_f1": { + "FAVOR": 0.6352459016393442, + "AGAINST": 0.6954177897574124, + "NONE": 0.695852534562212 + }, + "accuracy_pct": 67.78846153846155, + "macro_f1_pct": 67.55054086529896, + "weighted_f1_pct": 68.0190035766943, + "macro_precision_pct": 66.93785280741803, + "macro_recall_pct": 70.01392632524708, + "per_class_f1_pct": { + "FAVOR": 63.52459016393443, + "AGAINST": 69.54177897574124, + "NONE": 69.5852534562212 + } + }, + { + "after_task": 6, + "best_epoch": 15, + "best_dev_macro_f1": 0.6638493133102298, + "losses": { + "current_ce": 0.6931076322992643, + "replay_ce": 0.2496615366389354, + "dark_distillation": 0.33236186516781646, + "total": 1.1089501053094863, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6766321115636185, + "weighted_f1": 0.6950235326337785, + "macro_precision": 0.6731471227595258, + "macro_recall": 0.7787506246876562, + "per_class_f1": { + "FAVOR": 0.5675675675675675, + "AGAINST": 0.7123287671232876, + "NONE": 0.75 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 67.66321115636185, + "weighted_f1_pct": 69.50235326337784, + "macro_precision_pct": 67.31471227595259, + "macro_recall_pct": 77.87506246876562, + "per_class_f1_pct": { + "FAVOR": 56.75675675675676, + "AGAINST": 71.23287671232876, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8761061946902655, + "macro_f1": 0.7526873906998599, + "weighted_f1": 0.8817416480685216, + "macro_precision": 0.7355001309243256, + "macro_recall": 0.7880354325931318, + "per_class_f1": { + "FAVOR": 0.9104477611940298, + "AGAINST": 0.46153846153846156, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 87.61061946902655, + "macro_f1_pct": 75.268739069986, + "weighted_f1_pct": 88.17416480685216, + "macro_precision_pct": 73.55001309243256, + "macro_recall_pct": 78.80354325931319, + "per_class_f1_pct": { + "FAVOR": 91.04477611940298, + "AGAINST": 46.15384615384615, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5896460114605678, + "weighted_f1": 0.6112187711188604, + "macro_precision": 0.5821706121472979, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5084745762711864, + "AGAINST": 0.6737967914438503, + "NONE": 0.5866666666666667 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.964601146056786, + "weighted_f1_pct": 61.121877111886036, + "macro_precision_pct": 58.217061214729796, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 50.847457627118644, + "AGAINST": 67.37967914438502, + "NONE": 58.666666666666664 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6525647805393971, + "weighted_f1": 0.6767881817096262, + "macro_precision": 0.6480392156862745, + "macro_recall": 0.710381593714927, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.7103825136612022, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.25647805393972, + "weighted_f1_pct": 67.67881817096261, + "macro_precision_pct": 64.80392156862746, + "macro_recall_pct": 71.03815937149271, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 71.03825136612022, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6598984771573604, + "macro_f1": 0.6510334139823205, + "weighted_f1": 0.6648766556943265, + "macro_precision": 0.6380534670008354, + "macro_recall": 0.7056615083684319, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6802030456852792, + "NONE": 0.6728971962616822 + }, + "accuracy_pct": 65.98984771573603, + "macro_f1_pct": 65.10334139823205, + "weighted_f1_pct": 66.48766556943265, + "macro_precision_pct": 63.80534670008354, + "macro_recall_pct": 70.56615083684319, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.02030456852792, + "NONE": 67.28971962616822 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6408450704225352, + "macro_f1": 0.6472431951444642, + "weighted_f1": 0.6254764677873403, + "macro_precision": 0.6368563685636857, + "macro_recall": 0.6841880341880343, + "per_class_f1": { + "FAVOR": 0.7323943661971831, + "AGAINST": 0.49504950495049505, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 64.08450704225352, + "macro_f1_pct": 64.72431951444642, + "weighted_f1_pct": 62.547646778734034, + "macro_precision_pct": 63.68563685636857, + "macro_recall_pct": 68.41880341880344, + "per_class_f1_pct": { + "FAVOR": 73.23943661971832, + "AGAINST": 49.504950495049506, + "NONE": 71.42857142857143 + } + } + ], + "accuracy": 0.6765913757700205, + "macro_f1": 0.6789733506294212, + "weighted_f1": 0.6758886752653566, + "macro_precision": 0.6749714678269177, + "macro_recall": 0.7094032861078315, + "per_class_f1": { + "FAVOR": 0.6620689655172414, + "AGAINST": 0.6650544135429263, + "NONE": 0.7097966728280961 + }, + "accuracy_pct": 67.65913757700204, + "macro_f1_pct": 67.89733506294212, + "weighted_f1_pct": 67.58886752653565, + "macro_precision_pct": 67.49714678269177, + "macro_recall_pct": 70.94032861078315, + "per_class_f1_pct": { + "FAVOR": 66.20689655172414, + "AGAINST": 66.50544135429263, + "NONE": 70.97966728280961 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.6877609015188634, + "losses": { + "current_ce": 0.6289582167826001, + "replay_ce": 0.21940828836153423, + "dark_distillation": 0.31687278581990136, + "total": 1.0068028972262428, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7033333333333333, + "weighted_f1": 0.7306164383561644, + "macro_precision": 0.6924440880335027, + "macro_recall": 0.7968665667166417, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.76, + "NONE": 0.75 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.33333333333333, + "weighted_f1_pct": 73.06164383561644, + "macro_precision_pct": 69.24440880335027, + "macro_recall_pct": 79.68665667166417, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.0, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6857007290611626, + "weighted_f1": 0.8213385503838054, + "macro_precision": 0.6736694677871148, + "macro_recall": 0.804950855478704, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.36363636363636365, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.57007290611627, + "weighted_f1_pct": 82.13385503838055, + "macro_precision_pct": 67.36694677871148, + "macro_recall_pct": 80.4950855478704, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 36.36363636363637, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6135822214360529, + "weighted_f1": 0.6346977638008596, + "macro_precision": 0.6033031296189191, + "macro_recall": 0.6303558460421206, + "per_class_f1": { + "FAVOR": 0.5045045045045045, + "AGAINST": 0.7046632124352331, + "NONE": 0.631578947368421 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 61.35822214360529, + "weighted_f1_pct": 63.46977638008596, + "macro_precision_pct": 60.330312961891906, + "macro_recall_pct": 63.03558460421206, + "per_class_f1_pct": { + "FAVOR": 50.45045045045045, + "AGAINST": 70.46632124352331, + "NONE": 63.1578947368421 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6700329703135232, + "weighted_f1": 0.698103519303906, + "macro_precision": 0.661120393576534, + "macro_recall": 0.724354657687991, + "per_class_f1": { + "FAVOR": 0.5977011494252874, + "AGAINST": 0.7391304347826086, + "NONE": 0.6732673267326733 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.00329703135232, + "weighted_f1_pct": 69.8103519303906, + "macro_precision_pct": 66.1120393576534, + "macro_recall_pct": 72.4354657687991, + "per_class_f1_pct": { + "FAVOR": 59.77011494252874, + "AGAINST": 73.91304347826086, + "NONE": 67.32673267326733 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7411167512690355, + "macro_f1": 0.7313881929908913, + "weighted_f1": 0.7434558934077867, + "macro_precision": 0.7186318972033258, + "macro_recall": 0.7529012141193298, + "per_class_f1": { + "FAVOR": 0.7246376811594203, + "AGAINST": 0.7677725118483413, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 74.11167512690355, + "macro_f1_pct": 73.13881929908914, + "weighted_f1_pct": 74.34558934077867, + "macro_precision_pct": 71.86318972033257, + "macro_recall_pct": 75.29012141193297, + "per_class_f1_pct": { + "FAVOR": 72.46376811594203, + "AGAINST": 76.77725118483413, + "NONE": 70.17543859649122 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5713784521880473, + "weighted_f1": 0.5516787061014946, + "macro_precision": 0.5620993589743589, + "macro_recall": 0.6055555555555555, + "per_class_f1": { + "FAVOR": 0.6376811594202898, + "AGAINST": 0.40404040404040403, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.13784521880473, + "weighted_f1_pct": 55.167870610149464, + "macro_precision_pct": 56.20993589743589, + "macro_recall_pct": 60.55555555555555, + "per_class_f1_pct": { + "FAVOR": 63.76811594202898, + "AGAINST": 40.4040404040404, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6795023188656119, + "weighted_f1": 0.666788179467017, + "macro_precision": 0.6694764355141714, + "macro_recall": 0.700032427674704, + "per_class_f1": { + "FAVOR": 0.6357615894039735, + "AGAINST": 0.6580645161290323, + "NONE": 0.7446808510638298 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.95023188656118, + "weighted_f1_pct": 66.6788179467017, + "macro_precision_pct": 66.94764355141713, + "macro_recall_pct": 70.0032427674704, + "per_class_f1_pct": { + "FAVOR": 63.576158940397356, + "AGAINST": 65.80645161290323, + "NONE": 74.46808510638297 + } + } + ], + "accuracy": 0.6848381601362862, + "macro_f1": 0.6846530831357932, + "weighted_f1": 0.6846049690343968, + "macro_precision": 0.6775326474079937, + "macro_recall": 0.7027644557587713, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.6883629191321499, + "NONE": 0.7155963302752294 + }, + "accuracy_pct": 68.48381601362861, + "macro_f1_pct": 68.46530831357931, + "weighted_f1_pct": 68.46049690343968, + "macro_precision_pct": 67.75326474079937, + "macro_recall_pct": 70.27644557587713, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 68.83629191321499, + "NONE": 71.55963302752293 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.7352292768959435, + "losses": { + "current_ce": 0.5906174354363178, + "replay_ce": 0.2176864936299946, + "dark_distillation": 0.31402705337690273, + "total": 0.9653174545453942, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7074314574314574, + "weighted_f1": 0.7355823400343949, + "macro_precision": 0.6891082526729244, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.75 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.74314574314575, + "weighted_f1_pct": 73.5582340034395, + "macro_precision_pct": 68.91082526729244, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7203548525457949, + "weighted_f1": 0.8497291640455168, + "macro_precision": 0.7026196703616058, + "macro_recall": 0.8216963960684383, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.42105263157894735, + "NONE": 0.8717948717948718 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.03548525457948, + "weighted_f1_pct": 84.97291640455168, + "macro_precision_pct": 70.26196703616058, + "macro_recall_pct": 82.16963960684383, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 42.10526315789473, + "NONE": 87.17948717948718 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6013859309350852, + "weighted_f1": 0.6282230479847177, + "macro_precision": 0.5964056840345501, + "macro_recall": 0.6074800290486565, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.7035175879396985, + "NONE": 0.5915492957746479 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 60.13859309350852, + "weighted_f1_pct": 62.82230479847178, + "macro_precision_pct": 59.640568403455006, + "macro_recall_pct": 60.74800290486565, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 70.35175879396985, + "NONE": 59.154929577464785 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.6630824372759857, + "weighted_f1": 0.6850502948317725, + "macro_precision": 0.6557692307692308, + "macro_recall": 0.7219416386083052, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.7204301075268817, + "NONE": 0.6451612903225806 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 66.30824372759857, + "weighted_f1_pct": 68.50502948317725, + "macro_precision_pct": 65.57692307692308, + "macro_recall_pct": 72.19416386083051, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 72.04301075268818, + "NONE": 64.51612903225806 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6649746192893401, + "macro_f1": 0.6541619058790477, + "weighted_f1": 0.6688058022731063, + "macro_precision": 0.6380442114595178, + "macro_recall": 0.7014391704657191, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.693069306930693, + "NONE": 0.6415094339622641 + }, + "accuracy_pct": 66.49746192893402, + "macro_f1_pct": 65.41619058790477, + "weighted_f1_pct": 66.88058022731063, + "macro_precision_pct": 63.804421145951785, + "macro_recall_pct": 70.14391704657191, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 69.3069306930693, + "NONE": 64.15094339622641 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6379917787175137, + "weighted_f1": 0.6194204348931706, + "macro_precision": 0.6307095652410157, + "macro_recall": 0.6683760683760683, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.49504950495049505, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 63.79917787175137, + "weighted_f1_pct": 61.94204348931706, + "macro_precision_pct": 63.070956524101575, + "macro_recall_pct": 66.83760683760683, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 49.504950495049506, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.673738353080374, + "weighted_f1": 0.6627401552374849, + "macro_precision": 0.6649572649572649, + "macro_recall": 0.6917440684378353, + "per_class_f1": { + "FAVOR": 0.6490066225165563, + "AGAINST": 0.6410256410256411, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.3738353080374, + "weighted_f1_pct": 66.27401552374849, + "macro_precision_pct": 66.49572649572649, + "macro_recall_pct": 69.17440684378353, + "per_class_f1_pct": { + "FAVOR": 64.90066225165563, + "AGAINST": 64.1025641025641, + "NONE": 73.11827956989248 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7290427437266881, + "weighted_f1": 0.7341221350449818, + "macro_precision": 0.731765018607124, + "macro_recall": 0.7291628050664195, + "per_class_f1": { + "FAVOR": 0.7102803738317757, + "AGAINST": 0.7049180327868853, + "NONE": 0.7719298245614035 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.90427437266881, + "weighted_f1_pct": 73.41221350449818, + "macro_precision_pct": 73.17650186071239, + "macro_recall_pct": 72.91628050664195, + "per_class_f1_pct": { + "FAVOR": 71.02803738317756, + "AGAINST": 70.49180327868852, + "NONE": 77.19298245614034 + } + } + ], + "accuracy": 0.6885007278020379, + "macro_f1": 0.6896110617972729, + "weighted_f1": 0.6881317977474748, + "macro_precision": 0.6844711626263947, + "macro_recall": 0.7029191588022572, + "per_class_f1": { + "FAVOR": 0.6691358024691358, + "AGAINST": 0.6812993854258121, + "NONE": 0.718397997496871 + }, + "accuracy_pct": 68.85007278020379, + "macro_f1_pct": 68.9611061797273, + "weighted_f1_pct": 68.81317977474748, + "macro_precision_pct": 68.44711626263947, + "macro_recall_pct": 70.29191588022572, + "per_class_f1_pct": { + "FAVOR": 66.91358024691358, + "AGAINST": 68.12993854258121, + "NONE": 71.8397997496871 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6448730201517713, + "losses": { + "current_ce": 0.4456216856837273, + "replay_ce": 0.2269111292436719, + "dark_distillation": 0.33741367689023416, + "total": 0.8412396517892679, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7096626856420892, + "weighted_f1": 0.7393530500343373, + "macro_precision": 0.6874499470407399, + "macro_recall": 0.7828535732133933, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.7721518987341772, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 70.96626856420892, + "weighted_f1_pct": 73.93530500343373, + "macro_precision_pct": 68.74499470407399, + "macro_recall_pct": 78.28535732133933, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 77.21518987341773, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6924486803519061, + "weighted_f1": 0.8308560195157398, + "macro_precision": 0.6816702389457808, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.36363636363636365, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.2448680351906, + "weighted_f1_pct": 83.08560195157398, + "macro_precision_pct": 68.16702389457808, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 36.36363636363637, + "NONE": 87.5 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.5986009498138878, + "weighted_f1": 0.6307435705165879, + "macro_precision": 0.602020202020202, + "macro_recall": 0.5994916485112564, + "per_class_f1": { + "FAVOR": 0.46938775510204084, + "AGAINST": 0.7264150943396226, + "NONE": 0.6 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 59.86009498138878, + "weighted_f1_pct": 63.074357051658794, + "macro_precision_pct": 60.2020202020202, + "macro_recall_pct": 59.94916485112564, + "per_class_f1_pct": { + "FAVOR": 46.93877551020408, + "AGAINST": 72.64150943396226, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7088543640267778, + "weighted_f1": 0.7484262122749331, + "macro_precision": 0.6973811339944063, + "macro_recall": 0.7325476992143659, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.8078817733990148, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 70.88543640267778, + "weighted_f1_pct": 74.84262122749331, + "macro_precision_pct": 69.73811339944062, + "macro_recall_pct": 73.25476992143659, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 80.78817733990148, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6672700269191498, + "weighted_f1": 0.6829234677217579, + "macro_precision": 0.6503608177320149, + "macro_recall": 0.695986412800572, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7142857142857143, + "NONE": 0.6296296296296297 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.72700269191498, + "weighted_f1_pct": 68.29234677217579, + "macro_precision_pct": 65.03608177320149, + "macro_recall_pct": 69.59864128005721, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 71.42857142857143, + "NONE": 62.96296296296296 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5751948370960899, + "weighted_f1": 0.5629790803506644, + "macro_precision": 0.5746214194490057, + "macro_recall": 0.5824786324786325, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.45454545454545453, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.519483709608984, + "weighted_f1_pct": 56.297908035066435, + "macro_precision_pct": 57.462141944900566, + "macro_recall_pct": 58.24786324786325, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 45.45454545454545, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6754494932984635, + "weighted_f1": 0.662559333115397, + "macro_precision": 0.6848180456427878, + "macro_recall": 0.697442074135841, + "per_class_f1": { + "FAVOR": 0.6015037593984962, + "AGAINST": 0.6857142857142857, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.54494932984635, + "weighted_f1_pct": 66.2559333115397, + "macro_precision_pct": 68.48180456427878, + "macro_recall_pct": 69.7442074135841, + "per_class_f1_pct": { + "FAVOR": 60.150375939849624, + "AGAINST": 68.57142857142857, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7135238095238096, + "weighted_f1": 0.7224571428571429, + "macro_precision": 0.7194444444444444, + "macro_recall": 0.7104417670682731, + "per_class_f1": { + "FAVOR": 0.68, + "AGAINST": 0.672, + "NONE": 0.7885714285714286 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 71.35238095238095, + "weighted_f1_pct": 72.24571428571429, + "macro_precision_pct": 71.94444444444444, + "macro_recall_pct": 71.04417670682732, + "per_class_f1_pct": { + "FAVOR": 68.0, + "AGAINST": 67.2, + "NONE": 78.85714285714286 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.825, + "macro_f1": 0.8115569192038392, + "weighted_f1": 0.8263479663283769, + "macro_precision": 0.7991636798088412, + "macro_recall": 0.826579302941122, + "per_class_f1": { + "FAVOR": 0.8070175438596491, + "AGAINST": 0.7666666666666667, + "NONE": 0.8609865470852018 + }, + "accuracy_pct": 82.5, + "macro_f1_pct": 81.15569192038392, + "weighted_f1_pct": 82.63479663283769, + "macro_precision_pct": 79.91636798088412, + "macro_recall_pct": 82.6579302941122, + "per_class_f1_pct": { + "FAVOR": 80.7017543859649, + "AGAINST": 76.66666666666667, + "NONE": 86.09865470852019 + } + } + ], + "accuracy": 0.7090216010165185, + "macro_f1": 0.7049716462293852, + "weighted_f1": 0.7080214771442565, + "macro_precision": 0.7046060874516115, + "macro_recall": 0.7068630800080543, + "per_class_f1": { + "FAVOR": 0.6556122448979592, + "AGAINST": 0.7071161048689139, + "NONE": 0.7521865889212828 + }, + "accuracy_pct": 70.90216010165184, + "macro_f1_pct": 70.49716462293853, + "weighted_f1_pct": 70.80214771442564, + "macro_precision_pct": 70.46060874516115, + "macro_recall_pct": 70.68630800080543, + "per_class_f1_pct": { + "FAVOR": 65.56122448979592, + "AGAINST": 70.71161048689139, + "NONE": 75.21865889212827 + } + }, + { + "after_task": 10, + "best_epoch": 4, + "best_dev_macro_f1": 0.6724870537736035, + "losses": { + "current_ce": 0.6312720719056252, + "replay_ce": 0.22631243329781753, + "dark_distillation": 0.33656182885169983, + "total": 1.0258654163433956, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7042724542724542, + "weighted_f1": 0.7266949390237062, + "macro_precision": 0.6866757000903343, + "macro_recall": 0.7950274862568717, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.75, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.42724542724542, + "weighted_f1_pct": 72.66949390237062, + "macro_precision_pct": 68.66757000903343, + "macro_recall_pct": 79.50274862568718, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 75.0, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6958754456546625, + "weighted_f1": 0.8510294090829753, + "macro_precision": 0.6850158895999302, + "macro_recall": 0.7631598107025845, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.3333333333333333, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 69.58754456546626, + "weighted_f1_pct": 85.10294090829753, + "macro_precision_pct": 68.50158895999301, + "macro_recall_pct": 76.31598107025846, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 33.33333333333333, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.5980952380952381, + "weighted_f1": 0.6272481203007518, + "macro_precision": 0.5998389694041868, + "macro_recall": 0.599128540305011, + "per_class_f1": { + "FAVOR": 0.48, + "AGAINST": 0.7142857142857143, + "NONE": 0.6 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 59.80952380952381, + "weighted_f1_pct": 62.724812030075185, + "macro_precision_pct": 59.98389694041868, + "macro_recall_pct": 59.912854030501094, + "per_class_f1_pct": { + "FAVOR": 48.0, + "AGAINST": 71.42857142857143, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6813132250261148, + "weighted_f1": 0.7145526408197794, + "macro_precision": 0.6714050400034228, + "macro_recall": 0.7222783389450056, + "per_class_f1": { + "FAVOR": 0.5813953488372093, + "AGAINST": 0.7604166666666666, + "NONE": 0.7021276595744681 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 68.13132250261148, + "weighted_f1_pct": 71.45526408197794, + "macro_precision_pct": 67.14050400034228, + "macro_recall_pct": 72.22783389450056, + "per_class_f1_pct": { + "FAVOR": 58.139534883720934, + "AGAINST": 76.04166666666666, + "NONE": 70.2127659574468 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6622393075825238, + "weighted_f1": 0.6726386100423997, + "macro_precision": 0.6440631808278867, + "macro_recall": 0.7079751181781374, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.6896551724137931, + "NONE": 0.6542056074766355 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.22393075825238, + "weighted_f1_pct": 67.26386100423997, + "macro_precision_pct": 64.40631808278867, + "macro_recall_pct": 70.79751181781374, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 68.96551724137932, + "NONE": 65.42056074766354 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.574048555564355, + "weighted_f1": 0.5605857631998798, + "macro_precision": 0.5703183912861332, + "macro_recall": 0.5790598290598291, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.46956521739130436, + "NONE": 0.6296296296296297 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 57.4048555564355, + "weighted_f1_pct": 56.058576319987985, + "macro_precision_pct": 57.03183912861331, + "macro_recall_pct": 57.90598290598291, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 46.95652173913044, + "NONE": 62.96296296296296 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6861699501308447, + "weighted_f1": 0.6791919715597688, + "macro_precision": 0.6811619414217681, + "macro_recall": 0.6926782943043106, + "per_class_f1": { + "FAVOR": 0.6496815286624203, + "AGAINST": 0.6878980891719745, + "NONE": 0.7209302325581395 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.61699501308448, + "weighted_f1_pct": 67.91919715597689, + "macro_precision_pct": 68.11619414217681, + "macro_recall_pct": 69.26782943043106, + "per_class_f1_pct": { + "FAVOR": 64.96815286624204, + "AGAINST": 68.78980891719745, + "NONE": 72.09302325581395 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7051904438915098, + "weighted_f1": 0.7111515974209424, + "macro_precision": 0.706980056980057, + "macro_recall": 0.7052363299351251, + "per_class_f1": { + "FAVOR": 0.6862745098039216, + "AGAINST": 0.6715328467153284, + "NONE": 0.7577639751552795 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.51904438915098, + "weighted_f1_pct": 71.11515974209424, + "macro_precision_pct": 70.6980056980057, + "macro_recall_pct": 70.5236329935125, + "per_class_f1_pct": { + "FAVOR": 68.62745098039215, + "AGAINST": 67.15328467153284, + "NONE": 75.77639751552795 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7991414177492491, + "weighted_f1": 0.8181018154728497, + "macro_precision": 0.7838421800213219, + "macro_recall": 0.8207822014918467, + "per_class_f1": { + "FAVOR": 0.7931034482758621, + "AGAINST": 0.7419354838709677, + "NONE": 0.8623853211009175 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 79.9141417749249, + "weighted_f1_pct": 81.81018154728497, + "macro_precision_pct": 78.38421800213219, + "macro_recall_pct": 82.07822014918466, + "per_class_f1_pct": { + "FAVOR": 79.3103448275862, + "AGAINST": 74.19354838709677, + "NONE": 86.23853211009175 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.68, + "macro_f1": 0.6421550102590051, + "weighted_f1": 0.6791576032999622, + "macro_precision": 0.6450375824512963, + "macro_recall": 0.6395844716233067, + "per_class_f1": { + "FAVOR": 0.7464114832535885, + "AGAINST": 0.4578313253012048, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 64.21550102590051, + "weighted_f1_pct": 67.91576032999622, + "macro_precision_pct": 64.50375824512963, + "macro_recall_pct": 63.958447162330664, + "per_class_f1_pct": { + "FAVOR": 74.64114832535886, + "AGAINST": 45.78313253012048, + "NONE": 72.22222222222221 + } + } + ], + "accuracy": 0.6984216459977453, + "macro_f1": 0.700090864346493, + "weighted_f1": 0.6981148156216775, + "macro_precision": 0.6977710977628747, + "macro_recall": 0.7039091000431948, + "per_class_f1": { + "FAVOR": 0.6806083650190115, + "AGAINST": 0.6757728253055356, + "NONE": 0.7438914027149321 + }, + "accuracy_pct": 69.84216459977452, + "macro_f1_pct": 70.0090864346493, + "weighted_f1_pct": 69.81148156216776, + "macro_precision_pct": 69.77710977628747, + "macro_recall_pct": 70.39091000431948, + "per_class_f1_pct": { + "FAVOR": 68.06083650190115, + "AGAINST": 67.57728253055356, + "NONE": 74.38914027149322 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6748501048225958, + "average_forgetting": 0.03246032071069719, + "backward_transfer": 0.0018437183907823822, + "final_mean_target_macro_f1_pct": 67.48501048225958, + "average_forgetting_pct": 3.246032071069719, + "backward_transfer_pct": 0.18437183907823823, + "final_pooled_accuracy": 0.6984216459977453, + "final_pooled_accuracy_pct": 69.84216459977452, + "final_pooled_macro_f1": 0.700090864346493, + "final_pooled_macro_f1_pct": 70.0090864346493, + "final_pooled_weighted_f1": 0.6981148156216775, + "final_pooled_weighted_f1_pct": 69.81148156216776, + "final_pooled_macro_precision": 0.6977710977628747, + "final_pooled_macro_precision_pct": 69.77710977628747, + "final_pooled_macro_recall": 0.7039091000431948, + "final_pooled_macro_recall_pct": 70.39091000431948, + "final_pooled_per_class_f1": { + "FAVOR": 0.6806083650190115, + "AGAINST": 0.6757728253055356, + "NONE": 0.7438914027149321 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 68.06083650190115, + "AGAINST": 67.57728253055356, + "NONE": 74.38914027149322 + } + }, + "efficiency": { + "training_wall_time_seconds": 947.6958301349077, + "training_wall_time_minutes": 15.794930502248462, + "peak_gpu_memory_allocated_gb": 10.397007942199707, + "peak_gpu_memory_reserved_gb": 10.580078125 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s1__seed13/metrics.partial.json b/outputs/trace_s6_derpp__x__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..2dd14cbf0829bd590598e69589a5067e52771b52 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s1__seed13/metrics.partial.json @@ -0,0 +1,2169 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7086325439266616, + 0.5945165945165946, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7116641528406235, + 0.641371628745366, + 0.5594731164480781, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7037323701410388, + 0.7241850145075951, + 0.5746323527807488, + 0.7182016512153498, + null, + null, + null, + null, + null, + null + ], + [ + 0.67046883366121, + 0.7021442860715178, + 0.5939931528166822, + 0.6659498207885304, + 0.6349317422488153, + null, + null, + null, + null, + null + ], + [ + 0.6766321115636185, + 0.7526873906998599, + 0.5896460114605678, + 0.6525647805393971, + 0.6510334139823205, + 0.6472431951444642, + null, + null, + null, + null + ], + [ + 0.7033333333333333, + 0.6857007290611626, + 0.6135822214360529, + 0.6700329703135232, + 0.7313881929908913, + 0.5713784521880473, + 0.6795023188656119, + null, + null, + null + ], + [ + 0.7074314574314574, + 0.7203548525457949, + 0.6013859309350852, + 0.6630824372759857, + 0.6541619058790477, + 0.6379917787175137, + 0.673738353080374, + 0.7290427437266881, + null, + null + ], + [ + 0.7096626856420892, + 0.6924486803519061, + 0.5986009498138878, + 0.7088543640267778, + 0.6672700269191498, + 0.5751948370960899, + 0.6754494932984635, + 0.7135238095238096, + 0.8115569192038392, + null + ], + [ + 0.7042724542724542, + 0.6958754456546625, + 0.5980952380952381, + 0.6813132250261148, + 0.6622393075825238, + 0.574048555564355, + 0.6861699501308447, + 0.7051904438915098, + 0.7991414177492491, + 0.6421550102590051 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6840048840048839, + "losses": { + "current_ce": 0.742766341141292, + "replay_ce": 0.25674224751336233, + "dark_distillation": 0.361050540634564, + "total": 1.18003385918481, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7086325439266616, + "weighted_f1": 0.7367864206703852, + "macro_precision": 0.6917552558871426, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.7662337662337663, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.86325439266616, + "weighted_f1_pct": 73.67864206703852, + "macro_precision_pct": 69.17552558871427, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 76.62337662337663, + "NONE": 77.14285714285715 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5945165945165946, + "weighted_f1": 0.8493276634869555, + "macro_precision": 0.5920095408467502, + "macro_recall": 0.5975003033612426, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.0, + "NONE": 0.9047619047619048 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 59.45165945165945, + "weighted_f1_pct": 84.93276634869555, + "macro_precision_pct": 59.20095408467502, + "macro_recall_pct": 59.75003033612426, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 0.0, + "NONE": 90.47619047619048 + } + } + ], + "accuracy": 0.7799227799227799, + "macro_f1": 0.7812226797592651, + "weighted_f1": 0.7746856834661713, + "macro_precision": 0.7922095869857064, + "macro_recall": 0.7948816460694541, + "per_class_f1": { + "FAVOR": 0.78, + "AGAINST": 0.7195121951219512, + "NONE": 0.8441558441558441 + }, + "accuracy_pct": 77.99227799227799, + "macro_f1_pct": 78.1222679759265, + "weighted_f1_pct": 77.46856834661713, + "macro_precision_pct": 79.22095869857064, + "macro_recall_pct": 79.48816460694542, + "per_class_f1_pct": { + "FAVOR": 78.0, + "AGAINST": 71.95121951219512, + "NONE": 84.4155844155844 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6090658702505393, + "losses": { + "current_ce": 0.6782336643247894, + "replay_ce": 0.22404358504396496, + "dark_distillation": 0.3302716556823615, + "total": 1.0674130804610975, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7116641528406235, + "weighted_f1": 0.7454182937664, + "macro_precision": 0.694969315245478, + "macro_recall": 0.7865317341329335, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.782051282051282, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.16641528406234, + "weighted_f1_pct": 74.54182937664, + "macro_precision_pct": 69.4969315245478, + "macro_recall_pct": 78.65317341329336, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 78.2051282051282, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.641371628745366, + "weighted_f1": 0.8048368281004376, + "macro_precision": 0.6467673814165043, + "macro_recall": 0.7283339400558185, + "per_class_f1": { + "FAVOR": 0.8099173553719008, + "AGAINST": 0.25, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 64.1371628745366, + "weighted_f1_pct": 80.48368281004376, + "macro_precision_pct": 64.67673814165042, + "macro_recall_pct": 72.83339400558185, + "per_class_f1_pct": { + "FAVOR": 80.99173553719008, + "AGAINST": 25.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5594731164480781, + "weighted_f1": 0.5900438372374472, + "macro_precision": 0.5543874115302687, + "macro_recall": 0.5693536673928831, + "per_class_f1": { + "FAVOR": 0.4854368932038835, + "AGAINST": 0.6666666666666666, + "NONE": 0.5263157894736842 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 55.947311644807804, + "weighted_f1_pct": 59.00438372374472, + "macro_precision_pct": 55.43874115302687, + "macro_recall_pct": 56.935366739288305, + "per_class_f1_pct": { + "FAVOR": 48.54368932038835, + "AGAINST": 66.66666666666666, + "NONE": 52.63157894736842 + } + } + ], + "accuracy": 0.6815144766146993, + "macro_f1": 0.6838332194777023, + "weighted_f1": 0.6809023259620387, + "macro_precision": 0.6776791079406271, + "macro_recall": 0.6936577425236474, + "per_class_f1": { + "FAVOR": 0.6438356164383562, + "AGAINST": 0.6876640419947506, + "NONE": 0.72 + }, + "accuracy_pct": 68.15144766146993, + "macro_f1_pct": 68.38332194777023, + "weighted_f1_pct": 68.09023259620388, + "macro_precision_pct": 67.76791079406271, + "macro_recall_pct": 69.36577425236474, + "per_class_f1_pct": { + "FAVOR": 64.38356164383562, + "AGAINST": 68.76640419947506, + "NONE": 72.0 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.6498013902681231, + "losses": { + "current_ce": 0.6360149974172765, + "replay_ce": 0.20808983025225727, + "dark_distillation": 0.33606521649794147, + "total": 1.0121374368667602, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7037323701410388, + "weighted_f1": 0.7326734947764253, + "macro_precision": 0.6933806146572103, + "macro_recall": 0.7889955022488756, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.7631578947368421, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.37323701410388, + "weighted_f1_pct": 73.26734947764253, + "macro_precision_pct": 69.33806146572104, + "macro_recall_pct": 78.89955022488756, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 76.31578947368422, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7241850145075951, + "weighted_f1": 0.8470789210155467, + "macro_precision": 0.6974233436901737, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.8709677419354839, + "AGAINST": 0.4444444444444444, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.4185014507595, + "weighted_f1_pct": 84.70789210155468, + "macro_precision_pct": 69.74233436901737, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 87.09677419354838, + "AGAINST": 44.44444444444444, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6, + "macro_f1": 0.5746323527807488, + "weighted_f1": 0.5998972157087302, + "macro_precision": 0.5712692890910713, + "macro_recall": 0.5787944807552651, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.6699507389162561, + "NONE": 0.5633802816901409 + }, + "accuracy_pct": 60.0, + "macro_f1_pct": 57.463235278074876, + "weighted_f1_pct": 59.98972157087302, + "macro_precision_pct": 57.126928909107136, + "macro_recall_pct": 57.879448075526504, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 66.99507389162561, + "NONE": 56.33802816901409 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7182016512153498, + "weighted_f1": 0.746434472877248, + "macro_precision": 0.7022812667740205, + "macro_recall": 0.7455106621773289, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.79, + "NONE": 0.7070707070707071 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 71.82016512153498, + "weighted_f1_pct": 74.6434472877248, + "macro_precision_pct": 70.22812667740205, + "macro_recall_pct": 74.55106621773288, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 79.0, + "NONE": 70.70707070707071 + } + } + ], + "accuracy": 0.710236220472441, + "macro_f1": 0.7076228776661283, + "weighted_f1": 0.7111298601901942, + "macro_precision": 0.6999233219299281, + "macro_recall": 0.721446893544294, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7294938917975567, + "NONE": 0.7267080745341615 + }, + "accuracy_pct": 71.02362204724409, + "macro_f1_pct": 70.76228776661283, + "weighted_f1_pct": 71.11298601901942, + "macro_precision_pct": 69.99233219299282, + "macro_recall_pct": 72.1446893544294, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 72.94938917975567, + "NONE": 72.67080745341616 + } + }, + { + "after_task": 5, + "best_epoch": 4, + "best_dev_macro_f1": 0.687115687115687, + "losses": { + "current_ce": 0.5763537430402005, + "replay_ce": 0.1827620306701371, + "dark_distillation": 0.3391505776932745, + "total": 0.9286910642277111, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.67046883366121, + "weighted_f1": 0.691956730653933, + "macro_precision": 0.672269705603039, + "macro_recall": 0.767256371814093, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.7123287671232876, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 67.04688336612101, + "weighted_f1_pct": 69.19567306539331, + "macro_precision_pct": 67.2269705603039, + "macro_recall_pct": 76.7256371814093, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 71.23287671232876, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7021442860715178, + "weighted_f1": 0.8452115794435334, + "macro_precision": 0.6807726386673755, + "macro_recall": 0.7694697245479918, + "per_class_f1": { + "FAVOR": 0.8709677419354839, + "AGAINST": 0.375, + "NONE": 0.8604651162790697 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 70.21442860715177, + "weighted_f1_pct": 84.52115794435335, + "macro_precision_pct": 68.07726386673755, + "macro_recall_pct": 76.94697245479918, + "per_class_f1_pct": { + "FAVOR": 87.09677419354838, + "AGAINST": 37.5, + "NONE": 86.04651162790698 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5939931528166822, + "weighted_f1": 0.6105508870214753, + "macro_precision": 0.5882079203313323, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5042016806722689, + "AGAINST": 0.6666666666666666, + "NONE": 0.6111111111111112 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 59.39931528166822, + "weighted_f1_pct": 61.05508870214753, + "macro_precision_pct": 58.82079203313323, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 50.42016806722689, + "AGAINST": 66.66666666666666, + "NONE": 61.111111111111114 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6659498207885304, + "weighted_f1": 0.692299687825182, + "macro_precision": 0.6576433542377814, + "macro_recall": 0.7196408529741863, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7311827956989247, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 66.59498207885304, + "weighted_f1_pct": 69.2299687825182, + "macro_precision_pct": 65.76433542377814, + "macro_recall_pct": 71.96408529741863, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 73.11827956989248, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6548223350253807, + "macro_f1": 0.6349317422488153, + "weighted_f1": 0.6603065556389811, + "macro_precision": 0.6201288244766505, + "macro_recall": 0.6669348350764279, + "per_class_f1": { + "FAVOR": 0.5897435897435898, + "AGAINST": 0.7024390243902439, + "NONE": 0.6126126126126126 + }, + "accuracy_pct": 65.48223350253807, + "macro_f1_pct": 63.493174224881535, + "weighted_f1_pct": 66.03065556389811, + "macro_precision_pct": 62.01288244766505, + "macro_recall_pct": 66.6934835076428, + "per_class_f1_pct": { + "FAVOR": 58.97435897435898, + "AGAINST": 70.24390243902438, + "NONE": 61.261261261261254 + } + } + ], + "accuracy": 0.6778846153846154, + "macro_f1": 0.6755054086529896, + "weighted_f1": 0.680190035766943, + "macro_precision": 0.6693785280741803, + "macro_recall": 0.7001392632524709, + "per_class_f1": { + "FAVOR": 0.6352459016393442, + "AGAINST": 0.6954177897574124, + "NONE": 0.695852534562212 + }, + "accuracy_pct": 67.78846153846155, + "macro_f1_pct": 67.55054086529896, + "weighted_f1_pct": 68.0190035766943, + "macro_precision_pct": 66.93785280741803, + "macro_recall_pct": 70.01392632524708, + "per_class_f1_pct": { + "FAVOR": 63.52459016393443, + "AGAINST": 69.54177897574124, + "NONE": 69.5852534562212 + } + }, + { + "after_task": 6, + "best_epoch": 15, + "best_dev_macro_f1": 0.6638493133102298, + "losses": { + "current_ce": 0.6931076322992643, + "replay_ce": 0.2496615366389354, + "dark_distillation": 0.33236186516781646, + "total": 1.1089501053094863, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6766321115636185, + "weighted_f1": 0.6950235326337785, + "macro_precision": 0.6731471227595258, + "macro_recall": 0.7787506246876562, + "per_class_f1": { + "FAVOR": 0.5675675675675675, + "AGAINST": 0.7123287671232876, + "NONE": 0.75 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 67.66321115636185, + "weighted_f1_pct": 69.50235326337784, + "macro_precision_pct": 67.31471227595259, + "macro_recall_pct": 77.87506246876562, + "per_class_f1_pct": { + "FAVOR": 56.75675675675676, + "AGAINST": 71.23287671232876, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8761061946902655, + "macro_f1": 0.7526873906998599, + "weighted_f1": 0.8817416480685216, + "macro_precision": 0.7355001309243256, + "macro_recall": 0.7880354325931318, + "per_class_f1": { + "FAVOR": 0.9104477611940298, + "AGAINST": 0.46153846153846156, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 87.61061946902655, + "macro_f1_pct": 75.268739069986, + "weighted_f1_pct": 88.17416480685216, + "macro_precision_pct": 73.55001309243256, + "macro_recall_pct": 78.80354325931319, + "per_class_f1_pct": { + "FAVOR": 91.04477611940298, + "AGAINST": 46.15384615384615, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5896460114605678, + "weighted_f1": 0.6112187711188604, + "macro_precision": 0.5821706121472979, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5084745762711864, + "AGAINST": 0.6737967914438503, + "NONE": 0.5866666666666667 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.964601146056786, + "weighted_f1_pct": 61.121877111886036, + "macro_precision_pct": 58.217061214729796, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 50.847457627118644, + "AGAINST": 67.37967914438502, + "NONE": 58.666666666666664 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6525647805393971, + "weighted_f1": 0.6767881817096262, + "macro_precision": 0.6480392156862745, + "macro_recall": 0.710381593714927, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.7103825136612022, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.25647805393972, + "weighted_f1_pct": 67.67881817096261, + "macro_precision_pct": 64.80392156862746, + "macro_recall_pct": 71.03815937149271, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 71.03825136612022, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6598984771573604, + "macro_f1": 0.6510334139823205, + "weighted_f1": 0.6648766556943265, + "macro_precision": 0.6380534670008354, + "macro_recall": 0.7056615083684319, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6802030456852792, + "NONE": 0.6728971962616822 + }, + "accuracy_pct": 65.98984771573603, + "macro_f1_pct": 65.10334139823205, + "weighted_f1_pct": 66.48766556943265, + "macro_precision_pct": 63.80534670008354, + "macro_recall_pct": 70.56615083684319, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.02030456852792, + "NONE": 67.28971962616822 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6408450704225352, + "macro_f1": 0.6472431951444642, + "weighted_f1": 0.6254764677873403, + "macro_precision": 0.6368563685636857, + "macro_recall": 0.6841880341880343, + "per_class_f1": { + "FAVOR": 0.7323943661971831, + "AGAINST": 0.49504950495049505, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 64.08450704225352, + "macro_f1_pct": 64.72431951444642, + "weighted_f1_pct": 62.547646778734034, + "macro_precision_pct": 63.68563685636857, + "macro_recall_pct": 68.41880341880344, + "per_class_f1_pct": { + "FAVOR": 73.23943661971832, + "AGAINST": 49.504950495049506, + "NONE": 71.42857142857143 + } + } + ], + "accuracy": 0.6765913757700205, + "macro_f1": 0.6789733506294212, + "weighted_f1": 0.6758886752653566, + "macro_precision": 0.6749714678269177, + "macro_recall": 0.7094032861078315, + "per_class_f1": { + "FAVOR": 0.6620689655172414, + "AGAINST": 0.6650544135429263, + "NONE": 0.7097966728280961 + }, + "accuracy_pct": 67.65913757700204, + "macro_f1_pct": 67.89733506294212, + "weighted_f1_pct": 67.58886752653565, + "macro_precision_pct": 67.49714678269177, + "macro_recall_pct": 70.94032861078315, + "per_class_f1_pct": { + "FAVOR": 66.20689655172414, + "AGAINST": 66.50544135429263, + "NONE": 70.97966728280961 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.6877609015188634, + "losses": { + "current_ce": 0.6289582167826001, + "replay_ce": 0.21940828836153423, + "dark_distillation": 0.31687278581990136, + "total": 1.0068028972262428, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7033333333333333, + "weighted_f1": 0.7306164383561644, + "macro_precision": 0.6924440880335027, + "macro_recall": 0.7968665667166417, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.76, + "NONE": 0.75 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.33333333333333, + "weighted_f1_pct": 73.06164383561644, + "macro_precision_pct": 69.24440880335027, + "macro_recall_pct": 79.68665667166417, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.0, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6857007290611626, + "weighted_f1": 0.8213385503838054, + "macro_precision": 0.6736694677871148, + "macro_recall": 0.804950855478704, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.36363636363636365, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.57007290611627, + "weighted_f1_pct": 82.13385503838055, + "macro_precision_pct": 67.36694677871148, + "macro_recall_pct": 80.4950855478704, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 36.36363636363637, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6135822214360529, + "weighted_f1": 0.6346977638008596, + "macro_precision": 0.6033031296189191, + "macro_recall": 0.6303558460421206, + "per_class_f1": { + "FAVOR": 0.5045045045045045, + "AGAINST": 0.7046632124352331, + "NONE": 0.631578947368421 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 61.35822214360529, + "weighted_f1_pct": 63.46977638008596, + "macro_precision_pct": 60.330312961891906, + "macro_recall_pct": 63.03558460421206, + "per_class_f1_pct": { + "FAVOR": 50.45045045045045, + "AGAINST": 70.46632124352331, + "NONE": 63.1578947368421 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6700329703135232, + "weighted_f1": 0.698103519303906, + "macro_precision": 0.661120393576534, + "macro_recall": 0.724354657687991, + "per_class_f1": { + "FAVOR": 0.5977011494252874, + "AGAINST": 0.7391304347826086, + "NONE": 0.6732673267326733 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.00329703135232, + "weighted_f1_pct": 69.8103519303906, + "macro_precision_pct": 66.1120393576534, + "macro_recall_pct": 72.4354657687991, + "per_class_f1_pct": { + "FAVOR": 59.77011494252874, + "AGAINST": 73.91304347826086, + "NONE": 67.32673267326733 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7411167512690355, + "macro_f1": 0.7313881929908913, + "weighted_f1": 0.7434558934077867, + "macro_precision": 0.7186318972033258, + "macro_recall": 0.7529012141193298, + "per_class_f1": { + "FAVOR": 0.7246376811594203, + "AGAINST": 0.7677725118483413, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 74.11167512690355, + "macro_f1_pct": 73.13881929908914, + "weighted_f1_pct": 74.34558934077867, + "macro_precision_pct": 71.86318972033257, + "macro_recall_pct": 75.29012141193297, + "per_class_f1_pct": { + "FAVOR": 72.46376811594203, + "AGAINST": 76.77725118483413, + "NONE": 70.17543859649122 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5713784521880473, + "weighted_f1": 0.5516787061014946, + "macro_precision": 0.5620993589743589, + "macro_recall": 0.6055555555555555, + "per_class_f1": { + "FAVOR": 0.6376811594202898, + "AGAINST": 0.40404040404040403, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.13784521880473, + "weighted_f1_pct": 55.167870610149464, + "macro_precision_pct": 56.20993589743589, + "macro_recall_pct": 60.55555555555555, + "per_class_f1_pct": { + "FAVOR": 63.76811594202898, + "AGAINST": 40.4040404040404, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6795023188656119, + "weighted_f1": 0.666788179467017, + "macro_precision": 0.6694764355141714, + "macro_recall": 0.700032427674704, + "per_class_f1": { + "FAVOR": 0.6357615894039735, + "AGAINST": 0.6580645161290323, + "NONE": 0.7446808510638298 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.95023188656118, + "weighted_f1_pct": 66.6788179467017, + "macro_precision_pct": 66.94764355141713, + "macro_recall_pct": 70.0032427674704, + "per_class_f1_pct": { + "FAVOR": 63.576158940397356, + "AGAINST": 65.80645161290323, + "NONE": 74.46808510638297 + } + } + ], + "accuracy": 0.6848381601362862, + "macro_f1": 0.6846530831357932, + "weighted_f1": 0.6846049690343968, + "macro_precision": 0.6775326474079937, + "macro_recall": 0.7027644557587713, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.6883629191321499, + "NONE": 0.7155963302752294 + }, + "accuracy_pct": 68.48381601362861, + "macro_f1_pct": 68.46530831357931, + "weighted_f1_pct": 68.46049690343968, + "macro_precision_pct": 67.75326474079937, + "macro_recall_pct": 70.27644557587713, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 68.83629191321499, + "NONE": 71.55963302752293 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.7352292768959435, + "losses": { + "current_ce": 0.5906174354363178, + "replay_ce": 0.2176864936299946, + "dark_distillation": 0.31402705337690273, + "total": 0.9653174545453942, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7074314574314574, + "weighted_f1": 0.7355823400343949, + "macro_precision": 0.6891082526729244, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.75 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.74314574314575, + "weighted_f1_pct": 73.5582340034395, + "macro_precision_pct": 68.91082526729244, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7203548525457949, + "weighted_f1": 0.8497291640455168, + "macro_precision": 0.7026196703616058, + "macro_recall": 0.8216963960684383, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.42105263157894735, + "NONE": 0.8717948717948718 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.03548525457948, + "weighted_f1_pct": 84.97291640455168, + "macro_precision_pct": 70.26196703616058, + "macro_recall_pct": 82.16963960684383, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 42.10526315789473, + "NONE": 87.17948717948718 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6013859309350852, + "weighted_f1": 0.6282230479847177, + "macro_precision": 0.5964056840345501, + "macro_recall": 0.6074800290486565, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.7035175879396985, + "NONE": 0.5915492957746479 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 60.13859309350852, + "weighted_f1_pct": 62.82230479847178, + "macro_precision_pct": 59.640568403455006, + "macro_recall_pct": 60.74800290486565, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 70.35175879396985, + "NONE": 59.154929577464785 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6774193548387096, + "macro_f1": 0.6630824372759857, + "weighted_f1": 0.6850502948317725, + "macro_precision": 0.6557692307692308, + "macro_recall": 0.7219416386083052, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.7204301075268817, + "NONE": 0.6451612903225806 + }, + "accuracy_pct": 67.74193548387096, + "macro_f1_pct": 66.30824372759857, + "weighted_f1_pct": 68.50502948317725, + "macro_precision_pct": 65.57692307692308, + "macro_recall_pct": 72.19416386083051, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 72.04301075268818, + "NONE": 64.51612903225806 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6649746192893401, + "macro_f1": 0.6541619058790477, + "weighted_f1": 0.6688058022731063, + "macro_precision": 0.6380442114595178, + "macro_recall": 0.7014391704657191, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.693069306930693, + "NONE": 0.6415094339622641 + }, + "accuracy_pct": 66.49746192893402, + "macro_f1_pct": 65.41619058790477, + "weighted_f1_pct": 66.88058022731063, + "macro_precision_pct": 63.804421145951785, + "macro_recall_pct": 70.14391704657191, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 69.3069306930693, + "NONE": 64.15094339622641 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6338028169014085, + "macro_f1": 0.6379917787175137, + "weighted_f1": 0.6194204348931706, + "macro_precision": 0.6307095652410157, + "macro_recall": 0.6683760683760683, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.49504950495049505, + "NONE": 0.7130434782608696 + }, + "accuracy_pct": 63.38028169014085, + "macro_f1_pct": 63.79917787175137, + "weighted_f1_pct": 61.94204348931706, + "macro_precision_pct": 63.070956524101575, + "macro_recall_pct": 66.83760683760683, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 49.504950495049506, + "NONE": 71.30434782608695 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.673738353080374, + "weighted_f1": 0.6627401552374849, + "macro_precision": 0.6649572649572649, + "macro_recall": 0.6917440684378353, + "per_class_f1": { + "FAVOR": 0.6490066225165563, + "AGAINST": 0.6410256410256411, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.3738353080374, + "weighted_f1_pct": 66.27401552374849, + "macro_precision_pct": 66.49572649572649, + "macro_recall_pct": 69.17440684378353, + "per_class_f1_pct": { + "FAVOR": 64.90066225165563, + "AGAINST": 64.1025641025641, + "NONE": 73.11827956989248 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7290427437266881, + "weighted_f1": 0.7341221350449818, + "macro_precision": 0.731765018607124, + "macro_recall": 0.7291628050664195, + "per_class_f1": { + "FAVOR": 0.7102803738317757, + "AGAINST": 0.7049180327868853, + "NONE": 0.7719298245614035 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.90427437266881, + "weighted_f1_pct": 73.41221350449818, + "macro_precision_pct": 73.17650186071239, + "macro_recall_pct": 72.91628050664195, + "per_class_f1_pct": { + "FAVOR": 71.02803738317756, + "AGAINST": 70.49180327868852, + "NONE": 77.19298245614034 + } + } + ], + "accuracy": 0.6885007278020379, + "macro_f1": 0.6896110617972729, + "weighted_f1": 0.6881317977474748, + "macro_precision": 0.6844711626263947, + "macro_recall": 0.7029191588022572, + "per_class_f1": { + "FAVOR": 0.6691358024691358, + "AGAINST": 0.6812993854258121, + "NONE": 0.718397997496871 + }, + "accuracy_pct": 68.85007278020379, + "macro_f1_pct": 68.9611061797273, + "weighted_f1_pct": 68.81317977474748, + "macro_precision_pct": 68.44711626263947, + "macro_recall_pct": 70.29191588022572, + "per_class_f1_pct": { + "FAVOR": 66.91358024691358, + "AGAINST": 68.12993854258121, + "NONE": 71.8397997496871 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6448730201517713, + "losses": { + "current_ce": 0.4456216856837273, + "replay_ce": 0.2269111292436719, + "dark_distillation": 0.33741367689023416, + "total": 0.8412396517892679, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7096626856420892, + "weighted_f1": 0.7393530500343373, + "macro_precision": 0.6874499470407399, + "macro_recall": 0.7828535732133933, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.7721518987341772, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 70.96626856420892, + "weighted_f1_pct": 73.93530500343373, + "macro_precision_pct": 68.74499470407399, + "macro_recall_pct": 78.28535732133933, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 77.21518987341773, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6924486803519061, + "weighted_f1": 0.8308560195157398, + "macro_precision": 0.6816702389457808, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.36363636363636365, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.2448680351906, + "weighted_f1_pct": 83.08560195157398, + "macro_precision_pct": 68.16702389457808, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 36.36363636363637, + "NONE": 87.5 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.5986009498138878, + "weighted_f1": 0.6307435705165879, + "macro_precision": 0.602020202020202, + "macro_recall": 0.5994916485112564, + "per_class_f1": { + "FAVOR": 0.46938775510204084, + "AGAINST": 0.7264150943396226, + "NONE": 0.6 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 59.86009498138878, + "weighted_f1_pct": 63.074357051658794, + "macro_precision_pct": 60.2020202020202, + "macro_recall_pct": 59.94916485112564, + "per_class_f1_pct": { + "FAVOR": 46.93877551020408, + "AGAINST": 72.64150943396226, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7088543640267778, + "weighted_f1": 0.7484262122749331, + "macro_precision": 0.6973811339944063, + "macro_recall": 0.7325476992143659, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.8078817733990148, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 70.88543640267778, + "weighted_f1_pct": 74.84262122749331, + "macro_precision_pct": 69.73811339944062, + "macro_recall_pct": 73.25476992143659, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 80.78817733990148, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6672700269191498, + "weighted_f1": 0.6829234677217579, + "macro_precision": 0.6503608177320149, + "macro_recall": 0.695986412800572, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7142857142857143, + "NONE": 0.6296296296296297 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.72700269191498, + "weighted_f1_pct": 68.29234677217579, + "macro_precision_pct": 65.03608177320149, + "macro_recall_pct": 69.59864128005721, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 71.42857142857143, + "NONE": 62.96296296296296 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5751948370960899, + "weighted_f1": 0.5629790803506644, + "macro_precision": 0.5746214194490057, + "macro_recall": 0.5824786324786325, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.45454545454545453, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.519483709608984, + "weighted_f1_pct": 56.297908035066435, + "macro_precision_pct": 57.462141944900566, + "macro_recall_pct": 58.24786324786325, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 45.45454545454545, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6754494932984635, + "weighted_f1": 0.662559333115397, + "macro_precision": 0.6848180456427878, + "macro_recall": 0.697442074135841, + "per_class_f1": { + "FAVOR": 0.6015037593984962, + "AGAINST": 0.6857142857142857, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.54494932984635, + "weighted_f1_pct": 66.2559333115397, + "macro_precision_pct": 68.48180456427878, + "macro_recall_pct": 69.7442074135841, + "per_class_f1_pct": { + "FAVOR": 60.150375939849624, + "AGAINST": 68.57142857142857, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7135238095238096, + "weighted_f1": 0.7224571428571429, + "macro_precision": 0.7194444444444444, + "macro_recall": 0.7104417670682731, + "per_class_f1": { + "FAVOR": 0.68, + "AGAINST": 0.672, + "NONE": 0.7885714285714286 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 71.35238095238095, + "weighted_f1_pct": 72.24571428571429, + "macro_precision_pct": 71.94444444444444, + "macro_recall_pct": 71.04417670682732, + "per_class_f1_pct": { + "FAVOR": 68.0, + "AGAINST": 67.2, + "NONE": 78.85714285714286 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.825, + "macro_f1": 0.8115569192038392, + "weighted_f1": 0.8263479663283769, + "macro_precision": 0.7991636798088412, + "macro_recall": 0.826579302941122, + "per_class_f1": { + "FAVOR": 0.8070175438596491, + "AGAINST": 0.7666666666666667, + "NONE": 0.8609865470852018 + }, + "accuracy_pct": 82.5, + "macro_f1_pct": 81.15569192038392, + "weighted_f1_pct": 82.63479663283769, + "macro_precision_pct": 79.91636798088412, + "macro_recall_pct": 82.6579302941122, + "per_class_f1_pct": { + "FAVOR": 80.7017543859649, + "AGAINST": 76.66666666666667, + "NONE": 86.09865470852019 + } + } + ], + "accuracy": 0.7090216010165185, + "macro_f1": 0.7049716462293852, + "weighted_f1": 0.7080214771442565, + "macro_precision": 0.7046060874516115, + "macro_recall": 0.7068630800080543, + "per_class_f1": { + "FAVOR": 0.6556122448979592, + "AGAINST": 0.7071161048689139, + "NONE": 0.7521865889212828 + }, + "accuracy_pct": 70.90216010165184, + "macro_f1_pct": 70.49716462293853, + "weighted_f1_pct": 70.80214771442564, + "macro_precision_pct": 70.46060874516115, + "macro_recall_pct": 70.68630800080543, + "per_class_f1_pct": { + "FAVOR": 65.56122448979592, + "AGAINST": 70.71161048689139, + "NONE": 75.21865889212827 + } + }, + { + "after_task": 10, + "best_epoch": 4, + "best_dev_macro_f1": 0.6724870537736035, + "losses": { + "current_ce": 0.6312720719056252, + "replay_ce": 0.22631243329781753, + "dark_distillation": 0.33656182885169983, + "total": 1.0258654163433956, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7042724542724542, + "weighted_f1": 0.7266949390237062, + "macro_precision": 0.6866757000903343, + "macro_recall": 0.7950274862568717, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.75, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.42724542724542, + "weighted_f1_pct": 72.66949390237062, + "macro_precision_pct": 68.66757000903343, + "macro_recall_pct": 79.50274862568718, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 75.0, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6958754456546625, + "weighted_f1": 0.8510294090829753, + "macro_precision": 0.6850158895999302, + "macro_recall": 0.7631598107025845, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.3333333333333333, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 69.58754456546626, + "weighted_f1_pct": 85.10294090829753, + "macro_precision_pct": 68.50158895999301, + "macro_recall_pct": 76.31598107025846, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 33.33333333333333, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.5980952380952381, + "weighted_f1": 0.6272481203007518, + "macro_precision": 0.5998389694041868, + "macro_recall": 0.599128540305011, + "per_class_f1": { + "FAVOR": 0.48, + "AGAINST": 0.7142857142857143, + "NONE": 0.6 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 59.80952380952381, + "weighted_f1_pct": 62.724812030075185, + "macro_precision_pct": 59.98389694041868, + "macro_recall_pct": 59.912854030501094, + "per_class_f1_pct": { + "FAVOR": 48.0, + "AGAINST": 71.42857142857143, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6813132250261148, + "weighted_f1": 0.7145526408197794, + "macro_precision": 0.6714050400034228, + "macro_recall": 0.7222783389450056, + "per_class_f1": { + "FAVOR": 0.5813953488372093, + "AGAINST": 0.7604166666666666, + "NONE": 0.7021276595744681 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 68.13132250261148, + "weighted_f1_pct": 71.45526408197794, + "macro_precision_pct": 67.14050400034228, + "macro_recall_pct": 72.22783389450056, + "per_class_f1_pct": { + "FAVOR": 58.139534883720934, + "AGAINST": 76.04166666666666, + "NONE": 70.2127659574468 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6622393075825238, + "weighted_f1": 0.6726386100423997, + "macro_precision": 0.6440631808278867, + "macro_recall": 0.7079751181781374, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.6896551724137931, + "NONE": 0.6542056074766355 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.22393075825238, + "weighted_f1_pct": 67.26386100423997, + "macro_precision_pct": 64.40631808278867, + "macro_recall_pct": 70.79751181781374, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 68.96551724137932, + "NONE": 65.42056074766354 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.574048555564355, + "weighted_f1": 0.5605857631998798, + "macro_precision": 0.5703183912861332, + "macro_recall": 0.5790598290598291, + "per_class_f1": { + "FAVOR": 0.6229508196721312, + "AGAINST": 0.46956521739130436, + "NONE": 0.6296296296296297 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 57.4048555564355, + "weighted_f1_pct": 56.058576319987985, + "macro_precision_pct": 57.03183912861331, + "macro_recall_pct": 57.90598290598291, + "per_class_f1_pct": { + "FAVOR": 62.295081967213115, + "AGAINST": 46.95652173913044, + "NONE": 62.96296296296296 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6861699501308447, + "weighted_f1": 0.6791919715597688, + "macro_precision": 0.6811619414217681, + "macro_recall": 0.6926782943043106, + "per_class_f1": { + "FAVOR": 0.6496815286624203, + "AGAINST": 0.6878980891719745, + "NONE": 0.7209302325581395 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.61699501308448, + "weighted_f1_pct": 67.91919715597689, + "macro_precision_pct": 68.11619414217681, + "macro_recall_pct": 69.26782943043106, + "per_class_f1_pct": { + "FAVOR": 64.96815286624204, + "AGAINST": 68.78980891719745, + "NONE": 72.09302325581395 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7051904438915098, + "weighted_f1": 0.7111515974209424, + "macro_precision": 0.706980056980057, + "macro_recall": 0.7052363299351251, + "per_class_f1": { + "FAVOR": 0.6862745098039216, + "AGAINST": 0.6715328467153284, + "NONE": 0.7577639751552795 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.51904438915098, + "weighted_f1_pct": 71.11515974209424, + "macro_precision_pct": 70.6980056980057, + "macro_recall_pct": 70.5236329935125, + "per_class_f1_pct": { + "FAVOR": 68.62745098039215, + "AGAINST": 67.15328467153284, + "NONE": 75.77639751552795 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7991414177492491, + "weighted_f1": 0.8181018154728497, + "macro_precision": 0.7838421800213219, + "macro_recall": 0.8207822014918467, + "per_class_f1": { + "FAVOR": 0.7931034482758621, + "AGAINST": 0.7419354838709677, + "NONE": 0.8623853211009175 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 79.9141417749249, + "weighted_f1_pct": 81.81018154728497, + "macro_precision_pct": 78.38421800213219, + "macro_recall_pct": 82.07822014918466, + "per_class_f1_pct": { + "FAVOR": 79.3103448275862, + "AGAINST": 74.19354838709677, + "NONE": 86.23853211009175 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.68, + "macro_f1": 0.6421550102590051, + "weighted_f1": 0.6791576032999622, + "macro_precision": 0.6450375824512963, + "macro_recall": 0.6395844716233067, + "per_class_f1": { + "FAVOR": 0.7464114832535885, + "AGAINST": 0.4578313253012048, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 64.21550102590051, + "weighted_f1_pct": 67.91576032999622, + "macro_precision_pct": 64.50375824512963, + "macro_recall_pct": 63.958447162330664, + "per_class_f1_pct": { + "FAVOR": 74.64114832535886, + "AGAINST": 45.78313253012048, + "NONE": 72.22222222222221 + } + } + ], + "accuracy": 0.6984216459977453, + "macro_f1": 0.700090864346493, + "weighted_f1": 0.6981148156216775, + "macro_precision": 0.6977710977628747, + "macro_recall": 0.7039091000431948, + "per_class_f1": { + "FAVOR": 0.6806083650190115, + "AGAINST": 0.6757728253055356, + "NONE": 0.7438914027149321 + }, + "accuracy_pct": 69.84216459977452, + "macro_f1_pct": 70.0090864346493, + "weighted_f1_pct": 69.81148156216776, + "macro_precision_pct": 69.77710977628747, + "macro_recall_pct": 70.39091000431948, + "per_class_f1_pct": { + "FAVOR": 68.06083650190115, + "AGAINST": 67.57728253055356, + "NONE": 74.38914027149322 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s1__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__x__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a2b04f1f071d843f67b3ac0b67df2160e71306ef --- /dev/null +++ b/outputs/trace_s6_derpp__x__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__x__s1__seed13/run.log b/outputs/trace_s6_derpp__x__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..96e1531697607d8bbfc79e24da8728f8da411a98 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 05:41:36,674 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__x__s1__seed13 device=cuda +2026-09-02 05:57:25,702 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6748501048225958, 'average_forgetting': 0.03246032071069719, 'backward_transfer': 0.0018437183907823822, 'final_mean_target_macro_f1_pct': 67.48501048225958, 'average_forgetting_pct': 3.246032071069719, 'backward_transfer_pct': 0.18437183907823823, 'final_pooled_accuracy': 0.6984216459977453, 'final_pooled_accuracy_pct': 69.84216459977452, 'final_pooled_macro_f1': 0.700090864346493, 'final_pooled_macro_f1_pct': 70.0090864346493, 'final_pooled_weighted_f1': 0.6981148156216775, 'final_pooled_weighted_f1_pct': 69.81148156216776, 'final_pooled_macro_precision': 0.6977710977628747, 'final_pooled_macro_precision_pct': 69.77710977628747, 'final_pooled_macro_recall': 0.7039091000431948, 'final_pooled_macro_recall_pct': 70.39091000431948, 'final_pooled_per_class_f1': {'FAVOR': 0.6806083650190115, 'AGAINST': 0.6757728253055356, 'NONE': 0.7438914027149321}, 'final_pooled_per_class_f1_pct': {'FAVOR': 68.06083650190115, 'AGAINST': 67.57728253055356, 'NONE': 74.38914027149322}} +2026-09-02 05:57:25,702 [INFO] __main__: efficiency={'training_wall_time_seconds': 947.6958301349077, 'training_wall_time_minutes': 15.794930502248462, 'peak_gpu_memory_allocated_gb': 10.397007942199707, 'peak_gpu_memory_reserved_gb': 10.580078125} diff --git a/outputs/trace_s6_derpp__x__s2__seed13/environment.json b/outputs/trace_s6_derpp__x__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..b47ad4d678e602869297691b5099233ceb871587 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T05:57:33.783745+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s2__seed13/metrics.json b/outputs/trace_s6_derpp__x__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..b0731e5d5afdd43f279d55632e88407d531eb6ec --- /dev/null +++ b/outputs/trace_s6_derpp__x__s2__seed13/metrics.json @@ -0,0 +1,2285 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.6855880496528508, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7209993223564046, + 0.6237140526745072, + 0.5632553606237817, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.76460983884648, + 0.6738486312399355, + 0.5863907860455616, + 0.7040694957733317, + null, + null, + null, + null, + null, + null + ], + [ + 0.751736308060755, + 0.6635325602140946, + 0.5842626966529406, + 0.7249169828143054, + 0.6405580153208237, + null, + null, + null, + null, + null + ], + [ + 0.6435897435897436, + 0.7008263272308216, + 0.5533002778765491, + 0.7166064773016645, + 0.6475320454733945, + 0.7256176220031643, + null, + null, + null, + null + ], + [ + 0.6221139971139972, + 0.626937859115123, + 0.5572363446108066, + 0.6939795322809879, + 0.6375222132146481, + 0.6762501172584546, + 0.5501464097955325, + null, + null, + null + ], + [ + 0.6959841269841269, + 0.7026415498520571, + 0.5988554140580411, + 0.7234930546705747, + 0.645052905326266, + 0.7231150793650793, + 0.5743890518084066, + 0.6980320761207496, + null, + null + ], + [ + 0.6751450476585164, + 0.6481411774229454, + 0.568864468864469, + 0.7017926066706556, + 0.660162353263135, + 0.7374999999999999, + 0.5693402727110591, + 0.6838579256883186, + 0.6735109092832671, + null + ], + [ + 0.6907074869983317, + 0.6761111258680786, + 0.5374294155042818, + 0.6727954497944394, + 0.6218952933168378, + 0.7358434400176685, + 0.5994877856579984, + 0.6858314369611097, + 0.6429926256013213, + 0.7789812424412296 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.5997024266560196, + "losses": { + "current_ce": 0.7091012922200289, + "replay_ce": 0.13691359695160027, + "dark_distillation": 0.35026023983955384, + "total": 1.0211450067433443, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6855880496528508, + "weighted_f1": 0.7172916256625695, + "macro_precision": 0.6708709493712247, + "macro_recall": 0.7226711560044894, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7628865979381443, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.55880496528508, + "weighted_f1_pct": 71.72916256625696, + "macro_precision_pct": 67.08709493712247, + "macro_recall_pct": 72.26711560044893, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.28865979381443, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7625418060200669, + "macro_f1": 0.7628319929959244, + "weighted_f1": 0.7616314361001101, + "macro_precision": 0.7644090923946791, + "macro_recall": 0.7679138368662962, + "per_class_f1": { + "FAVOR": 0.7735849056603774, + "AGAINST": 0.7439613526570048, + "NONE": 0.770949720670391 + }, + "accuracy_pct": 76.2541806020067, + "macro_f1_pct": 76.28319929959244, + "weighted_f1_pct": 76.163143610011, + "macro_precision_pct": 76.44090923946791, + "macro_recall_pct": 76.79138368662962, + "per_class_f1_pct": { + "FAVOR": 77.35849056603774, + "AGAINST": 74.39613526570048, + "NONE": 77.09497206703911 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.5960432482171613, + "losses": { + "current_ce": 0.7478645399212838, + "replay_ce": 0.2459234899530808, + "dark_distillation": 0.346903287867705, + "total": 1.1672396744290987, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7209993223564046, + "weighted_f1": 0.8484982249534481, + "macro_precision": 0.6996527777777777, + "macro_recall": 0.7681349350806941, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.42857142857142855, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 72.09993223564047, + "weighted_f1_pct": 84.84982249534481, + "macro_precision_pct": 69.96527777777777, + "macro_recall_pct": 76.81349350806941, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 42.857142857142854, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6236559139784946, + "macro_f1": 0.6237140526745072, + "weighted_f1": 0.6257743692375852, + "macro_precision": 0.6408730158730158, + "macro_recall": 0.7196969696969697, + "per_class_f1": { + "FAVOR": 0.5825242718446602, + "AGAINST": 0.6219512195121951, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.365591397849464, + "macro_f1_pct": 62.37140526745072, + "weighted_f1_pct": 62.577436923758526, + "macro_precision_pct": 64.08730158730158, + "macro_recall_pct": 71.96969696969697, + "per_class_f1_pct": { + "FAVOR": 58.252427184466015, + "AGAINST": 62.19512195121951, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5632553606237817, + "weighted_f1": 0.5381146528292562, + "macro_precision": 0.5607885304659498, + "macro_recall": 0.6277777777777778, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.35555555555555557, + "NONE": 0.6842105263157895 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 56.32553606237817, + "weighted_f1_pct": 53.81146528292562, + "macro_precision_pct": 56.07885304659498, + "macro_recall_pct": 62.77777777777778, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 35.55555555555556, + "NONE": 68.42105263157895 + } + } + ], + "accuracy": 0.6621315192743764, + "macro_f1": 0.6562666497660318, + "weighted_f1": 0.6444898260980724, + "macro_precision": 0.674604015406202, + "macro_recall": 0.687903373050395, + "per_class_f1": { + "FAVOR": 0.7197452229299363, + "AGAINST": 0.5223880597014925, + "NONE": 0.7266666666666667 + }, + "accuracy_pct": 66.21315192743764, + "macro_f1_pct": 65.62666497660318, + "weighted_f1_pct": 64.44898260980723, + "macro_precision_pct": 67.4604015406202, + "macro_recall_pct": 68.7903373050395, + "per_class_f1_pct": { + "FAVOR": 71.97452229299363, + "AGAINST": 52.23880597014925, + "NONE": 72.66666666666667 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7263591869208339, + "losses": { + "current_ce": 0.6619226377079452, + "replay_ce": 0.2405188981173695, + "dark_distillation": 0.3454159757797269, + "total": 1.0751495221386784, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.76460983884648, + "weighted_f1": 0.8661054065617331, + "macro_precision": 0.7345619658119658, + "macro_recall": 0.8397767261254702, + "per_class_f1": { + "FAVOR": 0.8854961832061069, + "AGAINST": 0.5333333333333333, + "NONE": 0.875 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 76.460983884648, + "weighted_f1_pct": 86.61054065617331, + "macro_precision_pct": 73.45619658119658, + "macro_recall_pct": 83.97767261254701, + "per_class_f1_pct": { + "FAVOR": 88.54961832061069, + "AGAINST": 53.333333333333336, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6738486312399355, + "weighted_f1": 0.6892706872370266, + "macro_precision": 0.6661987102665069, + "macro_recall": 0.7423120089786757, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.7111111111111111, + "NONE": 0.68 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 67.38486312399355, + "weighted_f1_pct": 68.92706872370266, + "macro_precision_pct": 66.61987102665069, + "macro_recall_pct": 74.23120089786757, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 71.11111111111111, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5863907860455616, + "weighted_f1": 0.5623457240987509, + "macro_precision": 0.6025953859804792, + "macro_recall": 0.6606837606837607, + "per_class_f1": { + "FAVOR": 0.6582278481012658, + "AGAINST": 0.35714285714285715, + "NONE": 0.743801652892562 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 58.63907860455616, + "weighted_f1_pct": 56.23457240987509, + "macro_precision_pct": 60.259538598047925, + "macro_recall_pct": 66.06837606837607, + "per_class_f1_pct": { + "FAVOR": 65.82278481012658, + "AGAINST": 35.714285714285715, + "NONE": 74.3801652892562 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7040694957733317, + "weighted_f1": 0.7120125525678602, + "macro_precision": 0.7107921956661453, + "macro_recall": 0.7124961383997529, + "per_class_f1": { + "FAVOR": 0.6949152542372882, + "AGAINST": 0.631578947368421, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.40694957733317, + "weighted_f1_pct": 71.20125525678603, + "macro_precision_pct": 71.07921956661453, + "macro_recall_pct": 71.2496138399753, + "per_class_f1_pct": { + "FAVOR": 69.49152542372882, + "AGAINST": 63.1578947368421, + "NONE": 78.57142857142857 + } + } + ], + "accuracy": 0.7067082683307332, + "macro_f1": 0.7021739720156036, + "weighted_f1": 0.6977168624660642, + "macro_precision": 0.7135357368753953, + "macro_recall": 0.7202111613876321, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.6055979643765903, + "NONE": 0.767590618336887 + }, + "accuracy_pct": 70.67082683307332, + "macro_f1_pct": 70.21739720156036, + "weighted_f1_pct": 69.77168624660642, + "macro_precision_pct": 71.35357368753952, + "macro_recall_pct": 72.0211161387632, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 60.55979643765903, + "NONE": 76.7590618336887 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6402269231537524, + "losses": { + "current_ce": 0.6802984879567073, + "replay_ce": 0.25908706325751085, + "dark_distillation": 0.3671783589399778, + "total": 1.1229747341229366, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.751736308060755, + "weighted_f1": 0.8481168964807573, + "macro_precision": 0.7189614476789928, + "macro_recall": 0.8298264773692513, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.5333333333333333, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 75.1736308060755, + "weighted_f1_pct": 84.81168964807573, + "macro_precision_pct": 71.89614476789929, + "macro_recall_pct": 82.98264773692513, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 53.333333333333336, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6635325602140946, + "weighted_f1": 0.6691246295070646, + "macro_precision": 0.6565217391304348, + "macro_recall": 0.7427609427609427, + "per_class_f1": { + "FAVOR": 0.6526315789473685, + "AGAINST": 0.6779661016949152, + "NONE": 0.66 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 66.35325602140945, + "weighted_f1_pct": 66.91246295070647, + "macro_precision_pct": 65.65217391304347, + "macro_recall_pct": 74.27609427609427, + "per_class_f1_pct": { + "FAVOR": 65.26315789473685, + "AGAINST": 67.79661016949152, + "NONE": 66.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5842626966529406, + "weighted_f1": 0.5564756699844299, + "macro_precision": 0.6065640373422361, + "macro_recall": 0.6615384615384615, + "per_class_f1": { + "FAVOR": 0.6753246753246753, + "AGAINST": 0.34146341463414637, + "NONE": 0.736 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 58.426269665294065, + "weighted_f1_pct": 55.64756699844299, + "macro_precision_pct": 60.65640373422361, + "macro_recall_pct": 66.15384615384615, + "per_class_f1_pct": { + "FAVOR": 67.53246753246754, + "AGAINST": 34.146341463414636, + "NONE": 73.6 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.74, + "macro_f1": 0.7249169828143054, + "weighted_f1": 0.733666571140042, + "macro_precision": 0.7435681953247514, + "macro_recall": 0.7314643188137163, + "per_class_f1": { + "FAVOR": 0.7130434782608696, + "AGAINST": 0.6481481481481481, + "NONE": 0.8135593220338984 + }, + "accuracy_pct": 74.0, + "macro_f1_pct": 72.49169828143054, + "weighted_f1_pct": 73.3666571140042, + "macro_precision_pct": 74.35681953247514, + "macro_recall_pct": 73.14643188137163, + "per_class_f1_pct": { + "FAVOR": 71.30434782608695, + "AGAINST": 64.81481481481481, + "NONE": 81.35593220338984 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.6405580153208237, + "weighted_f1": 0.6760420619860849, + "macro_precision": 0.6475247524752475, + "macro_recall": 0.6484904523739475, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.4444444444444444, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 64.05580153208237, + "weighted_f1_pct": 67.60420619860848, + "macro_precision_pct": 64.75247524752476, + "macro_recall_pct": 64.84904523739475, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 44.44444444444444, + "NONE": 74.19354838709677 + } + } + ], + "accuracy": 0.7015457788347206, + "macro_f1": 0.6900498368273961, + "weighted_f1": 0.6899914146767683, + "macro_precision": 0.7079188602066525, + "macro_recall": 0.701600922596346, + "per_class_f1": { + "FAVOR": 0.7387096774193549, + "AGAINST": 0.5682819383259912, + "NONE": 0.7631578947368421 + }, + "accuracy_pct": 70.15457788347206, + "macro_f1_pct": 69.0049836827396, + "weighted_f1_pct": 68.99914146767682, + "macro_precision_pct": 70.79188602066525, + "macro_recall_pct": 70.1600922596346, + "per_class_f1_pct": { + "FAVOR": 73.87096774193549, + "AGAINST": 56.82819383259912, + "NONE": 76.31578947368422 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.8459824489236253, + "losses": { + "current_ce": 0.4918774113886886, + "replay_ce": 0.2782205246112965, + "dark_distillation": 0.3717112809419632, + "total": 0.9559535768572931, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6435897435897436, + "weighted_f1": 0.8120490129339687, + "macro_precision": 0.6369186369186369, + "macro_recall": 0.678412813978886, + "per_class_f1": { + "FAVOR": 0.8307692307692308, + "AGAINST": 0.25, + "NONE": 0.85 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 64.35897435897436, + "weighted_f1_pct": 81.20490129339687, + "macro_precision_pct": 63.691863691863695, + "macro_recall_pct": 67.8412813978886, + "per_class_f1_pct": { + "FAVOR": 83.07692307692308, + "AGAINST": 25.0, + "NONE": 85.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.7008263272308216, + "weighted_f1": 0.7266881813366046, + "macro_precision": 0.6903866248693834, + "macro_recall": 0.7466329966329965, + "per_class_f1": { + "FAVOR": 0.6590909090909091, + "AGAINST": 0.7692307692307693, + "NONE": 0.6741573033707865 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 70.08263272308216, + "weighted_f1_pct": 72.66881813366047, + "macro_precision_pct": 69.03866248693834, + "macro_recall_pct": 74.66329966329965, + "per_class_f1_pct": { + "FAVOR": 65.9090909090909, + "AGAINST": 76.92307692307693, + "NONE": 67.41573033707866 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5533002778765491, + "weighted_f1": 0.5277433237566921, + "macro_precision": 0.5542027417027416, + "macro_recall": 0.6175213675213675, + "per_class_f1": { + "FAVOR": 0.6410256410256411, + "AGAINST": 0.3409090909090909, + "NONE": 0.6779661016949152 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.33002778765491, + "weighted_f1_pct": 52.77433237566921, + "macro_precision_pct": 55.42027417027416, + "macro_recall_pct": 61.75213675213676, + "per_class_f1_pct": { + "FAVOR": 64.1025641025641, + "AGAINST": 34.090909090909086, + "NONE": 67.79661016949152 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7166064773016645, + "weighted_f1": 0.7199929682616849, + "macro_precision": 0.72055614809238, + "macro_recall": 0.7268612913191226, + "per_class_f1": { + "FAVOR": 0.7107438016528925, + "AGAINST": 0.6890756302521008, + "NONE": 0.75 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 71.66064773016645, + "weighted_f1_pct": 71.99929682616849, + "macro_precision_pct": 72.055614809238, + "macro_recall_pct": 72.68612913191225, + "per_class_f1_pct": { + "FAVOR": 71.07438016528926, + "AGAINST": 68.90756302521008, + "NONE": 75.0 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.6475320454733945, + "weighted_f1": 0.6759216778834782, + "macro_precision": 0.6435395474804342, + "macro_recall": 0.6523459280740834, + "per_class_f1": { + "FAVOR": 0.7164179104477612, + "AGAINST": 0.46511627906976744, + "NONE": 0.7610619469026548 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 64.75320454733945, + "weighted_f1_pct": 67.59216778834782, + "macro_precision_pct": 64.35395474804342, + "macro_recall_pct": 65.23459280740835, + "per_class_f1_pct": { + "FAVOR": 71.64179104477611, + "AGAINST": 46.51162790697674, + "NONE": 76.10619469026548 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7256176220031643, + "weighted_f1": 0.7694739105863053, + "macro_precision": 0.702960102960103, + "macro_recall": 0.7737331334332834, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.8192771084337349, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 72.56176220031642, + "weighted_f1_pct": 76.94739105863053, + "macro_precision_pct": 70.29601029601031, + "macro_recall_pct": 77.37331334332835, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 81.92771084337349, + "NONE": 75.75757575757575 + } + } + ], + "accuracy": 0.7031408308004052, + "macro_f1": 0.7039433347741956, + "weighted_f1": 0.7009754038396915, + "macro_precision": 0.7064144918748888, + "macro_recall": 0.710720371349668, + "per_class_f1": { + "FAVOR": 0.7109144542772862, + "AGAINST": 0.6597014925373135, + "NONE": 0.7412140575079872 + }, + "accuracy_pct": 70.31408308004052, + "macro_f1_pct": 70.39433347741955, + "weighted_f1_pct": 70.09754038396915, + "macro_precision_pct": 70.64144918748887, + "macro_recall_pct": 71.0720371349668, + "per_class_f1_pct": { + "FAVOR": 71.09144542772862, + "AGAINST": 65.97014925373135, + "NONE": 74.12140575079871 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.5939530234882558, + "losses": { + "current_ce": 0.5948785619302229, + "replay_ce": 0.240766610882499, + "dark_distillation": 0.3915195183320479, + "total": 1.031404926560142, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6221139971139972, + "weighted_f1": 0.8268829253342528, + "macro_precision": 0.6233618233618233, + "macro_recall": 0.6298264773692512, + "per_class_f1": { + "FAVOR": 0.8484848484848485, + "AGAINST": 0.14285714285714285, + "NONE": 0.875 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 62.211399711399714, + "weighted_f1_pct": 82.68829253342528, + "macro_precision_pct": 62.33618233618233, + "macro_recall_pct": 62.982647736925124, + "per_class_f1_pct": { + "FAVOR": 84.84848484848484, + "AGAINST": 14.285714285714285, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.626937859115123, + "weighted_f1": 0.6372787594585096, + "macro_precision": 0.6423485693050911, + "macro_recall": 0.7081930415263749, + "per_class_f1": { + "FAVOR": 0.5740740740740741, + "AGAINST": 0.6473988439306358, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.6937859115123, + "weighted_f1_pct": 63.72787594585097, + "macro_precision_pct": 64.23485693050911, + "macro_recall_pct": 70.81930415263749, + "per_class_f1_pct": { + "FAVOR": 57.407407407407405, + "AGAINST": 64.73988439306359, + "NONE": 65.93406593406593 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5572363446108066, + "weighted_f1": 0.5287067535132683, + "macro_precision": 0.5648717948717948, + "macro_recall": 0.6341880341880342, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.32941176470588235, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 55.723634461080664, + "weighted_f1_pct": 52.87067535132684, + "macro_precision_pct": 56.48717948717948, + "macro_recall_pct": 63.418803418803414, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 32.94117647058823, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6939795322809879, + "weighted_f1": 0.7019042998249061, + "macro_precision": 0.7355272441115814, + "macro_recall": 0.7119864071671301, + "per_class_f1": { + "FAVOR": 0.6814814814814815, + "AGAINST": 0.6274509803921569, + "NONE": 0.7730061349693251 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.39795322809879, + "weighted_f1_pct": 70.19042998249061, + "macro_precision_pct": 73.55272441115814, + "macro_recall_pct": 71.19864071671302, + "per_class_f1_pct": { + "FAVOR": 68.14814814814815, + "AGAINST": 62.745098039215684, + "NONE": 77.30061349693251 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.6375222132146481, + "weighted_f1": 0.679235521815927, + "macro_precision": 0.6368191721132898, + "macro_recall": 0.6428417321621206, + "per_class_f1": { + "FAVOR": 0.751219512195122, + "AGAINST": 0.4155844155844156, + "NONE": 0.7457627118644068 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 63.75222132146481, + "weighted_f1_pct": 67.9235521815927, + "macro_precision_pct": 63.68191721132897, + "macro_recall_pct": 64.28417321621205, + "per_class_f1_pct": { + "FAVOR": 75.1219512195122, + "AGAINST": 41.55844155844156, + "NONE": 74.57627118644068 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6762501172584546, + "weighted_f1": 0.7156854838725379, + "macro_precision": 0.6710858585858586, + "macro_recall": 0.7369315342328836, + "per_class_f1": { + "FAVOR": 0.547945205479452, + "AGAINST": 0.759493670886076, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 67.62501172584547, + "weighted_f1_pct": 71.5685483872538, + "macro_precision_pct": 67.10858585858585, + "macro_recall_pct": 73.69315342328836, + "per_class_f1_pct": { + "FAVOR": 54.794520547945204, + "AGAINST": 75.9493670886076, + "NONE": 72.1311475409836 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5501464097955325, + "weighted_f1": 0.5667102365163307, + "macro_precision": 0.5405684754521963, + "macro_recall": 0.5682643427741466, + "per_class_f1": { + "FAVOR": 0.49122807017543857, + "AGAINST": 0.6137566137566137, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 55.014640979553256, + "weighted_f1_pct": 56.67102365163307, + "macro_precision_pct": 54.05684754521963, + "macro_recall_pct": 56.82643427741466, + "per_class_f1_pct": { + "FAVOR": 49.122807017543856, + "AGAINST": 61.37566137566137, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6601529311809685, + "macro_f1": 0.6622403014498427, + "weighted_f1": 0.6546056149499762, + "macro_precision": 0.6706817100231862, + "macro_recall": 0.6783931407825213, + "per_class_f1": { + "FAVOR": 0.6713780918727915, + "AGAINST": 0.5939849624060151, + "NONE": 0.7213578500707214 + }, + "accuracy_pct": 66.01529311809685, + "macro_f1_pct": 66.22403014498427, + "weighted_f1_pct": 65.46056149499762, + "macro_precision_pct": 67.06817100231862, + "macro_recall_pct": 67.83931407825213, + "per_class_f1_pct": { + "FAVOR": 67.13780918727915, + "AGAINST": 59.3984962406015, + "NONE": 72.13578500707214 + } + }, + { + "after_task": 8, + "best_epoch": 14, + "best_dev_macro_f1": 0.6854605776736924, + "losses": { + "current_ce": 0.565870771353895, + "replay_ce": 0.2280369180621523, + "dark_distillation": 0.36453325269800246, + "total": 0.9761743115656304, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6959841269841269, + "weighted_f1": 0.8276439106616097, + "macro_precision": 0.6813292071912761, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.832, + "AGAINST": 0.38095238095238093, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.59841269841269, + "weighted_f1_pct": 82.76439106616097, + "macro_precision_pct": 68.13292071912761, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.2, + "AGAINST": 38.095238095238095, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.7026415498520571, + "weighted_f1": 0.7262068107700501, + "macro_precision": 0.6878637933160686, + "macro_recall": 0.7509539842873177, + "per_class_f1": { + "FAVOR": 0.6744186046511628, + "AGAINST": 0.7668393782383419, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 70.2641549852057, + "weighted_f1_pct": 72.62068107700502, + "macro_precision_pct": 68.78637933160687, + "macro_recall_pct": 75.09539842873177, + "per_class_f1_pct": { + "FAVOR": 67.44186046511628, + "AGAINST": 76.68393782383419, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5988554140580411, + "weighted_f1": 0.570057835236827, + "macro_precision": 0.6059286146400483, + "macro_recall": 0.6495726495726496, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.38202247191011235, + "NONE": 0.7086614173228346 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.88554140580411, + "weighted_f1_pct": 57.005783523682695, + "macro_precision_pct": 60.592861464004834, + "macro_recall_pct": 64.95726495726495, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 38.20224719101123, + "NONE": 70.86614173228347 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.73, + "macro_f1": 0.7234930546705747, + "weighted_f1": 0.7294345503589483, + "macro_precision": 0.7278173594452997, + "macro_recall": 0.7277108433734939, + "per_class_f1": { + "FAVOR": 0.7017543859649122, + "AGAINST": 0.6949152542372882, + "NONE": 0.7738095238095238 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 72.34930546705746, + "weighted_f1_pct": 72.94345503589483, + "macro_precision_pct": 72.78173594452997, + "macro_recall_pct": 72.7710843373494, + "per_class_f1_pct": { + "FAVOR": 70.17543859649122, + "AGAINST": 69.49152542372882, + "NONE": 77.38095238095238 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.645052905326266, + "weighted_f1": 0.6763768536716097, + "macro_precision": 0.6381766381766382, + "macro_recall": 0.6579946482859104, + "per_class_f1": { + "FAVOR": 0.7253886010362695, + "AGAINST": 0.45977011494252873, + "NONE": 0.75 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 64.5052905326266, + "weighted_f1_pct": 67.63768536716097, + "macro_precision_pct": 63.817663817663814, + "macro_recall_pct": 65.79946482859104, + "per_class_f1_pct": { + "FAVOR": 72.53886010362694, + "AGAINST": 45.97701149425287, + "NONE": 75.0 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7231150793650793, + "weighted_f1": 0.7698242824527072, + "macro_precision": 0.7026315789473685, + "macro_recall": 0.7640229885057472, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.8214285714285714, + "NONE": 0.78125 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 72.31150793650794, + "weighted_f1_pct": 76.98242824527071, + "macro_precision_pct": 70.26315789473685, + "macro_recall_pct": 76.40229885057472, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 82.14285714285714, + "NONE": 78.125 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5743890518084066, + "weighted_f1": 0.5758666460873592, + "macro_precision": 0.5652901348553523, + "macro_recall": 0.6089324618736384, + "per_class_f1": { + "FAVOR": 0.532258064516129, + "AGAINST": 0.5909090909090909, + "NONE": 0.6 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 57.43890518084066, + "weighted_f1_pct": 57.586664608735916, + "macro_precision_pct": 56.52901348553523, + "macro_recall_pct": 60.893246187363836, + "per_class_f1_pct": { + "FAVOR": 53.2258064516129, + "AGAINST": 59.09090909090909, + "NONE": 60.0 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6980320761207496, + "weighted_f1": 0.7175603857500334, + "macro_precision": 0.6871572871572872, + "macro_recall": 0.71669847880155, + "per_class_f1": { + "FAVOR": 0.6470588235294118, + "AGAINST": 0.7452830188679245, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.80320761207496, + "weighted_f1_pct": 71.75603857500334, + "macro_precision_pct": 68.71572871572872, + "macro_recall_pct": 71.669847880155, + "per_class_f1_pct": { + "FAVOR": 64.70588235294117, + "AGAINST": 74.52830188679245, + "NONE": 70.17543859649122 + } + } + ], + "accuracy": 0.6935953420669578, + "macro_f1": 0.6956658843009055, + "weighted_f1": 0.6918128833366621, + "macro_precision": 0.6923164830314782, + "macro_recall": 0.7080891012593432, + "per_class_f1": { + "FAVOR": 0.684964200477327, + "AGAINST": 0.6691729323308271, + "NONE": 0.7328605200945626 + }, + "accuracy_pct": 69.35953420669578, + "macro_f1_pct": 69.56658843009055, + "weighted_f1_pct": 69.18128833366622, + "macro_precision_pct": 69.23164830314782, + "macro_recall_pct": 70.80891012593432, + "per_class_f1_pct": { + "FAVOR": 68.4964200477327, + "AGAINST": 66.9172932330827, + "NONE": 73.28605200945626 + } + }, + { + "after_task": 9, + "best_epoch": 14, + "best_dev_macro_f1": 0.6585250563984362, + "losses": { + "current_ce": 0.5479454851812786, + "replay_ce": 0.21342617401054928, + "dark_distillation": 0.3428357641375254, + "total": 0.9327895448321388, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6751450476585164, + "weighted_f1": 0.8263292673534579, + "macro_precision": 0.665018315018315, + "macro_recall": 0.7482344375682562, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.3157894736842105, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.51450476585164, + "weighted_f1_pct": 82.63292673534579, + "macro_precision_pct": 66.5018315018315, + "macro_recall_pct": 74.82344375682563, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 31.57894736842105, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6481411774229454, + "weighted_f1": 0.6637595696052299, + "macro_precision": 0.64975005852216, + "macro_recall": 0.7166105499438832, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6850828729281768, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 64.81411774229454, + "weighted_f1_pct": 66.37595696052298, + "macro_precision_pct": 64.975005852216, + "macro_recall_pct": 71.66105499438832, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.50828729281768, + "NONE": 65.93406593406593 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.568864468864469, + "weighted_f1": 0.5345921683949852, + "macro_precision": 0.5902027027027027, + "macro_recall": 0.6551282051282051, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.3, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 56.8864468864469, + "weighted_f1_pct": 53.459216839498524, + "macro_precision_pct": 59.02027027027027, + "macro_recall_pct": 65.51282051282051, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 30.0, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7017926066706556, + "weighted_f1": 0.7090290536022243, + "macro_precision": 0.7196196196196195, + "macro_recall": 0.7132684584491814, + "per_class_f1": { + "FAVOR": 0.6825396825396826, + "AGAINST": 0.6545454545454545, + "NONE": 0.7682926829268293 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.17926066706556, + "weighted_f1_pct": 70.90290536022243, + "macro_precision_pct": 71.96196196196195, + "macro_recall_pct": 71.32684584491814, + "per_class_f1_pct": { + "FAVOR": 68.25396825396825, + "AGAINST": 65.45454545454545, + "NONE": 76.82926829268293 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.7, + "macro_f1": 0.660162353263135, + "weighted_f1": 0.6974926253687315, + "macro_precision": 0.6591469015305641, + "macro_recall": 0.6619509939898289, + "per_class_f1": { + "FAVOR": 0.7572815533980582, + "AGAINST": 0.4444444444444444, + "NONE": 0.7787610619469026 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 66.0162353263135, + "weighted_f1_pct": 69.74926253687315, + "macro_precision_pct": 65.91469015305641, + "macro_recall_pct": 66.19509939898289, + "per_class_f1_pct": { + "FAVOR": 75.72815533980582, + "AGAINST": 44.44444444444444, + "NONE": 77.87610619469027 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7374999999999999, + "weighted_f1": 0.7835331050228311, + "macro_precision": 0.7165413533834587, + "macro_recall": 0.7791404297851074, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.8333333333333334, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.75, + "weighted_f1_pct": 78.35331050228311, + "macro_precision_pct": 71.65413533834587, + "macro_recall_pct": 77.91404297851075, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 83.33333333333334, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5693402727110591, + "weighted_f1": 0.5874227960094311, + "macro_precision": 0.5702878992352677, + "macro_recall": 0.5907770515613652, + "per_class_f1": { + "FAVOR": 0.5757575757575758, + "AGAINST": 0.6179775280898876, + "NONE": 0.5142857142857142 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 56.93402727110591, + "weighted_f1_pct": 58.742279600943114, + "macro_precision_pct": 57.028789923526766, + "macro_recall_pct": 59.07770515613652, + "per_class_f1_pct": { + "FAVOR": 57.57575757575758, + "AGAINST": 61.79775280898876, + "NONE": 51.42857142857142 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6852791878172588, + "macro_f1": 0.6838579256883186, + "weighted_f1": 0.6858666383075728, + "macro_precision": 0.6659991548933513, + "macro_recall": 0.7490574669127507, + "per_class_f1": { + "FAVOR": 0.6744186046511628, + "AGAINST": 0.6875, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 68.52791878172589, + "macro_f1_pct": 68.38579256883186, + "weighted_f1_pct": 68.58666383075727, + "macro_precision_pct": 66.59991548933513, + "macro_recall_pct": 74.90574669127507, + "per_class_f1_pct": { + "FAVOR": 67.44186046511628, + "AGAINST": 68.75, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6735109092832671, + "weighted_f1": 0.6574522083058669, + "macro_precision": 0.6942552136350586, + "macro_recall": 0.659741041854863, + "per_class_f1": { + "FAVOR": 0.6296296296296297, + "AGAINST": 0.6341463414634146, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 67.35109092832671, + "weighted_f1_pct": 65.74522083058669, + "macro_precision_pct": 69.42552136350587, + "macro_recall_pct": 65.97410418548631, + "per_class_f1_pct": { + "FAVOR": 62.96296296296296, + "AGAINST": 63.41463414634146, + "NONE": 75.67567567567568 + } + } + ], + "accuracy": 0.679161372299873, + "macro_f1": 0.6833621061724205, + "weighted_f1": 0.6764605363906485, + "macro_precision": 0.6839118747964701, + "macro_recall": 0.6971458196306716, + "per_class_f1": { + "FAVOR": 0.6796657381615598, + "AGAINST": 0.6376811594202898, + "NONE": 0.732739420935412 + }, + "accuracy_pct": 67.9161372299873, + "macro_f1_pct": 68.33621061724206, + "weighted_f1_pct": 67.64605363906485, + "macro_precision_pct": 68.39118747964702, + "macro_recall_pct": 69.71458196306716, + "per_class_f1_pct": { + "FAVOR": 67.96657381615599, + "AGAINST": 63.76811594202898, + "NONE": 73.2739420935412 + } + }, + { + "after_task": 10, + "best_epoch": 13, + "best_dev_macro_f1": 0.721058058314146, + "losses": { + "current_ce": 0.4208163469408949, + "replay_ce": 0.20656802905723454, + "dark_distillation": 0.3602398360768954, + "total": 0.8075042930742105, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6907074869983317, + "weighted_f1": 0.8418368454553427, + "macro_precision": 0.6796395455139089, + "macro_recall": 0.7581846863244751, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.3333333333333333, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 69.07074869983317, + "weighted_f1_pct": 84.18368454553426, + "macro_precision_pct": 67.9639545513909, + "macro_recall_pct": 75.8184686324475, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 33.33333333333333, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6761111258680786, + "weighted_f1": 0.7024772568949074, + "macro_precision": 0.6664672041656168, + "macro_recall": 0.7258136924803592, + "per_class_f1": { + "FAVOR": 0.6067415730337079, + "AGAINST": 0.7407407407407407, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 67.61111258680786, + "weighted_f1_pct": 70.24772568949074, + "macro_precision_pct": 66.64672041656168, + "macro_recall_pct": 72.58136924803593, + "per_class_f1_pct": { + "FAVOR": 60.67415730337079, + "AGAINST": 74.07407407407408, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5492957746478874, + "macro_f1": 0.5374294155042818, + "weighted_f1": 0.49773808326192154, + "macro_precision": 0.527403381642512, + "macro_recall": 0.6175213675213675, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.25882352941176473, + "NONE": 0.6611570247933884 + }, + "accuracy_pct": 54.929577464788736, + "macro_f1_pct": 53.74294155042818, + "weighted_f1_pct": 49.77380832619215, + "macro_precision_pct": 52.7403381642512, + "macro_recall_pct": 61.75213675213676, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 25.882352941176475, + "NONE": 66.11570247933885 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.68, + "macro_f1": 0.6727954497944394, + "weighted_f1": 0.6793584375653264, + "macro_precision": 0.6851334776334776, + "macro_recall": 0.6856657398826074, + "per_class_f1": { + "FAVOR": 0.6614173228346457, + "AGAINST": 0.6194690265486725, + "NONE": 0.7375 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 67.27954497944394, + "weighted_f1_pct": 67.93584375653265, + "macro_precision_pct": 68.51334776334777, + "macro_recall_pct": 68.56657398826074, + "per_class_f1_pct": { + "FAVOR": 66.14173228346458, + "AGAINST": 61.94690265486725, + "NONE": 73.75 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.645, + "macro_f1": 0.6218952933168378, + "weighted_f1": 0.6529679232064869, + "macro_precision": 0.6243016830901637, + "macro_recall": 0.6244553720281876, + "per_class_f1": { + "FAVOR": 0.696969696969697, + "AGAINST": 0.42105263157894735, + "NONE": 0.7476635514018691 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 62.18952933168378, + "weighted_f1_pct": 65.29679232064869, + "macro_precision_pct": 62.430168309016366, + "macro_recall_pct": 62.445537202818755, + "per_class_f1_pct": { + "FAVOR": 69.6969696969697, + "AGAINST": 42.10526315789473, + "NONE": 74.76635514018692 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7358434400176685, + "weighted_f1": 0.7816936452294332, + "macro_precision": 0.7160961814722052, + "macro_recall": 0.7694302848575711, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.8304093567251462, + "NONE": 0.8253968253968254 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.58434400176685, + "weighted_f1_pct": 78.16936452294331, + "macro_precision_pct": 71.60961814722052, + "macro_recall_pct": 76.94302848575711, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 83.04093567251462, + "NONE": 82.53968253968253 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5994877856579984, + "weighted_f1": 0.6192920243872091, + "macro_precision": 0.5925163507782846, + "macro_recall": 0.6147421931735657, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6595744680851063, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.94877856579984, + "weighted_f1_pct": 61.92920243872091, + "macro_precision_pct": 59.25163507782846, + "macro_recall_pct": 61.47421931735657, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 65.95744680851064, + "NONE": 55.55555555555556 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6858314369611097, + "weighted_f1": 0.6919016926464668, + "macro_precision": 0.6653784293236625, + "macro_recall": 0.7376629385999506, + "per_class_f1": { + "FAVOR": 0.6904761904761905, + "AGAINST": 0.7064676616915423, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 68.58314369611097, + "weighted_f1_pct": 69.19016926464668, + "macro_precision_pct": 66.53784293236626, + "macro_recall_pct": 73.76629385999506, + "per_class_f1_pct": { + "FAVOR": 69.04761904761905, + "AGAINST": 70.64676616915423, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6429926256013213, + "weighted_f1": 0.6310591359287011, + "macro_precision": 0.6615247300592129, + "macro_recall": 0.6312934781769471, + "per_class_f1": { + "FAVOR": 0.5838509316770186, + "AGAINST": 0.6424242424242425, + "NONE": 0.7027027027027027 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.29926256013214, + "weighted_f1_pct": 63.10591359287011, + "macro_precision_pct": 66.15247300592128, + "macro_recall_pct": 63.129347817694715, + "per_class_f1_pct": { + "FAVOR": 58.38509316770186, + "AGAINST": 64.24242424242425, + "NONE": 70.27027027027027 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7789812424412296, + "weighted_f1": 0.819386864861304, + "macro_precision": 0.762002967314218, + "macro_recall": 0.8047365206285746, + "per_class_f1": { + "FAVOR": 0.6885245901639344, + "AGAINST": 0.7768595041322314, + "NONE": 0.8715596330275229 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 77.89812424412295, + "weighted_f1_pct": 81.9386864861304, + "macro_precision_pct": 76.2002967314218, + "macro_recall_pct": 80.47365206285745, + "per_class_f1_pct": { + "FAVOR": 68.85245901639344, + "AGAINST": 77.68595041322314, + "NONE": 87.1559633027523 + } + } + ], + "accuracy": 0.6877113866967306, + "macro_f1": 0.6899933380976108, + "weighted_f1": 0.687378888819321, + "macro_precision": 0.6888985087023384, + "macro_recall": 0.6968406261081173, + "per_class_f1": { + "FAVOR": 0.6660633484162896, + "AGAINST": 0.6582466567607727, + "NONE": 0.7456700091157703 + }, + "accuracy_pct": 68.77113866967306, + "macro_f1_pct": 68.99933380976108, + "weighted_f1_pct": 68.7378888819321, + "macro_precision_pct": 68.88985087023384, + "macro_recall_pct": 69.68406261081172, + "per_class_f1_pct": { + "FAVOR": 66.60633484162895, + "AGAINST": 65.82466567607726, + "NONE": 74.56700091157703 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6642075302161297, + "average_forgetting": 0.030619168331382465, + "backward_transfer": -0.016260235150664635, + "final_mean_target_macro_f1_pct": 66.42075302161297, + "average_forgetting_pct": 3.0619168331382465, + "backward_transfer_pct": -1.6260235150664635, + "final_pooled_accuracy": 0.6877113866967306, + "final_pooled_accuracy_pct": 68.77113866967306, + "final_pooled_macro_f1": 0.6899933380976108, + "final_pooled_macro_f1_pct": 68.99933380976108, + "final_pooled_weighted_f1": 0.687378888819321, + "final_pooled_weighted_f1_pct": 68.7378888819321, + "final_pooled_macro_precision": 0.6888985087023384, + "final_pooled_macro_precision_pct": 68.88985087023384, + "final_pooled_macro_recall": 0.6968406261081173, + "final_pooled_macro_recall_pct": 69.68406261081172, + "final_pooled_per_class_f1": { + "FAVOR": 0.6660633484162896, + "AGAINST": 0.6582466567607727, + "NONE": 0.7456700091157703 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 66.60633484162895, + "AGAINST": 65.82466567607726, + "NONE": 74.56700091157703 + } + }, + "efficiency": { + "training_wall_time_seconds": 957.960110536078, + "training_wall_time_minutes": 15.966001842267966, + "peak_gpu_memory_allocated_gb": 10.398106575012207, + "peak_gpu_memory_reserved_gb": 10.62109375 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s2__seed13/metrics.partial.json b/outputs/trace_s6_derpp__x__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..a1bf248dfca6582eacf9be00b72261658fe7ce6d --- /dev/null +++ b/outputs/trace_s6_derpp__x__s2__seed13/metrics.partial.json @@ -0,0 +1,2169 @@ +{ + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.6855880496528508, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7209993223564046, + 0.6237140526745072, + 0.5632553606237817, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.76460983884648, + 0.6738486312399355, + 0.5863907860455616, + 0.7040694957733317, + null, + null, + null, + null, + null, + null + ], + [ + 0.751736308060755, + 0.6635325602140946, + 0.5842626966529406, + 0.7249169828143054, + 0.6405580153208237, + null, + null, + null, + null, + null + ], + [ + 0.6435897435897436, + 0.7008263272308216, + 0.5533002778765491, + 0.7166064773016645, + 0.6475320454733945, + 0.7256176220031643, + null, + null, + null, + null + ], + [ + 0.6221139971139972, + 0.626937859115123, + 0.5572363446108066, + 0.6939795322809879, + 0.6375222132146481, + 0.6762501172584546, + 0.5501464097955325, + null, + null, + null + ], + [ + 0.6959841269841269, + 0.7026415498520571, + 0.5988554140580411, + 0.7234930546705747, + 0.645052905326266, + 0.7231150793650793, + 0.5743890518084066, + 0.6980320761207496, + null, + null + ], + [ + 0.6751450476585164, + 0.6481411774229454, + 0.568864468864469, + 0.7017926066706556, + 0.660162353263135, + 0.7374999999999999, + 0.5693402727110591, + 0.6838579256883186, + 0.6735109092832671, + null + ], + [ + 0.6907074869983317, + 0.6761111258680786, + 0.5374294155042818, + 0.6727954497944394, + 0.6218952933168378, + 0.7358434400176685, + 0.5994877856579984, + 0.6858314369611097, + 0.6429926256013213, + 0.7789812424412296 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.5997024266560196, + "losses": { + "current_ce": 0.7091012922200289, + "replay_ce": 0.13691359695160027, + "dark_distillation": 0.35026023983955384, + "total": 1.0211450067433443, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6855880496528508, + "weighted_f1": 0.7172916256625695, + "macro_precision": 0.6708709493712247, + "macro_recall": 0.7226711560044894, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7628865979381443, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.55880496528508, + "weighted_f1_pct": 71.72916256625696, + "macro_precision_pct": 67.08709493712247, + "macro_recall_pct": 72.26711560044893, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.28865979381443, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7625418060200669, + "macro_f1": 0.7628319929959244, + "weighted_f1": 0.7616314361001101, + "macro_precision": 0.7644090923946791, + "macro_recall": 0.7679138368662962, + "per_class_f1": { + "FAVOR": 0.7735849056603774, + "AGAINST": 0.7439613526570048, + "NONE": 0.770949720670391 + }, + "accuracy_pct": 76.2541806020067, + "macro_f1_pct": 76.28319929959244, + "weighted_f1_pct": 76.163143610011, + "macro_precision_pct": 76.44090923946791, + "macro_recall_pct": 76.79138368662962, + "per_class_f1_pct": { + "FAVOR": 77.35849056603774, + "AGAINST": 74.39613526570048, + "NONE": 77.09497206703911 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.5960432482171613, + "losses": { + "current_ce": 0.7478645399212838, + "replay_ce": 0.2459234899530808, + "dark_distillation": 0.346903287867705, + "total": 1.1672396744290987, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7209993223564046, + "weighted_f1": 0.8484982249534481, + "macro_precision": 0.6996527777777777, + "macro_recall": 0.7681349350806941, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.42857142857142855, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 72.09993223564047, + "weighted_f1_pct": 84.84982249534481, + "macro_precision_pct": 69.96527777777777, + "macro_recall_pct": 76.81349350806941, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 42.857142857142854, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6236559139784946, + "macro_f1": 0.6237140526745072, + "weighted_f1": 0.6257743692375852, + "macro_precision": 0.6408730158730158, + "macro_recall": 0.7196969696969697, + "per_class_f1": { + "FAVOR": 0.5825242718446602, + "AGAINST": 0.6219512195121951, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.365591397849464, + "macro_f1_pct": 62.37140526745072, + "weighted_f1_pct": 62.577436923758526, + "macro_precision_pct": 64.08730158730158, + "macro_recall_pct": 71.96969696969697, + "per_class_f1_pct": { + "FAVOR": 58.252427184466015, + "AGAINST": 62.19512195121951, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5632553606237817, + "weighted_f1": 0.5381146528292562, + "macro_precision": 0.5607885304659498, + "macro_recall": 0.6277777777777778, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.35555555555555557, + "NONE": 0.6842105263157895 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 56.32553606237817, + "weighted_f1_pct": 53.81146528292562, + "macro_precision_pct": 56.07885304659498, + "macro_recall_pct": 62.77777777777778, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 35.55555555555556, + "NONE": 68.42105263157895 + } + } + ], + "accuracy": 0.6621315192743764, + "macro_f1": 0.6562666497660318, + "weighted_f1": 0.6444898260980724, + "macro_precision": 0.674604015406202, + "macro_recall": 0.687903373050395, + "per_class_f1": { + "FAVOR": 0.7197452229299363, + "AGAINST": 0.5223880597014925, + "NONE": 0.7266666666666667 + }, + "accuracy_pct": 66.21315192743764, + "macro_f1_pct": 65.62666497660318, + "weighted_f1_pct": 64.44898260980723, + "macro_precision_pct": 67.4604015406202, + "macro_recall_pct": 68.7903373050395, + "per_class_f1_pct": { + "FAVOR": 71.97452229299363, + "AGAINST": 52.23880597014925, + "NONE": 72.66666666666667 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.7263591869208339, + "losses": { + "current_ce": 0.6619226377079452, + "replay_ce": 0.2405188981173695, + "dark_distillation": 0.3454159757797269, + "total": 1.0751495221386784, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.76460983884648, + "weighted_f1": 0.8661054065617331, + "macro_precision": 0.7345619658119658, + "macro_recall": 0.8397767261254702, + "per_class_f1": { + "FAVOR": 0.8854961832061069, + "AGAINST": 0.5333333333333333, + "NONE": 0.875 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 76.460983884648, + "weighted_f1_pct": 86.61054065617331, + "macro_precision_pct": 73.45619658119658, + "macro_recall_pct": 83.97767261254701, + "per_class_f1_pct": { + "FAVOR": 88.54961832061069, + "AGAINST": 53.333333333333336, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6827956989247311, + "macro_f1": 0.6738486312399355, + "weighted_f1": 0.6892706872370266, + "macro_precision": 0.6661987102665069, + "macro_recall": 0.7423120089786757, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.7111111111111111, + "NONE": 0.68 + }, + "accuracy_pct": 68.27956989247312, + "macro_f1_pct": 67.38486312399355, + "weighted_f1_pct": 68.92706872370266, + "macro_precision_pct": 66.61987102665069, + "macro_recall_pct": 74.23120089786757, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 71.11111111111111, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5863907860455616, + "weighted_f1": 0.5623457240987509, + "macro_precision": 0.6025953859804792, + "macro_recall": 0.6606837606837607, + "per_class_f1": { + "FAVOR": 0.6582278481012658, + "AGAINST": 0.35714285714285715, + "NONE": 0.743801652892562 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 58.63907860455616, + "weighted_f1_pct": 56.23457240987509, + "macro_precision_pct": 60.259538598047925, + "macro_recall_pct": 66.06837606837607, + "per_class_f1_pct": { + "FAVOR": 65.82278481012658, + "AGAINST": 35.714285714285715, + "NONE": 74.3801652892562 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7040694957733317, + "weighted_f1": 0.7120125525678602, + "macro_precision": 0.7107921956661453, + "macro_recall": 0.7124961383997529, + "per_class_f1": { + "FAVOR": 0.6949152542372882, + "AGAINST": 0.631578947368421, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.40694957733317, + "weighted_f1_pct": 71.20125525678603, + "macro_precision_pct": 71.07921956661453, + "macro_recall_pct": 71.2496138399753, + "per_class_f1_pct": { + "FAVOR": 69.49152542372882, + "AGAINST": 63.1578947368421, + "NONE": 78.57142857142857 + } + } + ], + "accuracy": 0.7067082683307332, + "macro_f1": 0.7021739720156036, + "weighted_f1": 0.6977168624660642, + "macro_precision": 0.7135357368753953, + "macro_recall": 0.7202111613876321, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.6055979643765903, + "NONE": 0.767590618336887 + }, + "accuracy_pct": 70.67082683307332, + "macro_f1_pct": 70.21739720156036, + "weighted_f1_pct": 69.77168624660642, + "macro_precision_pct": 71.35357368753952, + "macro_recall_pct": 72.0211161387632, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 60.55979643765903, + "NONE": 76.7590618336887 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6402269231537524, + "losses": { + "current_ce": 0.6802984879567073, + "replay_ce": 0.25908706325751085, + "dark_distillation": 0.3671783589399778, + "total": 1.1229747341229366, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.751736308060755, + "weighted_f1": 0.8481168964807573, + "macro_precision": 0.7189614476789928, + "macro_recall": 0.8298264773692513, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.5333333333333333, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 75.1736308060755, + "weighted_f1_pct": 84.81168964807573, + "macro_precision_pct": 71.89614476789929, + "macro_recall_pct": 82.98264773692513, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 53.333333333333336, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6635325602140946, + "weighted_f1": 0.6691246295070646, + "macro_precision": 0.6565217391304348, + "macro_recall": 0.7427609427609427, + "per_class_f1": { + "FAVOR": 0.6526315789473685, + "AGAINST": 0.6779661016949152, + "NONE": 0.66 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 66.35325602140945, + "weighted_f1_pct": 66.91246295070647, + "macro_precision_pct": 65.65217391304347, + "macro_recall_pct": 74.27609427609427, + "per_class_f1_pct": { + "FAVOR": 65.26315789473685, + "AGAINST": 67.79661016949152, + "NONE": 66.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5842626966529406, + "weighted_f1": 0.5564756699844299, + "macro_precision": 0.6065640373422361, + "macro_recall": 0.6615384615384615, + "per_class_f1": { + "FAVOR": 0.6753246753246753, + "AGAINST": 0.34146341463414637, + "NONE": 0.736 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 58.426269665294065, + "weighted_f1_pct": 55.64756699844299, + "macro_precision_pct": 60.65640373422361, + "macro_recall_pct": 66.15384615384615, + "per_class_f1_pct": { + "FAVOR": 67.53246753246754, + "AGAINST": 34.146341463414636, + "NONE": 73.6 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.74, + "macro_f1": 0.7249169828143054, + "weighted_f1": 0.733666571140042, + "macro_precision": 0.7435681953247514, + "macro_recall": 0.7314643188137163, + "per_class_f1": { + "FAVOR": 0.7130434782608696, + "AGAINST": 0.6481481481481481, + "NONE": 0.8135593220338984 + }, + "accuracy_pct": 74.0, + "macro_f1_pct": 72.49169828143054, + "weighted_f1_pct": 73.3666571140042, + "macro_precision_pct": 74.35681953247514, + "macro_recall_pct": 73.14643188137163, + "per_class_f1_pct": { + "FAVOR": 71.30434782608695, + "AGAINST": 64.81481481481481, + "NONE": 81.35593220338984 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.6405580153208237, + "weighted_f1": 0.6760420619860849, + "macro_precision": 0.6475247524752475, + "macro_recall": 0.6484904523739475, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.4444444444444444, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 64.05580153208237, + "weighted_f1_pct": 67.60420619860848, + "macro_precision_pct": 64.75247524752476, + "macro_recall_pct": 64.84904523739475, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 44.44444444444444, + "NONE": 74.19354838709677 + } + } + ], + "accuracy": 0.7015457788347206, + "macro_f1": 0.6900498368273961, + "weighted_f1": 0.6899914146767683, + "macro_precision": 0.7079188602066525, + "macro_recall": 0.701600922596346, + "per_class_f1": { + "FAVOR": 0.7387096774193549, + "AGAINST": 0.5682819383259912, + "NONE": 0.7631578947368421 + }, + "accuracy_pct": 70.15457788347206, + "macro_f1_pct": 69.0049836827396, + "weighted_f1_pct": 68.99914146767682, + "macro_precision_pct": 70.79188602066525, + "macro_recall_pct": 70.1600922596346, + "per_class_f1_pct": { + "FAVOR": 73.87096774193549, + "AGAINST": 56.82819383259912, + "NONE": 76.31578947368422 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.8459824489236253, + "losses": { + "current_ce": 0.4918774113886886, + "replay_ce": 0.2782205246112965, + "dark_distillation": 0.3717112809419632, + "total": 0.9559535768572931, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6435897435897436, + "weighted_f1": 0.8120490129339687, + "macro_precision": 0.6369186369186369, + "macro_recall": 0.678412813978886, + "per_class_f1": { + "FAVOR": 0.8307692307692308, + "AGAINST": 0.25, + "NONE": 0.85 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 64.35897435897436, + "weighted_f1_pct": 81.20490129339687, + "macro_precision_pct": 63.691863691863695, + "macro_recall_pct": 67.8412813978886, + "per_class_f1_pct": { + "FAVOR": 83.07692307692308, + "AGAINST": 25.0, + "NONE": 85.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.7008263272308216, + "weighted_f1": 0.7266881813366046, + "macro_precision": 0.6903866248693834, + "macro_recall": 0.7466329966329965, + "per_class_f1": { + "FAVOR": 0.6590909090909091, + "AGAINST": 0.7692307692307693, + "NONE": 0.6741573033707865 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 70.08263272308216, + "weighted_f1_pct": 72.66881813366047, + "macro_precision_pct": 69.03866248693834, + "macro_recall_pct": 74.66329966329965, + "per_class_f1_pct": { + "FAVOR": 65.9090909090909, + "AGAINST": 76.92307692307693, + "NONE": 67.41573033707866 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5533002778765491, + "weighted_f1": 0.5277433237566921, + "macro_precision": 0.5542027417027416, + "macro_recall": 0.6175213675213675, + "per_class_f1": { + "FAVOR": 0.6410256410256411, + "AGAINST": 0.3409090909090909, + "NONE": 0.6779661016949152 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.33002778765491, + "weighted_f1_pct": 52.77433237566921, + "macro_precision_pct": 55.42027417027416, + "macro_recall_pct": 61.75213675213676, + "per_class_f1_pct": { + "FAVOR": 64.1025641025641, + "AGAINST": 34.090909090909086, + "NONE": 67.79661016949152 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7166064773016645, + "weighted_f1": 0.7199929682616849, + "macro_precision": 0.72055614809238, + "macro_recall": 0.7268612913191226, + "per_class_f1": { + "FAVOR": 0.7107438016528925, + "AGAINST": 0.6890756302521008, + "NONE": 0.75 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 71.66064773016645, + "weighted_f1_pct": 71.99929682616849, + "macro_precision_pct": 72.055614809238, + "macro_recall_pct": 72.68612913191225, + "per_class_f1_pct": { + "FAVOR": 71.07438016528926, + "AGAINST": 68.90756302521008, + "NONE": 75.0 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.6475320454733945, + "weighted_f1": 0.6759216778834782, + "macro_precision": 0.6435395474804342, + "macro_recall": 0.6523459280740834, + "per_class_f1": { + "FAVOR": 0.7164179104477612, + "AGAINST": 0.46511627906976744, + "NONE": 0.7610619469026548 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 64.75320454733945, + "weighted_f1_pct": 67.59216778834782, + "macro_precision_pct": 64.35395474804342, + "macro_recall_pct": 65.23459280740835, + "per_class_f1_pct": { + "FAVOR": 71.64179104477611, + "AGAINST": 46.51162790697674, + "NONE": 76.10619469026548 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7256176220031643, + "weighted_f1": 0.7694739105863053, + "macro_precision": 0.702960102960103, + "macro_recall": 0.7737331334332834, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.8192771084337349, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 72.56176220031642, + "weighted_f1_pct": 76.94739105863053, + "macro_precision_pct": 70.29601029601031, + "macro_recall_pct": 77.37331334332835, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 81.92771084337349, + "NONE": 75.75757575757575 + } + } + ], + "accuracy": 0.7031408308004052, + "macro_f1": 0.7039433347741956, + "weighted_f1": 0.7009754038396915, + "macro_precision": 0.7064144918748888, + "macro_recall": 0.710720371349668, + "per_class_f1": { + "FAVOR": 0.7109144542772862, + "AGAINST": 0.6597014925373135, + "NONE": 0.7412140575079872 + }, + "accuracy_pct": 70.31408308004052, + "macro_f1_pct": 70.39433347741955, + "weighted_f1_pct": 70.09754038396915, + "macro_precision_pct": 70.64144918748887, + "macro_recall_pct": 71.0720371349668, + "per_class_f1_pct": { + "FAVOR": 71.09144542772862, + "AGAINST": 65.97014925373135, + "NONE": 74.12140575079871 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.5939530234882558, + "losses": { + "current_ce": 0.5948785619302229, + "replay_ce": 0.240766610882499, + "dark_distillation": 0.3915195183320479, + "total": 1.031404926560142, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6221139971139972, + "weighted_f1": 0.8268829253342528, + "macro_precision": 0.6233618233618233, + "macro_recall": 0.6298264773692512, + "per_class_f1": { + "FAVOR": 0.8484848484848485, + "AGAINST": 0.14285714285714285, + "NONE": 0.875 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 62.211399711399714, + "weighted_f1_pct": 82.68829253342528, + "macro_precision_pct": 62.33618233618233, + "macro_recall_pct": 62.982647736925124, + "per_class_f1_pct": { + "FAVOR": 84.84848484848484, + "AGAINST": 14.285714285714285, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.626937859115123, + "weighted_f1": 0.6372787594585096, + "macro_precision": 0.6423485693050911, + "macro_recall": 0.7081930415263749, + "per_class_f1": { + "FAVOR": 0.5740740740740741, + "AGAINST": 0.6473988439306358, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.6937859115123, + "weighted_f1_pct": 63.72787594585097, + "macro_precision_pct": 64.23485693050911, + "macro_recall_pct": 70.81930415263749, + "per_class_f1_pct": { + "FAVOR": 57.407407407407405, + "AGAINST": 64.73988439306359, + "NONE": 65.93406593406593 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5572363446108066, + "weighted_f1": 0.5287067535132683, + "macro_precision": 0.5648717948717948, + "macro_recall": 0.6341880341880342, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.32941176470588235, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 55.723634461080664, + "weighted_f1_pct": 52.87067535132684, + "macro_precision_pct": 56.48717948717948, + "macro_recall_pct": 63.418803418803414, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 32.94117647058823, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6939795322809879, + "weighted_f1": 0.7019042998249061, + "macro_precision": 0.7355272441115814, + "macro_recall": 0.7119864071671301, + "per_class_f1": { + "FAVOR": 0.6814814814814815, + "AGAINST": 0.6274509803921569, + "NONE": 0.7730061349693251 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.39795322809879, + "weighted_f1_pct": 70.19042998249061, + "macro_precision_pct": 73.55272441115814, + "macro_recall_pct": 71.19864071671302, + "per_class_f1_pct": { + "FAVOR": 68.14814814814815, + "AGAINST": 62.745098039215684, + "NONE": 77.30061349693251 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.6375222132146481, + "weighted_f1": 0.679235521815927, + "macro_precision": 0.6368191721132898, + "macro_recall": 0.6428417321621206, + "per_class_f1": { + "FAVOR": 0.751219512195122, + "AGAINST": 0.4155844155844156, + "NONE": 0.7457627118644068 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 63.75222132146481, + "weighted_f1_pct": 67.9235521815927, + "macro_precision_pct": 63.68191721132897, + "macro_recall_pct": 64.28417321621205, + "per_class_f1_pct": { + "FAVOR": 75.1219512195122, + "AGAINST": 41.55844155844156, + "NONE": 74.57627118644068 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6762501172584546, + "weighted_f1": 0.7156854838725379, + "macro_precision": 0.6710858585858586, + "macro_recall": 0.7369315342328836, + "per_class_f1": { + "FAVOR": 0.547945205479452, + "AGAINST": 0.759493670886076, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 67.62501172584547, + "weighted_f1_pct": 71.5685483872538, + "macro_precision_pct": 67.10858585858585, + "macro_recall_pct": 73.69315342328836, + "per_class_f1_pct": { + "FAVOR": 54.794520547945204, + "AGAINST": 75.9493670886076, + "NONE": 72.1311475409836 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5501464097955325, + "weighted_f1": 0.5667102365163307, + "macro_precision": 0.5405684754521963, + "macro_recall": 0.5682643427741466, + "per_class_f1": { + "FAVOR": 0.49122807017543857, + "AGAINST": 0.6137566137566137, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 55.014640979553256, + "weighted_f1_pct": 56.67102365163307, + "macro_precision_pct": 54.05684754521963, + "macro_recall_pct": 56.82643427741466, + "per_class_f1_pct": { + "FAVOR": 49.122807017543856, + "AGAINST": 61.37566137566137, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6601529311809685, + "macro_f1": 0.6622403014498427, + "weighted_f1": 0.6546056149499762, + "macro_precision": 0.6706817100231862, + "macro_recall": 0.6783931407825213, + "per_class_f1": { + "FAVOR": 0.6713780918727915, + "AGAINST": 0.5939849624060151, + "NONE": 0.7213578500707214 + }, + "accuracy_pct": 66.01529311809685, + "macro_f1_pct": 66.22403014498427, + "weighted_f1_pct": 65.46056149499762, + "macro_precision_pct": 67.06817100231862, + "macro_recall_pct": 67.83931407825213, + "per_class_f1_pct": { + "FAVOR": 67.13780918727915, + "AGAINST": 59.3984962406015, + "NONE": 72.13578500707214 + } + }, + { + "after_task": 8, + "best_epoch": 14, + "best_dev_macro_f1": 0.6854605776736924, + "losses": { + "current_ce": 0.565870771353895, + "replay_ce": 0.2280369180621523, + "dark_distillation": 0.36453325269800246, + "total": 0.9761743115656304, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6959841269841269, + "weighted_f1": 0.8276439106616097, + "macro_precision": 0.6813292071912761, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.832, + "AGAINST": 0.38095238095238093, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.59841269841269, + "weighted_f1_pct": 82.76439106616097, + "macro_precision_pct": 68.13292071912761, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.2, + "AGAINST": 38.095238095238095, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.7026415498520571, + "weighted_f1": 0.7262068107700501, + "macro_precision": 0.6878637933160686, + "macro_recall": 0.7509539842873177, + "per_class_f1": { + "FAVOR": 0.6744186046511628, + "AGAINST": 0.7668393782383419, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 70.2641549852057, + "weighted_f1_pct": 72.62068107700502, + "macro_precision_pct": 68.78637933160687, + "macro_recall_pct": 75.09539842873177, + "per_class_f1_pct": { + "FAVOR": 67.44186046511628, + "AGAINST": 76.68393782383419, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5988554140580411, + "weighted_f1": 0.570057835236827, + "macro_precision": 0.6059286146400483, + "macro_recall": 0.6495726495726496, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.38202247191011235, + "NONE": 0.7086614173228346 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.88554140580411, + "weighted_f1_pct": 57.005783523682695, + "macro_precision_pct": 60.592861464004834, + "macro_recall_pct": 64.95726495726495, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 38.20224719101123, + "NONE": 70.86614173228347 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.73, + "macro_f1": 0.7234930546705747, + "weighted_f1": 0.7294345503589483, + "macro_precision": 0.7278173594452997, + "macro_recall": 0.7277108433734939, + "per_class_f1": { + "FAVOR": 0.7017543859649122, + "AGAINST": 0.6949152542372882, + "NONE": 0.7738095238095238 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 72.34930546705746, + "weighted_f1_pct": 72.94345503589483, + "macro_precision_pct": 72.78173594452997, + "macro_recall_pct": 72.7710843373494, + "per_class_f1_pct": { + "FAVOR": 70.17543859649122, + "AGAINST": 69.49152542372882, + "NONE": 77.38095238095238 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.645052905326266, + "weighted_f1": 0.6763768536716097, + "macro_precision": 0.6381766381766382, + "macro_recall": 0.6579946482859104, + "per_class_f1": { + "FAVOR": 0.7253886010362695, + "AGAINST": 0.45977011494252873, + "NONE": 0.75 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 64.5052905326266, + "weighted_f1_pct": 67.63768536716097, + "macro_precision_pct": 63.817663817663814, + "macro_recall_pct": 65.79946482859104, + "per_class_f1_pct": { + "FAVOR": 72.53886010362694, + "AGAINST": 45.97701149425287, + "NONE": 75.0 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7231150793650793, + "weighted_f1": 0.7698242824527072, + "macro_precision": 0.7026315789473685, + "macro_recall": 0.7640229885057472, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.8214285714285714, + "NONE": 0.78125 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 72.31150793650794, + "weighted_f1_pct": 76.98242824527071, + "macro_precision_pct": 70.26315789473685, + "macro_recall_pct": 76.40229885057472, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 82.14285714285714, + "NONE": 78.125 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5736842105263158, + "macro_f1": 0.5743890518084066, + "weighted_f1": 0.5758666460873592, + "macro_precision": 0.5652901348553523, + "macro_recall": 0.6089324618736384, + "per_class_f1": { + "FAVOR": 0.532258064516129, + "AGAINST": 0.5909090909090909, + "NONE": 0.6 + }, + "accuracy_pct": 57.36842105263158, + "macro_f1_pct": 57.43890518084066, + "weighted_f1_pct": 57.586664608735916, + "macro_precision_pct": 56.52901348553523, + "macro_recall_pct": 60.893246187363836, + "per_class_f1_pct": { + "FAVOR": 53.2258064516129, + "AGAINST": 59.09090909090909, + "NONE": 60.0 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6980320761207496, + "weighted_f1": 0.7175603857500334, + "macro_precision": 0.6871572871572872, + "macro_recall": 0.71669847880155, + "per_class_f1": { + "FAVOR": 0.6470588235294118, + "AGAINST": 0.7452830188679245, + "NONE": 0.7017543859649122 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.80320761207496, + "weighted_f1_pct": 71.75603857500334, + "macro_precision_pct": 68.71572871572872, + "macro_recall_pct": 71.669847880155, + "per_class_f1_pct": { + "FAVOR": 64.70588235294117, + "AGAINST": 74.52830188679245, + "NONE": 70.17543859649122 + } + } + ], + "accuracy": 0.6935953420669578, + "macro_f1": 0.6956658843009055, + "weighted_f1": 0.6918128833366621, + "macro_precision": 0.6923164830314782, + "macro_recall": 0.7080891012593432, + "per_class_f1": { + "FAVOR": 0.684964200477327, + "AGAINST": 0.6691729323308271, + "NONE": 0.7328605200945626 + }, + "accuracy_pct": 69.35953420669578, + "macro_f1_pct": 69.56658843009055, + "weighted_f1_pct": 69.18128833366622, + "macro_precision_pct": 69.23164830314782, + "macro_recall_pct": 70.80891012593432, + "per_class_f1_pct": { + "FAVOR": 68.4964200477327, + "AGAINST": 66.9172932330827, + "NONE": 73.28605200945626 + } + }, + { + "after_task": 9, + "best_epoch": 14, + "best_dev_macro_f1": 0.6585250563984362, + "losses": { + "current_ce": 0.5479454851812786, + "replay_ce": 0.21342617401054928, + "dark_distillation": 0.3428357641375254, + "total": 0.9327895448321388, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6751450476585164, + "weighted_f1": 0.8263292673534579, + "macro_precision": 0.665018315018315, + "macro_recall": 0.7482344375682562, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.3157894736842105, + "NONE": 0.875 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 67.51450476585164, + "weighted_f1_pct": 82.63292673534579, + "macro_precision_pct": 66.5018315018315, + "macro_recall_pct": 74.82344375682563, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 31.57894736842105, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6481411774229454, + "weighted_f1": 0.6637595696052299, + "macro_precision": 0.64975005852216, + "macro_recall": 0.7166105499438832, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6850828729281768, + "NONE": 0.6593406593406593 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 64.81411774229454, + "weighted_f1_pct": 66.37595696052298, + "macro_precision_pct": 64.975005852216, + "macro_recall_pct": 71.66105499438832, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 68.50828729281768, + "NONE": 65.93406593406593 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.568864468864469, + "weighted_f1": 0.5345921683949852, + "macro_precision": 0.5902027027027027, + "macro_recall": 0.6551282051282051, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.3, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 56.8864468864469, + "weighted_f1_pct": 53.459216839498524, + "macro_precision_pct": 59.02027027027027, + "macro_recall_pct": 65.51282051282051, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 30.0, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7017926066706556, + "weighted_f1": 0.7090290536022243, + "macro_precision": 0.7196196196196195, + "macro_recall": 0.7132684584491814, + "per_class_f1": { + "FAVOR": 0.6825396825396826, + "AGAINST": 0.6545454545454545, + "NONE": 0.7682926829268293 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.17926066706556, + "weighted_f1_pct": 70.90290536022243, + "macro_precision_pct": 71.96196196196195, + "macro_recall_pct": 71.32684584491814, + "per_class_f1_pct": { + "FAVOR": 68.25396825396825, + "AGAINST": 65.45454545454545, + "NONE": 76.82926829268293 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.7, + "macro_f1": 0.660162353263135, + "weighted_f1": 0.6974926253687315, + "macro_precision": 0.6591469015305641, + "macro_recall": 0.6619509939898289, + "per_class_f1": { + "FAVOR": 0.7572815533980582, + "AGAINST": 0.4444444444444444, + "NONE": 0.7787610619469026 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 66.0162353263135, + "weighted_f1_pct": 69.74926253687315, + "macro_precision_pct": 65.91469015305641, + "macro_recall_pct": 66.19509939898289, + "per_class_f1_pct": { + "FAVOR": 75.72815533980582, + "AGAINST": 44.44444444444444, + "NONE": 77.87610619469027 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7374999999999999, + "weighted_f1": 0.7835331050228311, + "macro_precision": 0.7165413533834587, + "macro_recall": 0.7791404297851074, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.8333333333333334, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.75, + "weighted_f1_pct": 78.35331050228311, + "macro_precision_pct": 71.65413533834587, + "macro_recall_pct": 77.91404297851075, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 83.33333333333334, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5693402727110591, + "weighted_f1": 0.5874227960094311, + "macro_precision": 0.5702878992352677, + "macro_recall": 0.5907770515613652, + "per_class_f1": { + "FAVOR": 0.5757575757575758, + "AGAINST": 0.6179775280898876, + "NONE": 0.5142857142857142 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 56.93402727110591, + "weighted_f1_pct": 58.742279600943114, + "macro_precision_pct": 57.028789923526766, + "macro_recall_pct": 59.07770515613652, + "per_class_f1_pct": { + "FAVOR": 57.57575757575758, + "AGAINST": 61.79775280898876, + "NONE": 51.42857142857142 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6852791878172588, + "macro_f1": 0.6838579256883186, + "weighted_f1": 0.6858666383075728, + "macro_precision": 0.6659991548933513, + "macro_recall": 0.7490574669127507, + "per_class_f1": { + "FAVOR": 0.6744186046511628, + "AGAINST": 0.6875, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 68.52791878172589, + "macro_f1_pct": 68.38579256883186, + "weighted_f1_pct": 68.58666383075727, + "macro_precision_pct": 66.59991548933513, + "macro_recall_pct": 74.90574669127507, + "per_class_f1_pct": { + "FAVOR": 67.44186046511628, + "AGAINST": 68.75, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.655, + "macro_f1": 0.6735109092832671, + "weighted_f1": 0.6574522083058669, + "macro_precision": 0.6942552136350586, + "macro_recall": 0.659741041854863, + "per_class_f1": { + "FAVOR": 0.6296296296296297, + "AGAINST": 0.6341463414634146, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 65.5, + "macro_f1_pct": 67.35109092832671, + "weighted_f1_pct": 65.74522083058669, + "macro_precision_pct": 69.42552136350587, + "macro_recall_pct": 65.97410418548631, + "per_class_f1_pct": { + "FAVOR": 62.96296296296296, + "AGAINST": 63.41463414634146, + "NONE": 75.67567567567568 + } + } + ], + "accuracy": 0.679161372299873, + "macro_f1": 0.6833621061724205, + "weighted_f1": 0.6764605363906485, + "macro_precision": 0.6839118747964701, + "macro_recall": 0.6971458196306716, + "per_class_f1": { + "FAVOR": 0.6796657381615598, + "AGAINST": 0.6376811594202898, + "NONE": 0.732739420935412 + }, + "accuracy_pct": 67.9161372299873, + "macro_f1_pct": 68.33621061724206, + "weighted_f1_pct": 67.64605363906485, + "macro_precision_pct": 68.39118747964702, + "macro_recall_pct": 69.71458196306716, + "per_class_f1_pct": { + "FAVOR": 67.96657381615599, + "AGAINST": 63.76811594202898, + "NONE": 73.2739420935412 + } + }, + { + "after_task": 10, + "best_epoch": 13, + "best_dev_macro_f1": 0.721058058314146, + "losses": { + "current_ce": 0.4208163469408949, + "replay_ce": 0.20656802905723454, + "dark_distillation": 0.3602398360768954, + "total": 0.8075042930742105, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6907074869983317, + "weighted_f1": 0.8418368454553427, + "macro_precision": 0.6796395455139089, + "macro_recall": 0.7581846863244751, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.3333333333333333, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 69.07074869983317, + "weighted_f1_pct": 84.18368454553426, + "macro_precision_pct": 67.9639545513909, + "macro_recall_pct": 75.8184686324475, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 33.33333333333333, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6761111258680786, + "weighted_f1": 0.7024772568949074, + "macro_precision": 0.6664672041656168, + "macro_recall": 0.7258136924803592, + "per_class_f1": { + "FAVOR": 0.6067415730337079, + "AGAINST": 0.7407407407407407, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 67.61111258680786, + "weighted_f1_pct": 70.24772568949074, + "macro_precision_pct": 66.64672041656168, + "macro_recall_pct": 72.58136924803593, + "per_class_f1_pct": { + "FAVOR": 60.67415730337079, + "AGAINST": 74.07407407407408, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5492957746478874, + "macro_f1": 0.5374294155042818, + "weighted_f1": 0.49773808326192154, + "macro_precision": 0.527403381642512, + "macro_recall": 0.6175213675213675, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.25882352941176473, + "NONE": 0.6611570247933884 + }, + "accuracy_pct": 54.929577464788736, + "macro_f1_pct": 53.74294155042818, + "weighted_f1_pct": 49.77380832619215, + "macro_precision_pct": 52.7403381642512, + "macro_recall_pct": 61.75213675213676, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 25.882352941176475, + "NONE": 66.11570247933885 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.68, + "macro_f1": 0.6727954497944394, + "weighted_f1": 0.6793584375653264, + "macro_precision": 0.6851334776334776, + "macro_recall": 0.6856657398826074, + "per_class_f1": { + "FAVOR": 0.6614173228346457, + "AGAINST": 0.6194690265486725, + "NONE": 0.7375 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 67.27954497944394, + "weighted_f1_pct": 67.93584375653265, + "macro_precision_pct": 68.51334776334777, + "macro_recall_pct": 68.56657398826074, + "per_class_f1_pct": { + "FAVOR": 66.14173228346458, + "AGAINST": 61.94690265486725, + "NONE": 73.75 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.645, + "macro_f1": 0.6218952933168378, + "weighted_f1": 0.6529679232064869, + "macro_precision": 0.6243016830901637, + "macro_recall": 0.6244553720281876, + "per_class_f1": { + "FAVOR": 0.696969696969697, + "AGAINST": 0.42105263157894735, + "NONE": 0.7476635514018691 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 62.18952933168378, + "weighted_f1_pct": 65.29679232064869, + "macro_precision_pct": 62.430168309016366, + "macro_recall_pct": 62.445537202818755, + "per_class_f1_pct": { + "FAVOR": 69.6969696969697, + "AGAINST": 42.10526315789473, + "NONE": 74.76635514018692 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7358434400176685, + "weighted_f1": 0.7816936452294332, + "macro_precision": 0.7160961814722052, + "macro_recall": 0.7694302848575711, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.8304093567251462, + "NONE": 0.8253968253968254 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.58434400176685, + "weighted_f1_pct": 78.16936452294331, + "macro_precision_pct": 71.60961814722052, + "macro_recall_pct": 76.94302848575711, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 83.04093567251462, + "NONE": 82.53968253968253 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5994877856579984, + "weighted_f1": 0.6192920243872091, + "macro_precision": 0.5925163507782846, + "macro_recall": 0.6147421931735657, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.6595744680851063, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.94877856579984, + "weighted_f1_pct": 61.92920243872091, + "macro_precision_pct": 59.25163507782846, + "macro_recall_pct": 61.47421931735657, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 65.95744680851064, + "NONE": 55.55555555555556 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6858314369611097, + "weighted_f1": 0.6919016926464668, + "macro_precision": 0.6653784293236625, + "macro_recall": 0.7376629385999506, + "per_class_f1": { + "FAVOR": 0.6904761904761905, + "AGAINST": 0.7064676616915423, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 68.58314369611097, + "weighted_f1_pct": 69.19016926464668, + "macro_precision_pct": 66.53784293236626, + "macro_recall_pct": 73.76629385999506, + "per_class_f1_pct": { + "FAVOR": 69.04761904761905, + "AGAINST": 70.64676616915423, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6429926256013213, + "weighted_f1": 0.6310591359287011, + "macro_precision": 0.6615247300592129, + "macro_recall": 0.6312934781769471, + "per_class_f1": { + "FAVOR": 0.5838509316770186, + "AGAINST": 0.6424242424242425, + "NONE": 0.7027027027027027 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.29926256013214, + "weighted_f1_pct": 63.10591359287011, + "macro_precision_pct": 66.15247300592128, + "macro_recall_pct": 63.129347817694715, + "per_class_f1_pct": { + "FAVOR": 58.38509316770186, + "AGAINST": 64.24242424242425, + "NONE": 70.27027027027027 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7789812424412296, + "weighted_f1": 0.819386864861304, + "macro_precision": 0.762002967314218, + "macro_recall": 0.8047365206285746, + "per_class_f1": { + "FAVOR": 0.6885245901639344, + "AGAINST": 0.7768595041322314, + "NONE": 0.8715596330275229 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 77.89812424412295, + "weighted_f1_pct": 81.9386864861304, + "macro_precision_pct": 76.2002967314218, + "macro_recall_pct": 80.47365206285745, + "per_class_f1_pct": { + "FAVOR": 68.85245901639344, + "AGAINST": 77.68595041322314, + "NONE": 87.1559633027523 + } + } + ], + "accuracy": 0.6877113866967306, + "macro_f1": 0.6899933380976108, + "weighted_f1": 0.687378888819321, + "macro_precision": 0.6888985087023384, + "macro_recall": 0.6968406261081173, + "per_class_f1": { + "FAVOR": 0.6660633484162896, + "AGAINST": 0.6582466567607727, + "NONE": 0.7456700091157703 + }, + "accuracy_pct": 68.77113866967306, + "macro_f1_pct": 68.99933380976108, + "weighted_f1_pct": 68.7378888819321, + "macro_precision_pct": 68.88985087023384, + "macro_recall_pct": 69.68406261081172, + "per_class_f1_pct": { + "FAVOR": 66.60633484162895, + "AGAINST": 65.82466567607726, + "NONE": 74.56700091157703 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s2__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__x__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..43ef93bdbd88cadb6203b85dae294deaa139cb09 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__x__s2__seed13/run.log b/outputs/trace_s6_derpp__x__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..27c9eb4b40fed3b88f36a6d90fd186caef0414dc --- /dev/null +++ b/outputs/trace_s6_derpp__x__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 05:57:33,833 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__x__s2__seed13 device=cuda +2026-09-02 06:13:33,188 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6642075302161297, 'average_forgetting': 0.030619168331382465, 'backward_transfer': -0.016260235150664635, 'final_mean_target_macro_f1_pct': 66.42075302161297, 'average_forgetting_pct': 3.0619168331382465, 'backward_transfer_pct': -1.6260235150664635, 'final_pooled_accuracy': 0.6877113866967306, 'final_pooled_accuracy_pct': 68.77113866967306, 'final_pooled_macro_f1': 0.6899933380976108, 'final_pooled_macro_f1_pct': 68.99933380976108, 'final_pooled_weighted_f1': 0.687378888819321, 'final_pooled_weighted_f1_pct': 68.7378888819321, 'final_pooled_macro_precision': 0.6888985087023384, 'final_pooled_macro_precision_pct': 68.88985087023384, 'final_pooled_macro_recall': 0.6968406261081173, 'final_pooled_macro_recall_pct': 69.68406261081172, 'final_pooled_per_class_f1': {'FAVOR': 0.6660633484162896, 'AGAINST': 0.6582466567607727, 'NONE': 0.7456700091157703}, 'final_pooled_per_class_f1_pct': {'FAVOR': 66.60633484162895, 'AGAINST': 65.82466567607726, 'NONE': 74.56700091157703}} +2026-09-02 06:13:33,189 [INFO] __main__: efficiency={'training_wall_time_seconds': 957.960110536078, 'training_wall_time_minutes': 15.966001842267966, 'peak_gpu_memory_allocated_gb': 10.398106575012207, 'peak_gpu_memory_reserved_gb': 10.62109375} diff --git a/outputs/trace_s6_derpp__x__s3__seed13/environment.json b/outputs/trace_s6_derpp__x__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..22ffaa19fb17b3359b5a60811a899e7343070ca7 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T06:13:41.643439+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s3__seed13/metrics.json b/outputs/trace_s6_derpp__x__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..80a7061c6b4a555dfa52180fd34b2c67ff9a2e4a --- /dev/null +++ b/outputs/trace_s6_derpp__x__s3__seed13/metrics.json @@ -0,0 +1,2285 @@ +{ + "experiment": "trace_s6_derpp", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 6, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": false, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 113910531, + "trainable_parameters": 4428291, + "trainable_fraction": 0.03887516774019779 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.6016841791489679, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7093435860466979, + 0.6161020036429873, + 0.6980752238363291, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7229955162713, + 0.6139385359182533, + 0.6574383881993164, + 0.670565066822749, + null, + null, + null, + null, + null, + null + ], + [ + 0.6961155336441379, + 0.6277700655714157, + 0.7145617667356797, + 0.6426875909887235, + 0.7906251215622203, + null, + null, + null, + null, + null + ], + [ + 0.6532146957520092, + 0.613289760348584, + 0.6959108536207355, + 0.6433780171397122, + 0.8002331002331003, + 0.6739929341168533, + null, + null, + null, + null + ], + [ + 0.6981721981721982, + 0.6342857142857143, + 0.717799266496919, + 0.6753747256922636, + 0.7735449735449734, + 0.6072401515439491, + 0.6829853049706481, + null, + null, + null + ], + [ + 0.6955803175901315, + 0.5882842795281281, + 0.6936507936507935, + 0.6555254441433302, + 0.7965408805031448, + 0.6684054684054684, + 0.6934202186339963, + 0.6186628186628186, + null, + null + ], + [ + 0.6955803175901315, + 0.5763509923517453, + 0.6707314701363362, + 0.6686274509803921, + 0.8080849148563699, + 0.6323855037833532, + 0.6857394354385301, + 0.5527339449541285, + 0.6931056201550388, + null + ], + [ + 0.6924690274119598, + 0.5840995625250645, + 0.6950100177631083, + 0.6417022674991785, + 0.8067380041136193, + 0.6399982832664075, + 0.6953016688061617, + 0.5591470258136925, + 0.6671717171717172, + 0.6359478520326857 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.62717150692173, + "losses": { + "current_ce": 0.7311763364257234, + "replay_ce": 0.21666293532559366, + "dark_distillation": 0.3519646299607826, + "total": 1.1238215887185299, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.6016841791489679, + "weighted_f1": 0.6250979561580007, + "macro_precision": 0.5981259036814593, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.6031746031746031, + "AGAINST": 0.6666666666666666, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 60.168417914896786, + "weighted_f1_pct": 62.509795615800066, + "macro_precision_pct": 59.81259036814593, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 60.317460317460316, + "AGAINST": 66.66666666666666, + "NONE": 53.52112676056338 + } + } + ], + "accuracy": 0.6726190476190477, + "macro_f1": 0.658125419259498, + "weighted_f1": 0.6801621558142686, + "macro_precision": 0.6473274758333041, + "macro_recall": 0.6927315824614532, + "per_class_f1": { + "FAVOR": 0.5894736842105263, + "AGAINST": 0.7230320699708455, + "NONE": 0.6618705035971223 + }, + "accuracy_pct": 67.26190476190477, + "macro_f1_pct": 65.8125419259498, + "weighted_f1_pct": 68.01621558142686, + "macro_precision_pct": 64.73274758333041, + "macro_recall_pct": 69.27315824614531, + "per_class_f1_pct": { + "FAVOR": 58.94736842105262, + "AGAINST": 72.30320699708454, + "NONE": 66.18705035971223 + } + }, + { + "after_task": 3, + "best_epoch": 13, + "best_dev_macro_f1": 0.7474842767295597, + "losses": { + "current_ce": 0.6197961993289717, + "replay_ce": 0.19484097415750676, + "dark_distillation": 0.3260704606771469, + "total": 0.9776724049539277, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7465753424657534, + "macro_f1": 0.7093435860466979, + "weighted_f1": 0.7548857021062281, + "macro_precision": 0.6866839122936684, + "macro_recall": 0.762808595702149, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.8048780487804879, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 74.65753424657534, + "macro_f1_pct": 70.9343586046698, + "weighted_f1_pct": 75.48857021062281, + "macro_precision_pct": 68.66839122936685, + "macro_recall_pct": 76.2808595702149, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 80.48780487804879, + "NONE": 77.14285714285715 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6161020036429873, + "weighted_f1": 0.6319286741443774, + "macro_precision": 0.6088025695071765, + "macro_recall": 0.6379811183732752, + "per_class_f1": { + "FAVOR": 0.5573770491803278, + "AGAINST": 0.6775956284153005, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 61.61020036429873, + "weighted_f1_pct": 63.192867414437735, + "macro_precision_pct": 60.88025695071765, + "macro_recall_pct": 63.79811183732752, + "per_class_f1_pct": { + "FAVOR": 55.73770491803278, + "AGAINST": 67.75956284153006, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6980752238363291, + "weighted_f1": 0.7129936874192117, + "macro_precision": 0.6934777642400244, + "macro_recall": 0.7137486262941755, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.7393364928909952, + "NONE": 0.6779661016949152 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.80752238363291, + "weighted_f1_pct": 71.29936874192117, + "macro_precision_pct": 69.34777642400243, + "macro_recall_pct": 71.37486262941755, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 73.93364928909952, + "NONE": 67.79661016949152 + } + } + ], + "accuracy": 0.6904315196998124, + "macro_f1": 0.6701723343029786, + "weighted_f1": 0.6951692687471527, + "macro_precision": 0.655365786856778, + "macro_recall": 0.7011135330182133, + "per_class_f1": { + "FAVOR": 0.5877551020408164, + "AGAINST": 0.7383512544802867, + "NONE": 0.6844106463878327 + }, + "accuracy_pct": 69.04315196998124, + "macro_f1_pct": 67.01723343029786, + "weighted_f1_pct": 69.51692687471527, + "macro_precision_pct": 65.53657868567781, + "macro_recall_pct": 70.11135330182134, + "per_class_f1_pct": { + "FAVOR": 58.77551020408164, + "AGAINST": 73.83512544802868, + "NONE": 68.44106463878326 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.6872455400103634, + "losses": { + "current_ce": 0.6491889049136449, + "replay_ce": 0.1964105042909819, + "dark_distillation": 0.31591956705328017, + "total": 1.0035591922109088, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7229955162713, + "weighted_f1": 0.7482915139708751, + "macro_precision": 0.7022841656987998, + "macro_recall": 0.8058970514742629, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7741935483870968, + "NONE": 0.7887323943661971 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 72.29955162713, + "weighted_f1_pct": 74.82915139708751, + "macro_precision_pct": 70.22841656987998, + "macro_recall_pct": 80.58970514742629, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 77.41935483870968, + "NONE": 78.87323943661971 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.6139385359182533, + "weighted_f1": 0.6202209397985733, + "macro_precision": 0.6238178503108561, + "macro_recall": 0.6517792302106028, + "per_class_f1": { + "FAVOR": 0.5899280575539568, + "AGAINST": 0.6385542168674698, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 61.393853591825334, + "weighted_f1_pct": 62.02209397985733, + "macro_precision_pct": 62.38178503108561, + "macro_recall_pct": 65.17792302106028, + "per_class_f1_pct": { + "FAVOR": 58.992805755395686, + "AGAINST": 63.85542168674698, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6574383881993164, + "weighted_f1": 0.6750066430894063, + "macro_precision": 0.6417624521072797, + "macro_recall": 0.7044100557895456, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.7, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 65.74383881993164, + "weighted_f1_pct": 67.50066430894063, + "macro_precision_pct": 64.17624521072797, + "macro_recall_pct": 70.44100557895456, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 70.0, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.670565066822749, + "weighted_f1": 0.6597241183786947, + "macro_precision": 0.6696519772362469, + "macro_recall": 0.6943344219766984, + "per_class_f1": { + "FAVOR": 0.6705882352941176, + "AGAINST": 0.6119402985074627, + "NONE": 0.7291666666666666 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.0565066822749, + "weighted_f1_pct": 65.97241183786947, + "macro_precision_pct": 66.9651977236247, + "macro_recall_pct": 69.43344219766983, + "per_class_f1_pct": { + "FAVOR": 67.05882352941175, + "AGAINST": 61.19402985074627, + "NONE": 72.91666666666666 + } + } + ], + "accuracy": 0.6684856753069577, + "macro_f1": 0.6684043724424299, + "weighted_f1": 0.6710945629369612, + "macro_precision": 0.6638014839726986, + "macro_recall": 0.7050096485496787, + "per_class_f1": { + "FAVOR": 0.6260869565217392, + "AGAINST": 0.683969465648855, + "NONE": 0.6951566951566952 + }, + "accuracy_pct": 66.84856753069577, + "macro_f1_pct": 66.84043724424299, + "weighted_f1_pct": 67.10945629369613, + "macro_precision_pct": 66.38014839726985, + "macro_recall_pct": 70.50096485496788, + "per_class_f1_pct": { + "FAVOR": 62.60869565217392, + "AGAINST": 68.3969465648855, + "NONE": 69.51566951566952 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6644970886821171, + "losses": { + "current_ce": 0.4574271487692992, + "replay_ce": 0.2013551927016427, + "dark_distillation": 0.33779942262917756, + "total": 0.8276820568988721, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6961155336441379, + "weighted_f1": 0.7317683323451168, + "macro_precision": 0.6735909822866345, + "macro_recall": 0.7598100949525238, + "per_class_f1": { + "FAVOR": 0.5614035087719298, + "AGAINST": 0.7701863354037267, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.6115533644138, + "weighted_f1_pct": 73.17683323451169, + "macro_precision_pct": 67.35909822866346, + "macro_recall_pct": 75.98100949525238, + "per_class_f1_pct": { + "FAVOR": 56.14035087719298, + "AGAINST": 77.01863354037268, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6421052631578947, + "macro_f1": 0.6277700655714157, + "weighted_f1": 0.6471213639766652, + "macro_precision": 0.6181170681170681, + "macro_recall": 0.6535947712418301, + "per_class_f1": { + "FAVOR": 0.6050420168067226, + "AGAINST": 0.6885245901639344, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 64.21052631578948, + "macro_f1_pct": 62.77700655714157, + "weighted_f1_pct": 64.71213639766653, + "macro_precision_pct": 61.81170681170681, + "macro_recall_pct": 65.359477124183, + "per_class_f1_pct": { + "FAVOR": 60.50420168067227, + "AGAINST": 68.85245901639344, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7145617667356797, + "weighted_f1": 0.7281920316129099, + "macro_precision": 0.7020332187857962, + "macro_recall": 0.7369004990035704, + "per_class_f1": { + "FAVOR": 0.6956521739130435, + "AGAINST": 0.7523809523809524, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 71.45617667356797, + "weighted_f1_pct": 72.81920316129099, + "macro_precision_pct": 70.20332187857962, + "macro_recall_pct": 73.69004990035704, + "per_class_f1_pct": { + "FAVOR": 69.56521739130434, + "AGAINST": 75.23809523809524, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6426875909887235, + "weighted_f1": 0.6336139610629951, + "macro_precision": 0.6348927777499206, + "macro_recall": 0.6546915896509392, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.631578947368421, + "NONE": 0.6888888888888889 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 64.26875909887235, + "weighted_f1_pct": 63.36139610629951, + "macro_precision_pct": 63.48927777499206, + "macro_recall_pct": 65.46915896509392, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 63.1578947368421, + "NONE": 68.88888888888889 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7906251215622203, + "weighted_f1": 0.817178783988797, + "macro_precision": 0.7778998778998779, + "macro_recall": 0.8075850963407185, + "per_class_f1": { + "FAVOR": 0.7368421052631579, + "AGAINST": 0.7804878048780488, + "NONE": 0.8545454545454545 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 79.06251215622203, + "weighted_f1_pct": 81.71787839887969, + "macro_precision_pct": 77.78998778998779, + "macro_recall_pct": 80.75850963407186, + "per_class_f1_pct": { + "FAVOR": 73.68421052631578, + "AGAINST": 78.04878048780488, + "NONE": 85.45454545454545 + } + } + ], + "accuracy": 0.7084673097534834, + "macro_f1": 0.7009661754952897, + "weighted_f1": 0.7089731626246859, + "macro_precision": 0.6949839336982592, + "macro_recall": 0.7120972706751488, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.7237635705669482, + "NONE": 0.7487001733102253 + }, + "accuracy_pct": 70.84673097534834, + "macro_f1_pct": 70.09661754952897, + "weighted_f1_pct": 70.89731626246859, + "macro_precision_pct": 69.49839336982592, + "macro_recall_pct": 71.20972706751488, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 72.37635705669481, + "NONE": 74.87001733102254 + } + }, + { + "after_task": 6, + "best_epoch": 4, + "best_dev_macro_f1": 0.7152560083594567, + "losses": { + "current_ce": 0.4044448312549364, + "replay_ce": 0.21253694224925268, + "dark_distillation": 0.37164135646252405, + "total": 0.8028024511677878, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6532146957520092, + "weighted_f1": 0.6985436358775144, + "macro_precision": 0.6375386996904026, + "macro_recall": 0.7084257871064468, + "per_class_f1": { + "FAVOR": 0.5230769230769231, + "AGAINST": 0.75, + "NONE": 0.6865671641791045 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 65.32146957520092, + "weighted_f1_pct": 69.85436358775145, + "macro_precision_pct": 63.75386996904025, + "macro_recall_pct": 70.84257871064467, + "per_class_f1_pct": { + "FAVOR": 52.307692307692314, + "AGAINST": 75.0, + "NONE": 68.65671641791045 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.613289760348584, + "weighted_f1": 0.6263157894736842, + "macro_precision": 0.6139511360099595, + "macro_recall": 0.6245461147421931, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.6666666666666666, + "NONE": 0.6176470588235294 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 61.32897603485839, + "weighted_f1_pct": 62.63157894736842, + "macro_precision_pct": 61.395113600995955, + "macro_recall_pct": 62.45461147421931, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 66.66666666666666, + "NONE": 61.76470588235294 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.6959108536207355, + "weighted_f1": 0.724050207000931, + "macro_precision": 0.699291756836514, + "macro_recall": 0.6938883893594983, + "per_class_f1": { + "FAVOR": 0.6865671641791045, + "AGAINST": 0.782608695652174, + "NONE": 0.6185567010309279 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 69.59108536207354, + "weighted_f1_pct": 72.4050207000931, + "macro_precision_pct": 69.92917568365141, + "macro_recall_pct": 69.38883893594982, + "per_class_f1_pct": { + "FAVOR": 68.65671641791045, + "AGAINST": 78.26086956521739, + "NONE": 61.855670103092784 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6433780171397122, + "weighted_f1": 0.6305571389260162, + "macro_precision": 0.6488660676080444, + "macro_recall": 0.6388483542955088, + "per_class_f1": { + "FAVOR": 0.5962732919254659, + "AGAINST": 0.625, + "NONE": 0.7088607594936709 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.33780171397122, + "weighted_f1_pct": 63.05571389260162, + "macro_precision_pct": 64.88660676080444, + "macro_recall_pct": 63.884835429550876, + "per_class_f1_pct": { + "FAVOR": 59.62732919254658, + "AGAINST": 62.5, + "NONE": 70.88607594936708 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.8002331002331003, + "weighted_f1": 0.81852331002331, + "macro_precision": 0.781969652937395, + "macro_recall": 0.8378736557647102, + "per_class_f1": { + "FAVOR": 0.7666666666666667, + "AGAINST": 0.7878787878787878, + "NONE": 0.8461538461538461 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 80.02331002331003, + "weighted_f1_pct": 81.852331002331, + "macro_precision_pct": 78.1969652937395, + "macro_recall_pct": 83.78736557647102, + "per_class_f1_pct": { + "FAVOR": 76.66666666666667, + "AGAINST": 78.78787878787878, + "NONE": 84.61538461538461 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6739929341168533, + "weighted_f1": 0.8158039237065612, + "macro_precision": 0.6676413255360624, + "macro_recall": 0.7936658172551875, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.34782608695652173, + "NONE": 0.8354430379746836 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 67.39929341168533, + "weighted_f1_pct": 81.58039237065611, + "macro_precision_pct": 66.76413255360623, + "macro_recall_pct": 79.36658172551874, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 34.78260869565217, + "NONE": 83.54430379746836 + } + } + ], + "accuracy": 0.7065009560229446, + "macro_f1": 0.7054468432361576, + "weighted_f1": 0.7076596142454398, + "macro_precision": 0.7071395733040599, + "macro_recall": 0.7055862331992636, + "per_class_f1": { + "FAVOR": 0.6567164179104478, + "AGAINST": 0.7138047138047138, + "NONE": 0.745819397993311 + }, + "accuracy_pct": 70.65009560229446, + "macro_f1_pct": 70.54468432361575, + "weighted_f1_pct": 70.76596142454397, + "macro_precision_pct": 70.71395733040599, + "macro_recall_pct": 70.55862331992635, + "per_class_f1_pct": { + "FAVOR": 65.67164179104478, + "AGAINST": 71.38047138047138, + "NONE": 74.5819397993311 + } + }, + { + "after_task": 7, + "best_epoch": 9, + "best_dev_macro_f1": 0.6233995358995359, + "losses": { + "current_ce": 0.5832350539438652, + "replay_ce": 0.23405812274325977, + "dark_distillation": 0.36451117116393467, + "total": 0.9995487657460299, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6981721981721982, + "weighted_f1": 0.7300648362292199, + "macro_precision": 0.6813563146884283, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.81721981721982, + "weighted_f1_pct": 73.00648362292199, + "macro_precision_pct": 68.13563146884283, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6342857142857143, + "weighted_f1": 0.6409022556390976, + "macro_precision": 0.6310147654468614, + "macro_recall": 0.6655773420479303, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6628571428571428, + "NONE": 0.64 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 63.42857142857143, + "weighted_f1_pct": 64.09022556390977, + "macro_precision_pct": 63.10147654468614, + "macro_recall_pct": 66.55773420479304, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 66.28571428571428, + "NONE": 64.0 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.717799266496919, + "weighted_f1": 0.7364897445263106, + "macro_precision": 0.7145040968570381, + "macro_recall": 0.7212836327880575, + "per_class_f1": { + "FAVOR": 0.7164179104477612, + "AGAINST": 0.7767857142857143, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.7799266496919, + "weighted_f1_pct": 73.64897445263105, + "macro_precision_pct": 71.4504096857038, + "macro_recall_pct": 72.12836327880575, + "per_class_f1_pct": { + "FAVOR": 71.64179104477611, + "AGAINST": 77.67857142857143, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6753747256922636, + "weighted_f1": 0.668157931467653, + "macro_precision": 0.6831881509896375, + "macro_recall": 0.678850207305492, + "per_class_f1": { + "FAVOR": 0.6741573033707865, + "AGAINST": 0.6376811594202898, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.53747256922637, + "weighted_f1_pct": 66.8157931467653, + "macro_precision_pct": 68.31881509896375, + "macro_recall_pct": 67.8850207305492, + "per_class_f1_pct": { + "FAVOR": 67.41573033707866, + "AGAINST": 63.76811594202898, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7735449735449734, + "weighted_f1": 0.8040952380952381, + "macro_precision": 0.7534436713753602, + "macro_recall": 0.8063893978936457, + "per_class_f1": { + "FAVOR": 0.7301587301587301, + "AGAINST": 0.7333333333333333, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 77.35449735449734, + "weighted_f1_pct": 80.4095238095238, + "macro_precision_pct": 75.34436713753601, + "macro_recall_pct": 80.63893978936457, + "per_class_f1_pct": { + "FAVOR": 73.01587301587301, + "AGAINST": 73.33333333333333, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6991150442477876, + "macro_f1": 0.6072401515439491, + "weighted_f1": 0.7601570428948048, + "macro_precision": 0.6518624273293203, + "macro_recall": 0.7470695303967965, + "per_class_f1": { + "FAVOR": 0.7387387387387387, + "AGAINST": 0.2222222222222222, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 69.91150442477876, + "macro_f1_pct": 60.724015154394905, + "weighted_f1_pct": 76.01570428948048, + "macro_precision_pct": 65.18624273293203, + "macro_recall_pct": 74.70695303967965, + "per_class_f1_pct": { + "FAVOR": 73.87387387387388, + "AGAINST": 22.22222222222222, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6829853049706481, + "weighted_f1": 0.716535017710578, + "macro_precision": 0.6687278582930757, + "macro_recall": 0.7140291806958473, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.7676767676767676, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.29853049706482, + "weighted_f1_pct": 71.6535017710578, + "macro_precision_pct": 66.87278582930757, + "macro_recall_pct": 71.40291806958473, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 76.76767676767676, + "NONE": 67.36842105263158 + } + } + ], + "accuracy": 0.711038961038961, + "macro_f1": 0.7102570742123664, + "weighted_f1": 0.7118356184163799, + "macro_precision": 0.7060735511721772, + "macro_recall": 0.7192645602619029, + "per_class_f1": { + "FAVOR": 0.6657060518731989, + "AGAINST": 0.7119617224880382, + "NONE": 0.7531034482758621 + }, + "accuracy_pct": 71.1038961038961, + "macro_f1_pct": 71.02570742123663, + "weighted_f1_pct": 71.18356184163798, + "macro_precision_pct": 70.60735511721772, + "macro_recall_pct": 71.9264560261903, + "per_class_f1_pct": { + "FAVOR": 66.57060518731988, + "AGAINST": 71.19617224880382, + "NONE": 75.3103448275862 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.6302859099469269, + "losses": { + "current_ce": 0.6469366321961085, + "replay_ce": 0.23819180826346079, + "dark_distillation": 0.3770508905251821, + "total": 1.0736538906892141, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6955803175901315, + "weighted_f1": 0.7217632049921158, + "macro_precision": 0.67994227994228, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.75, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.55803175901315, + "weighted_f1_pct": 72.17632049921157, + "macro_precision_pct": 67.994227994228, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.0, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5882842795281281, + "weighted_f1": 0.5864062129417099, + "macro_precision": 0.5997059409888358, + "macro_recall": 0.622730573710966, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.5903614457831325, + "NONE": 0.6111111111111112 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.82842795281281, + "weighted_f1_pct": 58.64062129417099, + "macro_precision_pct": 59.97059409888358, + "macro_recall_pct": 62.2730573710966, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 59.036144578313255, + "NONE": 61.111111111111114 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6936507936507935, + "weighted_f1": 0.702610587382161, + "macro_precision": 0.6766156413091982, + "macro_recall": 0.7436257420640553, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7142857142857143, + "NONE": 0.7 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 69.36507936507935, + "weighted_f1_pct": 70.2610587382161, + "macro_precision_pct": 67.66156413091981, + "macro_recall_pct": 74.36257420640553, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 71.42857142857143, + "NONE": 70.0 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6555254441433302, + "weighted_f1": 0.6461111111111112, + "macro_precision": 0.6690368916073061, + "macro_recall": 0.6577413352481103, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5925925925925926, + "NONE": 0.7073170731707317 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 65.55254441433303, + "weighted_f1_pct": 64.61111111111111, + "macro_precision_pct": 66.90368916073061, + "macro_recall_pct": 65.77413352481103, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 59.25925925925925, + "NONE": 70.73170731707317 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7965408805031448, + "weighted_f1": 0.8082830188679244, + "macro_precision": 0.7768375953942964, + "macro_recall": 0.8301293797545671, + "per_class_f1": { + "FAVOR": 0.8, + "AGAINST": 0.75, + "NONE": 0.839622641509434 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 79.65408805031447, + "weighted_f1_pct": 80.82830188679245, + "macro_precision_pct": 77.68375953942964, + "macro_recall_pct": 83.01293797545671, + "per_class_f1_pct": { + "FAVOR": 80.0, + "AGAINST": 75.0, + "NONE": 83.9622641509434 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6684054684054684, + "weighted_f1": 0.8028580842740135, + "macro_precision": 0.6671759612936082, + "macro_recall": 0.7792258221089673, + "per_class_f1": { + "FAVOR": 0.8307692307692308, + "AGAINST": 0.36363636363636365, + "NONE": 0.8108108108108109 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.84054684054685, + "weighted_f1_pct": 80.28580842740135, + "macro_precision_pct": 66.71759612936083, + "macro_recall_pct": 77.92258221089673, + "per_class_f1_pct": { + "FAVOR": 83.07692307692308, + "AGAINST": 36.36363636363637, + "NONE": 81.08108108108108 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6934202186339963, + "weighted_f1": 0.7172590275698526, + "macro_precision": 0.6799411086622826, + "macro_recall": 0.7437149270482605, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.7513227513227513, + "NONE": 0.7010309278350515 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 69.34202186339962, + "weighted_f1_pct": 71.72590275698526, + "macro_precision_pct": 67.99411086622827, + "macro_recall_pct": 74.37149270482605, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 75.13227513227513, + "NONE": 70.10309278350515 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6186628186628186, + "weighted_f1": 0.599109810377416, + "macro_precision": 0.6112617702448211, + "macro_recall": 0.6316239316239316, + "per_class_f1": { + "FAVOR": 0.6984126984126984, + "AGAINST": 0.4909090909090909, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 61.86628186628186, + "weighted_f1_pct": 59.91098103774159, + "macro_precision_pct": 61.12617702448211, + "macro_recall_pct": 63.16239316239316, + "per_class_f1_pct": { + "FAVOR": 69.84126984126983, + "AGAINST": 49.09090909090909, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6914119359534207, + "macro_f1": 0.6937545818682378, + "weighted_f1": 0.6903451710222985, + "macro_precision": 0.6907404418418599, + "macro_recall": 0.7122723028605381, + "per_class_f1": { + "FAVOR": 0.6773794808405439, + "AGAINST": 0.6666666666666666, + "NONE": 0.737217598097503 + }, + "accuracy_pct": 69.14119359534206, + "macro_f1_pct": 69.37545818682378, + "weighted_f1_pct": 69.03451710222986, + "macro_precision_pct": 69.074044184186, + "macro_recall_pct": 71.22723028605381, + "per_class_f1_pct": { + "FAVOR": 67.7379480840544, + "AGAINST": 66.66666666666666, + "NONE": 73.7217598097503 + } + }, + { + "after_task": 9, + "best_epoch": 3, + "best_dev_macro_f1": 0.7123018203170876, + "losses": { + "current_ce": 0.5786916189867517, + "replay_ce": 0.2072711657786715, + "dark_distillation": 0.35583230766697205, + "total": 0.9638789411904155, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6955803175901315, + "weighted_f1": 0.7217632049921158, + "macro_precision": 0.67994227994228, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.75, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.55803175901315, + "weighted_f1_pct": 72.17632049921157, + "macro_precision_pct": 67.994227994228, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.0, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5763509923517453, + "weighted_f1": 0.5882876975460348, + "macro_precision": 0.582468637748581, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5693430656934306, + "AGAINST": 0.611764705882353, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 57.63509923517452, + "weighted_f1_pct": 58.828769754603485, + "macro_precision_pct": 58.246863774858106, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 56.934306569343065, + "AGAINST": 61.1764705882353, + "NONE": 54.794520547945204 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6751269035532995, + "macro_f1": 0.6707314701363362, + "weighted_f1": 0.677278910597524, + "macro_precision": 0.6522130647130647, + "macro_recall": 0.7216832563006431, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6903553299492385, + "NONE": 0.6551724137931034 + }, + "accuracy_pct": 67.51269035532995, + "macro_f1_pct": 67.07314701363362, + "weighted_f1_pct": 67.7278910597524, + "macro_precision_pct": 65.22130647130648, + "macro_recall_pct": 72.16832563006432, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 69.03553299492386, + "NONE": 65.51724137931035 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6686274509803921, + "weighted_f1": 0.6619915966386555, + "macro_precision": 0.6712214804184605, + "macro_recall": 0.6745767030319877, + "per_class_f1": { + "FAVOR": 0.6857142857142857, + "AGAINST": 0.6142857142857143, + "NONE": 0.7058823529411765 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 66.86274509803921, + "weighted_f1_pct": 66.19915966386554, + "macro_precision_pct": 67.12214804184605, + "macro_recall_pct": 67.45767030319877, + "per_class_f1_pct": { + "FAVOR": 68.57142857142857, + "AGAINST": 61.42857142857143, + "NONE": 70.58823529411765 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.8080849148563699, + "weighted_f1": 0.8177772514323227, + "macro_precision": 0.7894557823129252, + "macro_recall": 0.8387750569159865, + "per_class_f1": { + "FAVOR": 0.8135593220338984, + "AGAINST": 0.765625, + "NONE": 0.8450704225352113 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 80.80849148563699, + "weighted_f1_pct": 81.77772514323227, + "macro_precision_pct": 78.94557823129253, + "macro_recall_pct": 83.87750569159866, + "per_class_f1_pct": { + "FAVOR": 81.35593220338984, + "AGAINST": 76.5625, + "NONE": 84.50704225352112 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7345132743362832, + "macro_f1": 0.6323855037833532, + "weighted_f1": 0.7720865712986752, + "macro_precision": 0.6452293836194766, + "macro_recall": 0.7575051571411237, + "per_class_f1": { + "FAVOR": 0.7741935483870968, + "AGAINST": 0.2962962962962963, + "NONE": 0.8266666666666667 + }, + "accuracy_pct": 73.45132743362832, + "macro_f1_pct": 63.238550378335326, + "weighted_f1_pct": 77.20865712986752, + "macro_precision_pct": 64.52293836194765, + "macro_recall_pct": 75.75051571411237, + "per_class_f1_pct": { + "FAVOR": 77.41935483870968, + "AGAINST": 29.629629629629626, + "NONE": 82.66666666666667 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6857394354385301, + "weighted_f1": 0.7159126427247006, + "macro_precision": 0.6702151157207337, + "macro_recall": 0.7183501683501684, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7614213197969543, + "NONE": 0.6804123711340206 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.573943543853, + "weighted_f1_pct": 71.59126427247006, + "macro_precision_pct": 67.02151157207336, + "macro_recall_pct": 71.83501683501684, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.14213197969542, + "NONE": 68.04123711340206 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5563380281690141, + "macro_f1": 0.5527339449541285, + "weighted_f1": 0.5546091226256622, + "macro_precision": 0.5803193882141251, + "macro_recall": 0.541025641025641, + "per_class_f1": { + "FAVOR": 0.52, + "AGAINST": 0.496, + "NONE": 0.6422018348623854 + }, + "accuracy_pct": 55.633802816901415, + "macro_f1_pct": 55.27339449541285, + "weighted_f1_pct": 55.46091226256622, + "macro_precision_pct": 58.03193882141251, + "macro_recall_pct": 54.1025641025641, + "per_class_f1_pct": { + "FAVOR": 52.0, + "AGAINST": 49.6, + "NONE": 64.22018348623854 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6931056201550388, + "weighted_f1": 0.7032477470930233, + "macro_precision": 0.6969264609714049, + "macro_recall": 0.6907012666048811, + "per_class_f1": { + "FAVOR": 0.64, + "AGAINST": 0.671875, + "NONE": 0.7674418604651163 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.31056201550389, + "weighted_f1_pct": 70.32477470930233, + "macro_precision_pct": 69.69264609714048, + "macro_recall_pct": 69.07012666048811, + "per_class_f1_pct": { + "FAVOR": 64.0, + "AGAINST": 67.1875, + "NONE": 76.74418604651163 + } + } + ], + "accuracy": 0.6855146124523507, + "macro_f1": 0.6857458677443778, + "weighted_f1": 0.6850817203301499, + "macro_precision": 0.68153158486139, + "macro_recall": 0.6957238162376642, + "per_class_f1": { + "FAVOR": 0.6590126291618829, + "AGAINST": 0.6677215189873418, + "NONE": 0.7305034550839091 + }, + "accuracy_pct": 68.55146124523507, + "macro_f1_pct": 68.57458677443779, + "weighted_f1_pct": 68.50817203301499, + "macro_precision_pct": 68.15315848613899, + "macro_recall_pct": 69.57238162376642, + "per_class_f1_pct": { + "FAVOR": 65.90126291618829, + "AGAINST": 66.77215189873418, + "NONE": 73.05034550839092 + } + }, + { + "after_task": 10, + "best_epoch": 15, + "best_dev_macro_f1": 0.6554365639205962, + "losses": { + "current_ce": 0.6387930411558885, + "replay_ce": 0.23395875646517827, + "dark_distillation": 0.3512316930752534, + "total": 1.0483676415223342, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6924690274119598, + "weighted_f1": 0.7234345997650746, + "macro_precision": 0.6772204370374405, + "macro_recall": 0.7756621689155422, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7581699346405228, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.24690274119598, + "weighted_f1_pct": 72.34345997650746, + "macro_precision_pct": 67.72204370374405, + "macro_recall_pct": 77.56621689155422, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.81699346405229, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5840995625250645, + "weighted_f1": 0.5894391707919451, + "macro_precision": 0.590239539246913, + "macro_recall": 0.620551924473493, + "per_class_f1": { + "FAVOR": 0.5441176470588235, + "AGAINST": 0.6107784431137725, + "NONE": 0.5974025974025974 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.40995625250645, + "weighted_f1_pct": 58.943917079194506, + "macro_precision_pct": 59.0239539246913, + "macro_recall_pct": 62.055192447349306, + "per_class_f1_pct": { + "FAVOR": 54.41176470588235, + "AGAINST": 61.07784431137725, + "NONE": 59.74025974025974 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6950100177631083, + "weighted_f1": 0.7029056449701236, + "macro_precision": 0.6812091503267975, + "macro_recall": 0.7364745844704199, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.7171717171717171, + "NONE": 0.6829268292682927 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 69.50100177631083, + "weighted_f1_pct": 70.29056449701237, + "macro_precision_pct": 68.12091503267975, + "macro_recall_pct": 73.647458447042, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 71.71717171717171, + "NONE": 68.29268292682927 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6417022674991785, + "weighted_f1": 0.6254035491291489, + "macro_precision": 0.6471168404665941, + "macro_recall": 0.6489935839529336, + "per_class_f1": { + "FAVOR": 0.6368715083798883, + "AGAINST": 0.5588235294117647, + "NONE": 0.7294117647058823 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.17022674991784, + "weighted_f1_pct": 62.540354912914886, + "macro_precision_pct": 64.71168404665941, + "macro_recall_pct": 64.89935839529336, + "per_class_f1_pct": { + "FAVOR": 63.687150837988824, + "AGAINST": 55.88235294117647, + "NONE": 72.94117647058823 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.825, + "macro_f1": 0.8067380041136193, + "weighted_f1": 0.8279270483480434, + "macro_precision": 0.789197118241236, + "macro_recall": 0.83227645436541, + "per_class_f1": { + "FAVOR": 0.7796610169491526, + "AGAINST": 0.7741935483870968, + "NONE": 0.8663594470046083 + }, + "accuracy_pct": 82.5, + "macro_f1_pct": 80.67380041136192, + "weighted_f1_pct": 82.79270483480434, + "macro_precision_pct": 78.9197118241236, + "macro_recall_pct": 83.227645436541, + "per_class_f1_pct": { + "FAVOR": 77.96610169491525, + "AGAINST": 77.41935483870968, + "NONE": 86.63594470046083 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6399982832664075, + "weighted_f1": 0.7773558661263988, + "macro_precision": 0.6430892044927133, + "macro_recall": 0.7687901953646402, + "per_class_f1": { + "FAVOR": 0.7768595041322314, + "AGAINST": 0.3076923076923077, + "NONE": 0.8354430379746836 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.99982832664075, + "weighted_f1_pct": 77.73558661263988, + "macro_precision_pct": 64.30892044927134, + "macro_recall_pct": 76.87901953646403, + "per_class_f1_pct": { + "FAVOR": 77.68595041322314, + "AGAINST": 30.76923076923077, + "NONE": 83.54430379746836 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6953016688061617, + "weighted_f1": 0.7153091225309537, + "macro_precision": 0.6783519616272229, + "macro_recall": 0.7437149270482605, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.7473684210526316, + "NONE": 0.68 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 69.53016688061618, + "weighted_f1_pct": 71.53091225309537, + "macro_precision_pct": 67.83519616272228, + "macro_recall_pct": 74.37149270482605, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 74.73684210526315, + "NONE": 68.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5422535211267606, + "macro_f1": 0.5591470258136925, + "weighted_f1": 0.5345473514487599, + "macro_precision": 0.5475132275132275, + "macro_recall": 0.5773504273504274, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.41818181818181815, + "NONE": 0.5925925925925926 + }, + "accuracy_pct": 54.22535211267606, + "macro_f1_pct": 55.91470258136925, + "weighted_f1_pct": 53.45473514487599, + "macro_precision_pct": 54.75132275132275, + "macro_recall_pct": 57.73504273504274, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 41.81818181818181, + "NONE": 59.25925925925925 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.68, + "macro_f1": 0.6671717171717172, + "weighted_f1": 0.6789924242424243, + "macro_precision": 0.6693225051214515, + "macro_recall": 0.6658325610132839, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6515151515151515, + "NONE": 0.75 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 66.71717171717172, + "weighted_f1_pct": 67.89924242424243, + "macro_precision_pct": 66.93225051214515, + "macro_recall_pct": 66.58325610132839, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 65.15151515151516, + "NONE": 75.0 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6359478520326857, + "weighted_f1": 0.6666661733292645, + "macro_precision": 0.6318534318178826, + "macro_recall": 0.641173174182883, + "per_class_f1": { + "FAVOR": 0.71, + "AGAINST": 0.4367816091954023, + "NONE": 0.7610619469026548 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.594785203268565, + "weighted_f1_pct": 66.66661733292645, + "macro_precision_pct": 63.185343181788255, + "macro_recall_pct": 64.1173174182883, + "per_class_f1_pct": { + "FAVOR": 71.0, + "AGAINST": 43.67816091954023, + "NONE": 76.10619469026548 + } + } + ], + "accuracy": 0.6798196166854565, + "macro_f1": 0.6814223764893832, + "weighted_f1": 0.6781140615127972, + "macro_precision": 0.6791610351207202, + "macro_recall": 0.6903883261428144, + "per_class_f1": { + "FAVOR": 0.6629732225300092, + "AGAINST": 0.6439909297052154, + "NONE": 0.7373029772329247 + }, + "accuracy_pct": 67.98196166854565, + "macro_f1_pct": 68.14223764893832, + "weighted_f1_pct": 67.81140615127971, + "macro_precision_pct": 67.91610351207203, + "macro_recall_pct": 69.03883261428145, + "per_class_f1_pct": { + "FAVOR": 66.29732225300093, + "AGAINST": 64.39909297052154, + "NONE": 73.73029772329247 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6617585426403595, + "average_forgetting": 0.028453794977818303, + "backward_transfer": -0.018149220665020598, + "final_mean_target_macro_f1_pct": 66.17585426403595, + "average_forgetting_pct": 2.8453794977818303, + "backward_transfer_pct": -1.8149220665020598, + "final_pooled_accuracy": 0.6798196166854565, + "final_pooled_accuracy_pct": 67.98196166854565, + "final_pooled_macro_f1": 0.6814223764893832, + "final_pooled_macro_f1_pct": 68.14223764893832, + "final_pooled_weighted_f1": 0.6781140615127972, + "final_pooled_weighted_f1_pct": 67.81140615127971, + "final_pooled_macro_precision": 0.6791610351207202, + "final_pooled_macro_precision_pct": 67.91610351207203, + "final_pooled_macro_recall": 0.6903883261428144, + "final_pooled_macro_recall_pct": 69.03883261428145, + "final_pooled_per_class_f1": { + "FAVOR": 0.6629732225300092, + "AGAINST": 0.6439909297052154, + "NONE": 0.7373029772329247 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 66.29732225300093, + "AGAINST": 64.39909297052154, + "NONE": 73.73029772329247 + } + }, + "efficiency": { + "training_wall_time_seconds": 965.2921556148212, + "training_wall_time_minutes": 16.088202593580355, + "peak_gpu_memory_allocated_gb": 10.397007942199707, + "peak_gpu_memory_reserved_gb": 10.580078125 + } +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s3__seed13/metrics.partial.json b/outputs/trace_s6_derpp__x__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..9d92b2b18ee8d4828e9008c5200a6a097ab50be4 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s3__seed13/metrics.partial.json @@ -0,0 +1,2169 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.6016841791489679, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7093435860466979, + 0.6161020036429873, + 0.6980752238363291, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7229955162713, + 0.6139385359182533, + 0.6574383881993164, + 0.670565066822749, + null, + null, + null, + null, + null, + null + ], + [ + 0.6961155336441379, + 0.6277700655714157, + 0.7145617667356797, + 0.6426875909887235, + 0.7906251215622203, + null, + null, + null, + null, + null + ], + [ + 0.6532146957520092, + 0.613289760348584, + 0.6959108536207355, + 0.6433780171397122, + 0.8002331002331003, + 0.6739929341168533, + null, + null, + null, + null + ], + [ + 0.6981721981721982, + 0.6342857142857143, + 0.717799266496919, + 0.6753747256922636, + 0.7735449735449734, + 0.6072401515439491, + 0.6829853049706481, + null, + null, + null + ], + [ + 0.6955803175901315, + 0.5882842795281281, + 0.6936507936507935, + 0.6555254441433302, + 0.7965408805031448, + 0.6684054684054684, + 0.6934202186339963, + 0.6186628186628186, + null, + null + ], + [ + 0.6955803175901315, + 0.5763509923517453, + 0.6707314701363362, + 0.6686274509803921, + 0.8080849148563699, + 0.6323855037833532, + 0.6857394354385301, + 0.5527339449541285, + 0.6931056201550388, + null + ], + [ + 0.6924690274119598, + 0.5840995625250645, + 0.6950100177631083, + 0.6417022674991785, + 0.8067380041136193, + 0.6399982832664075, + 0.6953016688061617, + 0.5591470258136925, + 0.6671717171717172, + 0.6359478520326857 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 15, + "best_dev_macro_f1": 0.62717150692173, + "losses": { + "current_ce": 0.7311763364257234, + "replay_ce": 0.21666293532559366, + "dark_distillation": 0.3519646299607826, + "total": 1.1238215887185299, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.6016841791489679, + "weighted_f1": 0.6250979561580007, + "macro_precision": 0.5981259036814593, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.6031746031746031, + "AGAINST": 0.6666666666666666, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 60.168417914896786, + "weighted_f1_pct": 62.509795615800066, + "macro_precision_pct": 59.81259036814593, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 60.317460317460316, + "AGAINST": 66.66666666666666, + "NONE": 53.52112676056338 + } + } + ], + "accuracy": 0.6726190476190477, + "macro_f1": 0.658125419259498, + "weighted_f1": 0.6801621558142686, + "macro_precision": 0.6473274758333041, + "macro_recall": 0.6927315824614532, + "per_class_f1": { + "FAVOR": 0.5894736842105263, + "AGAINST": 0.7230320699708455, + "NONE": 0.6618705035971223 + }, + "accuracy_pct": 67.26190476190477, + "macro_f1_pct": 65.8125419259498, + "weighted_f1_pct": 68.01621558142686, + "macro_precision_pct": 64.73274758333041, + "macro_recall_pct": 69.27315824614531, + "per_class_f1_pct": { + "FAVOR": 58.94736842105262, + "AGAINST": 72.30320699708454, + "NONE": 66.18705035971223 + } + }, + { + "after_task": 3, + "best_epoch": 13, + "best_dev_macro_f1": 0.7474842767295597, + "losses": { + "current_ce": 0.6197961993289717, + "replay_ce": 0.19484097415750676, + "dark_distillation": 0.3260704606771469, + "total": 0.9776724049539277, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7465753424657534, + "macro_f1": 0.7093435860466979, + "weighted_f1": 0.7548857021062281, + "macro_precision": 0.6866839122936684, + "macro_recall": 0.762808595702149, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.8048780487804879, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 74.65753424657534, + "macro_f1_pct": 70.9343586046698, + "weighted_f1_pct": 75.48857021062281, + "macro_precision_pct": 68.66839122936685, + "macro_recall_pct": 76.2808595702149, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 80.48780487804879, + "NONE": 77.14285714285715 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6161020036429873, + "weighted_f1": 0.6319286741443774, + "macro_precision": 0.6088025695071765, + "macro_recall": 0.6379811183732752, + "per_class_f1": { + "FAVOR": 0.5573770491803278, + "AGAINST": 0.6775956284153005, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 61.61020036429873, + "weighted_f1_pct": 63.192867414437735, + "macro_precision_pct": 60.88025695071765, + "macro_recall_pct": 63.79811183732752, + "per_class_f1_pct": { + "FAVOR": 55.73770491803278, + "AGAINST": 67.75956284153006, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6980752238363291, + "weighted_f1": 0.7129936874192117, + "macro_precision": 0.6934777642400244, + "macro_recall": 0.7137486262941755, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.7393364928909952, + "NONE": 0.6779661016949152 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.80752238363291, + "weighted_f1_pct": 71.29936874192117, + "macro_precision_pct": 69.34777642400243, + "macro_recall_pct": 71.37486262941755, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 73.93364928909952, + "NONE": 67.79661016949152 + } + } + ], + "accuracy": 0.6904315196998124, + "macro_f1": 0.6701723343029786, + "weighted_f1": 0.6951692687471527, + "macro_precision": 0.655365786856778, + "macro_recall": 0.7011135330182133, + "per_class_f1": { + "FAVOR": 0.5877551020408164, + "AGAINST": 0.7383512544802867, + "NONE": 0.6844106463878327 + }, + "accuracy_pct": 69.04315196998124, + "macro_f1_pct": 67.01723343029786, + "weighted_f1_pct": 69.51692687471527, + "macro_precision_pct": 65.53657868567781, + "macro_recall_pct": 70.11135330182134, + "per_class_f1_pct": { + "FAVOR": 58.77551020408164, + "AGAINST": 73.83512544802868, + "NONE": 68.44106463878326 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.6872455400103634, + "losses": { + "current_ce": 0.6491889049136449, + "replay_ce": 0.1964105042909819, + "dark_distillation": 0.31591956705328017, + "total": 1.0035591922109088, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7229955162713, + "weighted_f1": 0.7482915139708751, + "macro_precision": 0.7022841656987998, + "macro_recall": 0.8058970514742629, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7741935483870968, + "NONE": 0.7887323943661971 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 72.29955162713, + "weighted_f1_pct": 74.82915139708751, + "macro_precision_pct": 70.22841656987998, + "macro_recall_pct": 80.58970514742629, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 77.41935483870968, + "NONE": 78.87323943661971 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.6139385359182533, + "weighted_f1": 0.6202209397985733, + "macro_precision": 0.6238178503108561, + "macro_recall": 0.6517792302106028, + "per_class_f1": { + "FAVOR": 0.5899280575539568, + "AGAINST": 0.6385542168674698, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 61.393853591825334, + "weighted_f1_pct": 62.02209397985733, + "macro_precision_pct": 62.38178503108561, + "macro_recall_pct": 65.17792302106028, + "per_class_f1_pct": { + "FAVOR": 58.992805755395686, + "AGAINST": 63.85542168674698, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6574383881993164, + "weighted_f1": 0.6750066430894063, + "macro_precision": 0.6417624521072797, + "macro_recall": 0.7044100557895456, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.7, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 65.74383881993164, + "weighted_f1_pct": 67.50066430894063, + "macro_precision_pct": 64.17624521072797, + "macro_recall_pct": 70.44100557895456, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 70.0, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.670565066822749, + "weighted_f1": 0.6597241183786947, + "macro_precision": 0.6696519772362469, + "macro_recall": 0.6943344219766984, + "per_class_f1": { + "FAVOR": 0.6705882352941176, + "AGAINST": 0.6119402985074627, + "NONE": 0.7291666666666666 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.0565066822749, + "weighted_f1_pct": 65.97241183786947, + "macro_precision_pct": 66.9651977236247, + "macro_recall_pct": 69.43344219766983, + "per_class_f1_pct": { + "FAVOR": 67.05882352941175, + "AGAINST": 61.19402985074627, + "NONE": 72.91666666666666 + } + } + ], + "accuracy": 0.6684856753069577, + "macro_f1": 0.6684043724424299, + "weighted_f1": 0.6710945629369612, + "macro_precision": 0.6638014839726986, + "macro_recall": 0.7050096485496787, + "per_class_f1": { + "FAVOR": 0.6260869565217392, + "AGAINST": 0.683969465648855, + "NONE": 0.6951566951566952 + }, + "accuracy_pct": 66.84856753069577, + "macro_f1_pct": 66.84043724424299, + "weighted_f1_pct": 67.10945629369613, + "macro_precision_pct": 66.38014839726985, + "macro_recall_pct": 70.50096485496788, + "per_class_f1_pct": { + "FAVOR": 62.60869565217392, + "AGAINST": 68.3969465648855, + "NONE": 69.51566951566952 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6644970886821171, + "losses": { + "current_ce": 0.4574271487692992, + "replay_ce": 0.2013551927016427, + "dark_distillation": 0.33779942262917756, + "total": 0.8276820568988721, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6961155336441379, + "weighted_f1": 0.7317683323451168, + "macro_precision": 0.6735909822866345, + "macro_recall": 0.7598100949525238, + "per_class_f1": { + "FAVOR": 0.5614035087719298, + "AGAINST": 0.7701863354037267, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.6115533644138, + "weighted_f1_pct": 73.17683323451169, + "macro_precision_pct": 67.35909822866346, + "macro_recall_pct": 75.98100949525238, + "per_class_f1_pct": { + "FAVOR": 56.14035087719298, + "AGAINST": 77.01863354037268, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6421052631578947, + "macro_f1": 0.6277700655714157, + "weighted_f1": 0.6471213639766652, + "macro_precision": 0.6181170681170681, + "macro_recall": 0.6535947712418301, + "per_class_f1": { + "FAVOR": 0.6050420168067226, + "AGAINST": 0.6885245901639344, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 64.21052631578948, + "macro_f1_pct": 62.77700655714157, + "weighted_f1_pct": 64.71213639766653, + "macro_precision_pct": 61.81170681170681, + "macro_recall_pct": 65.359477124183, + "per_class_f1_pct": { + "FAVOR": 60.50420168067227, + "AGAINST": 68.85245901639344, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7145617667356797, + "weighted_f1": 0.7281920316129099, + "macro_precision": 0.7020332187857962, + "macro_recall": 0.7369004990035704, + "per_class_f1": { + "FAVOR": 0.6956521739130435, + "AGAINST": 0.7523809523809524, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 71.45617667356797, + "weighted_f1_pct": 72.81920316129099, + "macro_precision_pct": 70.20332187857962, + "macro_recall_pct": 73.69004990035704, + "per_class_f1_pct": { + "FAVOR": 69.56521739130434, + "AGAINST": 75.23809523809524, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6426875909887235, + "weighted_f1": 0.6336139610629951, + "macro_precision": 0.6348927777499206, + "macro_recall": 0.6546915896509392, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.631578947368421, + "NONE": 0.6888888888888889 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 64.26875909887235, + "weighted_f1_pct": 63.36139610629951, + "macro_precision_pct": 63.48927777499206, + "macro_recall_pct": 65.46915896509392, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 63.1578947368421, + "NONE": 68.88888888888889 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7906251215622203, + "weighted_f1": 0.817178783988797, + "macro_precision": 0.7778998778998779, + "macro_recall": 0.8075850963407185, + "per_class_f1": { + "FAVOR": 0.7368421052631579, + "AGAINST": 0.7804878048780488, + "NONE": 0.8545454545454545 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 79.06251215622203, + "weighted_f1_pct": 81.71787839887969, + "macro_precision_pct": 77.78998778998779, + "macro_recall_pct": 80.75850963407186, + "per_class_f1_pct": { + "FAVOR": 73.68421052631578, + "AGAINST": 78.04878048780488, + "NONE": 85.45454545454545 + } + } + ], + "accuracy": 0.7084673097534834, + "macro_f1": 0.7009661754952897, + "weighted_f1": 0.7089731626246859, + "macro_precision": 0.6949839336982592, + "macro_recall": 0.7120972706751488, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.7237635705669482, + "NONE": 0.7487001733102253 + }, + "accuracy_pct": 70.84673097534834, + "macro_f1_pct": 70.09661754952897, + "weighted_f1_pct": 70.89731626246859, + "macro_precision_pct": 69.49839336982592, + "macro_recall_pct": 71.20972706751488, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 72.37635705669481, + "NONE": 74.87001733102254 + } + }, + { + "after_task": 6, + "best_epoch": 4, + "best_dev_macro_f1": 0.7152560083594567, + "losses": { + "current_ce": 0.4044448312549364, + "replay_ce": 0.21253694224925268, + "dark_distillation": 0.37164135646252405, + "total": 0.8028024511677878, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6532146957520092, + "weighted_f1": 0.6985436358775144, + "macro_precision": 0.6375386996904026, + "macro_recall": 0.7084257871064468, + "per_class_f1": { + "FAVOR": 0.5230769230769231, + "AGAINST": 0.75, + "NONE": 0.6865671641791045 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 65.32146957520092, + "weighted_f1_pct": 69.85436358775145, + "macro_precision_pct": 63.75386996904025, + "macro_recall_pct": 70.84257871064467, + "per_class_f1_pct": { + "FAVOR": 52.307692307692314, + "AGAINST": 75.0, + "NONE": 68.65671641791045 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.613289760348584, + "weighted_f1": 0.6263157894736842, + "macro_precision": 0.6139511360099595, + "macro_recall": 0.6245461147421931, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.6666666666666666, + "NONE": 0.6176470588235294 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 61.32897603485839, + "weighted_f1_pct": 62.63157894736842, + "macro_precision_pct": 61.395113600995955, + "macro_recall_pct": 62.45461147421931, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 66.66666666666666, + "NONE": 61.76470588235294 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.6959108536207355, + "weighted_f1": 0.724050207000931, + "macro_precision": 0.699291756836514, + "macro_recall": 0.6938883893594983, + "per_class_f1": { + "FAVOR": 0.6865671641791045, + "AGAINST": 0.782608695652174, + "NONE": 0.6185567010309279 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 69.59108536207354, + "weighted_f1_pct": 72.4050207000931, + "macro_precision_pct": 69.92917568365141, + "macro_recall_pct": 69.38883893594982, + "per_class_f1_pct": { + "FAVOR": 68.65671641791045, + "AGAINST": 78.26086956521739, + "NONE": 61.855670103092784 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6433780171397122, + "weighted_f1": 0.6305571389260162, + "macro_precision": 0.6488660676080444, + "macro_recall": 0.6388483542955088, + "per_class_f1": { + "FAVOR": 0.5962732919254659, + "AGAINST": 0.625, + "NONE": 0.7088607594936709 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.33780171397122, + "weighted_f1_pct": 63.05571389260162, + "macro_precision_pct": 64.88660676080444, + "macro_recall_pct": 63.884835429550876, + "per_class_f1_pct": { + "FAVOR": 59.62732919254658, + "AGAINST": 62.5, + "NONE": 70.88607594936708 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.8002331002331003, + "weighted_f1": 0.81852331002331, + "macro_precision": 0.781969652937395, + "macro_recall": 0.8378736557647102, + "per_class_f1": { + "FAVOR": 0.7666666666666667, + "AGAINST": 0.7878787878787878, + "NONE": 0.8461538461538461 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 80.02331002331003, + "weighted_f1_pct": 81.852331002331, + "macro_precision_pct": 78.1969652937395, + "macro_recall_pct": 83.78736557647102, + "per_class_f1_pct": { + "FAVOR": 76.66666666666667, + "AGAINST": 78.78787878787878, + "NONE": 84.61538461538461 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6739929341168533, + "weighted_f1": 0.8158039237065612, + "macro_precision": 0.6676413255360624, + "macro_recall": 0.7936658172551875, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.34782608695652173, + "NONE": 0.8354430379746836 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 67.39929341168533, + "weighted_f1_pct": 81.58039237065611, + "macro_precision_pct": 66.76413255360623, + "macro_recall_pct": 79.36658172551874, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 34.78260869565217, + "NONE": 83.54430379746836 + } + } + ], + "accuracy": 0.7065009560229446, + "macro_f1": 0.7054468432361576, + "weighted_f1": 0.7076596142454398, + "macro_precision": 0.7071395733040599, + "macro_recall": 0.7055862331992636, + "per_class_f1": { + "FAVOR": 0.6567164179104478, + "AGAINST": 0.7138047138047138, + "NONE": 0.745819397993311 + }, + "accuracy_pct": 70.65009560229446, + "macro_f1_pct": 70.54468432361575, + "weighted_f1_pct": 70.76596142454397, + "macro_precision_pct": 70.71395733040599, + "macro_recall_pct": 70.55862331992635, + "per_class_f1_pct": { + "FAVOR": 65.67164179104478, + "AGAINST": 71.38047138047138, + "NONE": 74.5819397993311 + } + }, + { + "after_task": 7, + "best_epoch": 9, + "best_dev_macro_f1": 0.6233995358995359, + "losses": { + "current_ce": 0.5832350539438652, + "replay_ce": 0.23405812274325977, + "dark_distillation": 0.36451117116393467, + "total": 0.9995487657460299, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6981721981721982, + "weighted_f1": 0.7300648362292199, + "macro_precision": 0.6813563146884283, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.81721981721982, + "weighted_f1_pct": 73.00648362292199, + "macro_precision_pct": 68.13563146884283, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6342857142857143, + "weighted_f1": 0.6409022556390976, + "macro_precision": 0.6310147654468614, + "macro_recall": 0.6655773420479303, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6628571428571428, + "NONE": 0.64 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 63.42857142857143, + "weighted_f1_pct": 64.09022556390977, + "macro_precision_pct": 63.10147654468614, + "macro_recall_pct": 66.55773420479304, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 66.28571428571428, + "NONE": 64.0 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.717799266496919, + "weighted_f1": 0.7364897445263106, + "macro_precision": 0.7145040968570381, + "macro_recall": 0.7212836327880575, + "per_class_f1": { + "FAVOR": 0.7164179104477612, + "AGAINST": 0.7767857142857143, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.7799266496919, + "weighted_f1_pct": 73.64897445263105, + "macro_precision_pct": 71.4504096857038, + "macro_recall_pct": 72.12836327880575, + "per_class_f1_pct": { + "FAVOR": 71.64179104477611, + "AGAINST": 77.67857142857143, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6753747256922636, + "weighted_f1": 0.668157931467653, + "macro_precision": 0.6831881509896375, + "macro_recall": 0.678850207305492, + "per_class_f1": { + "FAVOR": 0.6741573033707865, + "AGAINST": 0.6376811594202898, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.53747256922637, + "weighted_f1_pct": 66.8157931467653, + "macro_precision_pct": 68.31881509896375, + "macro_recall_pct": 67.8850207305492, + "per_class_f1_pct": { + "FAVOR": 67.41573033707866, + "AGAINST": 63.76811594202898, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.8, + "macro_f1": 0.7735449735449734, + "weighted_f1": 0.8040952380952381, + "macro_precision": 0.7534436713753602, + "macro_recall": 0.8063893978936457, + "per_class_f1": { + "FAVOR": 0.7301587301587301, + "AGAINST": 0.7333333333333333, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 80.0, + "macro_f1_pct": 77.35449735449734, + "weighted_f1_pct": 80.4095238095238, + "macro_precision_pct": 75.34436713753601, + "macro_recall_pct": 80.63893978936457, + "per_class_f1_pct": { + "FAVOR": 73.01587301587301, + "AGAINST": 73.33333333333333, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.6991150442477876, + "macro_f1": 0.6072401515439491, + "weighted_f1": 0.7601570428948048, + "macro_precision": 0.6518624273293203, + "macro_recall": 0.7470695303967965, + "per_class_f1": { + "FAVOR": 0.7387387387387387, + "AGAINST": 0.2222222222222222, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 69.91150442477876, + "macro_f1_pct": 60.724015154394905, + "weighted_f1_pct": 76.01570428948048, + "macro_precision_pct": 65.18624273293203, + "macro_recall_pct": 74.70695303967965, + "per_class_f1_pct": { + "FAVOR": 73.87387387387388, + "AGAINST": 22.22222222222222, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6829853049706481, + "weighted_f1": 0.716535017710578, + "macro_precision": 0.6687278582930757, + "macro_recall": 0.7140291806958473, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.7676767676767676, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.29853049706482, + "weighted_f1_pct": 71.6535017710578, + "macro_precision_pct": 66.87278582930757, + "macro_recall_pct": 71.40291806958473, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 76.76767676767676, + "NONE": 67.36842105263158 + } + } + ], + "accuracy": 0.711038961038961, + "macro_f1": 0.7102570742123664, + "weighted_f1": 0.7118356184163799, + "macro_precision": 0.7060735511721772, + "macro_recall": 0.7192645602619029, + "per_class_f1": { + "FAVOR": 0.6657060518731989, + "AGAINST": 0.7119617224880382, + "NONE": 0.7531034482758621 + }, + "accuracy_pct": 71.1038961038961, + "macro_f1_pct": 71.02570742123663, + "weighted_f1_pct": 71.18356184163798, + "macro_precision_pct": 70.60735511721772, + "macro_recall_pct": 71.9264560261903, + "per_class_f1_pct": { + "FAVOR": 66.57060518731988, + "AGAINST": 71.19617224880382, + "NONE": 75.3103448275862 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.6302859099469269, + "losses": { + "current_ce": 0.6469366321961085, + "replay_ce": 0.23819180826346079, + "dark_distillation": 0.3770508905251821, + "total": 1.0736538906892141, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6955803175901315, + "weighted_f1": 0.7217632049921158, + "macro_precision": 0.67994227994228, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.75, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.55803175901315, + "weighted_f1_pct": 72.17632049921157, + "macro_precision_pct": 67.994227994228, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.0, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5882842795281281, + "weighted_f1": 0.5864062129417099, + "macro_precision": 0.5997059409888358, + "macro_recall": 0.622730573710966, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.5903614457831325, + "NONE": 0.6111111111111112 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.82842795281281, + "weighted_f1_pct": 58.64062129417099, + "macro_precision_pct": 59.97059409888358, + "macro_recall_pct": 62.2730573710966, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 59.036144578313255, + "NONE": 61.111111111111114 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6936507936507935, + "weighted_f1": 0.702610587382161, + "macro_precision": 0.6766156413091982, + "macro_recall": 0.7436257420640553, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7142857142857143, + "NONE": 0.7 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 69.36507936507935, + "weighted_f1_pct": 70.2610587382161, + "macro_precision_pct": 67.66156413091981, + "macro_recall_pct": 74.36257420640553, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 71.42857142857143, + "NONE": 70.0 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6555254441433302, + "weighted_f1": 0.6461111111111112, + "macro_precision": 0.6690368916073061, + "macro_recall": 0.6577413352481103, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5925925925925926, + "NONE": 0.7073170731707317 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 65.55254441433303, + "weighted_f1_pct": 64.61111111111111, + "macro_precision_pct": 66.90368916073061, + "macro_recall_pct": 65.77413352481103, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 59.25925925925925, + "NONE": 70.73170731707317 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7965408805031448, + "weighted_f1": 0.8082830188679244, + "macro_precision": 0.7768375953942964, + "macro_recall": 0.8301293797545671, + "per_class_f1": { + "FAVOR": 0.8, + "AGAINST": 0.75, + "NONE": 0.839622641509434 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 79.65408805031447, + "weighted_f1_pct": 80.82830188679245, + "macro_precision_pct": 77.68375953942964, + "macro_recall_pct": 83.01293797545671, + "per_class_f1_pct": { + "FAVOR": 80.0, + "AGAINST": 75.0, + "NONE": 83.9622641509434 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6684054684054684, + "weighted_f1": 0.8028580842740135, + "macro_precision": 0.6671759612936082, + "macro_recall": 0.7792258221089673, + "per_class_f1": { + "FAVOR": 0.8307692307692308, + "AGAINST": 0.36363636363636365, + "NONE": 0.8108108108108109 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.84054684054685, + "weighted_f1_pct": 80.28580842740135, + "macro_precision_pct": 66.71759612936083, + "macro_recall_pct": 77.92258221089673, + "per_class_f1_pct": { + "FAVOR": 83.07692307692308, + "AGAINST": 36.36363636363637, + "NONE": 81.08108108108108 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6934202186339963, + "weighted_f1": 0.7172590275698526, + "macro_precision": 0.6799411086622826, + "macro_recall": 0.7437149270482605, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.7513227513227513, + "NONE": 0.7010309278350515 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 69.34202186339962, + "weighted_f1_pct": 71.72590275698526, + "macro_precision_pct": 67.99411086622827, + "macro_recall_pct": 74.37149270482605, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 75.13227513227513, + "NONE": 70.10309278350515 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.6186628186628186, + "weighted_f1": 0.599109810377416, + "macro_precision": 0.6112617702448211, + "macro_recall": 0.6316239316239316, + "per_class_f1": { + "FAVOR": 0.6984126984126984, + "AGAINST": 0.4909090909090909, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 61.86628186628186, + "weighted_f1_pct": 59.91098103774159, + "macro_precision_pct": 61.12617702448211, + "macro_recall_pct": 63.16239316239316, + "per_class_f1_pct": { + "FAVOR": 69.84126984126983, + "AGAINST": 49.09090909090909, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6914119359534207, + "macro_f1": 0.6937545818682378, + "weighted_f1": 0.6903451710222985, + "macro_precision": 0.6907404418418599, + "macro_recall": 0.7122723028605381, + "per_class_f1": { + "FAVOR": 0.6773794808405439, + "AGAINST": 0.6666666666666666, + "NONE": 0.737217598097503 + }, + "accuracy_pct": 69.14119359534206, + "macro_f1_pct": 69.37545818682378, + "weighted_f1_pct": 69.03451710222986, + "macro_precision_pct": 69.074044184186, + "macro_recall_pct": 71.22723028605381, + "per_class_f1_pct": { + "FAVOR": 67.7379480840544, + "AGAINST": 66.66666666666666, + "NONE": 73.7217598097503 + } + }, + { + "after_task": 9, + "best_epoch": 3, + "best_dev_macro_f1": 0.7123018203170876, + "losses": { + "current_ce": 0.5786916189867517, + "replay_ce": 0.2072711657786715, + "dark_distillation": 0.35583230766697205, + "total": 0.9638789411904155, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6955803175901315, + "weighted_f1": 0.7217632049921158, + "macro_precision": 0.67994227994228, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.75, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.55803175901315, + "weighted_f1_pct": 72.17632049921157, + "macro_precision_pct": 67.994227994228, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.0, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5763509923517453, + "weighted_f1": 0.5882876975460348, + "macro_precision": 0.582468637748581, + "macro_recall": 0.6067538126361657, + "per_class_f1": { + "FAVOR": 0.5693430656934306, + "AGAINST": 0.611764705882353, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 57.63509923517452, + "weighted_f1_pct": 58.828769754603485, + "macro_precision_pct": 58.246863774858106, + "macro_recall_pct": 60.675381263616565, + "per_class_f1_pct": { + "FAVOR": 56.934306569343065, + "AGAINST": 61.1764705882353, + "NONE": 54.794520547945204 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6751269035532995, + "macro_f1": 0.6707314701363362, + "weighted_f1": 0.677278910597524, + "macro_precision": 0.6522130647130647, + "macro_recall": 0.7216832563006431, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6903553299492385, + "NONE": 0.6551724137931034 + }, + "accuracy_pct": 67.51269035532995, + "macro_f1_pct": 67.07314701363362, + "weighted_f1_pct": 67.7278910597524, + "macro_precision_pct": 65.22130647130648, + "macro_recall_pct": 72.16832563006432, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 69.03553299492386, + "NONE": 65.51724137931035 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6686274509803921, + "weighted_f1": 0.6619915966386555, + "macro_precision": 0.6712214804184605, + "macro_recall": 0.6745767030319877, + "per_class_f1": { + "FAVOR": 0.6857142857142857, + "AGAINST": 0.6142857142857143, + "NONE": 0.7058823529411765 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 66.86274509803921, + "weighted_f1_pct": 66.19915966386554, + "macro_precision_pct": 67.12214804184605, + "macro_recall_pct": 67.45767030319877, + "per_class_f1_pct": { + "FAVOR": 68.57142857142857, + "AGAINST": 61.42857142857143, + "NONE": 70.58823529411765 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.8080849148563699, + "weighted_f1": 0.8177772514323227, + "macro_precision": 0.7894557823129252, + "macro_recall": 0.8387750569159865, + "per_class_f1": { + "FAVOR": 0.8135593220338984, + "AGAINST": 0.765625, + "NONE": 0.8450704225352113 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 80.80849148563699, + "weighted_f1_pct": 81.77772514323227, + "macro_precision_pct": 78.94557823129253, + "macro_recall_pct": 83.87750569159866, + "per_class_f1_pct": { + "FAVOR": 81.35593220338984, + "AGAINST": 76.5625, + "NONE": 84.50704225352112 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7345132743362832, + "macro_f1": 0.6323855037833532, + "weighted_f1": 0.7720865712986752, + "macro_precision": 0.6452293836194766, + "macro_recall": 0.7575051571411237, + "per_class_f1": { + "FAVOR": 0.7741935483870968, + "AGAINST": 0.2962962962962963, + "NONE": 0.8266666666666667 + }, + "accuracy_pct": 73.45132743362832, + "macro_f1_pct": 63.238550378335326, + "weighted_f1_pct": 77.20865712986752, + "macro_precision_pct": 64.52293836194765, + "macro_recall_pct": 75.75051571411237, + "per_class_f1_pct": { + "FAVOR": 77.41935483870968, + "AGAINST": 29.629629629629626, + "NONE": 82.66666666666667 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6857394354385301, + "weighted_f1": 0.7159126427247006, + "macro_precision": 0.6702151157207337, + "macro_recall": 0.7183501683501684, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7614213197969543, + "NONE": 0.6804123711340206 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.573943543853, + "weighted_f1_pct": 71.59126427247006, + "macro_precision_pct": 67.02151157207336, + "macro_recall_pct": 71.83501683501684, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.14213197969542, + "NONE": 68.04123711340206 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5563380281690141, + "macro_f1": 0.5527339449541285, + "weighted_f1": 0.5546091226256622, + "macro_precision": 0.5803193882141251, + "macro_recall": 0.541025641025641, + "per_class_f1": { + "FAVOR": 0.52, + "AGAINST": 0.496, + "NONE": 0.6422018348623854 + }, + "accuracy_pct": 55.633802816901415, + "macro_f1_pct": 55.27339449541285, + "weighted_f1_pct": 55.46091226256622, + "macro_precision_pct": 58.03193882141251, + "macro_recall_pct": 54.1025641025641, + "per_class_f1_pct": { + "FAVOR": 52.0, + "AGAINST": 49.6, + "NONE": 64.22018348623854 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6931056201550388, + "weighted_f1": 0.7032477470930233, + "macro_precision": 0.6969264609714049, + "macro_recall": 0.6907012666048811, + "per_class_f1": { + "FAVOR": 0.64, + "AGAINST": 0.671875, + "NONE": 0.7674418604651163 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.31056201550389, + "weighted_f1_pct": 70.32477470930233, + "macro_precision_pct": 69.69264609714048, + "macro_recall_pct": 69.07012666048811, + "per_class_f1_pct": { + "FAVOR": 64.0, + "AGAINST": 67.1875, + "NONE": 76.74418604651163 + } + } + ], + "accuracy": 0.6855146124523507, + "macro_f1": 0.6857458677443778, + "weighted_f1": 0.6850817203301499, + "macro_precision": 0.68153158486139, + "macro_recall": 0.6957238162376642, + "per_class_f1": { + "FAVOR": 0.6590126291618829, + "AGAINST": 0.6677215189873418, + "NONE": 0.7305034550839091 + }, + "accuracy_pct": 68.55146124523507, + "macro_f1_pct": 68.57458677443779, + "weighted_f1_pct": 68.50817203301499, + "macro_precision_pct": 68.15315848613899, + "macro_recall_pct": 69.57238162376642, + "per_class_f1_pct": { + "FAVOR": 65.90126291618829, + "AGAINST": 66.77215189873418, + "NONE": 73.05034550839092 + } + }, + { + "after_task": 10, + "best_epoch": 15, + "best_dev_macro_f1": 0.6554365639205962, + "losses": { + "current_ce": 0.6387930411558885, + "replay_ce": 0.23395875646517827, + "dark_distillation": 0.3512316930752534, + "total": 1.0483676415223342, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6924690274119598, + "weighted_f1": 0.7234345997650746, + "macro_precision": 0.6772204370374405, + "macro_recall": 0.7756621689155422, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7581699346405228, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.24690274119598, + "weighted_f1_pct": 72.34345997650746, + "macro_precision_pct": 67.72204370374405, + "macro_recall_pct": 77.56621689155422, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.81699346405229, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5840995625250645, + "weighted_f1": 0.5894391707919451, + "macro_precision": 0.590239539246913, + "macro_recall": 0.620551924473493, + "per_class_f1": { + "FAVOR": 0.5441176470588235, + "AGAINST": 0.6107784431137725, + "NONE": 0.5974025974025974 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.40995625250645, + "weighted_f1_pct": 58.943917079194506, + "macro_precision_pct": 59.0239539246913, + "macro_recall_pct": 62.055192447349306, + "per_class_f1_pct": { + "FAVOR": 54.41176470588235, + "AGAINST": 61.07784431137725, + "NONE": 59.74025974025974 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6950100177631083, + "weighted_f1": 0.7029056449701236, + "macro_precision": 0.6812091503267975, + "macro_recall": 0.7364745844704199, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.7171717171717171, + "NONE": 0.6829268292682927 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 69.50100177631083, + "weighted_f1_pct": 70.29056449701237, + "macro_precision_pct": 68.12091503267975, + "macro_recall_pct": 73.647458447042, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 71.71717171717171, + "NONE": 68.29268292682927 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.63, + "macro_f1": 0.6417022674991785, + "weighted_f1": 0.6254035491291489, + "macro_precision": 0.6471168404665941, + "macro_recall": 0.6489935839529336, + "per_class_f1": { + "FAVOR": 0.6368715083798883, + "AGAINST": 0.5588235294117647, + "NONE": 0.7294117647058823 + }, + "accuracy_pct": 63.0, + "macro_f1_pct": 64.17022674991784, + "weighted_f1_pct": 62.540354912914886, + "macro_precision_pct": 64.71168404665941, + "macro_recall_pct": 64.89935839529336, + "per_class_f1_pct": { + "FAVOR": 63.687150837988824, + "AGAINST": 55.88235294117647, + "NONE": 72.94117647058823 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.825, + "macro_f1": 0.8067380041136193, + "weighted_f1": 0.8279270483480434, + "macro_precision": 0.789197118241236, + "macro_recall": 0.83227645436541, + "per_class_f1": { + "FAVOR": 0.7796610169491526, + "AGAINST": 0.7741935483870968, + "NONE": 0.8663594470046083 + }, + "accuracy_pct": 82.5, + "macro_f1_pct": 80.67380041136192, + "weighted_f1_pct": 82.79270483480434, + "macro_precision_pct": 78.9197118241236, + "macro_recall_pct": 83.227645436541, + "per_class_f1_pct": { + "FAVOR": 77.96610169491525, + "AGAINST": 77.41935483870968, + "NONE": 86.63594470046083 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6399982832664075, + "weighted_f1": 0.7773558661263988, + "macro_precision": 0.6430892044927133, + "macro_recall": 0.7687901953646402, + "per_class_f1": { + "FAVOR": 0.7768595041322314, + "AGAINST": 0.3076923076923077, + "NONE": 0.8354430379746836 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.99982832664075, + "weighted_f1_pct": 77.73558661263988, + "macro_precision_pct": 64.30892044927134, + "macro_recall_pct": 76.87901953646403, + "per_class_f1_pct": { + "FAVOR": 77.68595041322314, + "AGAINST": 30.76923076923077, + "NONE": 83.54430379746836 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6953016688061617, + "weighted_f1": 0.7153091225309537, + "macro_precision": 0.6783519616272229, + "macro_recall": 0.7437149270482605, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.7473684210526316, + "NONE": 0.68 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 69.53016688061618, + "weighted_f1_pct": 71.53091225309537, + "macro_precision_pct": 67.83519616272228, + "macro_recall_pct": 74.37149270482605, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 74.73684210526315, + "NONE": 68.0 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5422535211267606, + "macro_f1": 0.5591470258136925, + "weighted_f1": 0.5345473514487599, + "macro_precision": 0.5475132275132275, + "macro_recall": 0.5773504273504274, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.41818181818181815, + "NONE": 0.5925925925925926 + }, + "accuracy_pct": 54.22535211267606, + "macro_f1_pct": 55.91470258136925, + "weighted_f1_pct": 53.45473514487599, + "macro_precision_pct": 54.75132275132275, + "macro_recall_pct": 57.73504273504274, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 41.81818181818181, + "NONE": 59.25925925925925 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.68, + "macro_f1": 0.6671717171717172, + "weighted_f1": 0.6789924242424243, + "macro_precision": 0.6693225051214515, + "macro_recall": 0.6658325610132839, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.6515151515151515, + "NONE": 0.75 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 66.71717171717172, + "weighted_f1_pct": 67.89924242424243, + "macro_precision_pct": 66.93225051214515, + "macro_recall_pct": 66.58325610132839, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 65.15151515151516, + "NONE": 75.0 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6359478520326857, + "weighted_f1": 0.6666661733292645, + "macro_precision": 0.6318534318178826, + "macro_recall": 0.641173174182883, + "per_class_f1": { + "FAVOR": 0.71, + "AGAINST": 0.4367816091954023, + "NONE": 0.7610619469026548 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.594785203268565, + "weighted_f1_pct": 66.66661733292645, + "macro_precision_pct": 63.185343181788255, + "macro_recall_pct": 64.1173174182883, + "per_class_f1_pct": { + "FAVOR": 71.0, + "AGAINST": 43.67816091954023, + "NONE": 76.10619469026548 + } + } + ], + "accuracy": 0.6798196166854565, + "macro_f1": 0.6814223764893832, + "weighted_f1": 0.6781140615127972, + "macro_precision": 0.6791610351207202, + "macro_recall": 0.6903883261428144, + "per_class_f1": { + "FAVOR": 0.6629732225300092, + "AGAINST": 0.6439909297052154, + "NONE": 0.7373029772329247 + }, + "accuracy_pct": 67.98196166854565, + "macro_f1_pct": 68.14223764893832, + "weighted_f1_pct": 67.81140615127971, + "macro_precision_pct": 67.91610351207203, + "macro_recall_pct": 69.03883261428145, + "per_class_f1_pct": { + "FAVOR": 66.29732225300093, + "AGAINST": 64.39909297052154, + "NONE": 73.73029772329247 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s6_derpp__x__s3__seed13/resolved_config.yaml b/outputs/trace_s6_derpp__x__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2ed64e060bd10400e4123a9002ad46b54a2e2343 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s6_derpp + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 6 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s6_derpp__x__s3__seed13/run.log b/outputs/trace_s6_derpp__x__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..ed829bc759124653f1e3347d34fa749051332a69 --- /dev/null +++ b/outputs/trace_s6_derpp__x__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 06:13:41,698 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s6_derpp__x__s3__seed13 device=cuda +2026-09-02 06:29:48,318 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6617585426403595, 'average_forgetting': 0.028453794977818303, 'backward_transfer': -0.018149220665020598, 'final_mean_target_macro_f1_pct': 66.17585426403595, 'average_forgetting_pct': 2.8453794977818303, 'backward_transfer_pct': -1.8149220665020598, 'final_pooled_accuracy': 0.6798196166854565, 'final_pooled_accuracy_pct': 67.98196166854565, 'final_pooled_macro_f1': 0.6814223764893832, 'final_pooled_macro_f1_pct': 68.14223764893832, 'final_pooled_weighted_f1': 0.6781140615127972, 'final_pooled_weighted_f1_pct': 67.81140615127971, 'final_pooled_macro_precision': 0.6791610351207202, 'final_pooled_macro_precision_pct': 67.91610351207203, 'final_pooled_macro_recall': 0.6903883261428144, 'final_pooled_macro_recall_pct': 69.03883261428145, 'final_pooled_per_class_f1': {'FAVOR': 0.6629732225300092, 'AGAINST': 0.6439909297052154, 'NONE': 0.7373029772329247}, 'final_pooled_per_class_f1_pct': {'FAVOR': 66.29732225300093, 'AGAINST': 64.39909297052154, 'NONE': 73.73029772329247}} +2026-09-02 06:29:48,319 [INFO] __main__: efficiency={'training_wall_time_seconds': 965.2921556148212, 'training_wall_time_minutes': 16.088202593580355, 'peak_gpu_memory_allocated_gb': 10.397007942199707, 'peak_gpu_memory_reserved_gb': 10.580078125} diff --git a/outputs/trace_s7_contrastive_replay__doc__s1__seed13/environment.json b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..f516902a790d785d405857a556ff0de4c9f0afd9 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T06:29:56.751977+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..e706323a0837c9ef2fb074e8fbf58d5360d734da --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.json @@ -0,0 +1,1601 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5529292676510033, + 0.7110802684224623, + null, + null, + null, + null, + null, + null + ], + [ + 0.5474857849051398, + 0.6842589420389547, + 0.577992794799215, + null, + null, + null, + null, + null + ], + [ + 0.5419910063445633, + 0.6883719911514116, + 0.5429839180967506, + 0.6005380584327953, + null, + null, + null, + null + ], + [ + 0.5538033990810041, + 0.6969398109319832, + 0.5496869997965116, + 0.5718693009118541, + 0.6821327640950252, + null, + null, + null + ], + [ + 0.5555344494972978, + 0.6745102861806241, + 0.5389436103907475, + 0.5480163594403195, + 0.6907029420611966, + 0.701716540815324, + null, + null + ], + [ + 0.5258408944107079, + 0.6999404499661249, + 0.5386593231724417, + 0.5647261126182014, + 0.6798903534079178, + 0.6696316482382692, + 0.675092389493213, + null + ], + [ + 0.5233170328550799, + 0.684901107041461, + 0.5226071184525704, + 0.5718081996599217, + 0.671946215673841, + 0.6378313442829572, + 0.6386522735421335, + 0.6784750789347859 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7294129145980998, + "losses": { + "current_ce": 0.5688029842149643, + "replay_ce": 0.2992020692995616, + "dark_distillation": 0.40858176557790665, + "supervised_contrastive": 4.246249643053328, + "total": 1.4969209099951244, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6060991105463787, + "macro_f1": 0.5529292676510033, + "weighted_f1": 0.6335090496033625, + "macro_precision": 0.5579641854647883, + "macro_recall": 0.5933235770926403, + "per_class_f1": { + "FAVOR": 0.4530386740331492, + "AGAINST": 0.4642857142857143, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.60991105463786, + "macro_f1_pct": 55.292926765100326, + "weighted_f1_pct": 63.35090496033625, + "macro_precision_pct": 55.79641854647883, + "macro_recall_pct": 59.33235770926403, + "per_class_f1_pct": { + "FAVOR": 45.30386740331492, + "AGAINST": 46.42857142857143, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7274220032840722, + "macro_f1": 0.7110802684224623, + "weighted_f1": 0.734623428825239, + "macro_precision": 0.710087355828214, + "macro_recall": 0.7251412600463428, + "per_class_f1": { + "FAVOR": 0.622356495468278, + "AGAINST": 0.7079646017699115, + "NONE": 0.8029197080291971 + }, + "accuracy_pct": 72.74220032840722, + "macro_f1_pct": 71.10802684224623, + "weighted_f1_pct": 73.4623428825239, + "macro_precision_pct": 71.0087355828214, + "macro_recall_pct": 72.51412600463428, + "per_class_f1_pct": { + "FAVOR": 62.2356495468278, + "AGAINST": 70.79646017699115, + "NONE": 80.2919708029197 + } + } + ], + "accuracy": 0.6590257879656161, + "macro_f1": 0.62576126230817, + "weighted_f1": 0.6748133435019595, + "macro_precision": 0.6249110667308727, + "macro_recall": 0.6555390610332643, + "per_class_f1": { + "FAVOR": 0.5339105339105339, + "AGAINST": 0.5772913816689467, + "NONE": 0.7660818713450293 + }, + "accuracy_pct": 65.90257879656161, + "macro_f1_pct": 62.576126230816996, + "weighted_f1_pct": 67.48133435019595, + "macro_precision_pct": 62.49110667308727, + "macro_recall_pct": 65.55390610332643, + "per_class_f1_pct": { + "FAVOR": 53.39105339105339, + "AGAINST": 57.729138166894664, + "NONE": 76.60818713450293 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.692250139586823, + "losses": { + "current_ce": 0.5507392211447633, + "replay_ce": 0.2011811656980879, + "dark_distillation": 0.32484443195992047, + "supervised_contrastive": 4.150537680917316, + "total": 1.329396380365841, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5474857849051398, + "weighted_f1": 0.6280694008611376, + "macro_precision": 0.5573604778697515, + "macro_recall": 0.5854527482523125, + "per_class_f1": { + "FAVOR": 0.432258064516129, + "AGAINST": 0.47555555555555556, + "NONE": 0.7346437346437347 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 54.74857849051398, + "weighted_f1_pct": 62.80694008611376, + "macro_precision_pct": 55.73604778697515, + "macro_recall_pct": 58.54527482523125, + "per_class_f1_pct": { + "FAVOR": 43.225806451612904, + "AGAINST": 47.55555555555556, + "NONE": 73.46437346437347 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6945812807881774, + "macro_f1": 0.6842589420389547, + "weighted_f1": 0.7027586455701511, + "macro_precision": 0.6896052313567691, + "macro_recall": 0.708929814487071, + "per_class_f1": { + "FAVOR": 0.6242424242424243, + "AGAINST": 0.6701846965699209, + "NONE": 0.7583497053045186 + }, + "accuracy_pct": 69.45812807881774, + "macro_f1_pct": 68.42589420389547, + "weighted_f1_pct": 70.27586455701511, + "macro_precision_pct": 68.96052313567691, + "macro_recall_pct": 70.8929814487071, + "per_class_f1_pct": { + "FAVOR": 62.42424242424243, + "AGAINST": 67.01846965699208, + "NONE": 75.83497053045186 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.612859097127223, + "macro_f1": 0.577992794799215, + "weighted_f1": 0.6244734584828522, + "macro_precision": 0.5889736836181085, + "macro_recall": 0.6069673129013604, + "per_class_f1": { + "FAVOR": 0.45188284518828453, + "AGAINST": 0.5961199294532628, + "NONE": 0.6859756097560976 + }, + "accuracy_pct": 61.2859097127223, + "macro_f1_pct": 57.799279479921495, + "weighted_f1_pct": 62.44734584828522, + "macro_precision_pct": 58.897368361810855, + "macro_recall_pct": 60.69673129013604, + "per_class_f1_pct": { + "FAVOR": 45.18828451882845, + "AGAINST": 59.61199294532628, + "NONE": 68.59756097560977 + } + } + ], + "accuracy": 0.6318758815232722, + "macro_f1": 0.6038807881862908, + "weighted_f1": 0.6469945502476803, + "macro_precision": 0.6110844398194826, + "macro_recall": 0.6361044872987649, + "per_class_f1": { + "FAVOR": 0.5096700796359499, + "AGAINST": 0.5773638968481375, + "NONE": 0.7246083880747852 + }, + "accuracy_pct": 63.18758815232722, + "macro_f1_pct": 60.388078818629076, + "weighted_f1_pct": 64.69945502476803, + "macro_precision_pct": 61.108443981948255, + "macro_recall_pct": 63.61044872987649, + "per_class_f1_pct": { + "FAVOR": 50.96700796359499, + "AGAINST": 57.73638968481375, + "NONE": 72.46083880747852 + } + }, + { + "after_task": 4, + "best_epoch": 8, + "best_dev_macro_f1": 0.6954337426167982, + "losses": { + "current_ce": 0.5132155825694402, + "replay_ce": 0.1793043177640229, + "dark_distillation": 0.3004567368772992, + "supervised_contrastive": 4.162839742292437, + "total": 1.2590322508100877, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6111817026683609, + "macro_f1": 0.5419910063445633, + "weighted_f1": 0.6377653542479916, + "macro_precision": 0.5439685972624501, + "macro_recall": 0.5691176470588234, + "per_class_f1": { + "FAVOR": 0.41194029850746267, + "AGAINST": 0.4489795918367347, + "NONE": 0.7650531286894924 + }, + "accuracy_pct": 61.11817026683609, + "macro_f1_pct": 54.19910063445633, + "weighted_f1_pct": 63.776535424799164, + "macro_precision_pct": 54.396859726245005, + "macro_recall_pct": 56.91176470588234, + "per_class_f1_pct": { + "FAVOR": 41.19402985074627, + "AGAINST": 44.89795918367347, + "NONE": 76.50531286894923 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.6883719911514116, + "weighted_f1": 0.714211914909319, + "macro_precision": 0.6915130023640662, + "macro_recall": 0.7048238894534057, + "per_class_f1": { + "FAVOR": 0.5964912280701754, + "AGAINST": 0.6783625730994152, + "NONE": 0.7902621722846442 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 68.83719911514116, + "weighted_f1_pct": 71.4211914909319, + "macro_precision_pct": 69.15130023640663, + "macro_recall_pct": 70.48238894534057, + "per_class_f1_pct": { + "FAVOR": 59.64912280701754, + "AGAINST": 67.83625730994152, + "NONE": 79.02621722846442 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5772913816689467, + "macro_f1": 0.5429839180967506, + "weighted_f1": 0.5996295679343902, + "macro_precision": 0.5576086445853888, + "macro_recall": 0.581597838670152, + "per_class_f1": { + "FAVOR": 0.39867109634551495, + "AGAINST": 0.5483234714003945, + "NONE": 0.6819571865443425 + }, + "accuracy_pct": 57.729138166894664, + "macro_f1_pct": 54.29839180967506, + "weighted_f1_pct": 59.96295679343901, + "macro_precision_pct": 55.76086445853888, + "macro_recall_pct": 58.1597838670152, + "per_class_f1_pct": { + "FAVOR": 39.8671096345515, + "AGAINST": 54.83234714003945, + "NONE": 68.19571865443424 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6502242152466368, + "macro_f1": 0.6005380584327953, + "weighted_f1": 0.6663660625426026, + "macro_precision": 0.601445382957988, + "macro_recall": 0.6201333150034015, + "per_class_f1": { + "FAVOR": 0.5473684210526316, + "AGAINST": 0.46853146853146854, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 65.02242152466367, + "macro_f1_pct": 60.05380584327953, + "weighted_f1_pct": 66.63660625426026, + "macro_precision_pct": 60.144538295798796, + "macro_recall_pct": 62.01333150034015, + "per_class_f1_pct": { + "FAVOR": 54.736842105263165, + "AGAINST": 46.85314685314685, + "NONE": 78.57142857142857 + } + } + ], + "accuracy": 0.6319742489270386, + "macro_f1": 0.5951508951232691, + "weighted_f1": 0.6501870936089699, + "macro_precision": 0.5979503615650604, + "macro_recall": 0.6214874381433705, + "per_class_f1": { + "FAVOR": 0.49337260677466865, + "AGAINST": 0.5370006548788474, + "NONE": 0.7550794237162911 + }, + "accuracy_pct": 63.19742489270386, + "macro_f1_pct": 59.51508951232691, + "weighted_f1_pct": 65.01870936089699, + "macro_precision_pct": 59.79503615650604, + "macro_recall_pct": 62.148743814337045, + "per_class_f1_pct": { + "FAVOR": 49.337260677466865, + "AGAINST": 53.70006548788474, + "NONE": 75.50794237162911 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.6846308749293825, + "losses": { + "current_ce": 0.4963967498924051, + "replay_ce": 0.19124230263488634, + "dark_distillation": 0.30105886097465245, + "supervised_contrastive": 4.159904560020991, + "total": 1.2541589464460101, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6099110546378653, + "macro_f1": 0.5538033990810041, + "weighted_f1": 0.6354003938481126, + "macro_precision": 0.562611434277827, + "macro_recall": 0.5929645034220198, + "per_class_f1": { + "FAVOR": 0.419672131147541, + "AGAINST": 0.5, + "NONE": 0.7417380660954712 + }, + "accuracy_pct": 60.99110546378653, + "macro_f1_pct": 55.38033990810041, + "weighted_f1_pct": 63.540039384811266, + "macro_precision_pct": 56.2611434277827, + "macro_recall_pct": 59.29645034220198, + "per_class_f1_pct": { + "FAVOR": 41.967213114754095, + "AGAINST": 50.0, + "NONE": 74.17380660954713 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6969398109319832, + "weighted_f1": 0.7146845301916495, + "macro_precision": 0.7017471851320963, + "macro_recall": 0.7211050584443113, + "per_class_f1": { + "FAVOR": 0.6349206349206349, + "AGAINST": 0.6887755102040817, + "NONE": 0.7671232876712328 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.69398109319832, + "weighted_f1_pct": 71.46845301916494, + "macro_precision_pct": 70.17471851320963, + "macro_recall_pct": 72.11050584443113, + "per_class_f1_pct": { + "FAVOR": 63.49206349206349, + "AGAINST": 68.87755102040816, + "NONE": 76.71232876712328 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5841313269493844, + "macro_f1": 0.5496869997965116, + "weighted_f1": 0.6009682579673599, + "macro_precision": 0.5653138956224256, + "macro_recall": 0.5845446711281997, + "per_class_f1": { + "FAVOR": 0.41044776119402987, + "AGAINST": 0.5698729582577132, + "NONE": 0.6687402799377916 + }, + "accuracy_pct": 58.41313269493844, + "macro_f1_pct": 54.968699979651156, + "weighted_f1_pct": 60.096825796735985, + "macro_precision_pct": 56.53138956224256, + "macro_recall_pct": 58.45446711281998, + "per_class_f1_pct": { + "FAVOR": 41.04477611940299, + "AGAINST": 56.98729582577132, + "NONE": 66.87402799377917 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6233183856502242, + "macro_f1": 0.5718693009118541, + "weighted_f1": 0.6430490093184492, + "macro_precision": 0.5803086523865745, + "macro_recall": 0.5916614085612532, + "per_class_f1": { + "FAVOR": 0.515, + "AGAINST": 0.42857142857142855, + "NONE": 0.7720364741641338 + }, + "accuracy_pct": 62.33183856502242, + "macro_f1_pct": 57.1869300911854, + "weighted_f1_pct": 64.30490093184493, + "macro_precision_pct": 58.03086523865745, + "macro_recall_pct": 59.16614085612532, + "per_class_f1_pct": { + "FAVOR": 51.5, + "AGAINST": 42.857142857142854, + "NONE": 77.20364741641338 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7203219315895373, + "macro_f1": 0.6821327640950252, + "weighted_f1": 0.7194463881381205, + "macro_precision": 0.6838318741893644, + "macro_recall": 0.6807691612407898, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6141732283464567, + "NONE": 0.8235294117647058 + }, + "accuracy_pct": 72.03219315895373, + "macro_f1_pct": 68.21327640950253, + "weighted_f1_pct": 71.94463881381205, + "macro_precision_pct": 68.38318741893644, + "macro_recall_pct": 68.07691612407898, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 61.417322834645674, + "NONE": 82.35294117647058 + } + } + ], + "accuracy": 0.6416641360461586, + "macro_f1": 0.6100041812359615, + "weighted_f1": 0.6569179529289879, + "macro_precision": 0.6114679739615544, + "macro_recall": 0.6352474679222415, + "per_class_f1": { + "FAVOR": 0.5164690382081687, + "AGAINST": 0.5629860031104199, + "NONE": 0.750557502389296 + }, + "accuracy_pct": 64.16641360461585, + "macro_f1_pct": 61.00041812359615, + "weighted_f1_pct": 65.69179529289879, + "macro_precision_pct": 61.14679739615544, + "macro_recall_pct": 63.52474679222415, + "per_class_f1_pct": { + "FAVOR": 51.64690382081687, + "AGAINST": 56.298600311041994, + "NONE": 75.05575023892959 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7237239504681366, + "losses": { + "current_ce": 0.40225261479970953, + "replay_ce": 0.15782791473681018, + "dark_distillation": 0.29148755002589455, + "supervised_contrastive": 4.079310572715032, + "total": 1.1137553697540648, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6086404066073697, + "macro_f1": 0.5555344494972978, + "weighted_f1": 0.6355547136828485, + "macro_precision": 0.5617699054579425, + "macro_recall": 0.5934869757854507, + "per_class_f1": { + "FAVOR": 0.4520123839009288, + "AGAINST": 0.4722222222222222, + "NONE": 0.7423687423687424 + }, + "accuracy_pct": 60.864040660736975, + "macro_f1_pct": 55.55344494972978, + "weighted_f1_pct": 63.555471368284856, + "macro_precision_pct": 56.17699054579425, + "macro_recall_pct": 59.34869757854507, + "per_class_f1_pct": { + "FAVOR": 45.20123839009288, + "AGAINST": 47.22222222222222, + "NONE": 74.23687423687424 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6830870279146142, + "macro_f1": 0.6745102861806241, + "weighted_f1": 0.6931303545509839, + "macro_precision": 0.685955642196344, + "macro_recall": 0.7017098877172362, + "per_class_f1": { + "FAVOR": 0.601123595505618, + "AGAINST": 0.6759002770083102, + "NONE": 0.7465069860279441 + }, + "accuracy_pct": 68.30870279146141, + "macro_f1_pct": 67.45102861806241, + "weighted_f1_pct": 69.31303545509839, + "macro_precision_pct": 68.5955642196344, + "macro_recall_pct": 70.17098877172361, + "per_class_f1_pct": { + "FAVOR": 60.1123595505618, + "AGAINST": 67.59002770083103, + "NONE": 74.6506986027944 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5663474692202463, + "macro_f1": 0.5389436103907475, + "weighted_f1": 0.5877168891759554, + "macro_precision": 0.5594265890584552, + "macro_recall": 0.5862269046902459, + "per_class_f1": { + "FAVOR": 0.41025641025641024, + "AGAINST": 0.5514563106796116, + "NONE": 0.6551181102362205 + }, + "accuracy_pct": 56.634746922024625, + "macro_f1_pct": 53.894361039074745, + "weighted_f1_pct": 58.77168891759554, + "macro_precision_pct": 55.94265890584552, + "macro_recall_pct": 58.62269046902458, + "per_class_f1_pct": { + "FAVOR": 41.02564102564102, + "AGAINST": 55.14563106796116, + "NONE": 65.51181102362204 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5964125560538116, + "macro_f1": 0.5480163594403195, + "weighted_f1": 0.6208773849476538, + "macro_precision": 0.5589503462410962, + "macro_recall": 0.5671551123472064, + "per_class_f1": { + "FAVOR": 0.47668393782383417, + "AGAINST": 0.4131147540983607, + "NONE": 0.7542503863987635 + }, + "accuracy_pct": 59.64125560538116, + "macro_f1_pct": 54.80163594403194, + "weighted_f1_pct": 62.087738494765375, + "macro_precision_pct": 55.89503462410962, + "macro_recall_pct": 56.71551123472064, + "per_class_f1_pct": { + "FAVOR": 47.66839378238342, + "AGAINST": 41.31147540983607, + "NONE": 75.42503863987635 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7203219315895373, + "macro_f1": 0.6907029420611966, + "weighted_f1": 0.7242014053190414, + "macro_precision": 0.686484966378293, + "macro_recall": 0.6974223407532176, + "per_class_f1": { + "FAVOR": 0.6065573770491803, + "AGAINST": 0.6490566037735849, + "NONE": 0.8164948453608247 + }, + "accuracy_pct": 72.03219315895373, + "macro_f1_pct": 69.07029420611967, + "weighted_f1_pct": 72.42014053190414, + "macro_precision_pct": 68.64849663782931, + "macro_recall_pct": 69.74223407532176, + "per_class_f1_pct": { + "FAVOR": 60.65573770491803, + "AGAINST": 64.90566037735849, + "NONE": 81.64948453608247 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7303822937625755, + "macro_f1": 0.701716540815324, + "weighted_f1": 0.7422749490709618, + "macro_precision": 0.7044804772874947, + "macro_recall": 0.7214754984869928, + "per_class_f1": { + "FAVOR": 0.6132404181184669, + "AGAINST": 0.6606334841628959, + "NONE": 0.831275720164609 + }, + "accuracy_pct": 73.03822937625755, + "macro_f1_pct": 70.1716540815324, + "weighted_f1_pct": 74.22749490709619, + "macro_precision_pct": 70.44804772874947, + "macro_recall_pct": 72.14754984869927, + "per_class_f1_pct": { + "FAVOR": 61.32404181184668, + "AGAINST": 66.06334841628959, + "NONE": 83.1275720164609 + } + } + ], + "accuracy": 0.6408970976253299, + "macro_f1": 0.6113428675554314, + "weighted_f1": 0.6576535812547676, + "macro_precision": 0.6148863062972453, + "macro_recall": 0.6393910158135551, + "per_class_f1": { + "FAVOR": 0.5220125786163522, + "AGAINST": 0.5602667937112911, + "NONE": 0.751749230338651 + }, + "accuracy_pct": 64.08970976253299, + "macro_f1_pct": 61.134286755543144, + "weighted_f1_pct": 65.76535812547675, + "macro_precision_pct": 61.488630629724526, + "macro_recall_pct": 63.93910158135551, + "per_class_f1_pct": { + "FAVOR": 52.20125786163522, + "AGAINST": 56.026679371129106, + "NONE": 75.1749230338651 + } + }, + { + "after_task": 7, + "best_epoch": 10, + "best_dev_macro_f1": 0.7522783521410902, + "losses": { + "current_ce": 0.34459639222520155, + "replay_ce": 0.1451279732027674, + "dark_distillation": 0.27080509992149787, + "supervised_contrastive": 4.003387960961194, + "total": 1.0254657183236222, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5806861499364676, + "macro_f1": 0.5258408944107079, + "weighted_f1": 0.6083763776384752, + "macro_precision": 0.5604840914146968, + "macro_recall": 0.5662520264372116, + "per_class_f1": { + "FAVOR": 0.3938223938223938, + "AGAINST": 0.4674329501915709, + "NONE": 0.7162673392181589 + }, + "accuracy_pct": 58.06861499364676, + "macro_f1_pct": 52.584089441070795, + "weighted_f1_pct": 60.83763776384752, + "macro_precision_pct": 56.048409141469676, + "macro_recall_pct": 56.62520264372116, + "per_class_f1_pct": { + "FAVOR": 39.38223938223938, + "AGAINST": 46.74329501915709, + "NONE": 71.62673392181588 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6999404499661249, + "weighted_f1": 0.7155048541649449, + "macro_precision": 0.7025989844314294, + "macro_recall": 0.7229464673585922, + "per_class_f1": { + "FAVOR": 0.6559485530546624, + "AGAINST": 0.680306905370844, + "NONE": 0.7635658914728682 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.99404499661249, + "weighted_f1_pct": 71.5504854164945, + "macro_precision_pct": 70.25989844314294, + "macro_recall_pct": 72.29464673585923, + "per_class_f1_pct": { + "FAVOR": 65.59485530546624, + "AGAINST": 68.03069053708441, + "NONE": 76.35658914728683 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5386593231724417, + "weighted_f1": 0.5865355989636675, + "macro_precision": 0.5607224877331329, + "macro_recall": 0.5735003195666069, + "per_class_f1": { + "FAVOR": 0.4166666666666667, + "AGAINST": 0.5432098765432098, + "NONE": 0.6561014263074485 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 53.86593231724417, + "weighted_f1_pct": 58.653559896366744, + "macro_precision_pct": 56.07224877331329, + "macro_recall_pct": 57.350031956660686, + "per_class_f1_pct": { + "FAVOR": 41.66666666666667, + "AGAINST": 54.32098765432099, + "NONE": 65.61014263074485 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6113602391629297, + "macro_f1": 0.5647261126182014, + "weighted_f1": 0.6336489083677446, + "macro_precision": 0.5762941388867896, + "macro_recall": 0.5896092787054369, + "per_class_f1": { + "FAVOR": 0.4625, + "AGAINST": 0.46831955922865015, + "NONE": 0.7633587786259542 + }, + "accuracy_pct": 61.13602391629297, + "macro_f1_pct": 56.47261126182014, + "weighted_f1_pct": 63.36489083677446, + "macro_precision_pct": 57.62941388867896, + "macro_recall_pct": 58.96092787054369, + "per_class_f1_pct": { + "FAVOR": 46.25, + "AGAINST": 46.83195592286501, + "NONE": 76.33587786259542 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6798903534079178, + "weighted_f1": 0.7145490817721899, + "macro_precision": 0.6807565738572475, + "macro_recall": 0.685405481278731, + "per_class_f1": { + "FAVOR": 0.5892857142857143, + "AGAINST": 0.6405693950177936, + "NONE": 0.8098159509202454 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.98903534079179, + "weighted_f1_pct": 71.454908177219, + "macro_precision_pct": 68.07565738572475, + "macro_recall_pct": 68.54054812787311, + "per_class_f1_pct": { + "FAVOR": 58.92857142857143, + "AGAINST": 64.05693950177937, + "NONE": 80.98159509202453 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6981891348088531, + "macro_f1": 0.6696316482382692, + "weighted_f1": 0.7125030897142438, + "macro_precision": 0.6690685265913713, + "macro_recall": 0.6962019490755122, + "per_class_f1": { + "FAVOR": 0.5692883895131086, + "AGAINST": 0.6328125, + "NONE": 0.8067940552016986 + }, + "accuracy_pct": 69.81891348088531, + "macro_f1_pct": 66.96316482382692, + "weighted_f1_pct": 71.25030897142437, + "macro_precision_pct": 66.90685265913713, + "macro_recall_pct": 69.62019490755122, + "per_class_f1_pct": { + "FAVOR": 56.92883895131086, + "AGAINST": 63.28125, + "NONE": 80.67940552016985 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7154811715481172, + "macro_f1": 0.675092389493213, + "weighted_f1": 0.7257851205305241, + "macro_precision": 0.665665572407237, + "macro_recall": 0.7097729373438751, + "per_class_f1": { + "FAVOR": 0.5617977528089888, + "AGAINST": 0.6666666666666666, + "NONE": 0.796812749003984 + }, + "accuracy_pct": 71.54811715481172, + "macro_f1_pct": 67.5092389493213, + "weighted_f1_pct": 72.5785120530524, + "macro_precision_pct": 66.5665572407237, + "macro_recall_pct": 70.97729373438752, + "per_class_f1_pct": { + "FAVOR": 56.17977528089888, + "AGAINST": 66.66666666666666, + "NONE": 79.6812749003984 + } + } + ], + "accuracy": 0.6496560905258487, + "macro_f1": 0.6172147762045124, + "weighted_f1": 0.6644349240609934, + "macro_precision": 0.6229001481618165, + "macro_recall": 0.6441414865925155, + "per_class_f1": { + "FAVOR": 0.5219665271966527, + "AGAINST": 0.5748031496062992, + "NONE": 0.754874651810585 + }, + "accuracy_pct": 64.96560905258487, + "macro_f1_pct": 61.72147762045124, + "weighted_f1_pct": 66.44349240609934, + "macro_precision_pct": 62.290014816181646, + "macro_recall_pct": 64.41414865925155, + "per_class_f1_pct": { + "FAVOR": 52.19665271966527, + "AGAINST": 57.48031496062992, + "NONE": 75.4874651810585 + } + }, + { + "after_task": 8, + "best_epoch": 11, + "best_dev_macro_f1": 0.7794668489405332, + "losses": { + "current_ce": 0.3272139158172935, + "replay_ce": 0.13582197868940876, + "dark_distillation": 0.2597892032242289, + "supervised_contrastive": 4.019440527990753, + "total": 0.9948745572099499, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.5233170328550799, + "weighted_f1": 0.6103155283478797, + "macro_precision": 0.5414929153728992, + "macro_recall": 0.5604122195080801, + "per_class_f1": { + "FAVOR": 0.4072948328267477, + "AGAINST": 0.4365256124721604, + "NONE": 0.7261306532663316 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 52.33170328550799, + "weighted_f1_pct": 61.03155283478797, + "macro_precision_pct": 54.149291537289926, + "macro_recall_pct": 56.04122195080801, + "per_class_f1_pct": { + "FAVOR": 40.72948328267477, + "AGAINST": 43.652561247216035, + "NONE": 72.61306532663316 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6962233169129721, + "macro_f1": 0.684901107041461, + "weighted_f1": 0.7047170003640627, + "macro_precision": 0.6883483856443037, + "macro_recall": 0.7076573151282276, + "per_class_f1": { + "FAVOR": 0.6217008797653959, + "AGAINST": 0.6685236768802229, + "NONE": 0.7644787644787645 + }, + "accuracy_pct": 69.6223316912972, + "macro_f1_pct": 68.4901107041461, + "weighted_f1_pct": 70.47170003640628, + "macro_precision_pct": 68.83483856443037, + "macro_recall_pct": 70.76573151282275, + "per_class_f1_pct": { + "FAVOR": 62.17008797653959, + "AGAINST": 66.85236768802228, + "NONE": 76.44787644787645 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5403556771545828, + "macro_f1": 0.5226071184525704, + "weighted_f1": 0.5639715359708796, + "macro_precision": 0.5573327829038535, + "macro_recall": 0.5918559549292727, + "per_class_f1": { + "FAVOR": 0.4273972602739726, + "AGAINST": 0.5083333333333333, + "NONE": 0.6320907617504052 + }, + "accuracy_pct": 54.03556771545828, + "macro_f1_pct": 52.260711845257035, + "weighted_f1_pct": 56.39715359708796, + "macro_precision_pct": 55.73327829038534, + "macro_recall_pct": 59.18559549292727, + "per_class_f1_pct": { + "FAVOR": 42.73972602739726, + "AGAINST": 50.83333333333333, + "NONE": 63.20907617504052 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6188340807174888, + "macro_f1": 0.5718081996599217, + "weighted_f1": 0.6376903544070702, + "macro_precision": 0.58041276309569, + "macro_recall": 0.5943239486432913, + "per_class_f1": { + "FAVOR": 0.5247524752475248, + "AGAINST": 0.43416370106761565, + "NONE": 0.7565084226646248 + }, + "accuracy_pct": 61.88340807174888, + "macro_f1_pct": 57.18081996599217, + "weighted_f1_pct": 63.76903544070702, + "macro_precision_pct": 58.041276309569, + "macro_recall_pct": 59.43239486432913, + "per_class_f1_pct": { + "FAVOR": 52.475247524752476, + "AGAINST": 43.41637010676156, + "NONE": 75.65084226646248 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.671946215673841, + "weighted_f1": 0.7138470353695248, + "macro_precision": 0.6727107103003803, + "macro_recall": 0.6716840298269923, + "per_class_f1": { + "FAVOR": 0.5892116182572614, + "AGAINST": 0.5959183673469388, + "NONE": 0.8307086614173228 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 67.1946215673841, + "weighted_f1_pct": 71.38470353695247, + "macro_precision_pct": 67.27107103003803, + "macro_recall_pct": 67.16840298269922, + "per_class_f1_pct": { + "FAVOR": 58.921161825726145, + "AGAINST": 59.591836734693885, + "NONE": 83.07086614173228 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.670020120724346, + "macro_f1": 0.6378313442829572, + "weighted_f1": 0.6888415450757244, + "macro_precision": 0.6560311207211477, + "macro_recall": 0.659939249594422, + "per_class_f1": { + "FAVOR": 0.5512820512820513, + "AGAINST": 0.5622119815668203, + "NONE": 0.8 + }, + "accuracy_pct": 67.0020120724346, + "macro_f1_pct": 63.78313442829572, + "weighted_f1_pct": 68.88415450757243, + "macro_precision_pct": 65.60311207211477, + "macro_recall_pct": 65.9939249594422, + "per_class_f1_pct": { + "FAVOR": 55.12820512820513, + "AGAINST": 56.22119815668203, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6386522735421335, + "weighted_f1": 0.6821897498335434, + "macro_precision": 0.6369362611383887, + "macro_recall": 0.6935759037074419, + "per_class_f1": { + "FAVOR": 0.5362776025236593, + "AGAINST": 0.6374695863746959, + "NONE": 0.7422096317280453 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 63.86522735421335, + "weighted_f1_pct": 68.21897498335434, + "macro_precision_pct": 63.693626113838874, + "macro_recall_pct": 69.35759037074419, + "per_class_f1_pct": { + "FAVOR": 53.62776025236593, + "AGAINST": 63.74695863746959, + "NONE": 74.22096317280453 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.717607973421927, + "macro_f1": 0.6784750789347859, + "weighted_f1": 0.7252138938001059, + "macro_precision": 0.668728505462934, + "macro_recall": 0.7091038201611278, + "per_class_f1": { + "FAVOR": 0.6323529411764706, + "AGAINST": 0.593103448275862, + "NONE": 0.8099688473520249 + }, + "accuracy_pct": 71.76079734219269, + "macro_f1_pct": 67.84750789347859, + "weighted_f1_pct": 72.52138938001059, + "macro_precision_pct": 66.8728505462934, + "macro_recall_pct": 70.91038201611278, + "per_class_f1_pct": { + "FAVOR": 63.23529411764706, + "AGAINST": 59.310344827586206, + "NONE": 80.99688473520249 + } + } + ], + "accuracy": 0.642982971227246, + "macro_f1": 0.6116104264633364, + "weighted_f1": 0.6600020141474947, + "macro_precision": 0.6143539861992621, + "macro_recall": 0.6435497052807193, + "per_class_f1": { + "FAVOR": 0.5308020147229756, + "AGAINST": 0.5505124450951684, + "NONE": 0.7535168195718654 + }, + "accuracy_pct": 64.2982971227246, + "macro_f1_pct": 61.16104264633364, + "weighted_f1_pct": 66.00020141474947, + "macro_precision_pct": 61.435398619926204, + "macro_recall_pct": 64.35497052807193, + "per_class_f1_pct": { + "FAVOR": 53.08020147229756, + "AGAINST": 55.05124450951684, + "NONE": 75.35168195718654 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6161922963053439, + "average_forgetting": 0.0423933477119672, + "backward_transfer": -0.04116903657394271, + "final_mean_target_macro_f1_pct": 61.61922963053439, + "average_forgetting_pct": 4.23933477119672, + "backward_transfer_pct": -4.116903657394271, + "final_pooled_accuracy": 0.642982971227246, + "final_pooled_accuracy_pct": 64.2982971227246, + "final_pooled_macro_f1": 0.6116104264633364, + "final_pooled_macro_f1_pct": 61.16104264633364, + "final_pooled_weighted_f1": 0.6600020141474947, + "final_pooled_weighted_f1_pct": 66.00020141474947, + "final_pooled_macro_precision": 0.6143539861992621, + "final_pooled_macro_precision_pct": 61.435398619926204, + "final_pooled_macro_recall": 0.6435497052807193, + "final_pooled_macro_recall_pct": 64.35497052807193, + "final_pooled_per_class_f1": { + "FAVOR": 0.5308020147229756, + "AGAINST": 0.5505124450951684, + "NONE": 0.7535168195718654 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 53.08020147229756, + "AGAINST": 55.05124450951684, + "NONE": 75.35168195718654 + } + }, + "efficiency": { + "training_wall_time_seconds": 1804.7932216611225, + "training_wall_time_minutes": 30.079887027685377, + "peak_gpu_memory_allocated_gb": 10.40872573852539, + "peak_gpu_memory_reserved_gb": 10.599609375 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..031ca977bac63aa724115143c481087e8900aced --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/metrics.partial.json @@ -0,0 +1,1495 @@ +{ + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5529292676510033, + 0.7110802684224623, + null, + null, + null, + null, + null, + null + ], + [ + 0.5474857849051398, + 0.6842589420389547, + 0.577992794799215, + null, + null, + null, + null, + null + ], + [ + 0.5419910063445633, + 0.6883719911514116, + 0.5429839180967506, + 0.6005380584327953, + null, + null, + null, + null + ], + [ + 0.5538033990810041, + 0.6969398109319832, + 0.5496869997965116, + 0.5718693009118541, + 0.6821327640950252, + null, + null, + null + ], + [ + 0.5555344494972978, + 0.6745102861806241, + 0.5389436103907475, + 0.5480163594403195, + 0.6907029420611966, + 0.701716540815324, + null, + null + ], + [ + 0.5258408944107079, + 0.6999404499661249, + 0.5386593231724417, + 0.5647261126182014, + 0.6798903534079178, + 0.6696316482382692, + 0.675092389493213, + null + ], + [ + 0.5233170328550799, + 0.684901107041461, + 0.5226071184525704, + 0.5718081996599217, + 0.671946215673841, + 0.6378313442829572, + 0.6386522735421335, + 0.6784750789347859 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7294129145980998, + "losses": { + "current_ce": 0.5688029842149643, + "replay_ce": 0.2992020692995616, + "dark_distillation": 0.40858176557790665, + "supervised_contrastive": 4.246249643053328, + "total": 1.4969209099951244, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6060991105463787, + "macro_f1": 0.5529292676510033, + "weighted_f1": 0.6335090496033625, + "macro_precision": 0.5579641854647883, + "macro_recall": 0.5933235770926403, + "per_class_f1": { + "FAVOR": 0.4530386740331492, + "AGAINST": 0.4642857142857143, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.60991105463786, + "macro_f1_pct": 55.292926765100326, + "weighted_f1_pct": 63.35090496033625, + "macro_precision_pct": 55.79641854647883, + "macro_recall_pct": 59.33235770926403, + "per_class_f1_pct": { + "FAVOR": 45.30386740331492, + "AGAINST": 46.42857142857143, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7274220032840722, + "macro_f1": 0.7110802684224623, + "weighted_f1": 0.734623428825239, + "macro_precision": 0.710087355828214, + "macro_recall": 0.7251412600463428, + "per_class_f1": { + "FAVOR": 0.622356495468278, + "AGAINST": 0.7079646017699115, + "NONE": 0.8029197080291971 + }, + "accuracy_pct": 72.74220032840722, + "macro_f1_pct": 71.10802684224623, + "weighted_f1_pct": 73.4623428825239, + "macro_precision_pct": 71.0087355828214, + "macro_recall_pct": 72.51412600463428, + "per_class_f1_pct": { + "FAVOR": 62.2356495468278, + "AGAINST": 70.79646017699115, + "NONE": 80.2919708029197 + } + } + ], + "accuracy": 0.6590257879656161, + "macro_f1": 0.62576126230817, + "weighted_f1": 0.6748133435019595, + "macro_precision": 0.6249110667308727, + "macro_recall": 0.6555390610332643, + "per_class_f1": { + "FAVOR": 0.5339105339105339, + "AGAINST": 0.5772913816689467, + "NONE": 0.7660818713450293 + }, + "accuracy_pct": 65.90257879656161, + "macro_f1_pct": 62.576126230816996, + "weighted_f1_pct": 67.48133435019595, + "macro_precision_pct": 62.49110667308727, + "macro_recall_pct": 65.55390610332643, + "per_class_f1_pct": { + "FAVOR": 53.39105339105339, + "AGAINST": 57.729138166894664, + "NONE": 76.60818713450293 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.692250139586823, + "losses": { + "current_ce": 0.5507392211447633, + "replay_ce": 0.2011811656980879, + "dark_distillation": 0.32484443195992047, + "supervised_contrastive": 4.150537680917316, + "total": 1.329396380365841, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5474857849051398, + "weighted_f1": 0.6280694008611376, + "macro_precision": 0.5573604778697515, + "macro_recall": 0.5854527482523125, + "per_class_f1": { + "FAVOR": 0.432258064516129, + "AGAINST": 0.47555555555555556, + "NONE": 0.7346437346437347 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 54.74857849051398, + "weighted_f1_pct": 62.80694008611376, + "macro_precision_pct": 55.73604778697515, + "macro_recall_pct": 58.54527482523125, + "per_class_f1_pct": { + "FAVOR": 43.225806451612904, + "AGAINST": 47.55555555555556, + "NONE": 73.46437346437347 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6945812807881774, + "macro_f1": 0.6842589420389547, + "weighted_f1": 0.7027586455701511, + "macro_precision": 0.6896052313567691, + "macro_recall": 0.708929814487071, + "per_class_f1": { + "FAVOR": 0.6242424242424243, + "AGAINST": 0.6701846965699209, + "NONE": 0.7583497053045186 + }, + "accuracy_pct": 69.45812807881774, + "macro_f1_pct": 68.42589420389547, + "weighted_f1_pct": 70.27586455701511, + "macro_precision_pct": 68.96052313567691, + "macro_recall_pct": 70.8929814487071, + "per_class_f1_pct": { + "FAVOR": 62.42424242424243, + "AGAINST": 67.01846965699208, + "NONE": 75.83497053045186 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.612859097127223, + "macro_f1": 0.577992794799215, + "weighted_f1": 0.6244734584828522, + "macro_precision": 0.5889736836181085, + "macro_recall": 0.6069673129013604, + "per_class_f1": { + "FAVOR": 0.45188284518828453, + "AGAINST": 0.5961199294532628, + "NONE": 0.6859756097560976 + }, + "accuracy_pct": 61.2859097127223, + "macro_f1_pct": 57.799279479921495, + "weighted_f1_pct": 62.44734584828522, + "macro_precision_pct": 58.897368361810855, + "macro_recall_pct": 60.69673129013604, + "per_class_f1_pct": { + "FAVOR": 45.18828451882845, + "AGAINST": 59.61199294532628, + "NONE": 68.59756097560977 + } + } + ], + "accuracy": 0.6318758815232722, + "macro_f1": 0.6038807881862908, + "weighted_f1": 0.6469945502476803, + "macro_precision": 0.6110844398194826, + "macro_recall": 0.6361044872987649, + "per_class_f1": { + "FAVOR": 0.5096700796359499, + "AGAINST": 0.5773638968481375, + "NONE": 0.7246083880747852 + }, + "accuracy_pct": 63.18758815232722, + "macro_f1_pct": 60.388078818629076, + "weighted_f1_pct": 64.69945502476803, + "macro_precision_pct": 61.108443981948255, + "macro_recall_pct": 63.61044872987649, + "per_class_f1_pct": { + "FAVOR": 50.96700796359499, + "AGAINST": 57.73638968481375, + "NONE": 72.46083880747852 + } + }, + { + "after_task": 4, + "best_epoch": 8, + "best_dev_macro_f1": 0.6954337426167982, + "losses": { + "current_ce": 0.5132155825694402, + "replay_ce": 0.1793043177640229, + "dark_distillation": 0.3004567368772992, + "supervised_contrastive": 4.162839742292437, + "total": 1.2590322508100877, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6111817026683609, + "macro_f1": 0.5419910063445633, + "weighted_f1": 0.6377653542479916, + "macro_precision": 0.5439685972624501, + "macro_recall": 0.5691176470588234, + "per_class_f1": { + "FAVOR": 0.41194029850746267, + "AGAINST": 0.4489795918367347, + "NONE": 0.7650531286894924 + }, + "accuracy_pct": 61.11817026683609, + "macro_f1_pct": 54.19910063445633, + "weighted_f1_pct": 63.776535424799164, + "macro_precision_pct": 54.396859726245005, + "macro_recall_pct": 56.91176470588234, + "per_class_f1_pct": { + "FAVOR": 41.19402985074627, + "AGAINST": 44.89795918367347, + "NONE": 76.50531286894923 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.6883719911514116, + "weighted_f1": 0.714211914909319, + "macro_precision": 0.6915130023640662, + "macro_recall": 0.7048238894534057, + "per_class_f1": { + "FAVOR": 0.5964912280701754, + "AGAINST": 0.6783625730994152, + "NONE": 0.7902621722846442 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 68.83719911514116, + "weighted_f1_pct": 71.4211914909319, + "macro_precision_pct": 69.15130023640663, + "macro_recall_pct": 70.48238894534057, + "per_class_f1_pct": { + "FAVOR": 59.64912280701754, + "AGAINST": 67.83625730994152, + "NONE": 79.02621722846442 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5772913816689467, + "macro_f1": 0.5429839180967506, + "weighted_f1": 0.5996295679343902, + "macro_precision": 0.5576086445853888, + "macro_recall": 0.581597838670152, + "per_class_f1": { + "FAVOR": 0.39867109634551495, + "AGAINST": 0.5483234714003945, + "NONE": 0.6819571865443425 + }, + "accuracy_pct": 57.729138166894664, + "macro_f1_pct": 54.29839180967506, + "weighted_f1_pct": 59.96295679343901, + "macro_precision_pct": 55.76086445853888, + "macro_recall_pct": 58.1597838670152, + "per_class_f1_pct": { + "FAVOR": 39.8671096345515, + "AGAINST": 54.83234714003945, + "NONE": 68.19571865443424 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6502242152466368, + "macro_f1": 0.6005380584327953, + "weighted_f1": 0.6663660625426026, + "macro_precision": 0.601445382957988, + "macro_recall": 0.6201333150034015, + "per_class_f1": { + "FAVOR": 0.5473684210526316, + "AGAINST": 0.46853146853146854, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 65.02242152466367, + "macro_f1_pct": 60.05380584327953, + "weighted_f1_pct": 66.63660625426026, + "macro_precision_pct": 60.144538295798796, + "macro_recall_pct": 62.01333150034015, + "per_class_f1_pct": { + "FAVOR": 54.736842105263165, + "AGAINST": 46.85314685314685, + "NONE": 78.57142857142857 + } + } + ], + "accuracy": 0.6319742489270386, + "macro_f1": 0.5951508951232691, + "weighted_f1": 0.6501870936089699, + "macro_precision": 0.5979503615650604, + "macro_recall": 0.6214874381433705, + "per_class_f1": { + "FAVOR": 0.49337260677466865, + "AGAINST": 0.5370006548788474, + "NONE": 0.7550794237162911 + }, + "accuracy_pct": 63.19742489270386, + "macro_f1_pct": 59.51508951232691, + "weighted_f1_pct": 65.01870936089699, + "macro_precision_pct": 59.79503615650604, + "macro_recall_pct": 62.148743814337045, + "per_class_f1_pct": { + "FAVOR": 49.337260677466865, + "AGAINST": 53.70006548788474, + "NONE": 75.50794237162911 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.6846308749293825, + "losses": { + "current_ce": 0.4963967498924051, + "replay_ce": 0.19124230263488634, + "dark_distillation": 0.30105886097465245, + "supervised_contrastive": 4.159904560020991, + "total": 1.2541589464460101, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6099110546378653, + "macro_f1": 0.5538033990810041, + "weighted_f1": 0.6354003938481126, + "macro_precision": 0.562611434277827, + "macro_recall": 0.5929645034220198, + "per_class_f1": { + "FAVOR": 0.419672131147541, + "AGAINST": 0.5, + "NONE": 0.7417380660954712 + }, + "accuracy_pct": 60.99110546378653, + "macro_f1_pct": 55.38033990810041, + "weighted_f1_pct": 63.540039384811266, + "macro_precision_pct": 56.2611434277827, + "macro_recall_pct": 59.29645034220198, + "per_class_f1_pct": { + "FAVOR": 41.967213114754095, + "AGAINST": 50.0, + "NONE": 74.17380660954713 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6969398109319832, + "weighted_f1": 0.7146845301916495, + "macro_precision": 0.7017471851320963, + "macro_recall": 0.7211050584443113, + "per_class_f1": { + "FAVOR": 0.6349206349206349, + "AGAINST": 0.6887755102040817, + "NONE": 0.7671232876712328 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.69398109319832, + "weighted_f1_pct": 71.46845301916494, + "macro_precision_pct": 70.17471851320963, + "macro_recall_pct": 72.11050584443113, + "per_class_f1_pct": { + "FAVOR": 63.49206349206349, + "AGAINST": 68.87755102040816, + "NONE": 76.71232876712328 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5841313269493844, + "macro_f1": 0.5496869997965116, + "weighted_f1": 0.6009682579673599, + "macro_precision": 0.5653138956224256, + "macro_recall": 0.5845446711281997, + "per_class_f1": { + "FAVOR": 0.41044776119402987, + "AGAINST": 0.5698729582577132, + "NONE": 0.6687402799377916 + }, + "accuracy_pct": 58.41313269493844, + "macro_f1_pct": 54.968699979651156, + "weighted_f1_pct": 60.096825796735985, + "macro_precision_pct": 56.53138956224256, + "macro_recall_pct": 58.45446711281998, + "per_class_f1_pct": { + "FAVOR": 41.04477611940299, + "AGAINST": 56.98729582577132, + "NONE": 66.87402799377917 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6233183856502242, + "macro_f1": 0.5718693009118541, + "weighted_f1": 0.6430490093184492, + "macro_precision": 0.5803086523865745, + "macro_recall": 0.5916614085612532, + "per_class_f1": { + "FAVOR": 0.515, + "AGAINST": 0.42857142857142855, + "NONE": 0.7720364741641338 + }, + "accuracy_pct": 62.33183856502242, + "macro_f1_pct": 57.1869300911854, + "weighted_f1_pct": 64.30490093184493, + "macro_precision_pct": 58.03086523865745, + "macro_recall_pct": 59.16614085612532, + "per_class_f1_pct": { + "FAVOR": 51.5, + "AGAINST": 42.857142857142854, + "NONE": 77.20364741641338 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7203219315895373, + "macro_f1": 0.6821327640950252, + "weighted_f1": 0.7194463881381205, + "macro_precision": 0.6838318741893644, + "macro_recall": 0.6807691612407898, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6141732283464567, + "NONE": 0.8235294117647058 + }, + "accuracy_pct": 72.03219315895373, + "macro_f1_pct": 68.21327640950253, + "weighted_f1_pct": 71.94463881381205, + "macro_precision_pct": 68.38318741893644, + "macro_recall_pct": 68.07691612407898, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 61.417322834645674, + "NONE": 82.35294117647058 + } + } + ], + "accuracy": 0.6416641360461586, + "macro_f1": 0.6100041812359615, + "weighted_f1": 0.6569179529289879, + "macro_precision": 0.6114679739615544, + "macro_recall": 0.6352474679222415, + "per_class_f1": { + "FAVOR": 0.5164690382081687, + "AGAINST": 0.5629860031104199, + "NONE": 0.750557502389296 + }, + "accuracy_pct": 64.16641360461585, + "macro_f1_pct": 61.00041812359615, + "weighted_f1_pct": 65.69179529289879, + "macro_precision_pct": 61.14679739615544, + "macro_recall_pct": 63.52474679222415, + "per_class_f1_pct": { + "FAVOR": 51.64690382081687, + "AGAINST": 56.298600311041994, + "NONE": 75.05575023892959 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.7237239504681366, + "losses": { + "current_ce": 0.40225261479970953, + "replay_ce": 0.15782791473681018, + "dark_distillation": 0.29148755002589455, + "supervised_contrastive": 4.079310572715032, + "total": 1.1137553697540648, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6086404066073697, + "macro_f1": 0.5555344494972978, + "weighted_f1": 0.6355547136828485, + "macro_precision": 0.5617699054579425, + "macro_recall": 0.5934869757854507, + "per_class_f1": { + "FAVOR": 0.4520123839009288, + "AGAINST": 0.4722222222222222, + "NONE": 0.7423687423687424 + }, + "accuracy_pct": 60.864040660736975, + "macro_f1_pct": 55.55344494972978, + "weighted_f1_pct": 63.555471368284856, + "macro_precision_pct": 56.17699054579425, + "macro_recall_pct": 59.34869757854507, + "per_class_f1_pct": { + "FAVOR": 45.20123839009288, + "AGAINST": 47.22222222222222, + "NONE": 74.23687423687424 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6830870279146142, + "macro_f1": 0.6745102861806241, + "weighted_f1": 0.6931303545509839, + "macro_precision": 0.685955642196344, + "macro_recall": 0.7017098877172362, + "per_class_f1": { + "FAVOR": 0.601123595505618, + "AGAINST": 0.6759002770083102, + "NONE": 0.7465069860279441 + }, + "accuracy_pct": 68.30870279146141, + "macro_f1_pct": 67.45102861806241, + "weighted_f1_pct": 69.31303545509839, + "macro_precision_pct": 68.5955642196344, + "macro_recall_pct": 70.17098877172361, + "per_class_f1_pct": { + "FAVOR": 60.1123595505618, + "AGAINST": 67.59002770083103, + "NONE": 74.6506986027944 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5663474692202463, + "macro_f1": 0.5389436103907475, + "weighted_f1": 0.5877168891759554, + "macro_precision": 0.5594265890584552, + "macro_recall": 0.5862269046902459, + "per_class_f1": { + "FAVOR": 0.41025641025641024, + "AGAINST": 0.5514563106796116, + "NONE": 0.6551181102362205 + }, + "accuracy_pct": 56.634746922024625, + "macro_f1_pct": 53.894361039074745, + "weighted_f1_pct": 58.77168891759554, + "macro_precision_pct": 55.94265890584552, + "macro_recall_pct": 58.62269046902458, + "per_class_f1_pct": { + "FAVOR": 41.02564102564102, + "AGAINST": 55.14563106796116, + "NONE": 65.51181102362204 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5964125560538116, + "macro_f1": 0.5480163594403195, + "weighted_f1": 0.6208773849476538, + "macro_precision": 0.5589503462410962, + "macro_recall": 0.5671551123472064, + "per_class_f1": { + "FAVOR": 0.47668393782383417, + "AGAINST": 0.4131147540983607, + "NONE": 0.7542503863987635 + }, + "accuracy_pct": 59.64125560538116, + "macro_f1_pct": 54.80163594403194, + "weighted_f1_pct": 62.087738494765375, + "macro_precision_pct": 55.89503462410962, + "macro_recall_pct": 56.71551123472064, + "per_class_f1_pct": { + "FAVOR": 47.66839378238342, + "AGAINST": 41.31147540983607, + "NONE": 75.42503863987635 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7203219315895373, + "macro_f1": 0.6907029420611966, + "weighted_f1": 0.7242014053190414, + "macro_precision": 0.686484966378293, + "macro_recall": 0.6974223407532176, + "per_class_f1": { + "FAVOR": 0.6065573770491803, + "AGAINST": 0.6490566037735849, + "NONE": 0.8164948453608247 + }, + "accuracy_pct": 72.03219315895373, + "macro_f1_pct": 69.07029420611967, + "weighted_f1_pct": 72.42014053190414, + "macro_precision_pct": 68.64849663782931, + "macro_recall_pct": 69.74223407532176, + "per_class_f1_pct": { + "FAVOR": 60.65573770491803, + "AGAINST": 64.90566037735849, + "NONE": 81.64948453608247 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7303822937625755, + "macro_f1": 0.701716540815324, + "weighted_f1": 0.7422749490709618, + "macro_precision": 0.7044804772874947, + "macro_recall": 0.7214754984869928, + "per_class_f1": { + "FAVOR": 0.6132404181184669, + "AGAINST": 0.6606334841628959, + "NONE": 0.831275720164609 + }, + "accuracy_pct": 73.03822937625755, + "macro_f1_pct": 70.1716540815324, + "weighted_f1_pct": 74.22749490709619, + "macro_precision_pct": 70.44804772874947, + "macro_recall_pct": 72.14754984869927, + "per_class_f1_pct": { + "FAVOR": 61.32404181184668, + "AGAINST": 66.06334841628959, + "NONE": 83.1275720164609 + } + } + ], + "accuracy": 0.6408970976253299, + "macro_f1": 0.6113428675554314, + "weighted_f1": 0.6576535812547676, + "macro_precision": 0.6148863062972453, + "macro_recall": 0.6393910158135551, + "per_class_f1": { + "FAVOR": 0.5220125786163522, + "AGAINST": 0.5602667937112911, + "NONE": 0.751749230338651 + }, + "accuracy_pct": 64.08970976253299, + "macro_f1_pct": 61.134286755543144, + "weighted_f1_pct": 65.76535812547675, + "macro_precision_pct": 61.488630629724526, + "macro_recall_pct": 63.93910158135551, + "per_class_f1_pct": { + "FAVOR": 52.20125786163522, + "AGAINST": 56.026679371129106, + "NONE": 75.1749230338651 + } + }, + { + "after_task": 7, + "best_epoch": 10, + "best_dev_macro_f1": 0.7522783521410902, + "losses": { + "current_ce": 0.34459639222520155, + "replay_ce": 0.1451279732027674, + "dark_distillation": 0.27080509992149787, + "supervised_contrastive": 4.003387960961194, + "total": 1.0254657183236222, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5806861499364676, + "macro_f1": 0.5258408944107079, + "weighted_f1": 0.6083763776384752, + "macro_precision": 0.5604840914146968, + "macro_recall": 0.5662520264372116, + "per_class_f1": { + "FAVOR": 0.3938223938223938, + "AGAINST": 0.4674329501915709, + "NONE": 0.7162673392181589 + }, + "accuracy_pct": 58.06861499364676, + "macro_f1_pct": 52.584089441070795, + "weighted_f1_pct": 60.83763776384752, + "macro_precision_pct": 56.048409141469676, + "macro_recall_pct": 56.62520264372116, + "per_class_f1_pct": { + "FAVOR": 39.38223938223938, + "AGAINST": 46.74329501915709, + "NONE": 71.62673392181588 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7093596059113301, + "macro_f1": 0.6999404499661249, + "weighted_f1": 0.7155048541649449, + "macro_precision": 0.7025989844314294, + "macro_recall": 0.7229464673585922, + "per_class_f1": { + "FAVOR": 0.6559485530546624, + "AGAINST": 0.680306905370844, + "NONE": 0.7635658914728682 + }, + "accuracy_pct": 70.93596059113301, + "macro_f1_pct": 69.99404499661249, + "weighted_f1_pct": 71.5504854164945, + "macro_precision_pct": 70.25989844314294, + "macro_recall_pct": 72.29464673585923, + "per_class_f1_pct": { + "FAVOR": 65.59485530546624, + "AGAINST": 68.03069053708441, + "NONE": 76.35658914728683 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5386593231724417, + "weighted_f1": 0.5865355989636675, + "macro_precision": 0.5607224877331329, + "macro_recall": 0.5735003195666069, + "per_class_f1": { + "FAVOR": 0.4166666666666667, + "AGAINST": 0.5432098765432098, + "NONE": 0.6561014263074485 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 53.86593231724417, + "weighted_f1_pct": 58.653559896366744, + "macro_precision_pct": 56.07224877331329, + "macro_recall_pct": 57.350031956660686, + "per_class_f1_pct": { + "FAVOR": 41.66666666666667, + "AGAINST": 54.32098765432099, + "NONE": 65.61014263074485 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6113602391629297, + "macro_f1": 0.5647261126182014, + "weighted_f1": 0.6336489083677446, + "macro_precision": 0.5762941388867896, + "macro_recall": 0.5896092787054369, + "per_class_f1": { + "FAVOR": 0.4625, + "AGAINST": 0.46831955922865015, + "NONE": 0.7633587786259542 + }, + "accuracy_pct": 61.13602391629297, + "macro_f1_pct": 56.47261126182014, + "weighted_f1_pct": 63.36489083677446, + "macro_precision_pct": 57.62941388867896, + "macro_recall_pct": 58.96092787054369, + "per_class_f1_pct": { + "FAVOR": 46.25, + "AGAINST": 46.83195592286501, + "NONE": 76.33587786259542 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6798903534079178, + "weighted_f1": 0.7145490817721899, + "macro_precision": 0.6807565738572475, + "macro_recall": 0.685405481278731, + "per_class_f1": { + "FAVOR": 0.5892857142857143, + "AGAINST": 0.6405693950177936, + "NONE": 0.8098159509202454 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.98903534079179, + "weighted_f1_pct": 71.454908177219, + "macro_precision_pct": 68.07565738572475, + "macro_recall_pct": 68.54054812787311, + "per_class_f1_pct": { + "FAVOR": 58.92857142857143, + "AGAINST": 64.05693950177937, + "NONE": 80.98159509202453 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.6981891348088531, + "macro_f1": 0.6696316482382692, + "weighted_f1": 0.7125030897142438, + "macro_precision": 0.6690685265913713, + "macro_recall": 0.6962019490755122, + "per_class_f1": { + "FAVOR": 0.5692883895131086, + "AGAINST": 0.6328125, + "NONE": 0.8067940552016986 + }, + "accuracy_pct": 69.81891348088531, + "macro_f1_pct": 66.96316482382692, + "weighted_f1_pct": 71.25030897142437, + "macro_precision_pct": 66.90685265913713, + "macro_recall_pct": 69.62019490755122, + "per_class_f1_pct": { + "FAVOR": 56.92883895131086, + "AGAINST": 63.28125, + "NONE": 80.67940552016985 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7154811715481172, + "macro_f1": 0.675092389493213, + "weighted_f1": 0.7257851205305241, + "macro_precision": 0.665665572407237, + "macro_recall": 0.7097729373438751, + "per_class_f1": { + "FAVOR": 0.5617977528089888, + "AGAINST": 0.6666666666666666, + "NONE": 0.796812749003984 + }, + "accuracy_pct": 71.54811715481172, + "macro_f1_pct": 67.5092389493213, + "weighted_f1_pct": 72.5785120530524, + "macro_precision_pct": 66.5665572407237, + "macro_recall_pct": 70.97729373438752, + "per_class_f1_pct": { + "FAVOR": 56.17977528089888, + "AGAINST": 66.66666666666666, + "NONE": 79.6812749003984 + } + } + ], + "accuracy": 0.6496560905258487, + "macro_f1": 0.6172147762045124, + "weighted_f1": 0.6644349240609934, + "macro_precision": 0.6229001481618165, + "macro_recall": 0.6441414865925155, + "per_class_f1": { + "FAVOR": 0.5219665271966527, + "AGAINST": 0.5748031496062992, + "NONE": 0.754874651810585 + }, + "accuracy_pct": 64.96560905258487, + "macro_f1_pct": 61.72147762045124, + "weighted_f1_pct": 66.44349240609934, + "macro_precision_pct": 62.290014816181646, + "macro_recall_pct": 64.41414865925155, + "per_class_f1_pct": { + "FAVOR": 52.19665271966527, + "AGAINST": 57.48031496062992, + "NONE": 75.4874651810585 + } + }, + { + "after_task": 8, + "best_epoch": 11, + "best_dev_macro_f1": 0.7794668489405332, + "losses": { + "current_ce": 0.3272139158172935, + "replay_ce": 0.13582197868940876, + "dark_distillation": 0.2597892032242289, + "supervised_contrastive": 4.019440527990753, + "total": 0.9948745572099499, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.5233170328550799, + "weighted_f1": 0.6103155283478797, + "macro_precision": 0.5414929153728992, + "macro_recall": 0.5604122195080801, + "per_class_f1": { + "FAVOR": 0.4072948328267477, + "AGAINST": 0.4365256124721604, + "NONE": 0.7261306532663316 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 52.33170328550799, + "weighted_f1_pct": 61.03155283478797, + "macro_precision_pct": 54.149291537289926, + "macro_recall_pct": 56.04122195080801, + "per_class_f1_pct": { + "FAVOR": 40.72948328267477, + "AGAINST": 43.652561247216035, + "NONE": 72.61306532663316 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6962233169129721, + "macro_f1": 0.684901107041461, + "weighted_f1": 0.7047170003640627, + "macro_precision": 0.6883483856443037, + "macro_recall": 0.7076573151282276, + "per_class_f1": { + "FAVOR": 0.6217008797653959, + "AGAINST": 0.6685236768802229, + "NONE": 0.7644787644787645 + }, + "accuracy_pct": 69.6223316912972, + "macro_f1_pct": 68.4901107041461, + "weighted_f1_pct": 70.47170003640628, + "macro_precision_pct": 68.83483856443037, + "macro_recall_pct": 70.76573151282275, + "per_class_f1_pct": { + "FAVOR": 62.17008797653959, + "AGAINST": 66.85236768802228, + "NONE": 76.44787644787645 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5403556771545828, + "macro_f1": 0.5226071184525704, + "weighted_f1": 0.5639715359708796, + "macro_precision": 0.5573327829038535, + "macro_recall": 0.5918559549292727, + "per_class_f1": { + "FAVOR": 0.4273972602739726, + "AGAINST": 0.5083333333333333, + "NONE": 0.6320907617504052 + }, + "accuracy_pct": 54.03556771545828, + "macro_f1_pct": 52.260711845257035, + "weighted_f1_pct": 56.39715359708796, + "macro_precision_pct": 55.73327829038534, + "macro_recall_pct": 59.18559549292727, + "per_class_f1_pct": { + "FAVOR": 42.73972602739726, + "AGAINST": 50.83333333333333, + "NONE": 63.20907617504052 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6188340807174888, + "macro_f1": 0.5718081996599217, + "weighted_f1": 0.6376903544070702, + "macro_precision": 0.58041276309569, + "macro_recall": 0.5943239486432913, + "per_class_f1": { + "FAVOR": 0.5247524752475248, + "AGAINST": 0.43416370106761565, + "NONE": 0.7565084226646248 + }, + "accuracy_pct": 61.88340807174888, + "macro_f1_pct": 57.18081996599217, + "weighted_f1_pct": 63.76903544070702, + "macro_precision_pct": 58.041276309569, + "macro_recall_pct": 59.43239486432913, + "per_class_f1_pct": { + "FAVOR": 52.475247524752476, + "AGAINST": 43.41637010676156, + "NONE": 75.65084226646248 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.671946215673841, + "weighted_f1": 0.7138470353695248, + "macro_precision": 0.6727107103003803, + "macro_recall": 0.6716840298269923, + "per_class_f1": { + "FAVOR": 0.5892116182572614, + "AGAINST": 0.5959183673469388, + "NONE": 0.8307086614173228 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 67.1946215673841, + "weighted_f1_pct": 71.38470353695247, + "macro_precision_pct": 67.27107103003803, + "macro_recall_pct": 67.16840298269922, + "per_class_f1_pct": { + "FAVOR": 58.921161825726145, + "AGAINST": 59.591836734693885, + "NONE": 83.07086614173228 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.670020120724346, + "macro_f1": 0.6378313442829572, + "weighted_f1": 0.6888415450757244, + "macro_precision": 0.6560311207211477, + "macro_recall": 0.659939249594422, + "per_class_f1": { + "FAVOR": 0.5512820512820513, + "AGAINST": 0.5622119815668203, + "NONE": 0.8 + }, + "accuracy_pct": 67.0020120724346, + "macro_f1_pct": 63.78313442829572, + "weighted_f1_pct": 68.88415450757243, + "macro_precision_pct": 65.60311207211477, + "macro_recall_pct": 65.9939249594422, + "per_class_f1_pct": { + "FAVOR": 55.12820512820513, + "AGAINST": 56.22119815668203, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6386522735421335, + "weighted_f1": 0.6821897498335434, + "macro_precision": 0.6369362611383887, + "macro_recall": 0.6935759037074419, + "per_class_f1": { + "FAVOR": 0.5362776025236593, + "AGAINST": 0.6374695863746959, + "NONE": 0.7422096317280453 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 63.86522735421335, + "weighted_f1_pct": 68.21897498335434, + "macro_precision_pct": 63.693626113838874, + "macro_recall_pct": 69.35759037074419, + "per_class_f1_pct": { + "FAVOR": 53.62776025236593, + "AGAINST": 63.74695863746959, + "NONE": 74.22096317280453 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.717607973421927, + "macro_f1": 0.6784750789347859, + "weighted_f1": 0.7252138938001059, + "macro_precision": 0.668728505462934, + "macro_recall": 0.7091038201611278, + "per_class_f1": { + "FAVOR": 0.6323529411764706, + "AGAINST": 0.593103448275862, + "NONE": 0.8099688473520249 + }, + "accuracy_pct": 71.76079734219269, + "macro_f1_pct": 67.84750789347859, + "weighted_f1_pct": 72.52138938001059, + "macro_precision_pct": 66.8728505462934, + "macro_recall_pct": 70.91038201611278, + "per_class_f1_pct": { + "FAVOR": 63.23529411764706, + "AGAINST": 59.310344827586206, + "NONE": 80.99688473520249 + } + } + ], + "accuracy": 0.642982971227246, + "macro_f1": 0.6116104264633364, + "weighted_f1": 0.6600020141474947, + "macro_precision": 0.6143539861992621, + "macro_recall": 0.6435497052807193, + "per_class_f1": { + "FAVOR": 0.5308020147229756, + "AGAINST": 0.5505124450951684, + "NONE": 0.7535168195718654 + }, + "accuracy_pct": 64.2982971227246, + "macro_f1_pct": 61.16104264633364, + "weighted_f1_pct": 66.00020141474947, + "macro_precision_pct": 61.435398619926204, + "macro_recall_pct": 64.35497052807193, + "per_class_f1_pct": { + "FAVOR": 53.08020147229756, + "AGAINST": 55.05124450951684, + "NONE": 75.35168195718654 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s1__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e973337d226e839e33d590286314db7ab99496a9 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__doc__s1__seed13/run.log b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..015923bca96d34b4a5d2fe13151516d34935f1b8 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 06:29:56,800 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__doc__s1__seed13 device=cuda +2026-09-02 07:00:02,969 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6161922963053439, 'average_forgetting': 0.0423933477119672, 'backward_transfer': -0.04116903657394271, 'final_mean_target_macro_f1_pct': 61.61922963053439, 'average_forgetting_pct': 4.23933477119672, 'backward_transfer_pct': -4.116903657394271, 'final_pooled_accuracy': 0.642982971227246, 'final_pooled_accuracy_pct': 64.2982971227246, 'final_pooled_macro_f1': 0.6116104264633364, 'final_pooled_macro_f1_pct': 61.16104264633364, 'final_pooled_weighted_f1': 0.6600020141474947, 'final_pooled_weighted_f1_pct': 66.00020141474947, 'final_pooled_macro_precision': 0.6143539861992621, 'final_pooled_macro_precision_pct': 61.435398619926204, 'final_pooled_macro_recall': 0.6435497052807193, 'final_pooled_macro_recall_pct': 64.35497052807193, 'final_pooled_per_class_f1': {'FAVOR': 0.5308020147229756, 'AGAINST': 0.5505124450951684, 'NONE': 0.7535168195718654}, 'final_pooled_per_class_f1_pct': {'FAVOR': 53.08020147229756, 'AGAINST': 55.05124450951684, 'NONE': 75.35168195718654}} +2026-09-02 07:00:02,970 [INFO] __main__: efficiency={'training_wall_time_seconds': 1804.7932216611225, 'training_wall_time_minutes': 30.079887027685377, 'peak_gpu_memory_allocated_gb': 10.40872573852539, 'peak_gpu_memory_reserved_gb': 10.599609375} diff --git a/outputs/trace_s7_contrastive_replay__doc__s2__seed13/environment.json b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..fd3c9e074f18029ddb24e6548c04849d149d2fed --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T07:00:10.932644+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..64a279e5b4db8ac1ccf99fe5b97b8d7564be7409 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.json @@ -0,0 +1,1601 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.6513687947181975, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6657005224117883, + 0.5816105653599061, + null, + null, + null, + null, + null, + null + ], + [ + 0.6276019493374833, + 0.5582475091703928, + 0.7122876181351957, + null, + null, + null, + null, + null + ], + [ + 0.621907743858059, + 0.5244320352207028, + 0.6554243264523638, + 0.70353502494036, + null, + null, + null, + null + ], + [ + 0.641281595319099, + 0.5279806862190077, + 0.6741750880316179, + 0.6873411293367263, + 0.6367242196064132, + null, + null, + null + ], + [ + 0.6443936204024284, + 0.5478229375793808, + 0.6702664665818635, + 0.677844790779011, + 0.6085078385250307, + 0.6112572119378699, + null, + null + ], + [ + 0.6358106366195218, + 0.5106926477593023, + 0.6645717883501646, + 0.675152914939907, + 0.6013594109459625, + 0.5654736920713259, + 0.6883212193874031, + null + ], + [ + 0.6460721300459262, + 0.5244981177407264, + 0.6444272626707189, + 0.6743276278959311, + 0.5740215240050243, + 0.5379263612821078, + 0.6902964959568733, + 0.6652238411372376 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6952735888792025, + "losses": { + "current_ce": 0.5369640307341303, + "total": 0.5369640307341303, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6600985221674877, + "macro_f1": 0.6513687947181975, + "weighted_f1": 0.6700885375939004, + "macro_precision": 0.6615219122816131, + "macro_recall": 0.6759561456438371, + "per_class_f1": { + "FAVOR": 0.5754189944134078, + "AGAINST": 0.655367231638418, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.13687947181975, + "weighted_f1_pct": 67.00885375939004, + "macro_precision_pct": 66.1521912281613, + "macro_recall_pct": 67.59561456438371, + "per_class_f1_pct": { + "FAVOR": 57.54189944134078, + "AGAINST": 65.5367231638418, + "NONE": 72.33201581027669 + } + } + ], + "accuracy": 0.6600985221674877, + "macro_f1": 0.6513687947181975, + "weighted_f1": 0.6700885375939004, + "macro_precision": 0.6615219122816131, + "macro_recall": 0.6759561456438371, + "per_class_f1": { + "FAVOR": 0.5754189944134078, + "AGAINST": 0.655367231638418, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.13687947181975, + "weighted_f1_pct": 67.00885375939004, + "macro_precision_pct": 66.1521912281613, + "macro_recall_pct": 67.59561456438371, + "per_class_f1_pct": { + "FAVOR": 57.54189944134078, + "AGAINST": 65.5367231638418, + "NONE": 72.33201581027669 + } + }, + { + "after_task": 2, + "best_epoch": 14, + "best_dev_macro_f1": 0.6851579400510347, + "losses": { + "current_ce": 0.5884127289056778, + "replay_ce": 0.10817521661520005, + "dark_distillation": 0.2354041861076104, + "supervised_contrastive": 4.256149932376125, + "total": 1.2399050376917187, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6699507389162561, + "macro_f1": 0.6657005224117883, + "weighted_f1": 0.6777906120558884, + "macro_precision": 0.680324074074074, + "macro_recall": 0.6982276415834284, + "per_class_f1": { + "FAVOR": 0.5994550408719346, + "AGAINST": 0.6888888888888889, + "NONE": 0.7087576374745418 + }, + "accuracy_pct": 66.99507389162561, + "macro_f1_pct": 66.57005224117883, + "weighted_f1_pct": 67.77906120558885, + "macro_precision_pct": 68.0324074074074, + "macro_recall_pct": 69.82276415834285, + "per_class_f1_pct": { + "FAVOR": 59.945504087193456, + "AGAINST": 68.88888888888889, + "NONE": 70.87576374745417 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6203288490284006, + "macro_f1": 0.5816105653599061, + "weighted_f1": 0.6384424317763885, + "macro_precision": 0.5922763510998804, + "macro_recall": 0.6115264720638915, + "per_class_f1": { + "FAVOR": 0.5520581113801453, + "AGAINST": 0.4529616724738676, + "NONE": 0.7398119122257053 + }, + "accuracy_pct": 62.03288490284006, + "macro_f1_pct": 58.16105653599061, + "weighted_f1_pct": 63.84424317763885, + "macro_precision_pct": 59.22763510998804, + "macro_recall_pct": 61.152647206389155, + "per_class_f1_pct": { + "FAVOR": 55.205811138014525, + "AGAINST": 45.29616724738676, + "NONE": 73.98119122257053 + } + } + ], + "accuracy": 0.6439749608763693, + "macro_f1": 0.6283001235687409, + "weighted_f1": 0.6571767478765632, + "macro_precision": 0.6399963624169466, + "macro_recall": 0.660062213977916, + "per_class_f1": { + "FAVOR": 0.5743589743589743, + "AGAINST": 0.5842349304482226, + "NONE": 0.7263064658990257 + }, + "accuracy_pct": 64.39749608763692, + "macro_f1_pct": 62.830012356874086, + "weighted_f1_pct": 65.71767478765632, + "macro_precision_pct": 63.99963624169466, + "macro_recall_pct": 66.00622139779159, + "per_class_f1_pct": { + "FAVOR": 57.43589743589743, + "AGAINST": 58.42349304482226, + "NONE": 72.63064658990257 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7539987422241418, + "losses": { + "current_ce": 0.4725877070355983, + "replay_ce": 0.13467356124449345, + "dark_distillation": 0.2609920618079957, + "supervised_contrastive": 4.164763221854256, + "total": 1.1542336293629238, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6272577996715928, + "macro_f1": 0.6276019493374833, + "weighted_f1": 0.640233686096964, + "macro_precision": 0.6711881442884482, + "macro_recall": 0.6608475640165782, + "per_class_f1": { + "FAVOR": 0.5566037735849056, + "AGAINST": 0.653968253968254, + "NONE": 0.6722338204592901 + }, + "accuracy_pct": 62.725779967159276, + "macro_f1_pct": 62.76019493374832, + "weighted_f1_pct": 64.02336860969639, + "macro_precision_pct": 67.11881442884481, + "macro_recall_pct": 66.08475640165781, + "per_class_f1_pct": { + "FAVOR": 55.660377358490564, + "AGAINST": 65.39682539682539, + "NONE": 67.22338204592901 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.593423019431988, + "macro_f1": 0.5582475091703928, + "weighted_f1": 0.6117358270549611, + "macro_precision": 0.5952476984866806, + "macro_recall": 0.5994516295426799, + "per_class_f1": { + "FAVOR": 0.5466101694915254, + "AGAINST": 0.4226415094339623, + "NONE": 0.7054908485856906 + }, + "accuracy_pct": 59.342301943198805, + "macro_f1_pct": 55.82475091703928, + "weighted_f1_pct": 61.17358270549611, + "macro_precision_pct": 59.52476984866806, + "macro_recall_pct": 59.94516295426799, + "per_class_f1_pct": { + "FAVOR": 54.66101694915254, + "AGAINST": 42.26415094339623, + "NONE": 70.54908485856906 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7424547283702213, + "macro_f1": 0.7122876181351957, + "weighted_f1": 0.7511279162497875, + "macro_precision": 0.7126256054362755, + "macro_recall": 0.7257343550446999, + "per_class_f1": { + "FAVOR": 0.6181818181818182, + "AGAINST": 0.6820276497695853, + "NONE": 0.8366533864541833 + }, + "accuracy_pct": 74.24547283702213, + "macro_f1_pct": 71.22876181351957, + "weighted_f1_pct": 75.11279162497874, + "macro_precision_pct": 71.26256054362756, + "macro_recall_pct": 72.57343550446998, + "per_class_f1_pct": { + "FAVOR": 61.81818181818181, + "AGAINST": 68.20276497695853, + "NONE": 83.66533864541833 + } + } + ], + "accuracy": 0.6467605633802817, + "macro_f1": 0.6295903457779158, + "weighted_f1": 0.661836131894537, + "macro_precision": 0.654346713584342, + "macro_recall": 0.6597464011846169, + "per_class_f1": { + "FAVOR": 0.5670367207514945, + "AGAINST": 0.5846925972396487, + "NONE": 0.7370417193426043 + }, + "accuracy_pct": 64.67605633802816, + "macro_f1_pct": 62.959034577791584, + "weighted_f1_pct": 66.18361318945371, + "macro_precision_pct": 65.4346713584342, + "macro_recall_pct": 65.97464011846169, + "per_class_f1_pct": { + "FAVOR": 56.70367207514945, + "AGAINST": 58.469259723964875, + "NONE": 73.70417193426043 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7614626117964672, + "losses": { + "current_ce": 0.3983876136149846, + "replay_ce": 0.11799092511920369, + "dark_distillation": 0.26059304852696025, + "supervised_contrastive": 4.09482425848643, + "total": 1.0561574926563337, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6206896551724138, + "macro_f1": 0.621907743858059, + "weighted_f1": 0.6300587521919014, + "macro_precision": 0.6585558965901629, + "macro_recall": 0.6599303176921058, + "per_class_f1": { + "FAVOR": 0.5481481481481482, + "AGAINST": 0.6724137931034483, + "NONE": 0.6451612903225806 + }, + "accuracy_pct": 62.06896551724138, + "macro_f1_pct": 62.1907743858059, + "weighted_f1_pct": 63.005875219190145, + "macro_precision_pct": 65.85558965901629, + "macro_recall_pct": 65.99303176921057, + "per_class_f1_pct": { + "FAVOR": 54.81481481481482, + "AGAINST": 67.24137931034483, + "NONE": 64.51612903225806 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5485799701046338, + "macro_f1": 0.5244320352207028, + "weighted_f1": 0.5674577373012576, + "macro_precision": 0.5689094262087265, + "macro_recall": 0.5737515377755218, + "per_class_f1": { + "FAVOR": 0.5336225596529284, + "AGAINST": 0.39869281045751637, + "NONE": 0.6409807355516638 + }, + "accuracy_pct": 54.85799701046338, + "macro_f1_pct": 52.44320352207028, + "weighted_f1_pct": 56.74577373012576, + "macro_precision_pct": 56.89094262087265, + "macro_recall_pct": 57.37515377755218, + "per_class_f1_pct": { + "FAVOR": 53.36225596529284, + "AGAINST": 39.869281045751634, + "NONE": 64.09807355516638 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6554243264523638, + "weighted_f1": 0.6897146254336122, + "macro_precision": 0.6786208980383738, + "macro_recall": 0.6860015187601395, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.616822429906542, + "NONE": 0.7648351648351648 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 65.54243264523637, + "weighted_f1_pct": 68.97146254336121, + "macro_precision_pct": 67.86208980383738, + "macro_recall_pct": 68.60015187601395, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 61.6822429906542, + "NONE": 76.48351648351648 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7425249169435216, + "macro_f1": 0.70353502494036, + "weighted_f1": 0.74722010272993, + "macro_precision": 0.6986346018322763, + "macro_recall": 0.7121155654486283, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6163522012578616, + "NONE": 0.8275862068965517 + }, + "accuracy_pct": 74.25249169435216, + "macro_f1_pct": 70.35350249403601, + "weighted_f1_pct": 74.722010272993, + "macro_precision_pct": 69.86346018322763, + "macro_recall_pct": 71.21155654486283, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 61.63522012578616, + "NONE": 82.75862068965517 + } + } + ], + "accuracy": 0.642406394615061, + "macro_f1": 0.6242043945190932, + "weighted_f1": 0.6563717573085469, + "macro_precision": 0.6350244251322466, + "macro_recall": 0.6612393250664178, + "per_class_f1": { + "FAVOR": 0.5702127659574469, + "AGAINST": 0.5767284991568297, + "NONE": 0.7256719184430028 + }, + "accuracy_pct": 64.2406394615061, + "macro_f1_pct": 62.42043945190932, + "weighted_f1_pct": 65.6371757308547, + "macro_precision_pct": 63.50244251322466, + "macro_recall_pct": 66.12393250664178, + "per_class_f1_pct": { + "FAVOR": 57.02127659574469, + "AGAINST": 57.67284991568297, + "NONE": 72.56719184430028 + } + }, + { + "after_task": 5, + "best_epoch": 15, + "best_dev_macro_f1": 0.6669100294419987, + "losses": { + "current_ce": 0.5040755270035179, + "replay_ce": 0.10346628655714017, + "dark_distillation": 0.23923212061325708, + "supervised_contrastive": 4.128013796806336, + "total": 1.1399592612407825, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.641281595319099, + "weighted_f1": 0.6516722922510936, + "macro_precision": 0.6650537190764609, + "macro_recall": 0.6774192161516105, + "per_class_f1": { + "FAVOR": 0.575197889182058, + "AGAINST": 0.6721763085399449, + "NONE": 0.6764705882352942 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 64.1281595319099, + "weighted_f1_pct": 65.16722922510935, + "macro_precision_pct": 66.5053719076461, + "macro_recall_pct": 67.74192161516105, + "per_class_f1_pct": { + "FAVOR": 57.519788918205805, + "AGAINST": 67.21763085399449, + "NONE": 67.64705882352942 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5530642750373692, + "macro_f1": 0.5279806862190077, + "weighted_f1": 0.5766384093974084, + "macro_precision": 0.566455363366736, + "macro_recall": 0.5714473713585416, + "per_class_f1": { + "FAVOR": 0.5386416861826698, + "AGAINST": 0.3855421686746988, + "NONE": 0.6597582037996546 + }, + "accuracy_pct": 55.30642750373692, + "macro_f1_pct": 52.798068621900775, + "weighted_f1_pct": 57.66384093974084, + "macro_precision_pct": 56.645536336673594, + "macro_recall_pct": 57.144737135854164, + "per_class_f1_pct": { + "FAVOR": 53.86416861826698, + "AGAINST": 38.55421686746988, + "NONE": 65.97582037996546 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6941649899396378, + "macro_f1": 0.6741750880316179, + "weighted_f1": 0.7075591808464828, + "macro_precision": 0.6784072174590126, + "macro_recall": 0.7029195862529196, + "per_class_f1": { + "FAVOR": 0.5918367346938775, + "AGAINST": 0.6495726495726496, + "NONE": 0.7811158798283262 + }, + "accuracy_pct": 69.41649899396378, + "macro_f1_pct": 67.41750880316178, + "weighted_f1_pct": 70.75591808464829, + "macro_precision_pct": 67.84072174590125, + "macro_recall_pct": 70.29195862529195, + "per_class_f1_pct": { + "FAVOR": 59.183673469387756, + "AGAINST": 64.95726495726495, + "NONE": 78.11158798283262 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7225913621262459, + "macro_f1": 0.6873411293367263, + "weighted_f1": 0.7297748597192825, + "macro_precision": 0.6787865781122738, + "macro_recall": 0.7040081673384452, + "per_class_f1": { + "FAVOR": 0.6491228070175439, + "AGAINST": 0.6055045871559633, + "NONE": 0.8073959938366718 + }, + "accuracy_pct": 72.25913621262458, + "macro_f1_pct": 68.73411293367263, + "weighted_f1_pct": 72.97748597192825, + "macro_precision_pct": 67.87865781122738, + "macro_recall_pct": 70.40081673384452, + "per_class_f1_pct": { + "FAVOR": 64.91228070175438, + "AGAINST": 60.550458715596335, + "NONE": 80.73959938366718 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6912325285895807, + "macro_f1": 0.6367242196064132, + "weighted_f1": 0.7057354329210346, + "macro_precision": 0.624157304057905, + "macro_recall": 0.6702713408595762, + "per_class_f1": { + "FAVOR": 0.5661538461538461, + "AGAINST": 0.5444743935309974, + "NONE": 0.7995444191343963 + }, + "accuracy_pct": 69.12325285895807, + "macro_f1_pct": 63.67242196064132, + "weighted_f1_pct": 70.57354329210345, + "macro_precision_pct": 62.415730405790505, + "macro_recall_pct": 67.02713408595761, + "per_class_f1_pct": { + "FAVOR": 56.61538461538461, + "AGAINST": 54.44743935309974, + "NONE": 79.95444191343964 + } + } + ], + "accuracy": 0.6592920353982301, + "macro_f1": 0.6325578562919998, + "weighted_f1": 0.6737547580140615, + "macro_precision": 0.6321609511961056, + "macro_recall": 0.6683400199659717, + "per_class_f1": { + "FAVOR": 0.5771324863883848, + "AGAINST": 0.5679164105716041, + "NONE": 0.7526246719160105 + }, + "accuracy_pct": 65.929203539823, + "macro_f1_pct": 63.25578562919998, + "weighted_f1_pct": 67.37547580140615, + "macro_precision_pct": 63.21609511961056, + "macro_recall_pct": 66.83400199659717, + "per_class_f1_pct": { + "FAVOR": 57.71324863883848, + "AGAINST": 56.79164105716041, + "NONE": 75.26246719160105 + } + }, + { + "after_task": 6, + "best_epoch": 13, + "best_dev_macro_f1": 0.7140302120523639, + "losses": { + "current_ce": 0.44926638738326136, + "replay_ce": 0.09680496590685041, + "dark_distillation": 0.22680248683761983, + "supervised_contrastive": 4.039770916104317, + "total": 1.063449694239904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6443936204024284, + "weighted_f1": 0.6590184032855443, + "macro_precision": 0.6780700240090013, + "macro_recall": 0.6787147074195451, + "per_class_f1": { + "FAVOR": 0.575682382133995, + "AGAINST": 0.658753709198813, + "NONE": 0.698744769874477 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.43936204024284, + "weighted_f1_pct": 65.90184032855443, + "macro_precision_pct": 67.80700240090013, + "macro_recall_pct": 67.87147074195451, + "per_class_f1_pct": { + "FAVOR": 57.5682382133995, + "AGAINST": 65.8753709198813, + "NONE": 69.8744769874477 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5754857997010463, + "macro_f1": 0.5478229375793808, + "weighted_f1": 0.5992509079365502, + "macro_precision": 0.5764053583056973, + "macro_recall": 0.5874889991034783, + "per_class_f1": { + "FAVOR": 0.5397590361445783, + "AGAINST": 0.4146341463414634, + "NONE": 0.6890756302521008 + }, + "accuracy_pct": 57.548579970104626, + "macro_f1_pct": 54.782293757938085, + "weighted_f1_pct": 59.92509079365502, + "macro_precision_pct": 57.640535830569725, + "macro_recall_pct": 58.74889991034783, + "per_class_f1_pct": { + "FAVOR": 53.975903614457835, + "AGAINST": 41.46341463414634, + "NONE": 68.90756302521008 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6921529175050302, + "macro_f1": 0.6702664665818635, + "weighted_f1": 0.708507930585382, + "macro_precision": 0.681149442810571, + "macro_recall": 0.6931017224120671, + "per_class_f1": { + "FAVOR": 0.5629139072847682, + "AGAINST": 0.6547085201793722, + "NONE": 0.7931769722814499 + }, + "accuracy_pct": 69.21529175050301, + "macro_f1_pct": 67.02664665818635, + "weighted_f1_pct": 70.8507930585382, + "macro_precision_pct": 68.1149442810571, + "macro_recall_pct": 69.31017224120671, + "per_class_f1_pct": { + "FAVOR": 56.29139072847682, + "AGAINST": 65.47085201793722, + "NONE": 79.317697228145 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.729235880398671, + "macro_f1": 0.677844790779011, + "weighted_f1": 0.7325971341249313, + "macro_precision": 0.6709621311538663, + "macro_recall": 0.6893357633012313, + "per_class_f1": { + "FAVOR": 0.6115702479338843, + "AGAINST": 0.5923344947735192, + "NONE": 0.8296296296296296 + }, + "accuracy_pct": 72.9235880398671, + "macro_f1_pct": 67.78447907790111, + "weighted_f1_pct": 73.25971341249313, + "macro_precision_pct": 67.09621311538663, + "macro_recall_pct": 68.93357633012313, + "per_class_f1_pct": { + "FAVOR": 61.15702479338842, + "AGAINST": 59.23344947735192, + "NONE": 82.96296296296296 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6658195679796697, + "macro_f1": 0.6085078385250307, + "weighted_f1": 0.6820972552305876, + "macro_precision": 0.6018332166815114, + "macro_recall": 0.6497562773706564, + "per_class_f1": { + "FAVOR": 0.5519125683060109, + "AGAINST": 0.4897959183673469, + "NONE": 0.7838150289017342 + }, + "accuracy_pct": 66.58195679796697, + "macro_f1_pct": 60.85078385250306, + "weighted_f1_pct": 68.20972552305876, + "macro_precision_pct": 60.18332166815114, + "macro_recall_pct": 64.97562773706565, + "per_class_f1_pct": { + "FAVOR": 55.19125683060109, + "AGAINST": 48.97959183673469, + "NONE": 78.38150289017342 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6511627906976745, + "macro_f1": 0.6112572119378699, + "weighted_f1": 0.6624740655454842, + "macro_precision": 0.6097908786264951, + "macro_recall": 0.6389242553872617, + "per_class_f1": { + "FAVOR": 0.4715447154471545, + "AGAINST": 0.6325757575757576, + "NONE": 0.7296511627906976 + }, + "accuracy_pct": 65.11627906976744, + "macro_f1_pct": 61.12572119378699, + "weighted_f1_pct": 66.24740655454842, + "macro_precision_pct": 60.97908786264951, + "macro_recall_pct": 63.89242553872617, + "per_class_f1_pct": { + "FAVOR": 47.15447154471545, + "AGAINST": 63.25757575757576, + "NONE": 72.96511627906976 + } + } + ], + "accuracy": 0.6577663671373556, + "macro_f1": 0.6285104099034471, + "weighted_f1": 0.6726580542382737, + "macro_precision": 0.6286847431738013, + "macro_recall": 0.662205497382199, + "per_class_f1": { + "FAVOR": 0.5531914893617021, + "AGAINST": 0.5747800586510264, + "NONE": 0.7575596816976128 + }, + "accuracy_pct": 65.77663671373556, + "macro_f1_pct": 62.85104099034471, + "weighted_f1_pct": 67.26580542382737, + "macro_precision_pct": 62.86847431738013, + "macro_recall_pct": 66.2205497382199, + "per_class_f1_pct": { + "FAVOR": 55.319148936170215, + "AGAINST": 57.478005865102645, + "NONE": 75.75596816976127 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6399541751666544, + "losses": { + "current_ce": 0.4209737642180352, + "replay_ce": 0.09143301009954441, + "dark_distillation": 0.22609852960421925, + "supervised_contrastive": 4.070898907525199, + "total": 1.0325459368172145, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.638752052545156, + "macro_f1": 0.6358106366195218, + "weighted_f1": 0.6438370517574524, + "macro_precision": 0.660949537039758, + "macro_recall": 0.6755927613245433, + "per_class_f1": { + "FAVOR": 0.5888888888888889, + "AGAINST": 0.6547314578005116, + "NONE": 0.6638115631691649 + }, + "accuracy_pct": 63.8752052545156, + "macro_f1_pct": 63.58106366195217, + "weighted_f1_pct": 64.38370517574525, + "macro_precision_pct": 66.0949537039758, + "macro_recall_pct": 67.55927613245433, + "per_class_f1_pct": { + "FAVOR": 58.88888888888889, + "AGAINST": 65.47314578005115, + "NONE": 66.38115631691649 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5440956651718983, + "macro_f1": 0.5106926477593023, + "weighted_f1": 0.5675782703339058, + "macro_precision": 0.5494565804760746, + "macro_recall": 0.5495154905547976, + "per_class_f1": { + "FAVOR": 0.5155555555555555, + "AGAINST": 0.3509933774834437, + "NONE": 0.6655290102389079 + }, + "accuracy_pct": 54.40956651718983, + "macro_f1_pct": 51.069264775930236, + "weighted_f1_pct": 56.75782703339058, + "macro_precision_pct": 54.94565804760746, + "macro_recall_pct": 54.95154905547977, + "per_class_f1_pct": { + "FAVOR": 51.55555555555556, + "AGAINST": 35.099337748344375, + "NONE": 66.55290102389078 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.682092555331992, + "macro_f1": 0.6645717883501646, + "weighted_f1": 0.6985498083668483, + "macro_precision": 0.6778151220442409, + "macro_recall": 0.6919753086419753, + "per_class_f1": { + "FAVOR": 0.564935064935065, + "AGAINST": 0.6548672566371682, + "NONE": 0.7739130434782608 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 66.45717883501646, + "weighted_f1_pct": 69.85498083668483, + "macro_precision_pct": 67.78151220442409, + "macro_recall_pct": 69.19753086419753, + "per_class_f1_pct": { + "FAVOR": 56.493506493506494, + "AGAINST": 65.48672566371681, + "NONE": 77.39130434782608 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.675152914939907, + "weighted_f1": 0.726561619593977, + "macro_precision": 0.6651089066736567, + "macro_recall": 0.6920545071932627, + "per_class_f1": { + "FAVOR": 0.6134453781512605, + "AGAINST": 0.5942492012779552, + "NONE": 0.8177641653905053 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 67.5152914939907, + "weighted_f1_pct": 72.6561619593977, + "macro_precision_pct": 66.51089066736567, + "macro_recall_pct": 69.20545071932627, + "per_class_f1_pct": { + "FAVOR": 61.34453781512605, + "AGAINST": 59.424920127795524, + "NONE": 81.77641653905053 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6518424396442185, + "macro_f1": 0.6013594109459625, + "weighted_f1": 0.6699970835107516, + "macro_precision": 0.5965729552837082, + "macro_recall": 0.6449223535280181, + "per_class_f1": { + "FAVOR": 0.5268817204301075, + "AGAINST": 0.5142857142857142, + "NONE": 0.7629107981220657 + }, + "accuracy_pct": 65.18424396442185, + "macro_f1_pct": 60.13594109459625, + "weighted_f1_pct": 66.99970835107516, + "macro_precision_pct": 59.657295528370824, + "macro_recall_pct": 64.49223535280181, + "per_class_f1_pct": { + "FAVOR": 52.68817204301075, + "AGAINST": 51.42857142857142, + "NONE": 76.29107981220658 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5909712722298222, + "macro_f1": 0.5654736920713259, + "weighted_f1": 0.6115723625476617, + "macro_precision": 0.5796937699439856, + "macro_recall": 0.6217531863665108, + "per_class_f1": { + "FAVOR": 0.4471299093655589, + "AGAINST": 0.5714285714285714, + "NONE": 0.6778625954198473 + }, + "accuracy_pct": 59.097127222982216, + "macro_f1_pct": 56.54736920713259, + "weighted_f1_pct": 61.15723625476617, + "macro_precision_pct": 57.96937699439856, + "macro_recall_pct": 62.17531863665108, + "per_class_f1_pct": { + "FAVOR": 44.71299093655589, + "AGAINST": 57.14285714285714, + "NONE": 67.78625954198473 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6883212193874031, + "weighted_f1": 0.7196208715289613, + "macro_precision": 0.6863355352303523, + "macro_recall": 0.690613543229608, + "per_class_f1": { + "FAVOR": 0.6058091286307054, + "AGAINST": 0.6535433070866141, + "NONE": 0.8056112224448898 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 68.83212193874031, + "weighted_f1_pct": 71.96208715289613, + "macro_precision_pct": 68.63355352303523, + "macro_recall_pct": 69.0613543229608, + "per_class_f1_pct": { + "FAVOR": 60.58091286307054, + "AGAINST": 65.35433070866141, + "NONE": 80.56112224448898 + } + } + ], + "accuracy": 0.6436703096539163, + "macro_f1": 0.6177498947878984, + "weighted_f1": 0.6584633587671117, + "macro_precision": 0.6203799364014558, + "macro_recall": 0.651298592128999, + "per_class_f1": { + "FAVOR": 0.5452173913043479, + "AGAINST": 0.5683391003460208, + "NONE": 0.7396931927133269 + }, + "accuracy_pct": 64.36703096539162, + "macro_f1_pct": 61.77498947878984, + "weighted_f1_pct": 65.84633587671117, + "macro_precision_pct": 62.03799364014559, + "macro_recall_pct": 65.1298592128999, + "per_class_f1_pct": { + "FAVOR": 54.52173913043479, + "AGAINST": 56.83391003460208, + "NONE": 73.9693192713327 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.744947209653092, + "losses": { + "current_ce": 0.31210518405447163, + "replay_ce": 0.0900462328933361, + "dark_distillation": 0.21146271038588468, + "supervised_contrastive": 3.969975049708917, + "total": 0.9048802844877165, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6502463054187192, + "macro_f1": 0.6460721300459262, + "weighted_f1": 0.6589852203137017, + "macro_precision": 0.6656183638799046, + "macro_recall": 0.6789851894720235, + "per_class_f1": { + "FAVOR": 0.5916230366492147, + "AGAINST": 0.6512968299711815, + "NONE": 0.6952965235173824 + }, + "accuracy_pct": 65.02463054187191, + "macro_f1_pct": 64.60721300459262, + "weighted_f1_pct": 65.89852203137016, + "macro_precision_pct": 66.56183638799045, + "macro_recall_pct": 67.89851894720235, + "per_class_f1_pct": { + "FAVOR": 59.16230366492147, + "AGAINST": 65.12968299711815, + "NONE": 69.52965235173824 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5500747384155455, + "macro_f1": 0.5244981177407264, + "weighted_f1": 0.5702442936789999, + "macro_precision": 0.561861888141887, + "macro_recall": 0.5672661494335934, + "per_class_f1": { + "FAVOR": 0.5120350109409191, + "AGAINST": 0.4107744107744108, + "NONE": 0.6506849315068494 + }, + "accuracy_pct": 55.00747384155456, + "macro_f1_pct": 52.44981177407264, + "weighted_f1_pct": 57.024429367899984, + "macro_precision_pct": 56.1861888141887, + "macro_recall_pct": 56.726614943359344, + "per_class_f1_pct": { + "FAVOR": 51.203501094091905, + "AGAINST": 41.07744107744108, + "NONE": 65.06849315068493 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6619718309859155, + "macro_f1": 0.6444272626707189, + "weighted_f1": 0.6795031063608753, + "macro_precision": 0.670064927173932, + "macro_recall": 0.6736771254012633, + "per_class_f1": { + "FAVOR": 0.5626911314984709, + "AGAINST": 0.6139534883720931, + "NONE": 0.7566371681415929 + }, + "accuracy_pct": 66.19718309859155, + "macro_f1_pct": 64.44272626707189, + "weighted_f1_pct": 67.95031063608754, + "macro_precision_pct": 67.0064927173932, + "macro_recall_pct": 67.36771254012633, + "per_class_f1_pct": { + "FAVOR": 56.269113149847094, + "AGAINST": 61.395348837209305, + "NONE": 75.66371681415929 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7159468438538206, + "macro_f1": 0.6743276278959311, + "weighted_f1": 0.7232982936269272, + "macro_precision": 0.6633250798479805, + "macro_recall": 0.6974282517812783, + "per_class_f1": { + "FAVOR": 0.6171875, + "AGAINST": 0.5953177257525084, + "NONE": 0.810477657935285 + }, + "accuracy_pct": 71.59468438538205, + "macro_f1_pct": 67.43276278959311, + "weighted_f1_pct": 72.32982936269272, + "macro_precision_pct": 66.33250798479806, + "macro_recall_pct": 69.74282517812783, + "per_class_f1_pct": { + "FAVOR": 61.71875, + "AGAINST": 59.53177257525084, + "NONE": 81.0477657935285 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6416772554002541, + "macro_f1": 0.5740215240050243, + "weighted_f1": 0.654807276313168, + "macro_precision": 0.5776291516721361, + "macro_recall": 0.6188636547024347, + "per_class_f1": { + "FAVOR": 0.5326633165829145, + "AGAINST": 0.42105263157894735, + "NONE": 0.768348623853211 + }, + "accuracy_pct": 64.16772554002542, + "macro_f1_pct": 57.40215240050242, + "weighted_f1_pct": 65.48072763131681, + "macro_precision_pct": 57.76291516721361, + "macro_recall_pct": 61.88636547024346, + "per_class_f1_pct": { + "FAVOR": 53.266331658291456, + "AGAINST": 42.10526315789473, + "NONE": 76.8348623853211 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.560875512995896, + "macro_f1": 0.5379263612821078, + "weighted_f1": 0.5835758170033439, + "macro_precision": 0.5601708976273436, + "macro_recall": 0.5984887937515992, + "per_class_f1": { + "FAVOR": 0.43103448275862066, + "AGAINST": 0.5254237288135594, + "NONE": 0.6573208722741433 + }, + "accuracy_pct": 56.0875512995896, + "macro_f1_pct": 53.79263612821078, + "weighted_f1_pct": 58.35758170033439, + "macro_precision_pct": 56.017089762734365, + "macro_recall_pct": 59.848879375159925, + "per_class_f1_pct": { + "FAVOR": 43.103448275862064, + "AGAINST": 52.54237288135594, + "NONE": 65.73208722741433 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.6902964959568733, + "weighted_f1": 0.7198149062087409, + "macro_precision": 0.6874763794406652, + "macro_recall": 0.7000371827195703, + "per_class_f1": { + "FAVOR": 0.6113207547169811, + "AGAINST": 0.6587301587301587, + "NONE": 0.80083857442348 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 69.02964959568733, + "weighted_f1_pct": 71.98149062087408, + "macro_precision_pct": 68.74763794406653, + "macro_recall_pct": 70.00371827195703, + "per_class_f1_pct": { + "FAVOR": 61.13207547169811, + "AGAINST": 65.87301587301587, + "NONE": 80.083857442348 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.700139470013947, + "macro_f1": 0.6652238411372376, + "weighted_f1": 0.7120629153569743, + "macro_precision": 0.6550242400865232, + "macro_recall": 0.7073660605271298, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.6633906633906634, + "NONE": 0.7767253044654939 + }, + "accuracy_pct": 70.01394700139471, + "macro_f1_pct": 66.52238411372376, + "weighted_f1_pct": 71.20629153569743, + "macro_precision_pct": 65.50242400865231, + "macro_recall_pct": 70.73660605271299, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 66.33906633906635, + "NONE": 77.67253044654939 + } + } + ], + "accuracy": 0.6451360344490116, + "macro_f1": 0.6181564085981478, + "weighted_f1": 0.6602848629863943, + "macro_precision": 0.6224089406734592, + "macro_recall": 0.6547748394254188, + "per_class_f1": { + "FAVOR": 0.5461227489893421, + "AGAINST": 0.5669109139992287, + "NONE": 0.7414355628058727 + }, + "accuracy_pct": 64.51360344490115, + "macro_f1_pct": 61.81564085981478, + "weighted_f1_pct": 66.02848629863944, + "macro_precision_pct": 62.24089406734592, + "macro_recall_pct": 65.47748394254188, + "per_class_f1_pct": { + "FAVOR": 54.61227489893421, + "AGAINST": 56.69109139992287, + "NONE": 74.14355628058728 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6195991700918182, + "average_forgetting": 0.043980980311661186, + "backward_transfer": -0.04193359064114821, + "final_mean_target_macro_f1_pct": 61.95991700918182, + "average_forgetting_pct": 4.398098031166119, + "backward_transfer_pct": -4.193359064114821, + "final_pooled_accuracy": 0.6451360344490116, + "final_pooled_accuracy_pct": 64.51360344490115, + "final_pooled_macro_f1": 0.6181564085981478, + "final_pooled_macro_f1_pct": 61.81564085981478, + "final_pooled_weighted_f1": 0.6602848629863943, + "final_pooled_weighted_f1_pct": 66.02848629863944, + "final_pooled_macro_precision": 0.6224089406734592, + "final_pooled_macro_precision_pct": 62.24089406734592, + "final_pooled_macro_recall": 0.6547748394254188, + "final_pooled_macro_recall_pct": 65.47748394254188, + "final_pooled_per_class_f1": { + "FAVOR": 0.5461227489893421, + "AGAINST": 0.5669109139992287, + "NONE": 0.7414355628058727 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 54.61227489893421, + "AGAINST": 56.69109139992287, + "NONE": 74.14355628058728 + } + }, + "efficiency": { + "training_wall_time_seconds": 1840.3666741380002, + "training_wall_time_minutes": 30.672777902300005, + "peak_gpu_memory_allocated_gb": 10.40872573852539, + "peak_gpu_memory_reserved_gb": 10.599609375 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..5c3e861de0e32a8f8b7e3094e16cb9a59f42f23d --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/metrics.partial.json @@ -0,0 +1,1495 @@ +{ + "matrix": [ + [ + 0.6513687947181975, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6657005224117883, + 0.5816105653599061, + null, + null, + null, + null, + null, + null + ], + [ + 0.6276019493374833, + 0.5582475091703928, + 0.7122876181351957, + null, + null, + null, + null, + null + ], + [ + 0.621907743858059, + 0.5244320352207028, + 0.6554243264523638, + 0.70353502494036, + null, + null, + null, + null + ], + [ + 0.641281595319099, + 0.5279806862190077, + 0.6741750880316179, + 0.6873411293367263, + 0.6367242196064132, + null, + null, + null + ], + [ + 0.6443936204024284, + 0.5478229375793808, + 0.6702664665818635, + 0.677844790779011, + 0.6085078385250307, + 0.6112572119378699, + null, + null + ], + [ + 0.6358106366195218, + 0.5106926477593023, + 0.6645717883501646, + 0.675152914939907, + 0.6013594109459625, + 0.5654736920713259, + 0.6883212193874031, + null + ], + [ + 0.6460721300459262, + 0.5244981177407264, + 0.6444272626707189, + 0.6743276278959311, + 0.5740215240050243, + 0.5379263612821078, + 0.6902964959568733, + 0.6652238411372376 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6952735888792025, + "losses": { + "current_ce": 0.5369640307341303, + "total": 0.5369640307341303, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6600985221674877, + "macro_f1": 0.6513687947181975, + "weighted_f1": 0.6700885375939004, + "macro_precision": 0.6615219122816131, + "macro_recall": 0.6759561456438371, + "per_class_f1": { + "FAVOR": 0.5754189944134078, + "AGAINST": 0.655367231638418, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.13687947181975, + "weighted_f1_pct": 67.00885375939004, + "macro_precision_pct": 66.1521912281613, + "macro_recall_pct": 67.59561456438371, + "per_class_f1_pct": { + "FAVOR": 57.54189944134078, + "AGAINST": 65.5367231638418, + "NONE": 72.33201581027669 + } + } + ], + "accuracy": 0.6600985221674877, + "macro_f1": 0.6513687947181975, + "weighted_f1": 0.6700885375939004, + "macro_precision": 0.6615219122816131, + "macro_recall": 0.6759561456438371, + "per_class_f1": { + "FAVOR": 0.5754189944134078, + "AGAINST": 0.655367231638418, + "NONE": 0.7233201581027668 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.13687947181975, + "weighted_f1_pct": 67.00885375939004, + "macro_precision_pct": 66.1521912281613, + "macro_recall_pct": 67.59561456438371, + "per_class_f1_pct": { + "FAVOR": 57.54189944134078, + "AGAINST": 65.5367231638418, + "NONE": 72.33201581027669 + } + }, + { + "after_task": 2, + "best_epoch": 14, + "best_dev_macro_f1": 0.6851579400510347, + "losses": { + "current_ce": 0.5884127289056778, + "replay_ce": 0.10817521661520005, + "dark_distillation": 0.2354041861076104, + "supervised_contrastive": 4.256149932376125, + "total": 1.2399050376917187, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6699507389162561, + "macro_f1": 0.6657005224117883, + "weighted_f1": 0.6777906120558884, + "macro_precision": 0.680324074074074, + "macro_recall": 0.6982276415834284, + "per_class_f1": { + "FAVOR": 0.5994550408719346, + "AGAINST": 0.6888888888888889, + "NONE": 0.7087576374745418 + }, + "accuracy_pct": 66.99507389162561, + "macro_f1_pct": 66.57005224117883, + "weighted_f1_pct": 67.77906120558885, + "macro_precision_pct": 68.0324074074074, + "macro_recall_pct": 69.82276415834285, + "per_class_f1_pct": { + "FAVOR": 59.945504087193456, + "AGAINST": 68.88888888888889, + "NONE": 70.87576374745417 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.6203288490284006, + "macro_f1": 0.5816105653599061, + "weighted_f1": 0.6384424317763885, + "macro_precision": 0.5922763510998804, + "macro_recall": 0.6115264720638915, + "per_class_f1": { + "FAVOR": 0.5520581113801453, + "AGAINST": 0.4529616724738676, + "NONE": 0.7398119122257053 + }, + "accuracy_pct": 62.03288490284006, + "macro_f1_pct": 58.16105653599061, + "weighted_f1_pct": 63.84424317763885, + "macro_precision_pct": 59.22763510998804, + "macro_recall_pct": 61.152647206389155, + "per_class_f1_pct": { + "FAVOR": 55.205811138014525, + "AGAINST": 45.29616724738676, + "NONE": 73.98119122257053 + } + } + ], + "accuracy": 0.6439749608763693, + "macro_f1": 0.6283001235687409, + "weighted_f1": 0.6571767478765632, + "macro_precision": 0.6399963624169466, + "macro_recall": 0.660062213977916, + "per_class_f1": { + "FAVOR": 0.5743589743589743, + "AGAINST": 0.5842349304482226, + "NONE": 0.7263064658990257 + }, + "accuracy_pct": 64.39749608763692, + "macro_f1_pct": 62.830012356874086, + "weighted_f1_pct": 65.71767478765632, + "macro_precision_pct": 63.99963624169466, + "macro_recall_pct": 66.00622139779159, + "per_class_f1_pct": { + "FAVOR": 57.43589743589743, + "AGAINST": 58.42349304482226, + "NONE": 72.63064658990257 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.7539987422241418, + "losses": { + "current_ce": 0.4725877070355983, + "replay_ce": 0.13467356124449345, + "dark_distillation": 0.2609920618079957, + "supervised_contrastive": 4.164763221854256, + "total": 1.1542336293629238, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6272577996715928, + "macro_f1": 0.6276019493374833, + "weighted_f1": 0.640233686096964, + "macro_precision": 0.6711881442884482, + "macro_recall": 0.6608475640165782, + "per_class_f1": { + "FAVOR": 0.5566037735849056, + "AGAINST": 0.653968253968254, + "NONE": 0.6722338204592901 + }, + "accuracy_pct": 62.725779967159276, + "macro_f1_pct": 62.76019493374832, + "weighted_f1_pct": 64.02336860969639, + "macro_precision_pct": 67.11881442884481, + "macro_recall_pct": 66.08475640165781, + "per_class_f1_pct": { + "FAVOR": 55.660377358490564, + "AGAINST": 65.39682539682539, + "NONE": 67.22338204592901 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.593423019431988, + "macro_f1": 0.5582475091703928, + "weighted_f1": 0.6117358270549611, + "macro_precision": 0.5952476984866806, + "macro_recall": 0.5994516295426799, + "per_class_f1": { + "FAVOR": 0.5466101694915254, + "AGAINST": 0.4226415094339623, + "NONE": 0.7054908485856906 + }, + "accuracy_pct": 59.342301943198805, + "macro_f1_pct": 55.82475091703928, + "weighted_f1_pct": 61.17358270549611, + "macro_precision_pct": 59.52476984866806, + "macro_recall_pct": 59.94516295426799, + "per_class_f1_pct": { + "FAVOR": 54.66101694915254, + "AGAINST": 42.26415094339623, + "NONE": 70.54908485856906 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7424547283702213, + "macro_f1": 0.7122876181351957, + "weighted_f1": 0.7511279162497875, + "macro_precision": 0.7126256054362755, + "macro_recall": 0.7257343550446999, + "per_class_f1": { + "FAVOR": 0.6181818181818182, + "AGAINST": 0.6820276497695853, + "NONE": 0.8366533864541833 + }, + "accuracy_pct": 74.24547283702213, + "macro_f1_pct": 71.22876181351957, + "weighted_f1_pct": 75.11279162497874, + "macro_precision_pct": 71.26256054362756, + "macro_recall_pct": 72.57343550446998, + "per_class_f1_pct": { + "FAVOR": 61.81818181818181, + "AGAINST": 68.20276497695853, + "NONE": 83.66533864541833 + } + } + ], + "accuracy": 0.6467605633802817, + "macro_f1": 0.6295903457779158, + "weighted_f1": 0.661836131894537, + "macro_precision": 0.654346713584342, + "macro_recall": 0.6597464011846169, + "per_class_f1": { + "FAVOR": 0.5670367207514945, + "AGAINST": 0.5846925972396487, + "NONE": 0.7370417193426043 + }, + "accuracy_pct": 64.67605633802816, + "macro_f1_pct": 62.959034577791584, + "weighted_f1_pct": 66.18361318945371, + "macro_precision_pct": 65.4346713584342, + "macro_recall_pct": 65.97464011846169, + "per_class_f1_pct": { + "FAVOR": 56.70367207514945, + "AGAINST": 58.469259723964875, + "NONE": 73.70417193426043 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7614626117964672, + "losses": { + "current_ce": 0.3983876136149846, + "replay_ce": 0.11799092511920369, + "dark_distillation": 0.26059304852696025, + "supervised_contrastive": 4.09482425848643, + "total": 1.0561574926563337, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6206896551724138, + "macro_f1": 0.621907743858059, + "weighted_f1": 0.6300587521919014, + "macro_precision": 0.6585558965901629, + "macro_recall": 0.6599303176921058, + "per_class_f1": { + "FAVOR": 0.5481481481481482, + "AGAINST": 0.6724137931034483, + "NONE": 0.6451612903225806 + }, + "accuracy_pct": 62.06896551724138, + "macro_f1_pct": 62.1907743858059, + "weighted_f1_pct": 63.005875219190145, + "macro_precision_pct": 65.85558965901629, + "macro_recall_pct": 65.99303176921057, + "per_class_f1_pct": { + "FAVOR": 54.81481481481482, + "AGAINST": 67.24137931034483, + "NONE": 64.51612903225806 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5485799701046338, + "macro_f1": 0.5244320352207028, + "weighted_f1": 0.5674577373012576, + "macro_precision": 0.5689094262087265, + "macro_recall": 0.5737515377755218, + "per_class_f1": { + "FAVOR": 0.5336225596529284, + "AGAINST": 0.39869281045751637, + "NONE": 0.6409807355516638 + }, + "accuracy_pct": 54.85799701046338, + "macro_f1_pct": 52.44320352207028, + "weighted_f1_pct": 56.74577373012576, + "macro_precision_pct": 56.89094262087265, + "macro_recall_pct": 57.37515377755218, + "per_class_f1_pct": { + "FAVOR": 53.36225596529284, + "AGAINST": 39.869281045751634, + "NONE": 64.09807355516638 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6740442655935613, + "macro_f1": 0.6554243264523638, + "weighted_f1": 0.6897146254336122, + "macro_precision": 0.6786208980383738, + "macro_recall": 0.6860015187601395, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.616822429906542, + "NONE": 0.7648351648351648 + }, + "accuracy_pct": 67.40442655935614, + "macro_f1_pct": 65.54243264523637, + "weighted_f1_pct": 68.97146254336121, + "macro_precision_pct": 67.86208980383738, + "macro_recall_pct": 68.60015187601395, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 61.6822429906542, + "NONE": 76.48351648351648 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7425249169435216, + "macro_f1": 0.70353502494036, + "weighted_f1": 0.74722010272993, + "macro_precision": 0.6986346018322763, + "macro_recall": 0.7121155654486283, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6163522012578616, + "NONE": 0.8275862068965517 + }, + "accuracy_pct": 74.25249169435216, + "macro_f1_pct": 70.35350249403601, + "weighted_f1_pct": 74.722010272993, + "macro_precision_pct": 69.86346018322763, + "macro_recall_pct": 71.21155654486283, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 61.63522012578616, + "NONE": 82.75862068965517 + } + } + ], + "accuracy": 0.642406394615061, + "macro_f1": 0.6242043945190932, + "weighted_f1": 0.6563717573085469, + "macro_precision": 0.6350244251322466, + "macro_recall": 0.6612393250664178, + "per_class_f1": { + "FAVOR": 0.5702127659574469, + "AGAINST": 0.5767284991568297, + "NONE": 0.7256719184430028 + }, + "accuracy_pct": 64.2406394615061, + "macro_f1_pct": 62.42043945190932, + "weighted_f1_pct": 65.6371757308547, + "macro_precision_pct": 63.50244251322466, + "macro_recall_pct": 66.12393250664178, + "per_class_f1_pct": { + "FAVOR": 57.02127659574469, + "AGAINST": 57.67284991568297, + "NONE": 72.56719184430028 + } + }, + { + "after_task": 5, + "best_epoch": 15, + "best_dev_macro_f1": 0.6669100294419987, + "losses": { + "current_ce": 0.5040755270035179, + "replay_ce": 0.10346628655714017, + "dark_distillation": 0.23923212061325708, + "supervised_contrastive": 4.128013796806336, + "total": 1.1399592612407825, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.641281595319099, + "weighted_f1": 0.6516722922510936, + "macro_precision": 0.6650537190764609, + "macro_recall": 0.6774192161516105, + "per_class_f1": { + "FAVOR": 0.575197889182058, + "AGAINST": 0.6721763085399449, + "NONE": 0.6764705882352942 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 64.1281595319099, + "weighted_f1_pct": 65.16722922510935, + "macro_precision_pct": 66.5053719076461, + "macro_recall_pct": 67.74192161516105, + "per_class_f1_pct": { + "FAVOR": 57.519788918205805, + "AGAINST": 67.21763085399449, + "NONE": 67.64705882352942 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5530642750373692, + "macro_f1": 0.5279806862190077, + "weighted_f1": 0.5766384093974084, + "macro_precision": 0.566455363366736, + "macro_recall": 0.5714473713585416, + "per_class_f1": { + "FAVOR": 0.5386416861826698, + "AGAINST": 0.3855421686746988, + "NONE": 0.6597582037996546 + }, + "accuracy_pct": 55.30642750373692, + "macro_f1_pct": 52.798068621900775, + "weighted_f1_pct": 57.66384093974084, + "macro_precision_pct": 56.645536336673594, + "macro_recall_pct": 57.144737135854164, + "per_class_f1_pct": { + "FAVOR": 53.86416861826698, + "AGAINST": 38.55421686746988, + "NONE": 65.97582037996546 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6941649899396378, + "macro_f1": 0.6741750880316179, + "weighted_f1": 0.7075591808464828, + "macro_precision": 0.6784072174590126, + "macro_recall": 0.7029195862529196, + "per_class_f1": { + "FAVOR": 0.5918367346938775, + "AGAINST": 0.6495726495726496, + "NONE": 0.7811158798283262 + }, + "accuracy_pct": 69.41649899396378, + "macro_f1_pct": 67.41750880316178, + "weighted_f1_pct": 70.75591808464829, + "macro_precision_pct": 67.84072174590125, + "macro_recall_pct": 70.29195862529195, + "per_class_f1_pct": { + "FAVOR": 59.183673469387756, + "AGAINST": 64.95726495726495, + "NONE": 78.11158798283262 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7225913621262459, + "macro_f1": 0.6873411293367263, + "weighted_f1": 0.7297748597192825, + "macro_precision": 0.6787865781122738, + "macro_recall": 0.7040081673384452, + "per_class_f1": { + "FAVOR": 0.6491228070175439, + "AGAINST": 0.6055045871559633, + "NONE": 0.8073959938366718 + }, + "accuracy_pct": 72.25913621262458, + "macro_f1_pct": 68.73411293367263, + "weighted_f1_pct": 72.97748597192825, + "macro_precision_pct": 67.87865781122738, + "macro_recall_pct": 70.40081673384452, + "per_class_f1_pct": { + "FAVOR": 64.91228070175438, + "AGAINST": 60.550458715596335, + "NONE": 80.73959938366718 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6912325285895807, + "macro_f1": 0.6367242196064132, + "weighted_f1": 0.7057354329210346, + "macro_precision": 0.624157304057905, + "macro_recall": 0.6702713408595762, + "per_class_f1": { + "FAVOR": 0.5661538461538461, + "AGAINST": 0.5444743935309974, + "NONE": 0.7995444191343963 + }, + "accuracy_pct": 69.12325285895807, + "macro_f1_pct": 63.67242196064132, + "weighted_f1_pct": 70.57354329210345, + "macro_precision_pct": 62.415730405790505, + "macro_recall_pct": 67.02713408595761, + "per_class_f1_pct": { + "FAVOR": 56.61538461538461, + "AGAINST": 54.44743935309974, + "NONE": 79.95444191343964 + } + } + ], + "accuracy": 0.6592920353982301, + "macro_f1": 0.6325578562919998, + "weighted_f1": 0.6737547580140615, + "macro_precision": 0.6321609511961056, + "macro_recall": 0.6683400199659717, + "per_class_f1": { + "FAVOR": 0.5771324863883848, + "AGAINST": 0.5679164105716041, + "NONE": 0.7526246719160105 + }, + "accuracy_pct": 65.929203539823, + "macro_f1_pct": 63.25578562919998, + "weighted_f1_pct": 67.37547580140615, + "macro_precision_pct": 63.21609511961056, + "macro_recall_pct": 66.83400199659717, + "per_class_f1_pct": { + "FAVOR": 57.71324863883848, + "AGAINST": 56.79164105716041, + "NONE": 75.26246719160105 + } + }, + { + "after_task": 6, + "best_epoch": 13, + "best_dev_macro_f1": 0.7140302120523639, + "losses": { + "current_ce": 0.44926638738326136, + "replay_ce": 0.09680496590685041, + "dark_distillation": 0.22680248683761983, + "supervised_contrastive": 4.039770916104317, + "total": 1.063449694239904, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6443936204024284, + "weighted_f1": 0.6590184032855443, + "macro_precision": 0.6780700240090013, + "macro_recall": 0.6787147074195451, + "per_class_f1": { + "FAVOR": 0.575682382133995, + "AGAINST": 0.658753709198813, + "NONE": 0.698744769874477 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.43936204024284, + "weighted_f1_pct": 65.90184032855443, + "macro_precision_pct": 67.80700240090013, + "macro_recall_pct": 67.87147074195451, + "per_class_f1_pct": { + "FAVOR": 57.5682382133995, + "AGAINST": 65.8753709198813, + "NONE": 69.8744769874477 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5754857997010463, + "macro_f1": 0.5478229375793808, + "weighted_f1": 0.5992509079365502, + "macro_precision": 0.5764053583056973, + "macro_recall": 0.5874889991034783, + "per_class_f1": { + "FAVOR": 0.5397590361445783, + "AGAINST": 0.4146341463414634, + "NONE": 0.6890756302521008 + }, + "accuracy_pct": 57.548579970104626, + "macro_f1_pct": 54.782293757938085, + "weighted_f1_pct": 59.92509079365502, + "macro_precision_pct": 57.640535830569725, + "macro_recall_pct": 58.74889991034783, + "per_class_f1_pct": { + "FAVOR": 53.975903614457835, + "AGAINST": 41.46341463414634, + "NONE": 68.90756302521008 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6921529175050302, + "macro_f1": 0.6702664665818635, + "weighted_f1": 0.708507930585382, + "macro_precision": 0.681149442810571, + "macro_recall": 0.6931017224120671, + "per_class_f1": { + "FAVOR": 0.5629139072847682, + "AGAINST": 0.6547085201793722, + "NONE": 0.7931769722814499 + }, + "accuracy_pct": 69.21529175050301, + "macro_f1_pct": 67.02664665818635, + "weighted_f1_pct": 70.8507930585382, + "macro_precision_pct": 68.1149442810571, + "macro_recall_pct": 69.31017224120671, + "per_class_f1_pct": { + "FAVOR": 56.29139072847682, + "AGAINST": 65.47085201793722, + "NONE": 79.317697228145 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.729235880398671, + "macro_f1": 0.677844790779011, + "weighted_f1": 0.7325971341249313, + "macro_precision": 0.6709621311538663, + "macro_recall": 0.6893357633012313, + "per_class_f1": { + "FAVOR": 0.6115702479338843, + "AGAINST": 0.5923344947735192, + "NONE": 0.8296296296296296 + }, + "accuracy_pct": 72.9235880398671, + "macro_f1_pct": 67.78447907790111, + "weighted_f1_pct": 73.25971341249313, + "macro_precision_pct": 67.09621311538663, + "macro_recall_pct": 68.93357633012313, + "per_class_f1_pct": { + "FAVOR": 61.15702479338842, + "AGAINST": 59.23344947735192, + "NONE": 82.96296296296296 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6658195679796697, + "macro_f1": 0.6085078385250307, + "weighted_f1": 0.6820972552305876, + "macro_precision": 0.6018332166815114, + "macro_recall": 0.6497562773706564, + "per_class_f1": { + "FAVOR": 0.5519125683060109, + "AGAINST": 0.4897959183673469, + "NONE": 0.7838150289017342 + }, + "accuracy_pct": 66.58195679796697, + "macro_f1_pct": 60.85078385250306, + "weighted_f1_pct": 68.20972552305876, + "macro_precision_pct": 60.18332166815114, + "macro_recall_pct": 64.97562773706565, + "per_class_f1_pct": { + "FAVOR": 55.19125683060109, + "AGAINST": 48.97959183673469, + "NONE": 78.38150289017342 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6511627906976745, + "macro_f1": 0.6112572119378699, + "weighted_f1": 0.6624740655454842, + "macro_precision": 0.6097908786264951, + "macro_recall": 0.6389242553872617, + "per_class_f1": { + "FAVOR": 0.4715447154471545, + "AGAINST": 0.6325757575757576, + "NONE": 0.7296511627906976 + }, + "accuracy_pct": 65.11627906976744, + "macro_f1_pct": 61.12572119378699, + "weighted_f1_pct": 66.24740655454842, + "macro_precision_pct": 60.97908786264951, + "macro_recall_pct": 63.89242553872617, + "per_class_f1_pct": { + "FAVOR": 47.15447154471545, + "AGAINST": 63.25757575757576, + "NONE": 72.96511627906976 + } + } + ], + "accuracy": 0.6577663671373556, + "macro_f1": 0.6285104099034471, + "weighted_f1": 0.6726580542382737, + "macro_precision": 0.6286847431738013, + "macro_recall": 0.662205497382199, + "per_class_f1": { + "FAVOR": 0.5531914893617021, + "AGAINST": 0.5747800586510264, + "NONE": 0.7575596816976128 + }, + "accuracy_pct": 65.77663671373556, + "macro_f1_pct": 62.85104099034471, + "weighted_f1_pct": 67.26580542382737, + "macro_precision_pct": 62.86847431738013, + "macro_recall_pct": 66.2205497382199, + "per_class_f1_pct": { + "FAVOR": 55.319148936170215, + "AGAINST": 57.478005865102645, + "NONE": 75.75596816976127 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6399541751666544, + "losses": { + "current_ce": 0.4209737642180352, + "replay_ce": 0.09143301009954441, + "dark_distillation": 0.22609852960421925, + "supervised_contrastive": 4.070898907525199, + "total": 1.0325459368172145, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.638752052545156, + "macro_f1": 0.6358106366195218, + "weighted_f1": 0.6438370517574524, + "macro_precision": 0.660949537039758, + "macro_recall": 0.6755927613245433, + "per_class_f1": { + "FAVOR": 0.5888888888888889, + "AGAINST": 0.6547314578005116, + "NONE": 0.6638115631691649 + }, + "accuracy_pct": 63.8752052545156, + "macro_f1_pct": 63.58106366195217, + "weighted_f1_pct": 64.38370517574525, + "macro_precision_pct": 66.0949537039758, + "macro_recall_pct": 67.55927613245433, + "per_class_f1_pct": { + "FAVOR": 58.88888888888889, + "AGAINST": 65.47314578005115, + "NONE": 66.38115631691649 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5440956651718983, + "macro_f1": 0.5106926477593023, + "weighted_f1": 0.5675782703339058, + "macro_precision": 0.5494565804760746, + "macro_recall": 0.5495154905547976, + "per_class_f1": { + "FAVOR": 0.5155555555555555, + "AGAINST": 0.3509933774834437, + "NONE": 0.6655290102389079 + }, + "accuracy_pct": 54.40956651718983, + "macro_f1_pct": 51.069264775930236, + "weighted_f1_pct": 56.75782703339058, + "macro_precision_pct": 54.94565804760746, + "macro_recall_pct": 54.95154905547977, + "per_class_f1_pct": { + "FAVOR": 51.55555555555556, + "AGAINST": 35.099337748344375, + "NONE": 66.55290102389078 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.682092555331992, + "macro_f1": 0.6645717883501646, + "weighted_f1": 0.6985498083668483, + "macro_precision": 0.6778151220442409, + "macro_recall": 0.6919753086419753, + "per_class_f1": { + "FAVOR": 0.564935064935065, + "AGAINST": 0.6548672566371682, + "NONE": 0.7739130434782608 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 66.45717883501646, + "weighted_f1_pct": 69.85498083668483, + "macro_precision_pct": 67.78151220442409, + "macro_recall_pct": 69.19753086419753, + "per_class_f1_pct": { + "FAVOR": 56.493506493506494, + "AGAINST": 65.48672566371681, + "NONE": 77.39130434782608 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7192691029900332, + "macro_f1": 0.675152914939907, + "weighted_f1": 0.726561619593977, + "macro_precision": 0.6651089066736567, + "macro_recall": 0.6920545071932627, + "per_class_f1": { + "FAVOR": 0.6134453781512605, + "AGAINST": 0.5942492012779552, + "NONE": 0.8177641653905053 + }, + "accuracy_pct": 71.92691029900332, + "macro_f1_pct": 67.5152914939907, + "weighted_f1_pct": 72.6561619593977, + "macro_precision_pct": 66.51089066736567, + "macro_recall_pct": 69.20545071932627, + "per_class_f1_pct": { + "FAVOR": 61.34453781512605, + "AGAINST": 59.424920127795524, + "NONE": 81.77641653905053 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6518424396442185, + "macro_f1": 0.6013594109459625, + "weighted_f1": 0.6699970835107516, + "macro_precision": 0.5965729552837082, + "macro_recall": 0.6449223535280181, + "per_class_f1": { + "FAVOR": 0.5268817204301075, + "AGAINST": 0.5142857142857142, + "NONE": 0.7629107981220657 + }, + "accuracy_pct": 65.18424396442185, + "macro_f1_pct": 60.13594109459625, + "weighted_f1_pct": 66.99970835107516, + "macro_precision_pct": 59.657295528370824, + "macro_recall_pct": 64.49223535280181, + "per_class_f1_pct": { + "FAVOR": 52.68817204301075, + "AGAINST": 51.42857142857142, + "NONE": 76.29107981220658 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.5909712722298222, + "macro_f1": 0.5654736920713259, + "weighted_f1": 0.6115723625476617, + "macro_precision": 0.5796937699439856, + "macro_recall": 0.6217531863665108, + "per_class_f1": { + "FAVOR": 0.4471299093655589, + "AGAINST": 0.5714285714285714, + "NONE": 0.6778625954198473 + }, + "accuracy_pct": 59.097127222982216, + "macro_f1_pct": 56.54736920713259, + "weighted_f1_pct": 61.15723625476617, + "macro_precision_pct": 57.96937699439856, + "macro_recall_pct": 62.17531863665108, + "per_class_f1_pct": { + "FAVOR": 44.71299093655589, + "AGAINST": 57.14285714285714, + "NONE": 67.78625954198473 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6883212193874031, + "weighted_f1": 0.7196208715289613, + "macro_precision": 0.6863355352303523, + "macro_recall": 0.690613543229608, + "per_class_f1": { + "FAVOR": 0.6058091286307054, + "AGAINST": 0.6535433070866141, + "NONE": 0.8056112224448898 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 68.83212193874031, + "weighted_f1_pct": 71.96208715289613, + "macro_precision_pct": 68.63355352303523, + "macro_recall_pct": 69.0613543229608, + "per_class_f1_pct": { + "FAVOR": 60.58091286307054, + "AGAINST": 65.35433070866141, + "NONE": 80.56112224448898 + } + } + ], + "accuracy": 0.6436703096539163, + "macro_f1": 0.6177498947878984, + "weighted_f1": 0.6584633587671117, + "macro_precision": 0.6203799364014558, + "macro_recall": 0.651298592128999, + "per_class_f1": { + "FAVOR": 0.5452173913043479, + "AGAINST": 0.5683391003460208, + "NONE": 0.7396931927133269 + }, + "accuracy_pct": 64.36703096539162, + "macro_f1_pct": 61.77498947878984, + "weighted_f1_pct": 65.84633587671117, + "macro_precision_pct": 62.03799364014559, + "macro_recall_pct": 65.1298592128999, + "per_class_f1_pct": { + "FAVOR": 54.52173913043479, + "AGAINST": 56.83391003460208, + "NONE": 73.9693192713327 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.744947209653092, + "losses": { + "current_ce": 0.31210518405447163, + "replay_ce": 0.0900462328933361, + "dark_distillation": 0.21146271038588468, + "supervised_contrastive": 3.969975049708917, + "total": 0.9048802844877165, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6502463054187192, + "macro_f1": 0.6460721300459262, + "weighted_f1": 0.6589852203137017, + "macro_precision": 0.6656183638799046, + "macro_recall": 0.6789851894720235, + "per_class_f1": { + "FAVOR": 0.5916230366492147, + "AGAINST": 0.6512968299711815, + "NONE": 0.6952965235173824 + }, + "accuracy_pct": 65.02463054187191, + "macro_f1_pct": 64.60721300459262, + "weighted_f1_pct": 65.89852203137016, + "macro_precision_pct": 66.56183638799045, + "macro_recall_pct": 67.89851894720235, + "per_class_f1_pct": { + "FAVOR": 59.16230366492147, + "AGAINST": 65.12968299711815, + "NONE": 69.52965235173824 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5500747384155455, + "macro_f1": 0.5244981177407264, + "weighted_f1": 0.5702442936789999, + "macro_precision": 0.561861888141887, + "macro_recall": 0.5672661494335934, + "per_class_f1": { + "FAVOR": 0.5120350109409191, + "AGAINST": 0.4107744107744108, + "NONE": 0.6506849315068494 + }, + "accuracy_pct": 55.00747384155456, + "macro_f1_pct": 52.44981177407264, + "weighted_f1_pct": 57.024429367899984, + "macro_precision_pct": 56.1861888141887, + "macro_recall_pct": 56.726614943359344, + "per_class_f1_pct": { + "FAVOR": 51.203501094091905, + "AGAINST": 41.07744107744108, + "NONE": 65.06849315068493 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6619718309859155, + "macro_f1": 0.6444272626707189, + "weighted_f1": 0.6795031063608753, + "macro_precision": 0.670064927173932, + "macro_recall": 0.6736771254012633, + "per_class_f1": { + "FAVOR": 0.5626911314984709, + "AGAINST": 0.6139534883720931, + "NONE": 0.7566371681415929 + }, + "accuracy_pct": 66.19718309859155, + "macro_f1_pct": 64.44272626707189, + "weighted_f1_pct": 67.95031063608754, + "macro_precision_pct": 67.0064927173932, + "macro_recall_pct": 67.36771254012633, + "per_class_f1_pct": { + "FAVOR": 56.269113149847094, + "AGAINST": 61.395348837209305, + "NONE": 75.66371681415929 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7159468438538206, + "macro_f1": 0.6743276278959311, + "weighted_f1": 0.7232982936269272, + "macro_precision": 0.6633250798479805, + "macro_recall": 0.6974282517812783, + "per_class_f1": { + "FAVOR": 0.6171875, + "AGAINST": 0.5953177257525084, + "NONE": 0.810477657935285 + }, + "accuracy_pct": 71.59468438538205, + "macro_f1_pct": 67.43276278959311, + "weighted_f1_pct": 72.32982936269272, + "macro_precision_pct": 66.33250798479806, + "macro_recall_pct": 69.74282517812783, + "per_class_f1_pct": { + "FAVOR": 61.71875, + "AGAINST": 59.53177257525084, + "NONE": 81.0477657935285 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6416772554002541, + "macro_f1": 0.5740215240050243, + "weighted_f1": 0.654807276313168, + "macro_precision": 0.5776291516721361, + "macro_recall": 0.6188636547024347, + "per_class_f1": { + "FAVOR": 0.5326633165829145, + "AGAINST": 0.42105263157894735, + "NONE": 0.768348623853211 + }, + "accuracy_pct": 64.16772554002542, + "macro_f1_pct": 57.40215240050242, + "weighted_f1_pct": 65.48072763131681, + "macro_precision_pct": 57.76291516721361, + "macro_recall_pct": 61.88636547024346, + "per_class_f1_pct": { + "FAVOR": 53.266331658291456, + "AGAINST": 42.10526315789473, + "NONE": 76.8348623853211 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.560875512995896, + "macro_f1": 0.5379263612821078, + "weighted_f1": 0.5835758170033439, + "macro_precision": 0.5601708976273436, + "macro_recall": 0.5984887937515992, + "per_class_f1": { + "FAVOR": 0.43103448275862066, + "AGAINST": 0.5254237288135594, + "NONE": 0.6573208722741433 + }, + "accuracy_pct": 56.0875512995896, + "macro_f1_pct": 53.79263612821078, + "weighted_f1_pct": 58.35758170033439, + "macro_precision_pct": 56.017089762734365, + "macro_recall_pct": 59.848879375159925, + "per_class_f1_pct": { + "FAVOR": 43.103448275862064, + "AGAINST": 52.54237288135594, + "NONE": 65.73208722741433 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.6902964959568733, + "weighted_f1": 0.7198149062087409, + "macro_precision": 0.6874763794406652, + "macro_recall": 0.7000371827195703, + "per_class_f1": { + "FAVOR": 0.6113207547169811, + "AGAINST": 0.6587301587301587, + "NONE": 0.80083857442348 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 69.02964959568733, + "weighted_f1_pct": 71.98149062087408, + "macro_precision_pct": 68.74763794406653, + "macro_recall_pct": 70.00371827195703, + "per_class_f1_pct": { + "FAVOR": 61.13207547169811, + "AGAINST": 65.87301587301587, + "NONE": 80.083857442348 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.700139470013947, + "macro_f1": 0.6652238411372376, + "weighted_f1": 0.7120629153569743, + "macro_precision": 0.6550242400865232, + "macro_recall": 0.7073660605271298, + "per_class_f1": { + "FAVOR": 0.5555555555555556, + "AGAINST": 0.6633906633906634, + "NONE": 0.7767253044654939 + }, + "accuracy_pct": 70.01394700139471, + "macro_f1_pct": 66.52238411372376, + "weighted_f1_pct": 71.20629153569743, + "macro_precision_pct": 65.50242400865231, + "macro_recall_pct": 70.73660605271299, + "per_class_f1_pct": { + "FAVOR": 55.55555555555556, + "AGAINST": 66.33906633906635, + "NONE": 77.67253044654939 + } + } + ], + "accuracy": 0.6451360344490116, + "macro_f1": 0.6181564085981478, + "weighted_f1": 0.6602848629863943, + "macro_precision": 0.6224089406734592, + "macro_recall": 0.6547748394254188, + "per_class_f1": { + "FAVOR": 0.5461227489893421, + "AGAINST": 0.5669109139992287, + "NONE": 0.7414355628058727 + }, + "accuracy_pct": 64.51360344490115, + "macro_f1_pct": 61.81564085981478, + "weighted_f1_pct": 66.02848629863944, + "macro_precision_pct": 62.24089406734592, + "macro_recall_pct": 65.47748394254188, + "per_class_f1_pct": { + "FAVOR": 54.61227489893421, + "AGAINST": 56.69109139992287, + "NONE": 74.14355628058728 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s2__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9a4d03d31c421aca1f525c9a0099c43232043298 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__doc__s2__seed13/run.log b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..d1da36fffdb45b79da37b4f5d5ff58d7e4bc6398 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 07:00:10,981 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__doc__s2__seed13 device=cuda +2026-09-02 07:30:52,727 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6195991700918182, 'average_forgetting': 0.043980980311661186, 'backward_transfer': -0.04193359064114821, 'final_mean_target_macro_f1_pct': 61.95991700918182, 'average_forgetting_pct': 4.398098031166119, 'backward_transfer_pct': -4.193359064114821, 'final_pooled_accuracy': 0.6451360344490116, 'final_pooled_accuracy_pct': 64.51360344490115, 'final_pooled_macro_f1': 0.6181564085981478, 'final_pooled_macro_f1_pct': 61.81564085981478, 'final_pooled_weighted_f1': 0.6602848629863943, 'final_pooled_weighted_f1_pct': 66.02848629863944, 'final_pooled_macro_precision': 0.6224089406734592, 'final_pooled_macro_precision_pct': 62.24089406734592, 'final_pooled_macro_recall': 0.6547748394254188, 'final_pooled_macro_recall_pct': 65.47748394254188, 'final_pooled_per_class_f1': {'FAVOR': 0.5461227489893421, 'AGAINST': 0.5669109139992287, 'NONE': 0.7414355628058727}, 'final_pooled_per_class_f1_pct': {'FAVOR': 54.61227489893421, 'AGAINST': 56.69109139992287, 'NONE': 74.14355628058728}} +2026-09-02 07:30:52,727 [INFO] __main__: efficiency={'training_wall_time_seconds': 1840.3666741380002, 'training_wall_time_minutes': 30.672777902300005, 'peak_gpu_memory_allocated_gb': 10.40872573852539, 'peak_gpu_memory_reserved_gb': 10.599609375} diff --git a/outputs/trace_s7_contrastive_replay__doc__s3__seed13/environment.json b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..74a36fbb28d5468a6ea0d82f443cfdbd5b81d195 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T07:31:00.825115+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..49e3f82fd226557cc03fd287b2bb865114e66b43 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.json @@ -0,0 +1,1601 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5883676774876555, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.528524953767758, + 0.5588175909233875, + null, + null, + null, + null, + null, + null + ], + [ + 0.513256095565033, + 0.5353330448267157, + 0.6394976851417354, + null, + null, + null, + null, + null + ], + [ + 0.5125697088573516, + 0.5068290882244371, + 0.6487337714313561, + 0.6443902498462676, + null, + null, + null, + null + ], + [ + 0.5156410556447129, + 0.5358485631445332, + 0.6514127421395416, + 0.6331450299191429, + 0.7045765045211908, + null, + null, + null + ], + [ + 0.5532441623491312, + 0.5173209227858777, + 0.6253456374089399, + 0.6331164780524762, + 0.6802673699225424, + 0.6076814687822151, + null, + null + ], + [ + 0.5330419163531069, + 0.5165240370300673, + 0.663338994136316, + 0.633673287841786, + 0.6998845787905729, + 0.5879611996255361, + 0.6885147545999609, + null + ], + [ + 0.549202098628991, + 0.5387839274369678, + 0.6710022687460286, + 0.6473503065776921, + 0.6962568067103975, + 0.584769003669174, + 0.7089966579521941, + 0.6982517140031105 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6032002407312393, + "total": 0.6032002407312393, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6429479034307497, + "macro_f1": 0.5883676774876555, + "weighted_f1": 0.6649549537129172, + "macro_precision": 0.6053504056767706, + "macro_recall": 0.6250440131159085, + "per_class_f1": { + "FAVOR": 0.4790874524714829, + "AGAINST": 0.5197505197505198, + "NONE": 0.7662650602409639 + }, + "accuracy_pct": 64.29479034307496, + "macro_f1_pct": 58.83676774876555, + "weighted_f1_pct": 66.49549537129172, + "macro_precision_pct": 60.53504056767706, + "macro_recall_pct": 62.504401311590854, + "per_class_f1_pct": { + "FAVOR": 47.90874524714829, + "AGAINST": 51.975051975051976, + "NONE": 76.62650602409639 + } + } + ], + "accuracy": 0.6429479034307497, + "macro_f1": 0.5883676774876555, + "weighted_f1": 0.6649549537129172, + "macro_precision": 0.6053504056767706, + "macro_recall": 0.6250440131159085, + "per_class_f1": { + "FAVOR": 0.4790874524714829, + "AGAINST": 0.5197505197505198, + "NONE": 0.7662650602409639 + }, + "accuracy_pct": 64.29479034307496, + "macro_f1_pct": 58.83676774876555, + "weighted_f1_pct": 66.49549537129172, + "macro_precision_pct": 60.53504056767706, + "macro_recall_pct": 62.504401311590854, + "per_class_f1_pct": { + "FAVOR": 47.90874524714829, + "AGAINST": 51.975051975051976, + "NONE": 76.62650602409639 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6666338985827803, + "losses": { + "current_ce": 0.5967546376976229, + "replay_ce": 0.2935830479546908, + "dark_distillation": 0.4028230265551616, + "supervised_contrastive": 4.222137814310808, + "total": 1.5139629886263892, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.528524953767758, + "weighted_f1": 0.616053225016387, + "macro_precision": 0.5624812393928811, + "macro_recall": 0.567602312155689, + "per_class_f1": { + "FAVOR": 0.37551020408163266, + "AGAINST": 0.4807692307692308, + "NONE": 0.7292954264524104 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 52.8524953767758, + "weighted_f1_pct": 61.6053225016387, + "macro_precision_pct": 56.24812393928811, + "macro_recall_pct": 56.7602312155689, + "per_class_f1_pct": { + "FAVOR": 37.55102040816327, + "AGAINST": 48.07692307692308, + "NONE": 72.92954264524104 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5978112175102599, + "macro_f1": 0.5588175909233875, + "weighted_f1": 0.6091564731811205, + "macro_precision": 0.5721921613297215, + "macro_recall": 0.5809609598946727, + "per_class_f1": { + "FAVOR": 0.4247787610619469, + "AGAINST": 0.5739130434782609, + "NONE": 0.6777609682299546 + }, + "accuracy_pct": 59.78112175102599, + "macro_f1_pct": 55.88175909233875, + "weighted_f1_pct": 60.91564731811205, + "macro_precision_pct": 57.21921613297215, + "macro_recall_pct": 58.096095989467265, + "per_class_f1_pct": { + "FAVOR": 42.47787610619469, + "AGAINST": 57.391304347826086, + "NONE": 67.77609682299546 + } + } + ], + "accuracy": 0.5948616600790514, + "macro_f1": 0.5449845191253945, + "weighted_f1": 0.6116469911881953, + "macro_precision": 0.5677660847957403, + "macro_recall": 0.5707398009669529, + "per_class_f1": { + "FAVOR": 0.3991507430997877, + "AGAINST": 0.5296803652968036, + "NONE": 0.7061224489795919 + }, + "accuracy_pct": 59.48616600790514, + "macro_f1_pct": 54.498451912539444, + "weighted_f1_pct": 61.164699118819534, + "macro_precision_pct": 56.77660847957403, + "macro_recall_pct": 57.07398009669529, + "per_class_f1_pct": { + "FAVOR": 39.91507430997877, + "AGAINST": 52.96803652968036, + "NONE": 70.61224489795919 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.660344875538868, + "losses": { + "current_ce": 0.5721369360174452, + "replay_ce": 0.19499383781637464, + "dark_distillation": 0.3282293052900405, + "supervised_contrastive": 4.2383809969538735, + "total": 1.3550835366759981, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.513256095565033, + "weighted_f1": 0.607930495253754, + "macro_precision": 0.5235779329755557, + "macro_recall": 0.5431499086727846, + "per_class_f1": { + "FAVOR": 0.3620178041543027, + "AGAINST": 0.4460431654676259, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 51.3256095565033, + "weighted_f1_pct": 60.793049525375395, + "macro_precision_pct": 52.35779329755557, + "macro_recall_pct": 54.31499086727845, + "per_class_f1_pct": { + "FAVOR": 36.20178041543027, + "AGAINST": 44.60431654676259, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5704514363885089, + "macro_f1": 0.5353330448267157, + "weighted_f1": 0.58465691417604, + "macro_precision": 0.5584902880898328, + "macro_recall": 0.5653303766377085, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5572649572649573, + "NONE": 0.6487341772151899 + }, + "accuracy_pct": 57.045143638850895, + "macro_f1_pct": 53.533304482671575, + "weighted_f1_pct": 58.465691417604006, + "macro_precision_pct": 55.84902880898328, + "macro_recall_pct": 56.53303766377085, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 55.72649572649573, + "NONE": 64.87341772151899 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.676056338028169, + "macro_f1": 0.6394976851417354, + "weighted_f1": 0.6802333264619174, + "macro_precision": 0.6367948221606758, + "macro_recall": 0.6442469754923698, + "per_class_f1": { + "FAVOR": 0.5394190871369294, + "AGAINST": 0.5864661654135338, + "NONE": 0.7926078028747433 + }, + "accuracy_pct": 67.6056338028169, + "macro_f1_pct": 63.94976851417354, + "weighted_f1_pct": 68.02333264619173, + "macro_precision_pct": 63.67948221606758, + "macro_recall_pct": 64.42469754923698, + "per_class_f1_pct": { + "FAVOR": 53.941908713692946, + "AGAINST": 58.64661654135338, + "NONE": 79.26078028747433 + } + } + ], + "accuracy": 0.5990074441687345, + "macro_f1": 0.5573486705940088, + "weighted_f1": 0.6176175837652687, + "macro_precision": 0.5640059441059339, + "macro_recall": 0.5812657929865005, + "per_class_f1": { + "FAVOR": 0.425273390036452, + "AGAINST": 0.526813880126183, + "NONE": 0.7199587416193914 + }, + "accuracy_pct": 59.90074441687345, + "macro_f1_pct": 55.734867059400884, + "weighted_f1_pct": 61.76175837652686, + "macro_precision_pct": 56.400594410593385, + "macro_recall_pct": 58.12657929865005, + "per_class_f1_pct": { + "FAVOR": 42.527339003645196, + "AGAINST": 52.6813880126183, + "NONE": 71.99587416193914 + } + }, + { + "after_task": 4, + "best_epoch": 13, + "best_dev_macro_f1": 0.7373233109787574, + "losses": { + "current_ce": 0.4128856910195777, + "replay_ce": 0.22053628183477292, + "dark_distillation": 0.31871484690080815, + "supervised_contrastive": 4.087906561052896, + "total": 1.2015700653316528, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5125697088573516, + "weighted_f1": 0.6036551912503414, + "macro_precision": 0.5255784787745075, + "macro_recall": 0.5452709740836102, + "per_class_f1": { + "FAVOR": 0.375, + "AGAINST": 0.4392523364485981, + "NONE": 0.7234567901234568 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 51.256970885735164, + "weighted_f1_pct": 60.36551912503414, + "macro_precision_pct": 52.55784787745075, + "macro_recall_pct": 54.52709740836102, + "per_class_f1_pct": { + "FAVOR": 37.5, + "AGAINST": 43.925233644859816, + "NONE": 72.34567901234567 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5294117647058824, + "macro_f1": 0.5068290882244371, + "weighted_f1": 0.5434764495305965, + "macro_precision": 0.5432621472082794, + "macro_recall": 0.5448955856973935, + "per_class_f1": { + "FAVOR": 0.4108527131782946, + "AGAINST": 0.5149501661129569, + "NONE": 0.5946843853820598 + }, + "accuracy_pct": 52.94117647058824, + "macro_f1_pct": 50.68290882244371, + "weighted_f1_pct": 54.34764495305965, + "macro_precision_pct": 54.326214720827934, + "macro_recall_pct": 54.48955856973935, + "per_class_f1_pct": { + "FAVOR": 41.08527131782946, + "AGAINST": 51.49501661129568, + "NONE": 59.46843853820598 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6861167002012073, + "macro_f1": 0.6487337714313561, + "weighted_f1": 0.6895412484915769, + "macro_precision": 0.6461413478175021, + "macro_recall": 0.6525212825876053, + "per_class_f1": { + "FAVOR": 0.5537190082644629, + "AGAINST": 0.5900383141762452, + "NONE": 0.8024439918533605 + }, + "accuracy_pct": 68.61167002012073, + "macro_f1_pct": 64.87337714313561, + "weighted_f1_pct": 68.9541248491577, + "macro_precision_pct": 64.61413478175021, + "macro_recall_pct": 65.25212825876054, + "per_class_f1_pct": { + "FAVOR": 55.371900826446286, + "AGAINST": 59.00383141762452, + "NONE": 80.24439918533605 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.694560669456067, + "macro_f1": 0.6443902498462676, + "weighted_f1": 0.7056595506729789, + "macro_precision": 0.649821660766828, + "macro_recall": 0.6742248703519805, + "per_class_f1": { + "FAVOR": 0.5126050420168067, + "AGAINST": 0.6280623608017817, + "NONE": 0.7925033467202142 + }, + "accuracy_pct": 69.4560669456067, + "macro_f1_pct": 64.43902498462676, + "weighted_f1_pct": 70.56595506729789, + "macro_precision_pct": 64.9821660766828, + "macro_recall_pct": 67.42248703519806, + "per_class_f1_pct": { + "FAVOR": 51.26050420168067, + "AGAINST": 62.80623608017817, + "NONE": 79.25033467202141 + } + } + ], + "accuracy": 0.6134699853587116, + "macro_f1": 0.5731532199537419, + "weighted_f1": 0.6316234483112692, + "macro_precision": 0.5803521243055304, + "macro_recall": 0.6004305882253572, + "per_class_f1": { + "FAVOR": 0.4543761638733706, + "AGAINST": 0.5367816091954023, + "NONE": 0.7283018867924528 + }, + "accuracy_pct": 61.34699853587116, + "macro_f1_pct": 57.31532199537419, + "weighted_f1_pct": 63.162344831126916, + "macro_precision_pct": 58.03521243055304, + "macro_recall_pct": 60.04305882253572, + "per_class_f1_pct": { + "FAVOR": 45.43761638733706, + "AGAINST": 53.67816091954023, + "NONE": 72.83018867924528 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.7242757785506546, + "losses": { + "current_ce": 0.43770238122769767, + "replay_ce": 0.18766750875683058, + "dark_distillation": 0.3184248819521495, + "supervised_contrastive": 4.069203281175523, + "total": 1.1915026639756703, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5743329097839899, + "macro_f1": 0.5156410556447129, + "weighted_f1": 0.6046398781716698, + "macro_precision": 0.538175715392169, + "macro_recall": 0.5521060275962236, + "per_class_f1": { + "FAVOR": 0.3724137931034483, + "AGAINST": 0.4536082474226804, + "NONE": 0.72090112640801 + }, + "accuracy_pct": 57.433290978398986, + "macro_f1_pct": 51.564105564471284, + "weighted_f1_pct": 60.463987817166974, + "macro_precision_pct": 53.8175715392169, + "macro_recall_pct": 55.21060275962236, + "per_class_f1_pct": { + "FAVOR": 37.24137931034483, + "AGAINST": 45.36082474226804, + "NONE": 72.090112640801 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5622435020519836, + "macro_f1": 0.5358485631445332, + "weighted_f1": 0.5756934346943957, + "macro_precision": 0.5669728546440874, + "macro_recall": 0.5704528828064147, + "per_class_f1": { + "FAVOR": 0.43089430894308944, + "AGAINST": 0.5457570715474209, + "NONE": 0.6308943089430894 + }, + "accuracy_pct": 56.22435020519836, + "macro_f1_pct": 53.58485631445332, + "weighted_f1_pct": 57.56934346943957, + "macro_precision_pct": 56.69728546440874, + "macro_recall_pct": 57.04528828064147, + "per_class_f1_pct": { + "FAVOR": 43.08943089430895, + "AGAINST": 54.57570715474209, + "NONE": 63.08943089430894 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6514127421395416, + "weighted_f1": 0.6939568285563769, + "macro_precision": 0.6500317795943539, + "macro_recall": 0.655946239955083, + "per_class_f1": { + "FAVOR": 0.5493562231759657, + "AGAINST": 0.5934065934065934, + "NONE": 0.8114754098360656 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 65.14127421395416, + "weighted_f1_pct": 69.39568285563769, + "macro_precision_pct": 65.00317795943539, + "macro_recall_pct": 65.5946239955083, + "per_class_f1_pct": { + "FAVOR": 54.93562231759657, + "AGAINST": 59.34065934065934, + "NONE": 81.14754098360656 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6582984658298466, + "macro_f1": 0.6331450299191429, + "weighted_f1": 0.6712573046685724, + "macro_precision": 0.6400545696610743, + "macro_recall": 0.6884113120723243, + "per_class_f1": { + "FAVOR": 0.5519713261648745, + "AGAINST": 0.6220302375809935, + "NONE": 0.7254335260115607 + }, + "accuracy_pct": 65.82984658298466, + "macro_f1_pct": 63.31450299191429, + "weighted_f1_pct": 67.12573046685723, + "macro_precision_pct": 64.00545696610743, + "macro_recall_pct": 68.84113120723244, + "per_class_f1_pct": { + "FAVOR": 55.19713261648745, + "AGAINST": 62.20302375809935, + "NONE": 72.54335260115607 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.715927750410509, + "macro_f1": 0.7045765045211908, + "weighted_f1": 0.7224902686685917, + "macro_precision": 0.7062979276941438, + "macro_recall": 0.7279117850918219, + "per_class_f1": { + "FAVOR": 0.6332288401253918, + "AGAINST": 0.7068062827225131, + "NONE": 0.7736943907156673 + }, + "accuracy_pct": 71.5927750410509, + "macro_f1_pct": 70.45765045211908, + "weighted_f1_pct": 72.24902686685917, + "macro_precision_pct": 70.62979276941438, + "macro_recall_pct": 72.79117850918219, + "per_class_f1_pct": { + "FAVOR": 63.322884012539184, + "AGAINST": 70.68062827225131, + "NONE": 77.36943907156673 + } + } + ], + "accuracy": 0.6327446872193954, + "macro_f1": 0.604340723224825, + "weighted_f1": 0.6476976322312762, + "macro_precision": 0.614297102511561, + "macro_recall": 0.6354048906800545, + "per_class_f1": { + "FAVOR": 0.5106071689831748, + "AGAINST": 0.5753176043557169, + "NONE": 0.7270973963355835 + }, + "accuracy_pct": 63.274468721939535, + "macro_f1_pct": 60.4340723224825, + "weighted_f1_pct": 64.76976322312763, + "macro_precision_pct": 61.4297102511561, + "macro_recall_pct": 63.54048906800544, + "per_class_f1_pct": { + "FAVOR": 51.06071689831748, + "AGAINST": 57.53176043557169, + "NONE": 72.70973963355834 + } + }, + { + "after_task": 6, + "best_epoch": 7, + "best_dev_macro_f1": 0.6843380527929387, + "losses": { + "current_ce": 0.4687780435801598, + "replay_ce": 0.16411057466216256, + "dark_distillation": 0.30473792803914923, + "supervised_contrastive": 4.099669701174686, + "total": 1.1952245625487545, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6200762388818297, + "macro_f1": 0.5532441623491312, + "weighted_f1": 0.6446768595538065, + "macro_precision": 0.5620411493155046, + "macro_recall": 0.5851921539285373, + "per_class_f1": { + "FAVOR": 0.40559440559440557, + "AGAINST": 0.4900662251655629, + "NONE": 0.7640718562874251 + }, + "accuracy_pct": 62.00762388818297, + "macro_f1_pct": 55.32441623491312, + "weighted_f1_pct": 64.46768595538065, + "macro_precision_pct": 56.204114931550464, + "macro_recall_pct": 58.51921539285373, + "per_class_f1_pct": { + "FAVOR": 40.55944055944056, + "AGAINST": 49.00662251655629, + "NONE": 76.40718562874251 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5173209227858777, + "weighted_f1": 0.5695453628153941, + "macro_precision": 0.5407624633431085, + "macro_recall": 0.5437930245429409, + "per_class_f1": { + "FAVOR": 0.3952569169960474, + "AGAINST": 0.5026178010471204, + "NONE": 0.6540880503144654 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 51.73209227858777, + "weighted_f1_pct": 56.95453628153941, + "macro_precision_pct": 54.07624633431085, + "macro_recall_pct": 54.379302454294084, + "per_class_f1_pct": { + "FAVOR": 39.52569169960474, + "AGAINST": 50.26178010471204, + "NONE": 65.40880503144653 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6253456374089399, + "weighted_f1": 0.6793755866541941, + "macro_precision": 0.6417550507329486, + "macro_recall": 0.6276969434670244, + "per_class_f1": { + "FAVOR": 0.47715736040609136, + "AGAINST": 0.5714285714285714, + "NONE": 0.8274509803921568 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 62.53456374089399, + "weighted_f1_pct": 67.9375586654194, + "macro_precision_pct": 64.17550507329486, + "macro_recall_pct": 62.769694346702444, + "per_class_f1_pct": { + "FAVOR": 47.71573604060914, + "AGAINST": 57.14285714285714, + "NONE": 82.74509803921568 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6750348675034867, + "macro_f1": 0.6331164780524762, + "weighted_f1": 0.6886882822836864, + "macro_precision": 0.6322762931007907, + "macro_recall": 0.6698621941491526, + "per_class_f1": { + "FAVOR": 0.5151515151515151, + "AGAINST": 0.6164383561643836, + "NONE": 0.76775956284153 + }, + "accuracy_pct": 67.50348675034867, + "macro_f1_pct": 63.31164780524762, + "weighted_f1_pct": 68.86882822836864, + "macro_precision_pct": 63.22762931007907, + "macro_recall_pct": 66.98621941491525, + "per_class_f1_pct": { + "FAVOR": 51.515151515151516, + "AGAINST": 61.64383561643836, + "NONE": 76.775956284153 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6912972085385879, + "macro_f1": 0.6802673699225424, + "weighted_f1": 0.6988269863248631, + "macro_precision": 0.6901823321159202, + "macro_recall": 0.704663133015858, + "per_class_f1": { + "FAVOR": 0.6168831168831169, + "AGAINST": 0.6699507389162561, + "NONE": 0.753968253968254 + }, + "accuracy_pct": 69.12972085385879, + "macro_f1_pct": 68.02673699225424, + "weighted_f1_pct": 69.88269863248631, + "macro_precision_pct": 69.01823321159202, + "macro_recall_pct": 70.4663133015858, + "per_class_f1_pct": { + "FAVOR": 61.68831168831169, + "AGAINST": 66.99507389162561, + "NONE": 75.39682539682539 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6442451420029895, + "macro_f1": 0.6076814687822151, + "weighted_f1": 0.6613195410456303, + "macro_precision": 0.6142845005334477, + "macro_recall": 0.6402938899052601, + "per_class_f1": { + "FAVOR": 0.5870646766169154, + "AGAINST": 0.4797297297297297, + "NONE": 0.75625 + }, + "accuracy_pct": 64.42451420029896, + "macro_f1_pct": 60.76814687822151, + "weighted_f1_pct": 66.13195410456302, + "macro_precision_pct": 61.428450053344775, + "macro_recall_pct": 64.02938899052602, + "per_class_f1_pct": { + "FAVOR": 58.70646766169154, + "AGAINST": 47.97297297297297, + "NONE": 75.625 + } + } + ], + "accuracy": 0.6399002493765586, + "macro_f1": 0.6053158615300013, + "weighted_f1": 0.6559717432239228, + "macro_precision": 0.6097155178596244, + "macro_recall": 0.6303435425835722, + "per_class_f1": { + "FAVOR": 0.5099415204678363, + "AGAINST": 0.553607827150428, + "NONE": 0.7523982369717397 + }, + "accuracy_pct": 63.99002493765587, + "macro_f1_pct": 60.53158615300013, + "weighted_f1_pct": 65.59717432239228, + "macro_precision_pct": 60.97155178596244, + "macro_recall_pct": 63.03435425835722, + "per_class_f1_pct": { + "FAVOR": 50.994152046783626, + "AGAINST": 55.3607827150428, + "NONE": 75.23982369717397 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.732392809474041, + "losses": { + "current_ce": 0.39321027770638467, + "replay_ce": 0.16690240779093335, + "dark_distillation": 0.2781970035816942, + "supervised_contrastive": 4.059705282392956, + "total": 1.1051817282324745, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5330419163531069, + "weighted_f1": 0.6284987435476702, + "macro_precision": 0.5495274557220038, + "macro_recall": 0.5659028557176705, + "per_class_f1": { + "FAVOR": 0.3709090909090909, + "AGAINST": 0.47578947368421054, + "NONE": 0.7524271844660194 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.304191635310694, + "weighted_f1_pct": 62.849874354767024, + "macro_precision_pct": 54.95274557220038, + "macro_recall_pct": 56.590285571767055, + "per_class_f1_pct": { + "FAVOR": 37.09090909090909, + "AGAINST": 47.578947368421055, + "NONE": 75.24271844660194 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5165240370300673, + "weighted_f1": 0.5602231641199779, + "macro_precision": 0.5612544272088135, + "macro_recall": 0.5472769768886274, + "per_class_f1": { + "FAVOR": 0.3893805309734513, + "AGAINST": 0.5488958990536278, + "NONE": 0.6112956810631229 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 51.65240370300673, + "weighted_f1_pct": 56.022316411997785, + "macro_precision_pct": 56.125442720881345, + "macro_recall_pct": 54.72769768886274, + "per_class_f1_pct": { + "FAVOR": 38.93805309734513, + "AGAINST": 54.889589905362776, + "NONE": 61.12956810631229 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.704225352112676, + "macro_f1": 0.663338994136316, + "weighted_f1": 0.7055967401050682, + "macro_precision": 0.6717680150599771, + "macro_recall": 0.6657794564942686, + "per_class_f1": { + "FAVOR": 0.5660377358490566, + "AGAINST": 0.6013986013986014, + "NONE": 0.8225806451612904 + }, + "accuracy_pct": 70.4225352112676, + "macro_f1_pct": 66.3338994136316, + "weighted_f1_pct": 70.55967401050682, + "macro_precision_pct": 67.1768015059977, + "macro_recall_pct": 66.57794564942687, + "per_class_f1_pct": { + "FAVOR": 56.60377358490566, + "AGAINST": 60.13986013986013, + "NONE": 82.25806451612904 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6722454672245467, + "macro_f1": 0.633673287841786, + "weighted_f1": 0.6876244661838745, + "macro_precision": 0.6319979312972944, + "macro_recall": 0.6745307890305203, + "per_class_f1": { + "FAVOR": 0.5232974910394266, + "AGAINST": 0.6125290023201856, + "NONE": 0.7651933701657458 + }, + "accuracy_pct": 67.22454672245468, + "macro_f1_pct": 63.3673287841786, + "weighted_f1_pct": 68.76244661838746, + "macro_precision_pct": 63.19979312972944, + "macro_recall_pct": 67.45307890305203, + "per_class_f1_pct": { + "FAVOR": 52.32974910394266, + "AGAINST": 61.25290023201856, + "NONE": 76.51933701657458 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6998845787905729, + "weighted_f1": 0.7124202118811314, + "macro_precision": 0.7130264913283781, + "macro_recall": 0.7290714745552468, + "per_class_f1": { + "FAVOR": 0.6511627906976745, + "AGAINST": 0.7, + "NONE": 0.7484909456740443 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.9884578790573, + "weighted_f1_pct": 71.24202118811314, + "macro_precision_pct": 71.30264913283781, + "macro_recall_pct": 72.90714745552468, + "per_class_f1_pct": { + "FAVOR": 65.11627906976744, + "AGAINST": 70.0, + "NONE": 74.84909456740442 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6233183856502242, + "macro_f1": 0.5879611996255361, + "weighted_f1": 0.6433908229507698, + "macro_precision": 0.5988484265186446, + "macro_recall": 0.6214122977391909, + "per_class_f1": { + "FAVOR": 0.5728643216080402, + "AGAINST": 0.45016077170418006, + "NONE": 0.7408585055643879 + }, + "accuracy_pct": 62.33183856502242, + "macro_f1_pct": 58.796119962553604, + "weighted_f1_pct": 64.33908229507698, + "macro_precision_pct": 59.88484265186445, + "macro_recall_pct": 62.14122977391909, + "per_class_f1_pct": { + "FAVOR": 57.286432160804026, + "AGAINST": 45.01607717041801, + "NONE": 74.0858505564388 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6885147545999609, + "weighted_f1": 0.7227362054459493, + "macro_precision": 0.6836179541611281, + "macro_recall": 0.7130389009699355, + "per_class_f1": { + "FAVOR": 0.5977859778597786, + "AGAINST": 0.6694214876033058, + "NONE": 0.7983367983367984 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 68.8514754599961, + "weighted_f1_pct": 72.27362054459493, + "macro_precision_pct": 68.36179541611281, + "macro_recall_pct": 71.30389009699356, + "per_class_f1_pct": { + "FAVOR": 59.77859778597786, + "AGAINST": 66.94214876033058, + "NONE": 79.83367983367984 + } + } + ], + "accuracy": 0.6456623030840914, + "macro_f1": 0.6168978492921153, + "weighted_f1": 0.6609638311161464, + "macro_precision": 0.6234776305297557, + "macro_recall": 0.6470145911415978, + "per_class_f1": { + "FAVOR": 0.5310907237512742, + "AGAINST": 0.5737763486959628, + "NONE": 0.7458264754291088 + }, + "accuracy_pct": 64.56623030840915, + "macro_f1_pct": 61.689784929211534, + "weighted_f1_pct": 66.09638311161464, + "macro_precision_pct": 62.347763052975566, + "macro_recall_pct": 64.70145911415977, + "per_class_f1_pct": { + "FAVOR": 53.10907237512742, + "AGAINST": 57.377634869596285, + "NONE": 74.58264754291089 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.7643056682921495, + "losses": { + "current_ce": 0.32467017376831936, + "replay_ce": 0.1549365407988137, + "dark_distillation": 0.27835183240035, + "supervised_contrastive": 4.00868492827696, + "total": 1.0196511304845997, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6226175349428208, + "macro_f1": 0.549202098628991, + "weighted_f1": 0.6453645950592148, + "macro_precision": 0.5594283781498394, + "macro_recall": 0.5777383493614431, + "per_class_f1": { + "FAVOR": 0.39114391143911437, + "AGAINST": 0.48577680525164113, + "NONE": 0.7706855791962175 + }, + "accuracy_pct": 62.26175349428208, + "macro_f1_pct": 54.92020986289911, + "weighted_f1_pct": 64.53645950592148, + "macro_precision_pct": 55.94283781498394, + "macro_recall_pct": 57.77383493614431, + "per_class_f1_pct": { + "FAVOR": 39.11439114391143, + "AGAINST": 48.57768052516411, + "NONE": 77.06855791962175 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5677154582763337, + "macro_f1": 0.5387839274369678, + "weighted_f1": 0.5835819066645362, + "macro_precision": 0.5594784363345936, + "macro_recall": 0.5732535994923005, + "per_class_f1": { + "FAVOR": 0.4247104247104247, + "AGAINST": 0.542907180385289, + "NONE": 0.6487341772151899 + }, + "accuracy_pct": 56.771545827633375, + "macro_f1_pct": 53.87839274369678, + "weighted_f1_pct": 58.35819066645362, + "macro_precision_pct": 55.94784363345936, + "macro_recall_pct": 57.32535994923005, + "per_class_f1_pct": { + "FAVOR": 42.471042471042466, + "AGAINST": 54.290718038528894, + "NONE": 64.87341772151899 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6710022687460286, + "weighted_f1": 0.7151002330394286, + "macro_precision": 0.6787058789283483, + "macro_recall": 0.6706594669011766, + "per_class_f1": { + "FAVOR": 0.5497630331753555, + "AGAINST": 0.6273062730627307, + "NONE": 0.8359375 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 67.10022687460287, + "weighted_f1_pct": 71.51002330394286, + "macro_precision_pct": 67.87058789283483, + "macro_recall_pct": 67.06594669011766, + "per_class_f1_pct": { + "FAVOR": 54.976303317535546, + "AGAINST": 62.73062730627307, + "NONE": 83.59375 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6875871687587168, + "macro_f1": 0.6473503065776921, + "weighted_f1": 0.7008597546288428, + "macro_precision": 0.6467662259392671, + "macro_recall": 0.6889718991950096, + "per_class_f1": { + "FAVOR": 0.5243445692883895, + "AGAINST": 0.6425339366515838, + "NONE": 0.7751724137931034 + }, + "accuracy_pct": 68.75871687587168, + "macro_f1_pct": 64.73503065776922, + "weighted_f1_pct": 70.08597546288428, + "macro_precision_pct": 64.67662259392671, + "macro_recall_pct": 68.89718991950096, + "per_class_f1_pct": { + "FAVOR": 52.43445692883895, + "AGAINST": 64.25339366515837, + "NONE": 77.51724137931035 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6962568067103975, + "weighted_f1": 0.7121906301119916, + "macro_precision": 0.7021514161220043, + "macro_recall": 0.7194313222329143, + "per_class_f1": { + "FAVOR": 0.6305084745762712, + "AGAINST": 0.7009803921568627, + "NONE": 0.7572815533980582 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.62568067103975, + "weighted_f1_pct": 71.21906301119915, + "macro_precision_pct": 70.21514161220043, + "macro_recall_pct": 71.94313222329143, + "per_class_f1_pct": { + "FAVOR": 63.05084745762712, + "AGAINST": 70.09803921568627, + "NONE": 75.72815533980582 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6218236173393124, + "macro_f1": 0.584769003669174, + "weighted_f1": 0.6418128696142593, + "macro_precision": 0.5973270336246949, + "macro_recall": 0.6181129100294102, + "per_class_f1": { + "FAVOR": 0.5700245700245701, + "AGAINST": 0.44224422442244227, + "NONE": 0.7420382165605095 + }, + "accuracy_pct": 62.18236173393124, + "macro_f1_pct": 58.47690036691739, + "weighted_f1_pct": 64.18128696142593, + "macro_precision_pct": 59.732703362469486, + "macro_recall_pct": 61.81129100294102, + "per_class_f1_pct": { + "FAVOR": 57.00245700245701, + "AGAINST": 44.224422442244226, + "NONE": 74.20382165605095 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7323943661971831, + "macro_f1": 0.7089966579521941, + "weighted_f1": 0.7410066973101214, + "macro_precision": 0.7022231930488813, + "macro_recall": 0.731940561250906, + "per_class_f1": { + "FAVOR": 0.6319702602230484, + "AGAINST": 0.6835443037974683, + "NONE": 0.8114754098360656 + }, + "accuracy_pct": 73.23943661971832, + "macro_f1_pct": 70.89966579521942, + "weighted_f1_pct": 74.10066973101213, + "macro_precision_pct": 70.22231930488813, + "macro_recall_pct": 73.1940561250906, + "per_class_f1_pct": { + "FAVOR": 63.19702602230484, + "AGAINST": 68.35443037974683, + "NONE": 81.14754098360656 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7308970099667774, + "macro_f1": 0.6982517140031105, + "weighted_f1": 0.7380301169718292, + "macro_precision": 0.6848284896230101, + "macro_recall": 0.7294661495258437, + "per_class_f1": { + "FAVOR": 0.6615969581749049, + "AGAINST": 0.6225165562913907, + "NONE": 0.810641627543036 + }, + "accuracy_pct": 73.08970099667775, + "macro_f1_pct": 69.82517140031105, + "weighted_f1_pct": 73.80301169718292, + "macro_precision_pct": 68.482848962301, + "macro_recall_pct": 72.94661495258437, + "per_class_f1_pct": { + "FAVOR": 66.15969581749049, + "AGAINST": 62.251655629139066, + "NONE": 81.06416275430361 + } + } + ], + "accuracy": 0.6666666666666666, + "macro_f1": 0.634732048260949, + "weighted_f1": 0.6801802466001565, + "macro_precision": 0.6332046070397764, + "macro_recall": 0.661880342913349, + "per_class_f1": { + "FAVOR": 0.5504014272970562, + "AGAINST": 0.587094617184888, + "NONE": 0.7667001003009027 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 63.4732048260949, + "weighted_f1_pct": 68.01802466001566, + "macro_precision_pct": 63.320460703977645, + "macro_recall_pct": 66.1880342913349, + "per_class_f1_pct": { + "FAVOR": 55.04014272970562, + "AGAINST": 58.7094617184888, + "NONE": 76.67001003009027 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6368265979655694, + "average_forgetting": 0.00847516722507832, + "backward_transfer": -0.0050692659401382345, + "final_mean_target_macro_f1_pct": 63.68265979655694, + "average_forgetting_pct": 0.8475167225078319, + "backward_transfer_pct": -0.5069265940138235, + "final_pooled_accuracy": 0.6666666666666666, + "final_pooled_accuracy_pct": 66.66666666666666, + "final_pooled_macro_f1": 0.634732048260949, + "final_pooled_macro_f1_pct": 63.4732048260949, + "final_pooled_weighted_f1": 0.6801802466001565, + "final_pooled_weighted_f1_pct": 68.01802466001566, + "final_pooled_macro_precision": 0.6332046070397764, + "final_pooled_macro_precision_pct": 63.320460703977645, + "final_pooled_macro_recall": 0.661880342913349, + "final_pooled_macro_recall_pct": 66.1880342913349, + "final_pooled_per_class_f1": { + "FAVOR": 0.5504014272970562, + "AGAINST": 0.587094617184888, + "NONE": 0.7667001003009027 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 55.04014272970562, + "AGAINST": 58.7094617184888, + "NONE": 76.67001003009027 + } + }, + "efficiency": { + "training_wall_time_seconds": 1813.4751372640021, + "training_wall_time_minutes": 30.2245856210667, + "peak_gpu_memory_allocated_gb": 10.40872573852539, + "peak_gpu_memory_reserved_gb": 10.599609375 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..e4d6b39d353689bb1c1beed95b06d1cee7e437ae --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/metrics.partial.json @@ -0,0 +1,1495 @@ +{ + "matrix": [ + [ + 0.5883676774876555, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.528524953767758, + 0.5588175909233875, + null, + null, + null, + null, + null, + null + ], + [ + 0.513256095565033, + 0.5353330448267157, + 0.6394976851417354, + null, + null, + null, + null, + null + ], + [ + 0.5125697088573516, + 0.5068290882244371, + 0.6487337714313561, + 0.6443902498462676, + null, + null, + null, + null + ], + [ + 0.5156410556447129, + 0.5358485631445332, + 0.6514127421395416, + 0.6331450299191429, + 0.7045765045211908, + null, + null, + null + ], + [ + 0.5532441623491312, + 0.5173209227858777, + 0.6253456374089399, + 0.6331164780524762, + 0.6802673699225424, + 0.6076814687822151, + null, + null + ], + [ + 0.5330419163531069, + 0.5165240370300673, + 0.663338994136316, + 0.633673287841786, + 0.6998845787905729, + 0.5879611996255361, + 0.6885147545999609, + null + ], + [ + 0.549202098628991, + 0.5387839274369678, + 0.6710022687460286, + 0.6473503065776921, + 0.6962568067103975, + 0.584769003669174, + 0.7089966579521941, + 0.6982517140031105 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6032002407312393, + "total": 0.6032002407312393, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6429479034307497, + "macro_f1": 0.5883676774876555, + "weighted_f1": 0.6649549537129172, + "macro_precision": 0.6053504056767706, + "macro_recall": 0.6250440131159085, + "per_class_f1": { + "FAVOR": 0.4790874524714829, + "AGAINST": 0.5197505197505198, + "NONE": 0.7662650602409639 + }, + "accuracy_pct": 64.29479034307496, + "macro_f1_pct": 58.83676774876555, + "weighted_f1_pct": 66.49549537129172, + "macro_precision_pct": 60.53504056767706, + "macro_recall_pct": 62.504401311590854, + "per_class_f1_pct": { + "FAVOR": 47.90874524714829, + "AGAINST": 51.975051975051976, + "NONE": 76.62650602409639 + } + } + ], + "accuracy": 0.6429479034307497, + "macro_f1": 0.5883676774876555, + "weighted_f1": 0.6649549537129172, + "macro_precision": 0.6053504056767706, + "macro_recall": 0.6250440131159085, + "per_class_f1": { + "FAVOR": 0.4790874524714829, + "AGAINST": 0.5197505197505198, + "NONE": 0.7662650602409639 + }, + "accuracy_pct": 64.29479034307496, + "macro_f1_pct": 58.83676774876555, + "weighted_f1_pct": 66.49549537129172, + "macro_precision_pct": 60.53504056767706, + "macro_recall_pct": 62.504401311590854, + "per_class_f1_pct": { + "FAVOR": 47.90874524714829, + "AGAINST": 51.975051975051976, + "NONE": 76.62650602409639 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6666338985827803, + "losses": { + "current_ce": 0.5967546376976229, + "replay_ce": 0.2935830479546908, + "dark_distillation": 0.4028230265551616, + "supervised_contrastive": 4.222137814310808, + "total": 1.5139629886263892, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.528524953767758, + "weighted_f1": 0.616053225016387, + "macro_precision": 0.5624812393928811, + "macro_recall": 0.567602312155689, + "per_class_f1": { + "FAVOR": 0.37551020408163266, + "AGAINST": 0.4807692307692308, + "NONE": 0.7292954264524104 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 52.8524953767758, + "weighted_f1_pct": 61.6053225016387, + "macro_precision_pct": 56.24812393928811, + "macro_recall_pct": 56.7602312155689, + "per_class_f1_pct": { + "FAVOR": 37.55102040816327, + "AGAINST": 48.07692307692308, + "NONE": 72.92954264524104 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5978112175102599, + "macro_f1": 0.5588175909233875, + "weighted_f1": 0.6091564731811205, + "macro_precision": 0.5721921613297215, + "macro_recall": 0.5809609598946727, + "per_class_f1": { + "FAVOR": 0.4247787610619469, + "AGAINST": 0.5739130434782609, + "NONE": 0.6777609682299546 + }, + "accuracy_pct": 59.78112175102599, + "macro_f1_pct": 55.88175909233875, + "weighted_f1_pct": 60.91564731811205, + "macro_precision_pct": 57.21921613297215, + "macro_recall_pct": 58.096095989467265, + "per_class_f1_pct": { + "FAVOR": 42.47787610619469, + "AGAINST": 57.391304347826086, + "NONE": 67.77609682299546 + } + } + ], + "accuracy": 0.5948616600790514, + "macro_f1": 0.5449845191253945, + "weighted_f1": 0.6116469911881953, + "macro_precision": 0.5677660847957403, + "macro_recall": 0.5707398009669529, + "per_class_f1": { + "FAVOR": 0.3991507430997877, + "AGAINST": 0.5296803652968036, + "NONE": 0.7061224489795919 + }, + "accuracy_pct": 59.48616600790514, + "macro_f1_pct": 54.498451912539444, + "weighted_f1_pct": 61.164699118819534, + "macro_precision_pct": 56.77660847957403, + "macro_recall_pct": 57.07398009669529, + "per_class_f1_pct": { + "FAVOR": 39.91507430997877, + "AGAINST": 52.96803652968036, + "NONE": 70.61224489795919 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.660344875538868, + "losses": { + "current_ce": 0.5721369360174452, + "replay_ce": 0.19499383781637464, + "dark_distillation": 0.3282293052900405, + "supervised_contrastive": 4.2383809969538735, + "total": 1.3550835366759981, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.576874205844981, + "macro_f1": 0.513256095565033, + "weighted_f1": 0.607930495253754, + "macro_precision": 0.5235779329755557, + "macro_recall": 0.5431499086727846, + "per_class_f1": { + "FAVOR": 0.3620178041543027, + "AGAINST": 0.4460431654676259, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 57.68742058449809, + "macro_f1_pct": 51.3256095565033, + "weighted_f1_pct": 60.793049525375395, + "macro_precision_pct": 52.35779329755557, + "macro_recall_pct": 54.31499086727845, + "per_class_f1_pct": { + "FAVOR": 36.20178041543027, + "AGAINST": 44.60431654676259, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5704514363885089, + "macro_f1": 0.5353330448267157, + "weighted_f1": 0.58465691417604, + "macro_precision": 0.5584902880898328, + "macro_recall": 0.5653303766377085, + "per_class_f1": { + "FAVOR": 0.4, + "AGAINST": 0.5572649572649573, + "NONE": 0.6487341772151899 + }, + "accuracy_pct": 57.045143638850895, + "macro_f1_pct": 53.533304482671575, + "weighted_f1_pct": 58.465691417604006, + "macro_precision_pct": 55.84902880898328, + "macro_recall_pct": 56.53303766377085, + "per_class_f1_pct": { + "FAVOR": 40.0, + "AGAINST": 55.72649572649573, + "NONE": 64.87341772151899 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.676056338028169, + "macro_f1": 0.6394976851417354, + "weighted_f1": 0.6802333264619174, + "macro_precision": 0.6367948221606758, + "macro_recall": 0.6442469754923698, + "per_class_f1": { + "FAVOR": 0.5394190871369294, + "AGAINST": 0.5864661654135338, + "NONE": 0.7926078028747433 + }, + "accuracy_pct": 67.6056338028169, + "macro_f1_pct": 63.94976851417354, + "weighted_f1_pct": 68.02333264619173, + "macro_precision_pct": 63.67948221606758, + "macro_recall_pct": 64.42469754923698, + "per_class_f1_pct": { + "FAVOR": 53.941908713692946, + "AGAINST": 58.64661654135338, + "NONE": 79.26078028747433 + } + } + ], + "accuracy": 0.5990074441687345, + "macro_f1": 0.5573486705940088, + "weighted_f1": 0.6176175837652687, + "macro_precision": 0.5640059441059339, + "macro_recall": 0.5812657929865005, + "per_class_f1": { + "FAVOR": 0.425273390036452, + "AGAINST": 0.526813880126183, + "NONE": 0.7199587416193914 + }, + "accuracy_pct": 59.90074441687345, + "macro_f1_pct": 55.734867059400884, + "weighted_f1_pct": 61.76175837652686, + "macro_precision_pct": 56.400594410593385, + "macro_recall_pct": 58.12657929865005, + "per_class_f1_pct": { + "FAVOR": 42.527339003645196, + "AGAINST": 52.6813880126183, + "NONE": 71.99587416193914 + } + }, + { + "after_task": 4, + "best_epoch": 13, + "best_dev_macro_f1": 0.7373233109787574, + "losses": { + "current_ce": 0.4128856910195777, + "replay_ce": 0.22053628183477292, + "dark_distillation": 0.31871484690080815, + "supervised_contrastive": 4.087906561052896, + "total": 1.2015700653316528, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5125697088573516, + "weighted_f1": 0.6036551912503414, + "macro_precision": 0.5255784787745075, + "macro_recall": 0.5452709740836102, + "per_class_f1": { + "FAVOR": 0.375, + "AGAINST": 0.4392523364485981, + "NONE": 0.7234567901234568 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 51.256970885735164, + "weighted_f1_pct": 60.36551912503414, + "macro_precision_pct": 52.55784787745075, + "macro_recall_pct": 54.52709740836102, + "per_class_f1_pct": { + "FAVOR": 37.5, + "AGAINST": 43.925233644859816, + "NONE": 72.34567901234567 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5294117647058824, + "macro_f1": 0.5068290882244371, + "weighted_f1": 0.5434764495305965, + "macro_precision": 0.5432621472082794, + "macro_recall": 0.5448955856973935, + "per_class_f1": { + "FAVOR": 0.4108527131782946, + "AGAINST": 0.5149501661129569, + "NONE": 0.5946843853820598 + }, + "accuracy_pct": 52.94117647058824, + "macro_f1_pct": 50.68290882244371, + "weighted_f1_pct": 54.34764495305965, + "macro_precision_pct": 54.326214720827934, + "macro_recall_pct": 54.48955856973935, + "per_class_f1_pct": { + "FAVOR": 41.08527131782946, + "AGAINST": 51.49501661129568, + "NONE": 59.46843853820598 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6861167002012073, + "macro_f1": 0.6487337714313561, + "weighted_f1": 0.6895412484915769, + "macro_precision": 0.6461413478175021, + "macro_recall": 0.6525212825876053, + "per_class_f1": { + "FAVOR": 0.5537190082644629, + "AGAINST": 0.5900383141762452, + "NONE": 0.8024439918533605 + }, + "accuracy_pct": 68.61167002012073, + "macro_f1_pct": 64.87337714313561, + "weighted_f1_pct": 68.9541248491577, + "macro_precision_pct": 64.61413478175021, + "macro_recall_pct": 65.25212825876054, + "per_class_f1_pct": { + "FAVOR": 55.371900826446286, + "AGAINST": 59.00383141762452, + "NONE": 80.24439918533605 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.694560669456067, + "macro_f1": 0.6443902498462676, + "weighted_f1": 0.7056595506729789, + "macro_precision": 0.649821660766828, + "macro_recall": 0.6742248703519805, + "per_class_f1": { + "FAVOR": 0.5126050420168067, + "AGAINST": 0.6280623608017817, + "NONE": 0.7925033467202142 + }, + "accuracy_pct": 69.4560669456067, + "macro_f1_pct": 64.43902498462676, + "weighted_f1_pct": 70.56595506729789, + "macro_precision_pct": 64.9821660766828, + "macro_recall_pct": 67.42248703519806, + "per_class_f1_pct": { + "FAVOR": 51.26050420168067, + "AGAINST": 62.80623608017817, + "NONE": 79.25033467202141 + } + } + ], + "accuracy": 0.6134699853587116, + "macro_f1": 0.5731532199537419, + "weighted_f1": 0.6316234483112692, + "macro_precision": 0.5803521243055304, + "macro_recall": 0.6004305882253572, + "per_class_f1": { + "FAVOR": 0.4543761638733706, + "AGAINST": 0.5367816091954023, + "NONE": 0.7283018867924528 + }, + "accuracy_pct": 61.34699853587116, + "macro_f1_pct": 57.31532199537419, + "weighted_f1_pct": 63.162344831126916, + "macro_precision_pct": 58.03521243055304, + "macro_recall_pct": 60.04305882253572, + "per_class_f1_pct": { + "FAVOR": 45.43761638733706, + "AGAINST": 53.67816091954023, + "NONE": 72.83018867924528 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.7242757785506546, + "losses": { + "current_ce": 0.43770238122769767, + "replay_ce": 0.18766750875683058, + "dark_distillation": 0.3184248819521495, + "supervised_contrastive": 4.069203281175523, + "total": 1.1915026639756703, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5743329097839899, + "macro_f1": 0.5156410556447129, + "weighted_f1": 0.6046398781716698, + "macro_precision": 0.538175715392169, + "macro_recall": 0.5521060275962236, + "per_class_f1": { + "FAVOR": 0.3724137931034483, + "AGAINST": 0.4536082474226804, + "NONE": 0.72090112640801 + }, + "accuracy_pct": 57.433290978398986, + "macro_f1_pct": 51.564105564471284, + "weighted_f1_pct": 60.463987817166974, + "macro_precision_pct": 53.8175715392169, + "macro_recall_pct": 55.21060275962236, + "per_class_f1_pct": { + "FAVOR": 37.24137931034483, + "AGAINST": 45.36082474226804, + "NONE": 72.090112640801 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5622435020519836, + "macro_f1": 0.5358485631445332, + "weighted_f1": 0.5756934346943957, + "macro_precision": 0.5669728546440874, + "macro_recall": 0.5704528828064147, + "per_class_f1": { + "FAVOR": 0.43089430894308944, + "AGAINST": 0.5457570715474209, + "NONE": 0.6308943089430894 + }, + "accuracy_pct": 56.22435020519836, + "macro_f1_pct": 53.58485631445332, + "weighted_f1_pct": 57.56934346943957, + "macro_precision_pct": 56.69728546440874, + "macro_recall_pct": 57.04528828064147, + "per_class_f1_pct": { + "FAVOR": 43.08943089430895, + "AGAINST": 54.57570715474209, + "NONE": 63.08943089430894 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6514127421395416, + "weighted_f1": 0.6939568285563769, + "macro_precision": 0.6500317795943539, + "macro_recall": 0.655946239955083, + "per_class_f1": { + "FAVOR": 0.5493562231759657, + "AGAINST": 0.5934065934065934, + "NONE": 0.8114754098360656 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 65.14127421395416, + "weighted_f1_pct": 69.39568285563769, + "macro_precision_pct": 65.00317795943539, + "macro_recall_pct": 65.5946239955083, + "per_class_f1_pct": { + "FAVOR": 54.93562231759657, + "AGAINST": 59.34065934065934, + "NONE": 81.14754098360656 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6582984658298466, + "macro_f1": 0.6331450299191429, + "weighted_f1": 0.6712573046685724, + "macro_precision": 0.6400545696610743, + "macro_recall": 0.6884113120723243, + "per_class_f1": { + "FAVOR": 0.5519713261648745, + "AGAINST": 0.6220302375809935, + "NONE": 0.7254335260115607 + }, + "accuracy_pct": 65.82984658298466, + "macro_f1_pct": 63.31450299191429, + "weighted_f1_pct": 67.12573046685723, + "macro_precision_pct": 64.00545696610743, + "macro_recall_pct": 68.84113120723244, + "per_class_f1_pct": { + "FAVOR": 55.19713261648745, + "AGAINST": 62.20302375809935, + "NONE": 72.54335260115607 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.715927750410509, + "macro_f1": 0.7045765045211908, + "weighted_f1": 0.7224902686685917, + "macro_precision": 0.7062979276941438, + "macro_recall": 0.7279117850918219, + "per_class_f1": { + "FAVOR": 0.6332288401253918, + "AGAINST": 0.7068062827225131, + "NONE": 0.7736943907156673 + }, + "accuracy_pct": 71.5927750410509, + "macro_f1_pct": 70.45765045211908, + "weighted_f1_pct": 72.24902686685917, + "macro_precision_pct": 70.62979276941438, + "macro_recall_pct": 72.79117850918219, + "per_class_f1_pct": { + "FAVOR": 63.322884012539184, + "AGAINST": 70.68062827225131, + "NONE": 77.36943907156673 + } + } + ], + "accuracy": 0.6327446872193954, + "macro_f1": 0.604340723224825, + "weighted_f1": 0.6476976322312762, + "macro_precision": 0.614297102511561, + "macro_recall": 0.6354048906800545, + "per_class_f1": { + "FAVOR": 0.5106071689831748, + "AGAINST": 0.5753176043557169, + "NONE": 0.7270973963355835 + }, + "accuracy_pct": 63.274468721939535, + "macro_f1_pct": 60.4340723224825, + "weighted_f1_pct": 64.76976322312763, + "macro_precision_pct": 61.4297102511561, + "macro_recall_pct": 63.54048906800544, + "per_class_f1_pct": { + "FAVOR": 51.06071689831748, + "AGAINST": 57.53176043557169, + "NONE": 72.70973963355834 + } + }, + { + "after_task": 6, + "best_epoch": 7, + "best_dev_macro_f1": 0.6843380527929387, + "losses": { + "current_ce": 0.4687780435801598, + "replay_ce": 0.16411057466216256, + "dark_distillation": 0.30473792803914923, + "supervised_contrastive": 4.099669701174686, + "total": 1.1952245625487545, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6200762388818297, + "macro_f1": 0.5532441623491312, + "weighted_f1": 0.6446768595538065, + "macro_precision": 0.5620411493155046, + "macro_recall": 0.5851921539285373, + "per_class_f1": { + "FAVOR": 0.40559440559440557, + "AGAINST": 0.4900662251655629, + "NONE": 0.7640718562874251 + }, + "accuracy_pct": 62.00762388818297, + "macro_f1_pct": 55.32441623491312, + "weighted_f1_pct": 64.46768595538065, + "macro_precision_pct": 56.204114931550464, + "macro_recall_pct": 58.51921539285373, + "per_class_f1_pct": { + "FAVOR": 40.55944055944056, + "AGAINST": 49.00662251655629, + "NONE": 76.40718562874251 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5173209227858777, + "weighted_f1": 0.5695453628153941, + "macro_precision": 0.5407624633431085, + "macro_recall": 0.5437930245429409, + "per_class_f1": { + "FAVOR": 0.3952569169960474, + "AGAINST": 0.5026178010471204, + "NONE": 0.6540880503144654 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 51.73209227858777, + "weighted_f1_pct": 56.95453628153941, + "macro_precision_pct": 54.07624633431085, + "macro_recall_pct": 54.379302454294084, + "per_class_f1_pct": { + "FAVOR": 39.52569169960474, + "AGAINST": 50.26178010471204, + "NONE": 65.40880503144653 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6253456374089399, + "weighted_f1": 0.6793755866541941, + "macro_precision": 0.6417550507329486, + "macro_recall": 0.6276969434670244, + "per_class_f1": { + "FAVOR": 0.47715736040609136, + "AGAINST": 0.5714285714285714, + "NONE": 0.8274509803921568 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 62.53456374089399, + "weighted_f1_pct": 67.9375586654194, + "macro_precision_pct": 64.17550507329486, + "macro_recall_pct": 62.769694346702444, + "per_class_f1_pct": { + "FAVOR": 47.71573604060914, + "AGAINST": 57.14285714285714, + "NONE": 82.74509803921568 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6750348675034867, + "macro_f1": 0.6331164780524762, + "weighted_f1": 0.6886882822836864, + "macro_precision": 0.6322762931007907, + "macro_recall": 0.6698621941491526, + "per_class_f1": { + "FAVOR": 0.5151515151515151, + "AGAINST": 0.6164383561643836, + "NONE": 0.76775956284153 + }, + "accuracy_pct": 67.50348675034867, + "macro_f1_pct": 63.31164780524762, + "weighted_f1_pct": 68.86882822836864, + "macro_precision_pct": 63.22762931007907, + "macro_recall_pct": 66.98621941491525, + "per_class_f1_pct": { + "FAVOR": 51.515151515151516, + "AGAINST": 61.64383561643836, + "NONE": 76.775956284153 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6912972085385879, + "macro_f1": 0.6802673699225424, + "weighted_f1": 0.6988269863248631, + "macro_precision": 0.6901823321159202, + "macro_recall": 0.704663133015858, + "per_class_f1": { + "FAVOR": 0.6168831168831169, + "AGAINST": 0.6699507389162561, + "NONE": 0.753968253968254 + }, + "accuracy_pct": 69.12972085385879, + "macro_f1_pct": 68.02673699225424, + "weighted_f1_pct": 69.88269863248631, + "macro_precision_pct": 69.01823321159202, + "macro_recall_pct": 70.4663133015858, + "per_class_f1_pct": { + "FAVOR": 61.68831168831169, + "AGAINST": 66.99507389162561, + "NONE": 75.39682539682539 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6442451420029895, + "macro_f1": 0.6076814687822151, + "weighted_f1": 0.6613195410456303, + "macro_precision": 0.6142845005334477, + "macro_recall": 0.6402938899052601, + "per_class_f1": { + "FAVOR": 0.5870646766169154, + "AGAINST": 0.4797297297297297, + "NONE": 0.75625 + }, + "accuracy_pct": 64.42451420029896, + "macro_f1_pct": 60.76814687822151, + "weighted_f1_pct": 66.13195410456302, + "macro_precision_pct": 61.428450053344775, + "macro_recall_pct": 64.02938899052602, + "per_class_f1_pct": { + "FAVOR": 58.70646766169154, + "AGAINST": 47.97297297297297, + "NONE": 75.625 + } + } + ], + "accuracy": 0.6399002493765586, + "macro_f1": 0.6053158615300013, + "weighted_f1": 0.6559717432239228, + "macro_precision": 0.6097155178596244, + "macro_recall": 0.6303435425835722, + "per_class_f1": { + "FAVOR": 0.5099415204678363, + "AGAINST": 0.553607827150428, + "NONE": 0.7523982369717397 + }, + "accuracy_pct": 63.99002493765587, + "macro_f1_pct": 60.53158615300013, + "weighted_f1_pct": 65.59717432239228, + "macro_precision_pct": 60.97155178596244, + "macro_recall_pct": 63.03435425835722, + "per_class_f1_pct": { + "FAVOR": 50.994152046783626, + "AGAINST": 55.3607827150428, + "NONE": 75.23982369717397 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.732392809474041, + "losses": { + "current_ce": 0.39321027770638467, + "replay_ce": 0.16690240779093335, + "dark_distillation": 0.2781970035816942, + "supervised_contrastive": 4.059705282392956, + "total": 1.1051817282324745, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.602287166454892, + "macro_f1": 0.5330419163531069, + "weighted_f1": 0.6284987435476702, + "macro_precision": 0.5495274557220038, + "macro_recall": 0.5659028557176705, + "per_class_f1": { + "FAVOR": 0.3709090909090909, + "AGAINST": 0.47578947368421054, + "NONE": 0.7524271844660194 + }, + "accuracy_pct": 60.2287166454892, + "macro_f1_pct": 53.304191635310694, + "weighted_f1_pct": 62.849874354767024, + "macro_precision_pct": 54.95274557220038, + "macro_recall_pct": 56.590285571767055, + "per_class_f1_pct": { + "FAVOR": 37.09090909090909, + "AGAINST": 47.578947368421055, + "NONE": 75.24271844660194 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5499316005471956, + "macro_f1": 0.5165240370300673, + "weighted_f1": 0.5602231641199779, + "macro_precision": 0.5612544272088135, + "macro_recall": 0.5472769768886274, + "per_class_f1": { + "FAVOR": 0.3893805309734513, + "AGAINST": 0.5488958990536278, + "NONE": 0.6112956810631229 + }, + "accuracy_pct": 54.99316005471956, + "macro_f1_pct": 51.65240370300673, + "weighted_f1_pct": 56.022316411997785, + "macro_precision_pct": 56.125442720881345, + "macro_recall_pct": 54.72769768886274, + "per_class_f1_pct": { + "FAVOR": 38.93805309734513, + "AGAINST": 54.889589905362776, + "NONE": 61.12956810631229 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.704225352112676, + "macro_f1": 0.663338994136316, + "weighted_f1": 0.7055967401050682, + "macro_precision": 0.6717680150599771, + "macro_recall": 0.6657794564942686, + "per_class_f1": { + "FAVOR": 0.5660377358490566, + "AGAINST": 0.6013986013986014, + "NONE": 0.8225806451612904 + }, + "accuracy_pct": 70.4225352112676, + "macro_f1_pct": 66.3338994136316, + "weighted_f1_pct": 70.55967401050682, + "macro_precision_pct": 67.1768015059977, + "macro_recall_pct": 66.57794564942687, + "per_class_f1_pct": { + "FAVOR": 56.60377358490566, + "AGAINST": 60.13986013986013, + "NONE": 82.25806451612904 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6722454672245467, + "macro_f1": 0.633673287841786, + "weighted_f1": 0.6876244661838745, + "macro_precision": 0.6319979312972944, + "macro_recall": 0.6745307890305203, + "per_class_f1": { + "FAVOR": 0.5232974910394266, + "AGAINST": 0.6125290023201856, + "NONE": 0.7651933701657458 + }, + "accuracy_pct": 67.22454672245468, + "macro_f1_pct": 63.3673287841786, + "weighted_f1_pct": 68.76244661838746, + "macro_precision_pct": 63.19979312972944, + "macro_recall_pct": 67.45307890305203, + "per_class_f1_pct": { + "FAVOR": 52.32974910394266, + "AGAINST": 61.25290023201856, + "NONE": 76.51933701657458 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6998845787905729, + "weighted_f1": 0.7124202118811314, + "macro_precision": 0.7130264913283781, + "macro_recall": 0.7290714745552468, + "per_class_f1": { + "FAVOR": 0.6511627906976745, + "AGAINST": 0.7, + "NONE": 0.7484909456740443 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.9884578790573, + "weighted_f1_pct": 71.24202118811314, + "macro_precision_pct": 71.30264913283781, + "macro_recall_pct": 72.90714745552468, + "per_class_f1_pct": { + "FAVOR": 65.11627906976744, + "AGAINST": 70.0, + "NONE": 74.84909456740442 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6233183856502242, + "macro_f1": 0.5879611996255361, + "weighted_f1": 0.6433908229507698, + "macro_precision": 0.5988484265186446, + "macro_recall": 0.6214122977391909, + "per_class_f1": { + "FAVOR": 0.5728643216080402, + "AGAINST": 0.45016077170418006, + "NONE": 0.7408585055643879 + }, + "accuracy_pct": 62.33183856502242, + "macro_f1_pct": 58.796119962553604, + "weighted_f1_pct": 64.33908229507698, + "macro_precision_pct": 59.88484265186445, + "macro_recall_pct": 62.14122977391909, + "per_class_f1_pct": { + "FAVOR": 57.286432160804026, + "AGAINST": 45.01607717041801, + "NONE": 74.0858505564388 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6885147545999609, + "weighted_f1": 0.7227362054459493, + "macro_precision": 0.6836179541611281, + "macro_recall": 0.7130389009699355, + "per_class_f1": { + "FAVOR": 0.5977859778597786, + "AGAINST": 0.6694214876033058, + "NONE": 0.7983367983367984 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 68.8514754599961, + "weighted_f1_pct": 72.27362054459493, + "macro_precision_pct": 68.36179541611281, + "macro_recall_pct": 71.30389009699356, + "per_class_f1_pct": { + "FAVOR": 59.77859778597786, + "AGAINST": 66.94214876033058, + "NONE": 79.83367983367984 + } + } + ], + "accuracy": 0.6456623030840914, + "macro_f1": 0.6168978492921153, + "weighted_f1": 0.6609638311161464, + "macro_precision": 0.6234776305297557, + "macro_recall": 0.6470145911415978, + "per_class_f1": { + "FAVOR": 0.5310907237512742, + "AGAINST": 0.5737763486959628, + "NONE": 0.7458264754291088 + }, + "accuracy_pct": 64.56623030840915, + "macro_f1_pct": 61.689784929211534, + "weighted_f1_pct": 66.09638311161464, + "macro_precision_pct": 62.347763052975566, + "macro_recall_pct": 64.70145911415977, + "per_class_f1_pct": { + "FAVOR": 53.10907237512742, + "AGAINST": 57.377634869596285, + "NONE": 74.58264754291089 + } + }, + { + "after_task": 8, + "best_epoch": 6, + "best_dev_macro_f1": 0.7643056682921495, + "losses": { + "current_ce": 0.32467017376831936, + "replay_ce": 0.1549365407988137, + "dark_distillation": 0.27835183240035, + "supervised_contrastive": 4.00868492827696, + "total": 1.0196511304845997, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6226175349428208, + "macro_f1": 0.549202098628991, + "weighted_f1": 0.6453645950592148, + "macro_precision": 0.5594283781498394, + "macro_recall": 0.5777383493614431, + "per_class_f1": { + "FAVOR": 0.39114391143911437, + "AGAINST": 0.48577680525164113, + "NONE": 0.7706855791962175 + }, + "accuracy_pct": 62.26175349428208, + "macro_f1_pct": 54.92020986289911, + "weighted_f1_pct": 64.53645950592148, + "macro_precision_pct": 55.94283781498394, + "macro_recall_pct": 57.77383493614431, + "per_class_f1_pct": { + "FAVOR": 39.11439114391143, + "AGAINST": 48.57768052516411, + "NONE": 77.06855791962175 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5677154582763337, + "macro_f1": 0.5387839274369678, + "weighted_f1": 0.5835819066645362, + "macro_precision": 0.5594784363345936, + "macro_recall": 0.5732535994923005, + "per_class_f1": { + "FAVOR": 0.4247104247104247, + "AGAINST": 0.542907180385289, + "NONE": 0.6487341772151899 + }, + "accuracy_pct": 56.771545827633375, + "macro_f1_pct": 53.87839274369678, + "weighted_f1_pct": 58.35819066645362, + "macro_precision_pct": 55.94784363345936, + "macro_recall_pct": 57.32535994923005, + "per_class_f1_pct": { + "FAVOR": 42.471042471042466, + "AGAINST": 54.290718038528894, + "NONE": 64.87341772151899 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7183098591549296, + "macro_f1": 0.6710022687460286, + "weighted_f1": 0.7151002330394286, + "macro_precision": 0.6787058789283483, + "macro_recall": 0.6706594669011766, + "per_class_f1": { + "FAVOR": 0.5497630331753555, + "AGAINST": 0.6273062730627307, + "NONE": 0.8359375 + }, + "accuracy_pct": 71.83098591549296, + "macro_f1_pct": 67.10022687460287, + "weighted_f1_pct": 71.51002330394286, + "macro_precision_pct": 67.87058789283483, + "macro_recall_pct": 67.06594669011766, + "per_class_f1_pct": { + "FAVOR": 54.976303317535546, + "AGAINST": 62.73062730627307, + "NONE": 83.59375 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6875871687587168, + "macro_f1": 0.6473503065776921, + "weighted_f1": 0.7008597546288428, + "macro_precision": 0.6467662259392671, + "macro_recall": 0.6889718991950096, + "per_class_f1": { + "FAVOR": 0.5243445692883895, + "AGAINST": 0.6425339366515838, + "NONE": 0.7751724137931034 + }, + "accuracy_pct": 68.75871687587168, + "macro_f1_pct": 64.73503065776922, + "weighted_f1_pct": 70.08597546288428, + "macro_precision_pct": 64.67662259392671, + "macro_recall_pct": 68.89718991950096, + "per_class_f1_pct": { + "FAVOR": 52.43445692883895, + "AGAINST": 64.25339366515837, + "NONE": 77.51724137931035 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7077175697865353, + "macro_f1": 0.6962568067103975, + "weighted_f1": 0.7121906301119916, + "macro_precision": 0.7021514161220043, + "macro_recall": 0.7194313222329143, + "per_class_f1": { + "FAVOR": 0.6305084745762712, + "AGAINST": 0.7009803921568627, + "NONE": 0.7572815533980582 + }, + "accuracy_pct": 70.77175697865353, + "macro_f1_pct": 69.62568067103975, + "weighted_f1_pct": 71.21906301119915, + "macro_precision_pct": 70.21514161220043, + "macro_recall_pct": 71.94313222329143, + "per_class_f1_pct": { + "FAVOR": 63.05084745762712, + "AGAINST": 70.09803921568627, + "NONE": 75.72815533980582 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6218236173393124, + "macro_f1": 0.584769003669174, + "weighted_f1": 0.6418128696142593, + "macro_precision": 0.5973270336246949, + "macro_recall": 0.6181129100294102, + "per_class_f1": { + "FAVOR": 0.5700245700245701, + "AGAINST": 0.44224422442244227, + "NONE": 0.7420382165605095 + }, + "accuracy_pct": 62.18236173393124, + "macro_f1_pct": 58.47690036691739, + "weighted_f1_pct": 64.18128696142593, + "macro_precision_pct": 59.732703362469486, + "macro_recall_pct": 61.81129100294102, + "per_class_f1_pct": { + "FAVOR": 57.00245700245701, + "AGAINST": 44.224422442244226, + "NONE": 74.20382165605095 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7323943661971831, + "macro_f1": 0.7089966579521941, + "weighted_f1": 0.7410066973101214, + "macro_precision": 0.7022231930488813, + "macro_recall": 0.731940561250906, + "per_class_f1": { + "FAVOR": 0.6319702602230484, + "AGAINST": 0.6835443037974683, + "NONE": 0.8114754098360656 + }, + "accuracy_pct": 73.23943661971832, + "macro_f1_pct": 70.89966579521942, + "weighted_f1_pct": 74.10066973101213, + "macro_precision_pct": 70.22231930488813, + "macro_recall_pct": 73.1940561250906, + "per_class_f1_pct": { + "FAVOR": 63.19702602230484, + "AGAINST": 68.35443037974683, + "NONE": 81.14754098360656 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7308970099667774, + "macro_f1": 0.6982517140031105, + "weighted_f1": 0.7380301169718292, + "macro_precision": 0.6848284896230101, + "macro_recall": 0.7294661495258437, + "per_class_f1": { + "FAVOR": 0.6615969581749049, + "AGAINST": 0.6225165562913907, + "NONE": 0.810641627543036 + }, + "accuracy_pct": 73.08970099667775, + "macro_f1_pct": 69.82517140031105, + "weighted_f1_pct": 73.80301169718292, + "macro_precision_pct": 68.482848962301, + "macro_recall_pct": 72.94661495258437, + "per_class_f1_pct": { + "FAVOR": 66.15969581749049, + "AGAINST": 62.251655629139066, + "NONE": 81.06416275430361 + } + } + ], + "accuracy": 0.6666666666666666, + "macro_f1": 0.634732048260949, + "weighted_f1": 0.6801802466001565, + "macro_precision": 0.6332046070397764, + "macro_recall": 0.661880342913349, + "per_class_f1": { + "FAVOR": 0.5504014272970562, + "AGAINST": 0.587094617184888, + "NONE": 0.7667001003009027 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 63.4732048260949, + "weighted_f1_pct": 68.01802466001566, + "macro_precision_pct": 63.320460703977645, + "macro_recall_pct": 66.1880342913349, + "per_class_f1_pct": { + "FAVOR": 55.04014272970562, + "AGAINST": 58.7094617184888, + "NONE": 76.67001003009027 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__doc__s3__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d07a74aa57420387e9cae38f686cfa74c6c1b4a1 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__doc__s3__seed13/run.log b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..a0cd26e5b95a12073886c2984842f69a85341a21 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__doc__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 07:31:00,873 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__doc__s3__seed13 device=cuda +2026-09-02 08:01:15,722 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6368265979655694, 'average_forgetting': 0.00847516722507832, 'backward_transfer': -0.0050692659401382345, 'final_mean_target_macro_f1_pct': 63.68265979655694, 'average_forgetting_pct': 0.8475167225078319, 'backward_transfer_pct': -0.5069265940138235, 'final_pooled_accuracy': 0.6666666666666666, 'final_pooled_accuracy_pct': 66.66666666666666, 'final_pooled_macro_f1': 0.634732048260949, 'final_pooled_macro_f1_pct': 63.4732048260949, 'final_pooled_weighted_f1': 0.6801802466001565, 'final_pooled_weighted_f1_pct': 68.01802466001566, 'final_pooled_macro_precision': 0.6332046070397764, 'final_pooled_macro_precision_pct': 63.320460703977645, 'final_pooled_macro_recall': 0.661880342913349, 'final_pooled_macro_recall_pct': 66.1880342913349, 'final_pooled_per_class_f1': {'FAVOR': 0.5504014272970562, 'AGAINST': 0.587094617184888, 'NONE': 0.7667001003009027}, 'final_pooled_per_class_f1_pct': {'FAVOR': 55.04014272970562, 'AGAINST': 58.7094617184888, 'NONE': 76.67001003009027}} +2026-09-02 08:01:15,722 [INFO] __main__: efficiency={'training_wall_time_seconds': 1813.4751372640021, 'training_wall_time_minutes': 30.2245856210667, 'peak_gpu_memory_allocated_gb': 10.40872573852539, 'peak_gpu_memory_reserved_gb': 10.599609375} diff --git a/outputs/trace_s7_contrastive_replay__x__s1__seed13/environment.json b/outputs/trace_s7_contrastive_replay__x__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..3a2420cac2837f63f2bad73a21ce74190c681a38 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T08:01:23.824101+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..1c2678fda1f73204006234cfde293ac3e56f73b9 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.json @@ -0,0 +1,2294 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7079373572961969, + 0.5945165945165946, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7116641528406235, + 0.666954732510288, + 0.5669638128861431, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7201661029068008, + 0.6759769681778743, + 0.5906450380912179, + 0.7117957746478872, + null, + null, + null, + null, + null, + null + ], + [ + 0.6864819298434707, + 0.7389676723454205, + 0.5858019917682743, + 0.6540830531288225, + 0.6611416361416361, + null, + null, + null, + null, + null + ], + [ + 0.701728286671219, + 0.6663298387436318, + 0.5559698868185851, + 0.697909950569525, + 0.6855451461913765, + 0.6244159021406728, + null, + null, + null, + null + ], + [ + 0.7002020955756615, + 0.6653991769547325, + 0.6141117449528665, + 0.6683952393214759, + 0.7260027172307875, + 0.5914560578277394, + 0.67870808346248, + null, + null, + null + ], + [ + 0.6874207394400546, + 0.7137290417499446, + 0.5948412144883061, + 0.6751449275362319, + 0.6994851532289418, + 0.5910934744268078, + 0.6728463725126684, + 0.7269883998287251, + null, + null + ], + [ + 0.7037506337913563, + 0.7107142857142857, + 0.6037833081950729, + 0.7018011005967254, + 0.6758963329246149, + 0.556099099099099, + 0.6801431683784624, + 0.7042177604096483, + 0.82064772715952, + null + ], + [ + 0.715996840996841, + 0.6607837525572126, + 0.5667685542685542, + 0.7148772494087, + 0.707284192215699, + 0.5740859252968563, + 0.6231884057971014, + 0.6847423248398331, + 0.7858608646245885, + 0.628658391759107 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6840048840048839, + "losses": { + "current_ce": 0.7400040751411802, + "replay_ce": 0.2537589294569833, + "dark_distillation": 0.3645425006037667, + "supervised_contrastive": 4.125260809489659, + "total": 1.5885603552772884, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7079373572961969, + "weighted_f1": 0.7361316277808484, + "macro_precision": 0.6902201740911419, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7662337662337663, + "NONE": 0.7605633802816901 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.79373572961968, + "weighted_f1_pct": 73.61316277808484, + "macro_precision_pct": 69.02201740911418, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 76.62337662337663, + "NONE": 76.05633802816901 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5945165945165946, + "weighted_f1": 0.8493276634869555, + "macro_precision": 0.5920095408467502, + "macro_recall": 0.5975003033612426, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.0, + "NONE": 0.9047619047619048 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 59.45165945165945, + "weighted_f1_pct": 84.93276634869555, + "macro_precision_pct": 59.20095408467502, + "macro_recall_pct": 59.75003033612426, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 0.0, + "NONE": 90.47619047619048 + } + } + ], + "accuracy": 0.7799227799227799, + "macro_f1": 0.7807138235104186, + "weighted_f1": 0.774606036143859, + "macro_precision": 0.7914249532983789, + "macro_recall": 0.7948816460694541, + "per_class_f1": { + "FAVOR": 0.7839195979899497, + "AGAINST": 0.7195121951219512, + "NONE": 0.8387096774193549 + }, + "accuracy_pct": 77.99227799227799, + "macro_f1_pct": 78.07138235104186, + "weighted_f1_pct": 77.4606036143859, + "macro_precision_pct": 79.14249532983789, + "macro_recall_pct": 79.48816460694542, + "per_class_f1_pct": { + "FAVOR": 78.39195979899498, + "AGAINST": 71.95121951219512, + "NONE": 83.87096774193549 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6090658702505393, + "losses": { + "current_ce": 0.6735803604125976, + "replay_ce": 0.22019005584897416, + "dark_distillation": 0.33510870933532716, + "supervised_contrastive": 4.312815257274743, + "total": 1.4926063060760497, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7116641528406235, + "weighted_f1": 0.7454182937664, + "macro_precision": 0.694969315245478, + "macro_recall": 0.7865317341329335, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.782051282051282, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.16641528406234, + "weighted_f1_pct": 74.54182937664, + "macro_precision_pct": 69.4969315245478, + "macro_recall_pct": 78.65317341329336, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 78.2051282051282, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.666954732510288, + "weighted_f1": 0.8119359772752103, + "macro_precision": 0.6665094339622641, + "macro_recall": 0.7950006067224851, + "per_class_f1": { + "FAVOR": 0.8166666666666667, + "AGAINST": 0.32, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.6954732510288, + "weighted_f1_pct": 81.19359772752102, + "macro_precision_pct": 66.65094339622641, + "macro_recall_pct": 79.50006067224851, + "per_class_f1_pct": { + "FAVOR": 81.66666666666667, + "AGAINST": 32.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5669638128861431, + "weighted_f1": 0.5952581409392855, + "macro_precision": 0.560749881347888, + "macro_recall": 0.5791575889615105, + "per_class_f1": { + "FAVOR": 0.4854368932038835, + "AGAINST": 0.67, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 56.696381288614305, + "weighted_f1_pct": 59.52581409392855, + "macro_precision_pct": 56.074988134788796, + "macro_recall_pct": 57.91575889615105, + "per_class_f1_pct": { + "FAVOR": 48.54368932038835, + "AGAINST": 67.0, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6859688195991092, + "macro_f1": 0.6882084042021889, + "weighted_f1": 0.6852601656641045, + "macro_precision": 0.6818939116282249, + "macro_recall": 0.6985379126048225, + "per_class_f1": { + "FAVOR": 0.6460481099656358, + "AGAINST": 0.6929133858267716, + "NONE": 0.7256637168141593 + }, + "accuracy_pct": 68.59688195991092, + "macro_f1_pct": 68.82084042021889, + "weighted_f1_pct": 68.52601656641045, + "macro_precision_pct": 68.18939116282249, + "macro_recall_pct": 69.85379126048224, + "per_class_f1_pct": { + "FAVOR": 64.60481099656357, + "AGAINST": 69.29133858267717, + "NONE": 72.56637168141593 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.6773407986345991, + "losses": { + "current_ce": 0.6309270640214284, + "replay_ce": 0.20339152113054737, + "dark_distillation": 0.34281918198773353, + "supervised_contrastive": 4.208024680975711, + "total": 1.426530661727443, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7201661029068008, + "weighted_f1": 0.7521011309478737, + "macro_precision": 0.704064454064454, + "macro_recall": 0.7998650674662668, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.7870967741935484, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 72.01661029068008, + "weighted_f1_pct": 75.21011309478737, + "macro_precision_pct": 70.4064454064454, + "macro_recall_pct": 79.98650674662669, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 78.70967741935485, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6759769681778743, + "weighted_f1": 0.8151418695683974, + "macro_precision": 0.6672688949111714, + "macro_recall": 0.7999757311005946, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.34782608695652173, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 67.59769681778744, + "weighted_f1_pct": 81.51418695683974, + "macro_precision_pct": 66.72688949111713, + "macro_recall_pct": 79.99757311005946, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 34.78260869565217, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.5906450380912179, + "weighted_f1": 0.6191959425791156, + "macro_precision": 0.5906273620559335, + "macro_recall": 0.5918663761801017, + "per_class_f1": { + "FAVOR": 0.5048543689320388, + "AGAINST": 0.6956521739130435, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 59.064503809121796, + "weighted_f1_pct": 61.919594257911555, + "macro_precision_pct": 59.062736205593346, + "macro_recall_pct": 59.186637618010174, + "per_class_f1_pct": { + "FAVOR": 50.48543689320388, + "AGAINST": 69.56521739130434, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7117957746478872, + "weighted_f1": 0.7457945252158109, + "macro_precision": 0.6992276020554722, + "macro_recall": 0.7298540965207633, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.8, + "NONE": 0.6875 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 71.17957746478872, + "weighted_f1_pct": 74.57945252158109, + "macro_precision_pct": 69.92276020554722, + "macro_recall_pct": 72.98540965207633, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 80.0, + "NONE": 68.75 + } + } + ], + "accuracy": 0.7133858267716535, + "macro_f1": 0.7077441760127878, + "weighted_f1": 0.7140736719876744, + "macro_precision": 0.7017116564311926, + "macro_recall": 0.7162822604844629, + "per_class_f1": { + "FAVOR": 0.6593406593406593, + "AGAINST": 0.7423728813559322, + "NONE": 0.7215189873417721 + }, + "accuracy_pct": 71.33858267716535, + "macro_f1_pct": 70.77441760127878, + "weighted_f1_pct": 71.40736719876743, + "macro_precision_pct": 70.17116564311927, + "macro_recall_pct": 71.62822604844628, + "per_class_f1_pct": { + "FAVOR": 65.93406593406593, + "AGAINST": 74.23728813559322, + "NONE": 72.15189873417721 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6965141612200436, + "losses": { + "current_ce": 0.573601751797127, + "replay_ce": 0.17971050359986046, + "dark_distillation": 0.34934047018036696, + "supervised_contrastive": 4.282232322114886, + "total": 1.3562057328946662, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6864819298434707, + "weighted_f1": 0.7120803064397646, + "macro_precision": 0.6811612364243943, + "macro_recall": 0.7781259370314842, + "per_class_f1": { + "FAVOR": 0.5675675675675675, + "AGAINST": 0.738255033557047, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 68.64819298434706, + "weighted_f1_pct": 71.20803064397646, + "macro_precision_pct": 68.11612364243943, + "macro_recall_pct": 77.81259370314842, + "per_class_f1_pct": { + "FAVOR": 56.75675675675676, + "AGAINST": 73.8255033557047, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8672566371681416, + "macro_f1": 0.7389676723454205, + "weighted_f1": 0.8754256538624482, + "macro_precision": 0.7211589580010633, + "macro_recall": 0.7830603082150224, + "per_class_f1": { + "FAVOR": 0.9022556390977443, + "AGAINST": 0.42857142857142855, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 86.72566371681415, + "macro_f1_pct": 73.89676723454205, + "weighted_f1_pct": 87.54256538624482, + "macro_precision_pct": 72.11589580010633, + "macro_recall_pct": 78.30603082150223, + "per_class_f1_pct": { + "FAVOR": 90.22556390977444, + "AGAINST": 42.857142857142854, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5858019917682743, + "weighted_f1": 0.6099381331683551, + "macro_precision": 0.5865800865800866, + "macro_recall": 0.5900508351488744, + "per_class_f1": { + "FAVOR": 0.5166666666666667, + "AGAINST": 0.6735751295336787, + "NONE": 0.5671641791044776 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.58019917682743, + "weighted_f1_pct": 60.993813316835514, + "macro_precision_pct": 58.65800865800866, + "macro_recall_pct": 59.00508351488744, + "per_class_f1_pct": { + "FAVOR": 51.66666666666667, + "AGAINST": 67.35751295336787, + "NONE": 56.71641791044776 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6540830531288225, + "weighted_f1": 0.6819988970120269, + "macro_precision": 0.6519146352853274, + "macro_recall": 0.7005050505050505, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.7225130890052356, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 65.40830531288225, + "weighted_f1_pct": 68.1998897012027, + "macro_precision_pct": 65.19146352853275, + "macro_recall_pct": 70.05050505050505, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 72.25130890052355, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6611416361416361, + "weighted_f1": 0.6841668061211209, + "macro_precision": 0.6460205192300529, + "macro_recall": 0.6924003175955283, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7238095238095238, + "NONE": 0.6346153846153846 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.11416361416362, + "weighted_f1_pct": 68.41668061211209, + "macro_precision_pct": 64.60205192300529, + "macro_recall_pct": 69.24003175955283, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 72.38095238095238, + "NONE": 63.46153846153846 + } + } + ], + "accuracy": 0.6899038461538461, + "macro_f1": 0.68728227398769, + "weighted_f1": 0.6923247015756923, + "macro_precision": 0.6835623360426611, + "macro_recall": 0.7072267145852051, + "per_class_f1": { + "FAVOR": 0.656, + "AGAINST": 0.7080581241743725, + "NONE": 0.6977886977886978 + }, + "accuracy_pct": 68.99038461538461, + "macro_f1_pct": 68.72822739876901, + "weighted_f1_pct": 69.23247015756922, + "macro_precision_pct": 68.35623360426611, + "macro_recall_pct": 70.72267145852051, + "per_class_f1_pct": { + "FAVOR": 65.60000000000001, + "AGAINST": 70.80581241743725, + "NONE": 69.77886977886978 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.7202971968929416, + "losses": { + "current_ce": 0.6812616641322772, + "replay_ce": 0.21127971162398657, + "dark_distillation": 0.35512676077584426, + "supervised_contrastive": 4.360985952615738, + "total": 1.506203364332517, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.701728286671219, + "weighted_f1": 0.7289521035702496, + "macro_precision": 0.6849723750219366, + "macro_recall": 0.7871564217891054, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7581699346405228, + "NONE": 0.75 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.1728286671219, + "weighted_f1_pct": 72.89521035702496, + "macro_precision_pct": 68.49723750219367, + "macro_recall_pct": 78.71564217891054, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.81699346405229, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6663298387436318, + "weighted_f1": 0.8137222354494249, + "macro_precision": 0.6700416600748826, + "macro_recall": 0.8013105205678923, + "per_class_f1": { + "FAVOR": 0.8103448275862069, + "AGAINST": 0.3076923076923077, + "NONE": 0.8809523809523809 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.63298387436318, + "weighted_f1_pct": 81.37222354494249, + "macro_precision_pct": 67.00416600748827, + "macro_recall_pct": 80.13105205678923, + "per_class_f1_pct": { + "FAVOR": 81.03448275862068, + "AGAINST": 30.76923076923077, + "NONE": 88.09523809523809 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5559698868185851, + "weighted_f1": 0.5887711125993911, + "macro_precision": 0.5577903238453698, + "macro_recall": 0.5610021786492375, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.6824644549763034, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 55.596988681858505, + "weighted_f1_pct": 58.877111259939106, + "macro_precision_pct": 55.779032384536976, + "macro_recall_pct": 56.10021786492375, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 68.24644549763033, + "NONE": 54.794520547945204 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.697909950569525, + "weighted_f1": 0.7330871342110189, + "macro_precision": 0.6838809282578279, + "macro_recall": 0.7303030303030303, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7878787878787878, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 69.7909950569525, + "weighted_f1_pct": 73.30871342110188, + "macro_precision_pct": 68.38809282578279, + "macro_recall_pct": 73.03030303030303, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 78.78787878787878, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.6855451461913765, + "weighted_f1": 0.7067791132489866, + "macro_precision": 0.6770630279402209, + "macro_recall": 0.6964333601501743, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.7420814479638009, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 68.55451461913765, + "weighted_f1_pct": 70.67791132489866, + "macro_precision_pct": 67.7063027940221, + "macro_recall_pct": 69.64333601501743, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 74.2081447963801, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6267605633802817, + "macro_f1": 0.6244159021406728, + "weighted_f1": 0.6254709910841193, + "macro_precision": 0.6568376068376068, + "macro_recall": 0.6102564102564103, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.576, + "NONE": 0.6972477064220184 + }, + "accuracy_pct": 62.676056338028175, + "macro_f1_pct": 62.441590214067276, + "weighted_f1_pct": 62.54709910841193, + "macro_precision_pct": 65.68376068376068, + "macro_recall_pct": 61.02564102564103, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 57.599999999999994, + "NONE": 69.72477064220183 + } + } + ], + "accuracy": 0.686858316221766, + "macro_f1": 0.6799748774952081, + "weighted_f1": 0.6867380195931585, + "macro_precision": 0.6759975649584357, + "macro_recall": 0.6858328960601687, + "per_class_f1": { + "FAVOR": 0.6291666666666667, + "AGAINST": 0.7066381156316917, + "NONE": 0.704119850187266 + }, + "accuracy_pct": 68.6858316221766, + "macro_f1_pct": 67.99748774952081, + "weighted_f1_pct": 68.67380195931585, + "macro_precision_pct": 67.59975649584356, + "macro_recall_pct": 68.58328960601688, + "per_class_f1_pct": { + "FAVOR": 62.916666666666664, + "AGAINST": 70.66381156316916, + "NONE": 70.4119850187266 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.681862967409186, + "losses": { + "current_ce": 0.624606605985808, + "replay_ce": 0.19730890766968803, + "dark_distillation": 0.3246166924635569, + "supervised_contrastive": 4.30782457230583, + "total": 1.4150063240338886, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7002020955756615, + "weighted_f1": 0.7322940637565756, + "macro_precision": 0.689588377723971, + "macro_recall": 0.7889955022488756, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7682119205298014, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.02020955756615, + "weighted_f1_pct": 73.22940637565756, + "macro_precision_pct": 68.9588377723971, + "macro_recall_pct": 78.89955022488756, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.82119205298014, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6653991769547325, + "weighted_f1": 0.8201424669507266, + "macro_precision": 0.657183908045977, + "macro_recall": 0.7432593131901468, + "per_class_f1": { + "FAVOR": 0.832, + "AGAINST": 0.3, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 66.53991769547325, + "weighted_f1_pct": 82.01424669507266, + "macro_precision_pct": 65.7183908045977, + "macro_recall_pct": 74.32593131901469, + "per_class_f1_pct": { + "FAVOR": 83.2, + "AGAINST": 30.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6141117449528665, + "weighted_f1": 0.6384436026344537, + "macro_precision": 0.6040847827372339, + "macro_recall": 0.630718954248366, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.7142857142857143, + "NONE": 0.6233766233766234 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 61.411174495286645, + "weighted_f1_pct": 63.84436026344537, + "macro_precision_pct": 60.40847827372339, + "macro_recall_pct": 63.0718954248366, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 71.42857142857143, + "NONE": 62.33766233766234 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6683952393214759, + "weighted_f1": 0.6970118818364385, + "macro_precision": 0.6567160613282207, + "macro_recall": 0.715712682379349, + "per_class_f1": { + "FAVOR": 0.6046511627906976, + "AGAINST": 0.7407407407407407, + "NONE": 0.6597938144329897 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.8395239321476, + "weighted_f1_pct": 69.70118818364385, + "macro_precision_pct": 65.67160613282206, + "macro_recall_pct": 71.5712682379349, + "per_class_f1_pct": { + "FAVOR": 60.46511627906976, + "AGAINST": 74.07407407407408, + "NONE": 65.97938144329896 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7260027172307875, + "weighted_f1": 0.7383331039489202, + "macro_precision": 0.7062732589048378, + "macro_recall": 0.7606465487777304, + "per_class_f1": { + "FAVOR": 0.7272727272727273, + "AGAINST": 0.7655502392344498, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 72.60027172307875, + "weighted_f1_pct": 73.83331039489201, + "macro_precision_pct": 70.62732589048379, + "macro_recall_pct": 76.06465487777304, + "per_class_f1_pct": { + "FAVOR": 72.72727272727273, + "AGAINST": 76.55502392344498, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5914560578277394, + "weighted_f1": 0.5780391509545194, + "macro_precision": 0.5886960248580306, + "macro_recall": 0.5991452991452991, + "per_class_f1": { + "FAVOR": 0.6333333333333333, + "AGAINST": 0.46846846846846846, + "NONE": 0.672566371681416 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 59.145605782773934, + "weighted_f1_pct": 57.80391509545194, + "macro_precision_pct": 58.869602485803064, + "macro_recall_pct": 59.914529914529915, + "per_class_f1_pct": { + "FAVOR": 63.33333333333333, + "AGAINST": 46.846846846846844, + "NONE": 67.2566371681416 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.67870808346248, + "weighted_f1": 0.6663767131594907, + "macro_precision": 0.667989417989418, + "macro_recall": 0.703730726765957, + "per_class_f1": { + "FAVOR": 0.6405228758169934, + "AGAINST": 0.6533333333333333, + "NONE": 0.7422680412371134 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.870808346248, + "weighted_f1_pct": 66.63767131594906, + "macro_precision_pct": 66.7989417989418, + "macro_recall_pct": 70.37307267659571, + "per_class_f1_pct": { + "FAVOR": 64.05228758169935, + "AGAINST": 65.33333333333333, + "NONE": 74.22680412371135 + } + } + ], + "accuracy": 0.686541737649063, + "macro_f1": 0.6856329686028019, + "weighted_f1": 0.6867257517704946, + "macro_precision": 0.6785361743108221, + "macro_recall": 0.701152109892079, + "per_class_f1": { + "FAVOR": 0.6548672566371682, + "AGAINST": 0.6939571150097466, + "NONE": 0.7080745341614907 + }, + "accuracy_pct": 68.6541737649063, + "macro_f1_pct": 68.56329686028019, + "weighted_f1_pct": 68.67257517704945, + "macro_precision_pct": 67.85361743108221, + "macro_recall_pct": 70.1152109892079, + "per_class_f1_pct": { + "FAVOR": 65.48672566371681, + "AGAINST": 69.39571150097466, + "NONE": 70.80745341614907 + } + }, + { + "after_task": 8, + "best_epoch": 4, + "best_dev_macro_f1": 0.7372919199734719, + "losses": { + "current_ce": 0.5858170489902081, + "replay_ce": 0.20176352707372194, + "dark_distillation": 0.31748082050378773, + "supervised_contrastive": 4.291883395374685, + "total": 1.3755093379297119, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6874207394400546, + "weighted_f1": 0.7156035896264047, + "macro_precision": 0.6715713862111118, + "macro_recall": 0.770199900049975, + "per_class_f1": { + "FAVOR": 0.5671641791044776, + "AGAINST": 0.7450980392156863, + "NONE": 0.75 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.74207394400545, + "weighted_f1_pct": 71.56035896264046, + "macro_precision_pct": 67.15713862111117, + "macro_recall_pct": 77.0199900049975, + "per_class_f1_pct": { + "FAVOR": 56.71641791044776, + "AGAINST": 74.50980392156863, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7137290417499446, + "weighted_f1": 0.8404825433486821, + "macro_precision": 0.6940301573606146, + "macro_recall": 0.8167212716903288, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.42105263157894735, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.37290417499446, + "weighted_f1_pct": 84.0482543348682, + "macro_precision_pct": 69.40301573606146, + "macro_recall_pct": 81.67212716903288, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 42.10526315789473, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.5948412144883061, + "weighted_f1": 0.621739188954392, + "macro_precision": 0.5897120827607422, + "macro_recall": 0.6013071895424836, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.6965174129353234, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 59.48412144883061, + "weighted_f1_pct": 62.17391889543919, + "macro_precision_pct": 58.971208276074215, + "macro_recall_pct": 60.130718954248366, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 69.65174129353234, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6751449275362319, + "weighted_f1": 0.7105084151472651, + "macro_precision": 0.6619642306506321, + "macro_recall": 0.7049943883277217, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.77, + "NONE": 0.6304347826086957 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 67.51449275362319, + "weighted_f1_pct": 71.05084151472651, + "macro_precision_pct": 66.19642306506321, + "macro_recall_pct": 70.49943883277217, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 77.0, + "NONE": 63.04347826086957 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6994851532289418, + "weighted_f1": 0.7183242832784892, + "macro_precision": 0.6829611612220309, + "macro_recall": 0.7309376955394654, + "per_class_f1": { + "FAVOR": 0.6835443037974683, + "AGAINST": 0.7547169811320755, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.94851532289418, + "weighted_f1_pct": 71.83242832784892, + "macro_precision_pct": 68.29611612220309, + "macro_recall_pct": 73.09376955394654, + "per_class_f1_pct": { + "FAVOR": 68.35443037974683, + "AGAINST": 75.47169811320755, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5910934744268078, + "weighted_f1": 0.58994708994709, + "macro_precision": 0.5975885225885226, + "macro_recall": 0.5871794871794872, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.5166666666666667, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 59.10934744268078, + "weighted_f1_pct": 58.994708994709, + "macro_precision_pct": 59.75885225885226, + "macro_recall_pct": 58.71794871794872, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 51.66666666666667, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6728463725126684, + "weighted_f1": 0.6619614386355209, + "macro_precision": 0.6640515371858656, + "macro_recall": 0.6909526787575567, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6206896551724138, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.28463725126684, + "weighted_f1_pct": 66.1961438635521, + "macro_precision_pct": 66.40515371858656, + "macro_recall_pct": 69.09526787575567, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.06896551724138, + "NONE": 73.11827956989248 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7269883998287251, + "weighted_f1": 0.7329902931201485, + "macro_precision": 0.7324736991403658, + "macro_recall": 0.7269385233240655, + "per_class_f1": { + "FAVOR": 0.7102803738317757, + "AGAINST": 0.6890756302521008, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.69883998287251, + "weighted_f1_pct": 73.29902931201485, + "macro_precision_pct": 73.24736991403657, + "macro_recall_pct": 72.69385233240655, + "per_class_f1_pct": { + "FAVOR": 71.02803738317756, + "AGAINST": 68.90756302521008, + "NONE": 78.16091954022988 + } + } + ], + "accuracy": 0.6914119359534207, + "macro_f1": 0.6910964636931798, + "weighted_f1": 0.6914402967826289, + "macro_precision": 0.6859499646613022, + "macro_recall": 0.7004664685859999, + "per_class_f1": { + "FAVOR": 0.6641221374045801, + "AGAINST": 0.6928999144568007, + "NONE": 0.7162673392181589 + }, + "accuracy_pct": 69.14119359534206, + "macro_f1_pct": 69.10964636931797, + "weighted_f1_pct": 69.14402967826288, + "macro_precision_pct": 68.59499646613023, + "macro_recall_pct": 70.0466468586, + "per_class_f1_pct": { + "FAVOR": 66.41221374045801, + "AGAINST": 69.28999144568007, + "NONE": 71.62673392181588 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6443877648811235, + "losses": { + "current_ce": 0.44664672889436285, + "replay_ce": 0.2207125659411152, + "dark_distillation": 0.33955786377191544, + "supervised_contrastive": 4.0738591939210895, + "total": 1.24452415133516, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7037506337913563, + "weighted_f1": 0.733003531809461, + "macro_precision": 0.6817663817663817, + "macro_recall": 0.7792303848075962, + "per_class_f1": { + "FAVOR": 0.5901639344262295, + "AGAINST": 0.7643312101910829, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.37506337913563, + "weighted_f1_pct": 73.3003531809461, + "macro_precision_pct": 68.17663817663816, + "macro_recall_pct": 77.92303848075962, + "per_class_f1_pct": { + "FAVOR": 59.01639344262295, + "AGAINST": 76.43312101910828, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7107142857142857, + "weighted_f1": 0.843394437420986, + "macro_precision": 0.6931189337968999, + "macro_recall": 0.8198762286130323, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.4, + "NONE": 0.875 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.07142857142857, + "weighted_f1_pct": 84.3394437420986, + "macro_precision_pct": 69.31189337968999, + "macro_recall_pct": 81.98762286130324, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 40.0, + "NONE": 87.5 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6037833081950729, + "weighted_f1": 0.6338330554894022, + "macro_precision": 0.6039046121593291, + "macro_recall": 0.6053013798111838, + "per_class_f1": { + "FAVOR": 0.49019607843137253, + "AGAINST": 0.7211538461538461, + "NONE": 0.6 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 60.37833081950728, + "weighted_f1_pct": 63.383305548940214, + "macro_precision_pct": 60.39046121593291, + "macro_recall_pct": 60.53013798111838, + "per_class_f1_pct": { + "FAVOR": 49.01960784313725, + "AGAINST": 72.11538461538461, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7365591397849462, + "macro_f1": 0.7018011005967254, + "weighted_f1": 0.7414795815212867, + "macro_precision": 0.6936310756186531, + "macro_recall": 0.7235129068462401, + "per_class_f1": { + "FAVOR": 0.6329113924050633, + "AGAINST": 0.8058252427184466, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 73.65591397849462, + "macro_f1_pct": 70.18011005967254, + "weighted_f1_pct": 74.14795815212867, + "macro_precision_pct": 69.3631075618653, + "macro_recall_pct": 72.35129068462402, + "per_class_f1_pct": { + "FAVOR": 63.29113924050633, + "AGAINST": 80.58252427184466, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6758963329246149, + "weighted_f1": 0.6942550588726983, + "macro_precision": 0.6603767529925227, + "macro_recall": 0.7161884330025923, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.7246376811594203, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 67.5896332924615, + "weighted_f1_pct": 69.42550588726984, + "macro_precision_pct": 66.03767529925227, + "macro_recall_pct": 71.61884330025923, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 72.46376811594203, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5563380281690141, + "macro_f1": 0.556099099099099, + "weighted_f1": 0.5549492450196675, + "macro_precision": 0.59745521270945, + "macro_recall": 0.541025641025641, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.496, + "NONE": 0.6306306306306306 + }, + "accuracy_pct": 55.633802816901415, + "macro_f1_pct": 55.6099099099099, + "weighted_f1_pct": 55.49492450196675, + "macro_precision_pct": 59.745521270945, + "macro_recall_pct": 54.1025641025641, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 49.6, + "NONE": 63.06306306306306 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6801431683784624, + "weighted_f1": 0.665204014939309, + "macro_precision": 0.6807521092445822, + "macro_recall": 0.696967240327674, + "per_class_f1": { + "FAVOR": 0.6142857142857143, + "AGAINST": 0.6705882352941176, + "NONE": 0.7555555555555555 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 68.01431683784625, + "weighted_f1_pct": 66.5204014939309, + "macro_precision_pct": 68.07521092445822, + "macro_recall_pct": 69.69672403276739, + "per_class_f1_pct": { + "FAVOR": 61.42857142857143, + "AGAINST": 67.05882352941175, + "NONE": 75.55555555555556 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7042177604096483, + "weighted_f1": 0.7148811930108252, + "macro_precision": 0.7172072590791804, + "macro_recall": 0.6994130367624344, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6504065040650406, + "NONE": 0.7955801104972375 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 70.42177604096483, + "weighted_f1_pct": 71.48811930108252, + "macro_precision_pct": 71.72072590791804, + "macro_recall_pct": 69.94130367624344, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 65.04065040650406, + "NONE": 79.55801104972376 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.84, + "macro_f1": 0.82064772715952, + "weighted_f1": 0.8415257616886248, + "macro_precision": 0.8066115648523545, + "macro_recall": 0.838123530827179, + "per_class_f1": { + "FAVOR": 0.7931034482758621, + "AGAINST": 0.7899159663865546, + "NONE": 0.8789237668161435 + }, + "accuracy_pct": 84.0, + "macro_f1_pct": 82.06477271595199, + "weighted_f1_pct": 84.15257616886248, + "macro_precision_pct": 80.66115648523545, + "macro_recall_pct": 83.8123530827179, + "per_class_f1_pct": { + "FAVOR": 79.3103448275862, + "AGAINST": 78.99159663865547, + "NONE": 87.89237668161435 + } + } + ], + "accuracy": 0.7102922490470139, + "macro_f1": 0.7065444797479171, + "weighted_f1": 0.7095329891622508, + "macro_precision": 0.7053461501364516, + "macro_recall": 0.7086717555711756, + "per_class_f1": { + "FAVOR": 0.654911838790932, + "AGAINST": 0.7071161048689139, + "NONE": 0.7576054955839058 + }, + "accuracy_pct": 71.0292249047014, + "macro_f1_pct": 70.65444797479171, + "weighted_f1_pct": 70.95329891622508, + "macro_precision_pct": 70.53461501364517, + "macro_recall_pct": 70.86717555711756, + "per_class_f1_pct": { + "FAVOR": 65.4911838790932, + "AGAINST": 70.71161048689139, + "NONE": 75.76054955839058 + } + }, + { + "after_task": 10, + "best_epoch": 8, + "best_dev_macro_f1": 0.6643372500515358, + "losses": { + "current_ce": 0.6257232047044314, + "replay_ce": 0.21735582458667266, + "dark_distillation": 0.3414711534212797, + "supervised_contrastive": 4.221410020192464, + "total": 1.4359556185893523, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.715996840996841, + "weighted_f1": 0.7401674824620031, + "macro_precision": 0.6955518794228471, + "macro_recall": 0.8022738630684657, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7662337662337663, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.5996840996841, + "weighted_f1_pct": 74.01674824620031, + "macro_precision_pct": 69.55518794228472, + "macro_recall_pct": 80.22738630684657, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 76.62337662337663, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.6607837525572126, + "weighted_f1": 0.8045251583574667, + "macro_precision": 0.6633645328049919, + "macro_recall": 0.7931804392670793, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.3076923076923077, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 66.07837525572126, + "weighted_f1_pct": 80.45251583574668, + "macro_precision_pct": 66.33645328049919, + "macro_recall_pct": 79.31804392670793, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 30.76923076923077, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5667685542685542, + "weighted_f1": 0.599139910587279, + "macro_precision": 0.5669753086419753, + "macro_recall": 0.5740740740740741, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.6952380952380952, + "NONE": 0.5675675675675675 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 56.67685542685542, + "weighted_f1_pct": 59.9139910587279, + "macro_precision_pct": 56.69753086419753, + "macro_recall_pct": 57.407407407407405, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 69.52380952380952, + "NONE": 56.75675675675676 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7473118279569892, + "macro_f1": 0.7148772494087, + "weighted_f1": 0.7523877174628759, + "macro_precision": 0.703153937296516, + "macro_recall": 0.7356341189674523, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.8097560975609757, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 74.73118279569893, + "macro_f1_pct": 71.48772494087, + "weighted_f1_pct": 75.2387717462876, + "macro_precision_pct": 70.3153937296516, + "macro_recall_pct": 73.56341189674524, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 80.97560975609757, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.707284192215699, + "weighted_f1": 0.7180188145072367, + "macro_precision": 0.6902634593356244, + "macro_recall": 0.734544823560961, + "per_class_f1": { + "FAVOR": 0.7123287671232876, + "AGAINST": 0.7428571428571429, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 70.72841922156991, + "weighted_f1_pct": 71.80188145072367, + "macro_precision_pct": 69.02634593356244, + "macro_recall_pct": 73.4544823560961, + "per_class_f1_pct": { + "FAVOR": 71.23287671232876, + "AGAINST": 74.28571428571429, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5740859252968563, + "weighted_f1": 0.5736110235974926, + "macro_precision": 0.6226268285091815, + "macro_recall": 0.5602564102564103, + "per_class_f1": { + "FAVOR": 0.5531914893617021, + "AGAINST": 0.5081967213114754, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 57.408592529685635, + "weighted_f1_pct": 57.36110235974926, + "macro_precision_pct": 62.26268285091815, + "macro_recall_pct": 56.02564102564103, + "per_class_f1_pct": { + "FAVOR": 55.319148936170215, + "AGAINST": 50.81967213114754, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.62, + "macro_f1": 0.6231884057971014, + "weighted_f1": 0.6039130434782609, + "macro_precision": 0.6446755973122026, + "macro_recall": 0.6522749557979911, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.6521739130434783, + "NONE": 0.717391304347826 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 62.31884057971014, + "weighted_f1_pct": 60.39130434782609, + "macro_precision_pct": 64.46755973122026, + "macro_recall_pct": 65.22749557979911, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 65.21739130434783, + "NONE": 71.73913043478261 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6847423248398331, + "weighted_f1": 0.6980273087808234, + "macro_precision": 0.6961810422701928, + "macro_recall": 0.6806147667593452, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.6423357664233577, + "NONE": 0.788235294117647 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 68.47423248398331, + "weighted_f1_pct": 69.80273087808234, + "macro_precision_pct": 69.61810422701929, + "macro_recall_pct": 68.06147667593451, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 64.23357664233576, + "NONE": 78.82352941176471 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7858608646245885, + "weighted_f1": 0.8080307565047375, + "macro_precision": 0.7676691061096402, + "macro_recall": 0.8121365243304274, + "per_class_f1": { + "FAVOR": 0.7666666666666667, + "AGAINST": 0.7377049180327869, + "NONE": 0.8532110091743119 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 78.58608646245885, + "weighted_f1_pct": 80.80307565047376, + "macro_precision_pct": 76.76691061096402, + "macro_recall_pct": 81.21365243304274, + "per_class_f1_pct": { + "FAVOR": 76.66666666666667, + "AGAINST": 73.77049180327869, + "NONE": 85.3211009174312 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.65, + "macro_f1": 0.628658391759107, + "weighted_f1": 0.6548312616741586, + "macro_precision": 0.6237509693142794, + "macro_recall": 0.6389890584065342, + "per_class_f1": { + "FAVOR": 0.6839378238341969, + "AGAINST": 0.43010752688172044, + "NONE": 0.7719298245614035 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 62.8658391759107, + "weighted_f1_pct": 65.48312616741586, + "macro_precision_pct": 62.37509693142794, + "macro_recall_pct": 63.89890584065342, + "per_class_f1_pct": { + "FAVOR": 68.39378238341969, + "AGAINST": 43.01075268817204, + "NONE": 77.19298245614034 + } + } + ], + "accuracy": 0.6910935738444194, + "macro_f1": 0.6889970040284997, + "weighted_f1": 0.6892762929935321, + "macro_precision": 0.6920004411294777, + "macro_recall": 0.6890624154294503, + "per_class_f1": { + "FAVOR": 0.635593220338983, + "AGAINST": 0.6794258373205742, + "NONE": 0.7519719544259421 + }, + "accuracy_pct": 69.10935738444194, + "macro_f1_pct": 68.89970040284997, + "weighted_f1_pct": 68.92762929935321, + "macro_precision_pct": 69.20004411294777, + "macro_recall_pct": 68.90624154294504, + "per_class_f1_pct": { + "FAVOR": 63.559322033898304, + "AGAINST": 67.94258373205741, + "NONE": 75.19719544259421 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6662246501764493, + "average_forgetting": 0.03662789995397295, + "backward_transfer": -0.007430456873193559, + "final_mean_target_macro_f1_pct": 66.62246501764493, + "average_forgetting_pct": 3.662789995397295, + "backward_transfer_pct": -0.7430456873193559, + "final_pooled_accuracy": 0.6910935738444194, + "final_pooled_accuracy_pct": 69.10935738444194, + "final_pooled_macro_f1": 0.6889970040284997, + "final_pooled_macro_f1_pct": 68.89970040284997, + "final_pooled_weighted_f1": 0.6892762929935321, + "final_pooled_weighted_f1_pct": 68.92762929935321, + "final_pooled_macro_precision": 0.6920004411294777, + "final_pooled_macro_precision_pct": 69.20004411294777, + "final_pooled_macro_recall": 0.6890624154294503, + "final_pooled_macro_recall_pct": 68.90624154294504, + "final_pooled_per_class_f1": { + "FAVOR": 0.635593220338983, + "AGAINST": 0.6794258373205742, + "NONE": 0.7519719544259421 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 63.559322033898304, + "AGAINST": 67.94258373205741, + "NONE": 75.19719544259421 + } + }, + "efficiency": { + "training_wall_time_seconds": 976.5617710498627, + "training_wall_time_minutes": 16.27602951749771, + "peak_gpu_memory_allocated_gb": 10.40872573852539, + "peak_gpu_memory_reserved_gb": 10.587890625 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e41330bcca058e632e09ef58d0d92d0b197792 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s1__seed13/metrics.partial.json @@ -0,0 +1,2178 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7079373572961969, + 0.5945165945165946, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7116641528406235, + 0.666954732510288, + 0.5669638128861431, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7201661029068008, + 0.6759769681778743, + 0.5906450380912179, + 0.7117957746478872, + null, + null, + null, + null, + null, + null + ], + [ + 0.6864819298434707, + 0.7389676723454205, + 0.5858019917682743, + 0.6540830531288225, + 0.6611416361416361, + null, + null, + null, + null, + null + ], + [ + 0.701728286671219, + 0.6663298387436318, + 0.5559698868185851, + 0.697909950569525, + 0.6855451461913765, + 0.6244159021406728, + null, + null, + null, + null + ], + [ + 0.7002020955756615, + 0.6653991769547325, + 0.6141117449528665, + 0.6683952393214759, + 0.7260027172307875, + 0.5914560578277394, + 0.67870808346248, + null, + null, + null + ], + [ + 0.6874207394400546, + 0.7137290417499446, + 0.5948412144883061, + 0.6751449275362319, + 0.6994851532289418, + 0.5910934744268078, + 0.6728463725126684, + 0.7269883998287251, + null, + null + ], + [ + 0.7037506337913563, + 0.7107142857142857, + 0.6037833081950729, + 0.7018011005967254, + 0.6758963329246149, + 0.556099099099099, + 0.6801431683784624, + 0.7042177604096483, + 0.82064772715952, + null + ], + [ + 0.715996840996841, + 0.6607837525572126, + 0.5667685542685542, + 0.7148772494087, + 0.707284192215699, + 0.5740859252968563, + 0.6231884057971014, + 0.6847423248398331, + 0.7858608646245885, + 0.628658391759107 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6840048840048839, + "losses": { + "current_ce": 0.7400040751411802, + "replay_ce": 0.2537589294569833, + "dark_distillation": 0.3645425006037667, + "supervised_contrastive": 4.125260809489659, + "total": 1.5885603552772884, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7079373572961969, + "weighted_f1": 0.7361316277808484, + "macro_precision": 0.6902201740911419, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7662337662337663, + "NONE": 0.7605633802816901 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.79373572961968, + "weighted_f1_pct": 73.61316277808484, + "macro_precision_pct": 69.02201740911418, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 76.62337662337663, + "NONE": 76.05633802816901 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.5945165945165946, + "weighted_f1": 0.8493276634869555, + "macro_precision": 0.5920095408467502, + "macro_recall": 0.5975003033612426, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.0, + "NONE": 0.9047619047619048 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 59.45165945165945, + "weighted_f1_pct": 84.93276634869555, + "macro_precision_pct": 59.20095408467502, + "macro_recall_pct": 59.75003033612426, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 0.0, + "NONE": 90.47619047619048 + } + } + ], + "accuracy": 0.7799227799227799, + "macro_f1": 0.7807138235104186, + "weighted_f1": 0.774606036143859, + "macro_precision": 0.7914249532983789, + "macro_recall": 0.7948816460694541, + "per_class_f1": { + "FAVOR": 0.7839195979899497, + "AGAINST": 0.7195121951219512, + "NONE": 0.8387096774193549 + }, + "accuracy_pct": 77.99227799227799, + "macro_f1_pct": 78.07138235104186, + "weighted_f1_pct": 77.4606036143859, + "macro_precision_pct": 79.14249532983789, + "macro_recall_pct": 79.48816460694542, + "per_class_f1_pct": { + "FAVOR": 78.39195979899498, + "AGAINST": 71.95121951219512, + "NONE": 83.87096774193549 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6090658702505393, + "losses": { + "current_ce": 0.6735803604125976, + "replay_ce": 0.22019005584897416, + "dark_distillation": 0.33510870933532716, + "supervised_contrastive": 4.312815257274743, + "total": 1.4926063060760497, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.7116641528406235, + "weighted_f1": 0.7454182937664, + "macro_precision": 0.694969315245478, + "macro_recall": 0.7865317341329335, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.782051282051282, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.16641528406234, + "weighted_f1_pct": 74.54182937664, + "macro_precision_pct": 69.4969315245478, + "macro_recall_pct": 78.65317341329336, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 78.2051282051282, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.666954732510288, + "weighted_f1": 0.8119359772752103, + "macro_precision": 0.6665094339622641, + "macro_recall": 0.7950006067224851, + "per_class_f1": { + "FAVOR": 0.8166666666666667, + "AGAINST": 0.32, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.6954732510288, + "weighted_f1_pct": 81.19359772752102, + "macro_precision_pct": 66.65094339622641, + "macro_recall_pct": 79.50006067224851, + "per_class_f1_pct": { + "FAVOR": 81.66666666666667, + "AGAINST": 32.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5669638128861431, + "weighted_f1": 0.5952581409392855, + "macro_precision": 0.560749881347888, + "macro_recall": 0.5791575889615105, + "per_class_f1": { + "FAVOR": 0.4854368932038835, + "AGAINST": 0.67, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 56.696381288614305, + "weighted_f1_pct": 59.52581409392855, + "macro_precision_pct": 56.074988134788796, + "macro_recall_pct": 57.91575889615105, + "per_class_f1_pct": { + "FAVOR": 48.54368932038835, + "AGAINST": 67.0, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6859688195991092, + "macro_f1": 0.6882084042021889, + "weighted_f1": 0.6852601656641045, + "macro_precision": 0.6818939116282249, + "macro_recall": 0.6985379126048225, + "per_class_f1": { + "FAVOR": 0.6460481099656358, + "AGAINST": 0.6929133858267716, + "NONE": 0.7256637168141593 + }, + "accuracy_pct": 68.59688195991092, + "macro_f1_pct": 68.82084042021889, + "weighted_f1_pct": 68.52601656641045, + "macro_precision_pct": 68.18939116282249, + "macro_recall_pct": 69.85379126048224, + "per_class_f1_pct": { + "FAVOR": 64.60481099656357, + "AGAINST": 69.29133858267717, + "NONE": 72.56637168141593 + } + }, + { + "after_task": 4, + "best_epoch": 15, + "best_dev_macro_f1": 0.6773407986345991, + "losses": { + "current_ce": 0.6309270640214284, + "replay_ce": 0.20339152113054737, + "dark_distillation": 0.34281918198773353, + "supervised_contrastive": 4.208024680975711, + "total": 1.426530661727443, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7201661029068008, + "weighted_f1": 0.7521011309478737, + "macro_precision": 0.704064454064454, + "macro_recall": 0.7998650674662668, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.7870967741935484, + "NONE": 0.7647058823529411 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 72.01661029068008, + "weighted_f1_pct": 75.21011309478737, + "macro_precision_pct": 70.4064454064454, + "macro_recall_pct": 79.98650674662669, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 78.70967741935485, + "NONE": 76.47058823529412 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6759769681778743, + "weighted_f1": 0.8151418695683974, + "macro_precision": 0.6672688949111714, + "macro_recall": 0.7999757311005946, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.34782608695652173, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 67.59769681778744, + "weighted_f1_pct": 81.51418695683974, + "macro_precision_pct": 66.72688949111713, + "macro_recall_pct": 79.99757311005946, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 34.78260869565217, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.5906450380912179, + "weighted_f1": 0.6191959425791156, + "macro_precision": 0.5906273620559335, + "macro_recall": 0.5918663761801017, + "per_class_f1": { + "FAVOR": 0.5048543689320388, + "AGAINST": 0.6956521739130435, + "NONE": 0.5714285714285714 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 59.064503809121796, + "weighted_f1_pct": 61.919594257911555, + "macro_precision_pct": 59.062736205593346, + "macro_recall_pct": 59.186637618010174, + "per_class_f1_pct": { + "FAVOR": 50.48543689320388, + "AGAINST": 69.56521739130434, + "NONE": 57.14285714285714 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7117957746478872, + "weighted_f1": 0.7457945252158109, + "macro_precision": 0.6992276020554722, + "macro_recall": 0.7298540965207633, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.8, + "NONE": 0.6875 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 71.17957746478872, + "weighted_f1_pct": 74.57945252158109, + "macro_precision_pct": 69.92276020554722, + "macro_recall_pct": 72.98540965207633, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 80.0, + "NONE": 68.75 + } + } + ], + "accuracy": 0.7133858267716535, + "macro_f1": 0.7077441760127878, + "weighted_f1": 0.7140736719876744, + "macro_precision": 0.7017116564311926, + "macro_recall": 0.7162822604844629, + "per_class_f1": { + "FAVOR": 0.6593406593406593, + "AGAINST": 0.7423728813559322, + "NONE": 0.7215189873417721 + }, + "accuracy_pct": 71.33858267716535, + "macro_f1_pct": 70.77441760127878, + "weighted_f1_pct": 71.40736719876743, + "macro_precision_pct": 70.17116564311927, + "macro_recall_pct": 71.62822604844628, + "per_class_f1_pct": { + "FAVOR": 65.93406593406593, + "AGAINST": 74.23728813559322, + "NONE": 72.15189873417721 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6965141612200436, + "losses": { + "current_ce": 0.573601751797127, + "replay_ce": 0.17971050359986046, + "dark_distillation": 0.34934047018036696, + "supervised_contrastive": 4.282232322114886, + "total": 1.3562057328946662, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6864819298434707, + "weighted_f1": 0.7120803064397646, + "macro_precision": 0.6811612364243943, + "macro_recall": 0.7781259370314842, + "per_class_f1": { + "FAVOR": 0.5675675675675675, + "AGAINST": 0.738255033557047, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 68.64819298434706, + "weighted_f1_pct": 71.20803064397646, + "macro_precision_pct": 68.11612364243943, + "macro_recall_pct": 77.81259370314842, + "per_class_f1_pct": { + "FAVOR": 56.75675675675676, + "AGAINST": 73.8255033557047, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8672566371681416, + "macro_f1": 0.7389676723454205, + "weighted_f1": 0.8754256538624482, + "macro_precision": 0.7211589580010633, + "macro_recall": 0.7830603082150224, + "per_class_f1": { + "FAVOR": 0.9022556390977443, + "AGAINST": 0.42857142857142855, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 86.72566371681415, + "macro_f1_pct": 73.89676723454205, + "weighted_f1_pct": 87.54256538624482, + "macro_precision_pct": 72.11589580010633, + "macro_recall_pct": 78.30603082150223, + "per_class_f1_pct": { + "FAVOR": 90.22556390977444, + "AGAINST": 42.857142857142854, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5858019917682743, + "weighted_f1": 0.6099381331683551, + "macro_precision": 0.5865800865800866, + "macro_recall": 0.5900508351488744, + "per_class_f1": { + "FAVOR": 0.5166666666666667, + "AGAINST": 0.6735751295336787, + "NONE": 0.5671641791044776 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 58.58019917682743, + "weighted_f1_pct": 60.993813316835514, + "macro_precision_pct": 58.65800865800866, + "macro_recall_pct": 59.00508351488744, + "per_class_f1_pct": { + "FAVOR": 51.66666666666667, + "AGAINST": 67.35751295336787, + "NONE": 56.71641791044776 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6540830531288225, + "weighted_f1": 0.6819988970120269, + "macro_precision": 0.6519146352853274, + "macro_recall": 0.7005050505050505, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.7225130890052356, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 65.40830531288225, + "weighted_f1_pct": 68.1998897012027, + "macro_precision_pct": 65.19146352853275, + "macro_recall_pct": 70.05050505050505, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 72.25130890052355, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6611416361416361, + "weighted_f1": 0.6841668061211209, + "macro_precision": 0.6460205192300529, + "macro_recall": 0.6924003175955283, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7238095238095238, + "NONE": 0.6346153846153846 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.11416361416362, + "weighted_f1_pct": 68.41668061211209, + "macro_precision_pct": 64.60205192300529, + "macro_recall_pct": 69.24003175955283, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 72.38095238095238, + "NONE": 63.46153846153846 + } + } + ], + "accuracy": 0.6899038461538461, + "macro_f1": 0.68728227398769, + "weighted_f1": 0.6923247015756923, + "macro_precision": 0.6835623360426611, + "macro_recall": 0.7072267145852051, + "per_class_f1": { + "FAVOR": 0.656, + "AGAINST": 0.7080581241743725, + "NONE": 0.6977886977886978 + }, + "accuracy_pct": 68.99038461538461, + "macro_f1_pct": 68.72822739876901, + "weighted_f1_pct": 69.23247015756922, + "macro_precision_pct": 68.35623360426611, + "macro_recall_pct": 70.72267145852051, + "per_class_f1_pct": { + "FAVOR": 65.60000000000001, + "AGAINST": 70.80581241743725, + "NONE": 69.77886977886978 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.7202971968929416, + "losses": { + "current_ce": 0.6812616641322772, + "replay_ce": 0.21127971162398657, + "dark_distillation": 0.35512676077584426, + "supervised_contrastive": 4.360985952615738, + "total": 1.506203364332517, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.701728286671219, + "weighted_f1": 0.7289521035702496, + "macro_precision": 0.6849723750219366, + "macro_recall": 0.7871564217891054, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7581699346405228, + "NONE": 0.75 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.1728286671219, + "weighted_f1_pct": 72.89521035702496, + "macro_precision_pct": 68.49723750219367, + "macro_recall_pct": 78.71564217891054, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 75.81699346405229, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7787610619469026, + "macro_f1": 0.6663298387436318, + "weighted_f1": 0.8137222354494249, + "macro_precision": 0.6700416600748826, + "macro_recall": 0.8013105205678923, + "per_class_f1": { + "FAVOR": 0.8103448275862069, + "AGAINST": 0.3076923076923077, + "NONE": 0.8809523809523809 + }, + "accuracy_pct": 77.87610619469027, + "macro_f1_pct": 66.63298387436318, + "weighted_f1_pct": 81.37222354494249, + "macro_precision_pct": 67.00416600748827, + "macro_recall_pct": 80.13105205678923, + "per_class_f1_pct": { + "FAVOR": 81.03448275862068, + "AGAINST": 30.76923076923077, + "NONE": 88.09523809523809 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5559698868185851, + "weighted_f1": 0.5887711125993911, + "macro_precision": 0.5577903238453698, + "macro_recall": 0.5610021786492375, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.6824644549763034, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 55.596988681858505, + "weighted_f1_pct": 58.877111259939106, + "macro_precision_pct": 55.779032384536976, + "macro_recall_pct": 56.10021786492375, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 68.24644549763033, + "NONE": 54.794520547945204 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.697909950569525, + "weighted_f1": 0.7330871342110189, + "macro_precision": 0.6838809282578279, + "macro_recall": 0.7303030303030303, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7878787878787878, + "NONE": 0.6808510638297872 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 69.7909950569525, + "weighted_f1_pct": 73.30871342110188, + "macro_precision_pct": 68.38809282578279, + "macro_recall_pct": 73.03030303030303, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 78.78787878787878, + "NONE": 68.08510638297872 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.6855451461913765, + "weighted_f1": 0.7067791132489866, + "macro_precision": 0.6770630279402209, + "macro_recall": 0.6964333601501743, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.7420814479638009, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 68.55451461913765, + "weighted_f1_pct": 70.67791132489866, + "macro_precision_pct": 67.7063027940221, + "macro_recall_pct": 69.64333601501743, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 74.2081447963801, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6267605633802817, + "macro_f1": 0.6244159021406728, + "weighted_f1": 0.6254709910841193, + "macro_precision": 0.6568376068376068, + "macro_recall": 0.6102564102564103, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.576, + "NONE": 0.6972477064220184 + }, + "accuracy_pct": 62.676056338028175, + "macro_f1_pct": 62.441590214067276, + "weighted_f1_pct": 62.54709910841193, + "macro_precision_pct": 65.68376068376068, + "macro_recall_pct": 61.02564102564103, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 57.599999999999994, + "NONE": 69.72477064220183 + } + } + ], + "accuracy": 0.686858316221766, + "macro_f1": 0.6799748774952081, + "weighted_f1": 0.6867380195931585, + "macro_precision": 0.6759975649584357, + "macro_recall": 0.6858328960601687, + "per_class_f1": { + "FAVOR": 0.6291666666666667, + "AGAINST": 0.7066381156316917, + "NONE": 0.704119850187266 + }, + "accuracy_pct": 68.6858316221766, + "macro_f1_pct": 67.99748774952081, + "weighted_f1_pct": 68.67380195931585, + "macro_precision_pct": 67.59975649584356, + "macro_recall_pct": 68.58328960601688, + "per_class_f1_pct": { + "FAVOR": 62.916666666666664, + "AGAINST": 70.66381156316916, + "NONE": 70.4119850187266 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.681862967409186, + "losses": { + "current_ce": 0.624606605985808, + "replay_ce": 0.19730890766968803, + "dark_distillation": 0.3246166924635569, + "supervised_contrastive": 4.30782457230583, + "total": 1.4150063240338886, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.7002020955756615, + "weighted_f1": 0.7322940637565756, + "macro_precision": 0.689588377723971, + "macro_recall": 0.7889955022488756, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7682119205298014, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 70.02020955756615, + "weighted_f1_pct": 73.22940637565756, + "macro_precision_pct": 68.9588377723971, + "macro_recall_pct": 78.89955022488756, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.82119205298014, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6653991769547325, + "weighted_f1": 0.8201424669507266, + "macro_precision": 0.657183908045977, + "macro_recall": 0.7432593131901468, + "per_class_f1": { + "FAVOR": 0.832, + "AGAINST": 0.3, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 66.53991769547325, + "weighted_f1_pct": 82.01424669507266, + "macro_precision_pct": 65.7183908045977, + "macro_recall_pct": 74.32593131901469, + "per_class_f1_pct": { + "FAVOR": 83.2, + "AGAINST": 30.0, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6141117449528665, + "weighted_f1": 0.6384436026344537, + "macro_precision": 0.6040847827372339, + "macro_recall": 0.630718954248366, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.7142857142857143, + "NONE": 0.6233766233766234 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 61.411174495286645, + "weighted_f1_pct": 63.84436026344537, + "macro_precision_pct": 60.40847827372339, + "macro_recall_pct": 63.0718954248366, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 71.42857142857143, + "NONE": 62.33766233766234 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6683952393214759, + "weighted_f1": 0.6970118818364385, + "macro_precision": 0.6567160613282207, + "macro_recall": 0.715712682379349, + "per_class_f1": { + "FAVOR": 0.6046511627906976, + "AGAINST": 0.7407407407407407, + "NONE": 0.6597938144329897 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 66.8395239321476, + "weighted_f1_pct": 69.70118818364385, + "macro_precision_pct": 65.67160613282206, + "macro_recall_pct": 71.5712682379349, + "per_class_f1_pct": { + "FAVOR": 60.46511627906976, + "AGAINST": 74.07407407407408, + "NONE": 65.97938144329896 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7260027172307875, + "weighted_f1": 0.7383331039489202, + "macro_precision": 0.7062732589048378, + "macro_recall": 0.7606465487777304, + "per_class_f1": { + "FAVOR": 0.7272727272727273, + "AGAINST": 0.7655502392344498, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 72.60027172307875, + "weighted_f1_pct": 73.83331039489201, + "macro_precision_pct": 70.62732589048379, + "macro_recall_pct": 76.06465487777304, + "per_class_f1_pct": { + "FAVOR": 72.72727272727273, + "AGAINST": 76.55502392344498, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5914560578277394, + "weighted_f1": 0.5780391509545194, + "macro_precision": 0.5886960248580306, + "macro_recall": 0.5991452991452991, + "per_class_f1": { + "FAVOR": 0.6333333333333333, + "AGAINST": 0.46846846846846846, + "NONE": 0.672566371681416 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 59.145605782773934, + "weighted_f1_pct": 57.80391509545194, + "macro_precision_pct": 58.869602485803064, + "macro_recall_pct": 59.914529914529915, + "per_class_f1_pct": { + "FAVOR": 63.33333333333333, + "AGAINST": 46.846846846846844, + "NONE": 67.2566371681416 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.67870808346248, + "weighted_f1": 0.6663767131594907, + "macro_precision": 0.667989417989418, + "macro_recall": 0.703730726765957, + "per_class_f1": { + "FAVOR": 0.6405228758169934, + "AGAINST": 0.6533333333333333, + "NONE": 0.7422680412371134 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 67.870808346248, + "weighted_f1_pct": 66.63767131594906, + "macro_precision_pct": 66.7989417989418, + "macro_recall_pct": 70.37307267659571, + "per_class_f1_pct": { + "FAVOR": 64.05228758169935, + "AGAINST": 65.33333333333333, + "NONE": 74.22680412371135 + } + } + ], + "accuracy": 0.686541737649063, + "macro_f1": 0.6856329686028019, + "weighted_f1": 0.6867257517704946, + "macro_precision": 0.6785361743108221, + "macro_recall": 0.701152109892079, + "per_class_f1": { + "FAVOR": 0.6548672566371682, + "AGAINST": 0.6939571150097466, + "NONE": 0.7080745341614907 + }, + "accuracy_pct": 68.6541737649063, + "macro_f1_pct": 68.56329686028019, + "weighted_f1_pct": 68.67257517704945, + "macro_precision_pct": 67.85361743108221, + "macro_recall_pct": 70.1152109892079, + "per_class_f1_pct": { + "FAVOR": 65.48672566371681, + "AGAINST": 69.39571150097466, + "NONE": 70.80745341614907 + } + }, + { + "after_task": 8, + "best_epoch": 4, + "best_dev_macro_f1": 0.7372919199734719, + "losses": { + "current_ce": 0.5858170489902081, + "replay_ce": 0.20176352707372194, + "dark_distillation": 0.31748082050378773, + "supervised_contrastive": 4.291883395374685, + "total": 1.3755093379297119, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6874207394400546, + "weighted_f1": 0.7156035896264047, + "macro_precision": 0.6715713862111118, + "macro_recall": 0.770199900049975, + "per_class_f1": { + "FAVOR": 0.5671641791044776, + "AGAINST": 0.7450980392156863, + "NONE": 0.75 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.74207394400545, + "weighted_f1_pct": 71.56035896264046, + "macro_precision_pct": 67.15713862111117, + "macro_recall_pct": 77.0199900049975, + "per_class_f1_pct": { + "FAVOR": 56.71641791044776, + "AGAINST": 74.50980392156863, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7137290417499446, + "weighted_f1": 0.8404825433486821, + "macro_precision": 0.6940301573606146, + "macro_recall": 0.8167212716903288, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.42105263157894735, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.37290417499446, + "weighted_f1_pct": 84.0482543348682, + "macro_precision_pct": 69.40301573606146, + "macro_recall_pct": 81.67212716903288, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 42.10526315789473, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6210526315789474, + "macro_f1": 0.5948412144883061, + "weighted_f1": 0.621739188954392, + "macro_precision": 0.5897120827607422, + "macro_recall": 0.6013071895424836, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.6965174129353234, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 62.10526315789474, + "macro_f1_pct": 59.48412144883061, + "weighted_f1_pct": 62.17391889543919, + "macro_precision_pct": 58.971208276074215, + "macro_recall_pct": 60.130718954248366, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 69.65174129353234, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6751449275362319, + "weighted_f1": 0.7105084151472651, + "macro_precision": 0.6619642306506321, + "macro_recall": 0.7049943883277217, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.77, + "NONE": 0.6304347826086957 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 67.51449275362319, + "weighted_f1_pct": 71.05084151472651, + "macro_precision_pct": 66.19642306506321, + "macro_recall_pct": 70.49943883277217, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 77.0, + "NONE": 63.04347826086957 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6994851532289418, + "weighted_f1": 0.7183242832784892, + "macro_precision": 0.6829611612220309, + "macro_recall": 0.7309376955394654, + "per_class_f1": { + "FAVOR": 0.6835443037974683, + "AGAINST": 0.7547169811320755, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.94851532289418, + "weighted_f1_pct": 71.83242832784892, + "macro_precision_pct": 68.29611612220309, + "macro_recall_pct": 73.09376955394654, + "per_class_f1_pct": { + "FAVOR": 68.35443037974683, + "AGAINST": 75.47169811320755, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5910934744268078, + "weighted_f1": 0.58994708994709, + "macro_precision": 0.5975885225885226, + "macro_recall": 0.5871794871794872, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.5166666666666667, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 59.10934744268078, + "weighted_f1_pct": 58.994708994709, + "macro_precision_pct": 59.75885225885226, + "macro_recall_pct": 58.71794871794872, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 51.66666666666667, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6728463725126684, + "weighted_f1": 0.6619614386355209, + "macro_precision": 0.6640515371858656, + "macro_recall": 0.6909526787575567, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6206896551724138, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.28463725126684, + "weighted_f1_pct": 66.1961438635521, + "macro_precision_pct": 66.40515371858656, + "macro_recall_pct": 69.09526787575567, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.06896551724138, + "NONE": 73.11827956989248 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.735, + "macro_f1": 0.7269883998287251, + "weighted_f1": 0.7329902931201485, + "macro_precision": 0.7324736991403658, + "macro_recall": 0.7269385233240655, + "per_class_f1": { + "FAVOR": 0.7102803738317757, + "AGAINST": 0.6890756302521008, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 73.5, + "macro_f1_pct": 72.69883998287251, + "weighted_f1_pct": 73.29902931201485, + "macro_precision_pct": 73.24736991403657, + "macro_recall_pct": 72.69385233240655, + "per_class_f1_pct": { + "FAVOR": 71.02803738317756, + "AGAINST": 68.90756302521008, + "NONE": 78.16091954022988 + } + } + ], + "accuracy": 0.6914119359534207, + "macro_f1": 0.6910964636931798, + "weighted_f1": 0.6914402967826289, + "macro_precision": 0.6859499646613022, + "macro_recall": 0.7004664685859999, + "per_class_f1": { + "FAVOR": 0.6641221374045801, + "AGAINST": 0.6928999144568007, + "NONE": 0.7162673392181589 + }, + "accuracy_pct": 69.14119359534206, + "macro_f1_pct": 69.10964636931797, + "weighted_f1_pct": 69.14402967826288, + "macro_precision_pct": 68.59499646613023, + "macro_recall_pct": 70.0466468586, + "per_class_f1_pct": { + "FAVOR": 66.41221374045801, + "AGAINST": 69.28999144568007, + "NONE": 71.62673392181588 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.6443877648811235, + "losses": { + "current_ce": 0.44664672889436285, + "replay_ce": 0.2207125659411152, + "dark_distillation": 0.33955786377191544, + "supervised_contrastive": 4.0738591939210895, + "total": 1.24452415133516, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7037506337913563, + "weighted_f1": 0.733003531809461, + "macro_precision": 0.6817663817663817, + "macro_recall": 0.7792303848075962, + "per_class_f1": { + "FAVOR": 0.5901639344262295, + "AGAINST": 0.7643312101910829, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.37506337913563, + "weighted_f1_pct": 73.3003531809461, + "macro_precision_pct": 68.17663817663816, + "macro_recall_pct": 77.92303848075962, + "per_class_f1_pct": { + "FAVOR": 59.01639344262295, + "AGAINST": 76.43312101910828, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7107142857142857, + "weighted_f1": 0.843394437420986, + "macro_precision": 0.6931189337968999, + "macro_recall": 0.8198762286130323, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.4, + "NONE": 0.875 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.07142857142857, + "weighted_f1_pct": 84.3394437420986, + "macro_precision_pct": 69.31189337968999, + "macro_recall_pct": 81.98762286130324, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 40.0, + "NONE": 87.5 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6037833081950729, + "weighted_f1": 0.6338330554894022, + "macro_precision": 0.6039046121593291, + "macro_recall": 0.6053013798111838, + "per_class_f1": { + "FAVOR": 0.49019607843137253, + "AGAINST": 0.7211538461538461, + "NONE": 0.6 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 60.37833081950728, + "weighted_f1_pct": 63.383305548940214, + "macro_precision_pct": 60.39046121593291, + "macro_recall_pct": 60.53013798111838, + "per_class_f1_pct": { + "FAVOR": 49.01960784313725, + "AGAINST": 72.11538461538461, + "NONE": 60.0 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7365591397849462, + "macro_f1": 0.7018011005967254, + "weighted_f1": 0.7414795815212867, + "macro_precision": 0.6936310756186531, + "macro_recall": 0.7235129068462401, + "per_class_f1": { + "FAVOR": 0.6329113924050633, + "AGAINST": 0.8058252427184466, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 73.65591397849462, + "macro_f1_pct": 70.18011005967254, + "weighted_f1_pct": 74.14795815212867, + "macro_precision_pct": 69.3631075618653, + "macro_recall_pct": 72.35129068462402, + "per_class_f1_pct": { + "FAVOR": 63.29113924050633, + "AGAINST": 80.58252427184466, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6758963329246149, + "weighted_f1": 0.6942550588726983, + "macro_precision": 0.6603767529925227, + "macro_recall": 0.7161884330025923, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.7246376811594203, + "NONE": 0.6601941747572816 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 67.5896332924615, + "weighted_f1_pct": 69.42550588726984, + "macro_precision_pct": 66.03767529925227, + "macro_recall_pct": 71.61884330025923, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 72.46376811594203, + "NONE": 66.01941747572816 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5563380281690141, + "macro_f1": 0.556099099099099, + "weighted_f1": 0.5549492450196675, + "macro_precision": 0.59745521270945, + "macro_recall": 0.541025641025641, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.496, + "NONE": 0.6306306306306306 + }, + "accuracy_pct": 55.633802816901415, + "macro_f1_pct": 55.6099099099099, + "weighted_f1_pct": 55.49492450196675, + "macro_precision_pct": 59.745521270945, + "macro_recall_pct": 54.1025641025641, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 49.6, + "NONE": 63.06306306306306 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.67, + "macro_f1": 0.6801431683784624, + "weighted_f1": 0.665204014939309, + "macro_precision": 0.6807521092445822, + "macro_recall": 0.696967240327674, + "per_class_f1": { + "FAVOR": 0.6142857142857143, + "AGAINST": 0.6705882352941176, + "NONE": 0.7555555555555555 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 68.01431683784625, + "weighted_f1_pct": 66.5204014939309, + "macro_precision_pct": 68.07521092445822, + "macro_recall_pct": 69.69672403276739, + "per_class_f1_pct": { + "FAVOR": 61.42857142857143, + "AGAINST": 67.05882352941175, + "NONE": 75.55555555555556 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7042177604096483, + "weighted_f1": 0.7148811930108252, + "macro_precision": 0.7172072590791804, + "macro_recall": 0.6994130367624344, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6504065040650406, + "NONE": 0.7955801104972375 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 70.42177604096483, + "weighted_f1_pct": 71.48811930108252, + "macro_precision_pct": 71.72072590791804, + "macro_recall_pct": 69.94130367624344, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 65.04065040650406, + "NONE": 79.55801104972376 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.84, + "macro_f1": 0.82064772715952, + "weighted_f1": 0.8415257616886248, + "macro_precision": 0.8066115648523545, + "macro_recall": 0.838123530827179, + "per_class_f1": { + "FAVOR": 0.7931034482758621, + "AGAINST": 0.7899159663865546, + "NONE": 0.8789237668161435 + }, + "accuracy_pct": 84.0, + "macro_f1_pct": 82.06477271595199, + "weighted_f1_pct": 84.15257616886248, + "macro_precision_pct": 80.66115648523545, + "macro_recall_pct": 83.8123530827179, + "per_class_f1_pct": { + "FAVOR": 79.3103448275862, + "AGAINST": 78.99159663865547, + "NONE": 87.89237668161435 + } + } + ], + "accuracy": 0.7102922490470139, + "macro_f1": 0.7065444797479171, + "weighted_f1": 0.7095329891622508, + "macro_precision": 0.7053461501364516, + "macro_recall": 0.7086717555711756, + "per_class_f1": { + "FAVOR": 0.654911838790932, + "AGAINST": 0.7071161048689139, + "NONE": 0.7576054955839058 + }, + "accuracy_pct": 71.0292249047014, + "macro_f1_pct": 70.65444797479171, + "weighted_f1_pct": 70.95329891622508, + "macro_precision_pct": 70.53461501364517, + "macro_recall_pct": 70.86717555711756, + "per_class_f1_pct": { + "FAVOR": 65.4911838790932, + "AGAINST": 70.71161048689139, + "NONE": 75.76054955839058 + } + }, + { + "after_task": 10, + "best_epoch": 8, + "best_dev_macro_f1": 0.6643372500515358, + "losses": { + "current_ce": 0.6257232047044314, + "replay_ce": 0.21735582458667266, + "dark_distillation": 0.3414711534212797, + "supervised_contrastive": 4.221410020192464, + "total": 1.4359556185893523, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.715996840996841, + "weighted_f1": 0.7401674824620031, + "macro_precision": 0.6955518794228471, + "macro_recall": 0.8022738630684657, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.7662337662337663, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.5996840996841, + "weighted_f1_pct": 74.01674824620031, + "macro_precision_pct": 69.55518794228472, + "macro_recall_pct": 80.22738630684657, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 76.62337662337663, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.6607837525572126, + "weighted_f1": 0.8045251583574667, + "macro_precision": 0.6633645328049919, + "macro_recall": 0.7931804392670793, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.3076923076923077, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 66.07837525572126, + "weighted_f1_pct": 80.45251583574668, + "macro_precision_pct": 66.33645328049919, + "macro_recall_pct": 79.31804392670793, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 30.76923076923077, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5667685542685542, + "weighted_f1": 0.599139910587279, + "macro_precision": 0.5669753086419753, + "macro_recall": 0.5740740740740741, + "per_class_f1": { + "FAVOR": 0.4375, + "AGAINST": 0.6952380952380952, + "NONE": 0.5675675675675675 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 56.67685542685542, + "weighted_f1_pct": 59.9139910587279, + "macro_precision_pct": 56.69753086419753, + "macro_recall_pct": 57.407407407407405, + "per_class_f1_pct": { + "FAVOR": 43.75, + "AGAINST": 69.52380952380952, + "NONE": 56.75675675675676 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7473118279569892, + "macro_f1": 0.7148772494087, + "weighted_f1": 0.7523877174628759, + "macro_precision": 0.703153937296516, + "macro_recall": 0.7356341189674523, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.8097560975609757, + "NONE": 0.7032967032967034 + }, + "accuracy_pct": 74.73118279569893, + "macro_f1_pct": 71.48772494087, + "weighted_f1_pct": 75.2387717462876, + "macro_precision_pct": 70.3153937296516, + "macro_recall_pct": 73.56341189674524, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 80.97560975609757, + "NONE": 70.32967032967034 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.707284192215699, + "weighted_f1": 0.7180188145072367, + "macro_precision": 0.6902634593356244, + "macro_recall": 0.734544823560961, + "per_class_f1": { + "FAVOR": 0.7123287671232876, + "AGAINST": 0.7428571428571429, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 70.72841922156991, + "weighted_f1_pct": 71.80188145072367, + "macro_precision_pct": 69.02634593356244, + "macro_recall_pct": 73.4544823560961, + "per_class_f1_pct": { + "FAVOR": 71.23287671232876, + "AGAINST": 74.28571428571429, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5740859252968563, + "weighted_f1": 0.5736110235974926, + "macro_precision": 0.6226268285091815, + "macro_recall": 0.5602564102564103, + "per_class_f1": { + "FAVOR": 0.5531914893617021, + "AGAINST": 0.5081967213114754, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 57.408592529685635, + "weighted_f1_pct": 57.36110235974926, + "macro_precision_pct": 62.26268285091815, + "macro_recall_pct": 56.02564102564103, + "per_class_f1_pct": { + "FAVOR": 55.319148936170215, + "AGAINST": 50.81967213114754, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.62, + "macro_f1": 0.6231884057971014, + "weighted_f1": 0.6039130434782609, + "macro_precision": 0.6446755973122026, + "macro_recall": 0.6522749557979911, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.6521739130434783, + "NONE": 0.717391304347826 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 62.31884057971014, + "weighted_f1_pct": 60.39130434782609, + "macro_precision_pct": 64.46755973122026, + "macro_recall_pct": 65.22749557979911, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 65.21739130434783, + "NONE": 71.73913043478261 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6847423248398331, + "weighted_f1": 0.6980273087808234, + "macro_precision": 0.6961810422701928, + "macro_recall": 0.6806147667593452, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.6423357664233577, + "NONE": 0.788235294117647 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 68.47423248398331, + "weighted_f1_pct": 69.80273087808234, + "macro_precision_pct": 69.61810422701929, + "macro_recall_pct": 68.06147667593451, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 64.23357664233576, + "NONE": 78.82352941176471 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7858608646245885, + "weighted_f1": 0.8080307565047375, + "macro_precision": 0.7676691061096402, + "macro_recall": 0.8121365243304274, + "per_class_f1": { + "FAVOR": 0.7666666666666667, + "AGAINST": 0.7377049180327869, + "NONE": 0.8532110091743119 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 78.58608646245885, + "weighted_f1_pct": 80.80307565047376, + "macro_precision_pct": 76.76691061096402, + "macro_recall_pct": 81.21365243304274, + "per_class_f1_pct": { + "FAVOR": 76.66666666666667, + "AGAINST": 73.77049180327869, + "NONE": 85.3211009174312 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.65, + "macro_f1": 0.628658391759107, + "weighted_f1": 0.6548312616741586, + "macro_precision": 0.6237509693142794, + "macro_recall": 0.6389890584065342, + "per_class_f1": { + "FAVOR": 0.6839378238341969, + "AGAINST": 0.43010752688172044, + "NONE": 0.7719298245614035 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 62.8658391759107, + "weighted_f1_pct": 65.48312616741586, + "macro_precision_pct": 62.37509693142794, + "macro_recall_pct": 63.89890584065342, + "per_class_f1_pct": { + "FAVOR": 68.39378238341969, + "AGAINST": 43.01075268817204, + "NONE": 77.19298245614034 + } + } + ], + "accuracy": 0.6910935738444194, + "macro_f1": 0.6889970040284997, + "weighted_f1": 0.6892762929935321, + "macro_precision": 0.6920004411294777, + "macro_recall": 0.6890624154294503, + "per_class_f1": { + "FAVOR": 0.635593220338983, + "AGAINST": 0.6794258373205742, + "NONE": 0.7519719544259421 + }, + "accuracy_pct": 69.10935738444194, + "macro_f1_pct": 68.89970040284997, + "weighted_f1_pct": 68.92762929935321, + "macro_precision_pct": 69.20004411294777, + "macro_recall_pct": 68.90624154294504, + "per_class_f1_pct": { + "FAVOR": 63.559322033898304, + "AGAINST": 67.94258373205741, + "NONE": 75.19719544259421 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s1__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__x__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fbf18947ea58926575fd3bf9563079afe29b4e0e --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__x__s1__seed13/run.log b/outputs/trace_s7_contrastive_replay__x__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..e9700bf0147b37907d464dd5304937f3b7ff466e --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 08:01:23,881 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__x__s1__seed13 device=cuda +2026-09-02 08:17:41,781 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6662246501764493, 'average_forgetting': 0.03662789995397295, 'backward_transfer': -0.007430456873193559, 'final_mean_target_macro_f1_pct': 66.62246501764493, 'average_forgetting_pct': 3.662789995397295, 'backward_transfer_pct': -0.7430456873193559, 'final_pooled_accuracy': 0.6910935738444194, 'final_pooled_accuracy_pct': 69.10935738444194, 'final_pooled_macro_f1': 0.6889970040284997, 'final_pooled_macro_f1_pct': 68.89970040284997, 'final_pooled_weighted_f1': 0.6892762929935321, 'final_pooled_weighted_f1_pct': 68.92762929935321, 'final_pooled_macro_precision': 0.6920004411294777, 'final_pooled_macro_precision_pct': 69.20004411294777, 'final_pooled_macro_recall': 0.6890624154294503, 'final_pooled_macro_recall_pct': 68.90624154294504, 'final_pooled_per_class_f1': {'FAVOR': 0.635593220338983, 'AGAINST': 0.6794258373205742, 'NONE': 0.7519719544259421}, 'final_pooled_per_class_f1_pct': {'FAVOR': 63.559322033898304, 'AGAINST': 67.94258373205741, 'NONE': 75.19719544259421}} +2026-09-02 08:17:41,781 [INFO] __main__: efficiency={'training_wall_time_seconds': 976.5617710498627, 'training_wall_time_minutes': 16.27602951749771, 'peak_gpu_memory_allocated_gb': 10.40872573852539, 'peak_gpu_memory_reserved_gb': 10.587890625} diff --git a/outputs/trace_s7_contrastive_replay__x__s2__seed13/environment.json b/outputs/trace_s7_contrastive_replay__x__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..c68588431f9d310d3a82bbc4fd4293f1c008240a --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T08:17:49.828262+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..cbd6b53733a591969a0c4a00498b6cf9875049fd --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.json @@ -0,0 +1,2294 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.6902344189866794, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.624808935920047, + 0.5588476221387614, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.743662833768158, + 0.6551389465153578, + 0.5764754932617199, + 0.7159593594019823, + null, + null, + null, + null, + null, + null + ], + [ + 0.7453606091030789, + 0.6692171717171718, + 0.5918684000040498, + 0.7285974948896298, + 0.6323761472037615, + null, + null, + null, + null, + null + ], + [ + 0.6682912162131839, + 0.6766054383903354, + 0.5781693054420326, + 0.6916519694297473, + 0.6346249709886074, + 0.7151625129542339, + null, + null, + null, + null + ], + [ + 0.6221139971139972, + 0.6239914344973965, + 0.5653664259116197, + 0.6751771291648593, + 0.6413422794185045, + 0.6906858340820605, + 0.5501464097955325, + null, + null, + null + ], + [ + 0.6939311817360597, + 0.7090851535295979, + 0.6141077192937089, + 0.6993348371797096, + 0.6387944622931766, + 0.7034157240608853, + 0.5643281951693166, + 0.7010562515352493, + null, + null + ], + [ + 0.7121565500514967, + 0.6619776949221012, + 0.5831333671446511, + 0.7029181628644596, + 0.6672176469118737, + 0.730108162550359, + 0.5808845297396096, + 0.672404474610357, + 0.6359586160347753, + null + ], + [ + 0.6974358974358975, + 0.6572943546479982, + 0.5719683420700997, + 0.6545612640583898, + 0.6284076461785983, + 0.7339459760512392, + 0.5886601661249549, + 0.6629735389301352, + 0.6537829304639944, + 0.7800320850253172 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.5997024266560196, + "losses": { + "current_ce": 0.7054417070114252, + "replay_ce": 0.13542797391613323, + "dark_distillation": 0.35343743282737156, + "supervised_contrastive": 4.253260811892423, + "total": 1.4429144855701561, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6902344189866794, + "weighted_f1": 0.7223228267969563, + "macro_precision": 0.6751058433047626, + "macro_recall": 0.7257575757575757, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.7692307692307693, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 69.02344189866794, + "weighted_f1_pct": 72.23228267969563, + "macro_precision_pct": 67.51058433047626, + "macro_recall_pct": 72.57575757575758, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 76.92307692307693, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7658862876254181, + "macro_f1": 0.7660669685015048, + "weighted_f1": 0.7651397808062407, + "macro_precision": 0.7672422847295511, + "macro_recall": 0.7708636893736709, + "per_class_f1": { + "FAVOR": 0.7772511848341233, + "AGAINST": 0.75, + "NONE": 0.770949720670391 + }, + "accuracy_pct": 76.58862876254182, + "macro_f1_pct": 76.60669685015048, + "weighted_f1_pct": 76.51397808062407, + "macro_precision_pct": 76.72422847295512, + "macro_recall_pct": 77.08636893736708, + "per_class_f1_pct": { + "FAVOR": 77.72511848341233, + "AGAINST": 75.0, + "NONE": 77.09497206703911 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.5960432482171613, + "losses": { + "current_ce": 0.7449606120586395, + "replay_ce": 0.24123859715958437, + "dark_distillation": 0.3530231888095538, + "supervised_contrastive": 4.430700246493021, + "total": 1.6057808419068655, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6236559139784946, + "macro_f1": 0.624808935920047, + "weighted_f1": 0.6244751664106503, + "macro_precision": 0.6466049382716049, + "macro_recall": 0.7267115600448935, + "per_class_f1": { + "FAVOR": 0.5904761904761905, + "AGAINST": 0.6172839506172839, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.365591397849464, + "macro_f1_pct": 62.480893592004705, + "weighted_f1_pct": 62.44751664106503, + "macro_precision_pct": 64.6604938271605, + "macro_recall_pct": 72.67115600448935, + "per_class_f1_pct": { + "FAVOR": 59.04761904761905, + "AGAINST": 61.72839506172839, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5588476221387614, + "weighted_f1": 0.5317774873845476, + "macro_precision": 0.5532148343208251, + "macro_recall": 0.6213675213675214, + "per_class_f1": { + "FAVOR": 0.6582278481012658, + "AGAINST": 0.3516483516483517, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.88476221387614, + "weighted_f1_pct": 53.17774873845475, + "macro_precision_pct": 55.32148343208251, + "macro_recall_pct": 62.136752136752136, + "per_class_f1_pct": { + "FAVOR": 65.82278481012658, + "AGAINST": 35.16483516483517, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6621315192743764, + "macro_f1": 0.6555483645823419, + "weighted_f1": 0.643383521233118, + "macro_precision": 0.6756272401433692, + "macro_recall": 0.688689336967799, + "per_class_f1": { + "FAVOR": 0.7278481012658228, + "AGAINST": 0.518796992481203, + "NONE": 0.72 + }, + "accuracy_pct": 66.21315192743764, + "macro_f1_pct": 65.55483645823419, + "weighted_f1_pct": 64.33835212331181, + "macro_precision_pct": 67.56272401433692, + "macro_recall_pct": 68.8689336967799, + "per_class_f1_pct": { + "FAVOR": 72.78481012658227, + "AGAINST": 51.8796992481203, + "NONE": 72.0 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.7218628612811903, + "losses": { + "current_ce": 0.6569286040637804, + "replay_ce": 0.23603768616482831, + "dark_distillation": 0.3524670530056608, + "supervised_contrastive": 4.381954227668651, + "total": 1.5073952502098635, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.743662833768158, + "weighted_f1": 0.8504113517872276, + "macro_precision": 0.716332735247209, + "macro_recall": 0.8266715204465478, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.5, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 74.3662833768158, + "weighted_f1_pct": 85.04113517872275, + "macro_precision_pct": 71.63327352472089, + "macro_recall_pct": 82.66715204465478, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 50.0, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6551389465153578, + "weighted_f1": 0.6759328854553298, + "macro_precision": 0.6565185518843345, + "macro_recall": 0.722783389450056, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7071823204419889, + "NONE": 0.6521739130434783 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.51389465153578, + "weighted_f1_pct": 67.59328854553299, + "macro_precision_pct": 65.65185518843344, + "macro_recall_pct": 72.2783389450056, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 70.71823204419888, + "NONE": 65.21739130434783 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5764754932617199, + "weighted_f1": 0.5511683202996074, + "macro_precision": 0.5884615384615385, + "macro_recall": 0.6525641025641026, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.35294117647058826, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 57.64754932617199, + "weighted_f1_pct": 55.11683202996074, + "macro_precision_pct": 58.84615384615385, + "macro_recall_pct": 65.25641025641026, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 35.294117647058826, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7159593594019823, + "weighted_f1": 0.7197524128876589, + "macro_precision": 0.7284469343292873, + "macro_recall": 0.7289928946555452, + "per_class_f1": { + "FAVOR": 0.7377049180327869, + "AGAINST": 0.6363636363636364, + "NONE": 0.7738095238095238 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 71.59593594019823, + "weighted_f1_pct": 71.97524128876589, + "macro_precision_pct": 72.84469343292874, + "macro_recall_pct": 72.89928946555452, + "per_class_f1_pct": { + "FAVOR": 73.77049180327869, + "AGAINST": 63.63636363636363, + "NONE": 77.38095238095238 + } + } + ], + "accuracy": 0.6989079563182528, + "macro_f1": 0.694919903445249, + "weighted_f1": 0.690156868947569, + "macro_precision": 0.7082819072180775, + "macro_recall": 0.7143934496875674, + "per_class_f1": { + "FAVOR": 0.7327188940092166, + "AGAINST": 0.6020408163265306, + "NONE": 0.75 + }, + "accuracy_pct": 69.89079563182527, + "macro_f1_pct": 69.49199034452491, + "weighted_f1_pct": 69.0156868947569, + "macro_precision_pct": 70.82819072180774, + "macro_recall_pct": 71.43934496875673, + "per_class_f1_pct": { + "FAVOR": 73.27188940092167, + "AGAINST": 60.204081632653065, + "NONE": 75.0 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6396795692570341, + "losses": { + "current_ce": 0.6728339387820317, + "replay_ce": 0.24794298800138326, + "dark_distillation": 0.38301720665051386, + "supervised_contrastive": 4.286247200843615, + "total": 1.5409102629392575, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7453606091030789, + "weighted_f1": 0.839142521413086, + "macro_precision": 0.7116575591985428, + "macro_recall": 0.8248513529911419, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.5333333333333333, + "NONE": 0.8433734939759037 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 74.5360609103079, + "weighted_f1_pct": 83.9142521413086, + "macro_precision_pct": 71.16575591985428, + "macro_recall_pct": 82.48513529911419, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 53.333333333333336, + "NONE": 84.33734939759037 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6692171717171718, + "weighted_f1": 0.674993890518084, + "macro_precision": 0.6641994171405936, + "macro_recall": 0.7501683501683502, + "per_class_f1": { + "FAVOR": 0.6458333333333334, + "AGAINST": 0.6818181818181818, + "NONE": 0.68 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.92171717171718, + "weighted_f1_pct": 67.49938905180841, + "macro_precision_pct": 66.41994171405936, + "macro_recall_pct": 75.01683501683502, + "per_class_f1_pct": { + "FAVOR": 64.58333333333334, + "AGAINST": 68.18181818181817, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5918684000040498, + "weighted_f1": 0.5701170849780516, + "macro_precision": 0.5996950696426622, + "macro_recall": 0.6495726495726496, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.38636363636363635, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.186840000404985, + "weighted_f1_pct": 57.01170849780516, + "macro_precision_pct": 59.969506964266216, + "macro_recall_pct": 64.95726495726495, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 38.63636363636363, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.745, + "macro_f1": 0.7285974948896298, + "weighted_f1": 0.7375361155698235, + "macro_precision": 0.7548582995951417, + "macro_recall": 0.7367624343527958, + "per_class_f1": { + "FAVOR": 0.717948717948718, + "AGAINST": 0.6476190476190476, + "NONE": 0.8202247191011236 + }, + "accuracy_pct": 74.5, + "macro_f1_pct": 72.85974948896298, + "weighted_f1_pct": 73.75361155698235, + "macro_precision_pct": 75.48582995951418, + "macro_recall_pct": 73.67624343527957, + "per_class_f1_pct": { + "FAVOR": 71.7948717948718, + "AGAINST": 64.76190476190476, + "NONE": 82.02247191011236 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.6323761472037615, + "weighted_f1": 0.6666825172928251, + "macro_precision": 0.636881875591553, + "macro_recall": 0.6420179604645625, + "per_class_f1": { + "FAVOR": 0.7227722772277227, + "AGAINST": 0.4383561643835616, + "NONE": 0.736 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 63.23761472037614, + "weighted_f1_pct": 66.66825172928252, + "macro_precision_pct": 63.688187559155296, + "macro_recall_pct": 64.20179604645625, + "per_class_f1_pct": { + "FAVOR": 72.27722772277228, + "AGAINST": 43.83561643835616, + "NONE": 73.6 + } + } + ], + "accuracy": 0.7003567181926278, + "macro_f1": 0.6896708240743553, + "weighted_f1": 0.6895544997484406, + "macro_precision": 0.7065236674023606, + "macro_recall": 0.7005112418727979, + "per_class_f1": { + "FAVOR": 0.7305194805194806, + "AGAINST": 0.5733041575492341, + "NONE": 0.7651888341543513 + }, + "accuracy_pct": 70.03567181926277, + "macro_f1_pct": 68.96708240743553, + "weighted_f1_pct": 68.95544997484406, + "macro_precision_pct": 70.65236674023606, + "macro_recall_pct": 70.05112418727978, + "per_class_f1_pct": { + "FAVOR": 73.05194805194806, + "AGAINST": 57.330415754923415, + "NONE": 76.51888341543514 + } + }, + { + "after_task": 6, + "best_epoch": 15, + "best_dev_macro_f1": 0.8457912457912458, + "losses": { + "current_ce": 0.4921357918668676, + "replay_ce": 0.2652038168024134, + "dark_distillation": 0.38115012123490927, + "supervised_contrastive": 3.7923592388629914, + "total": 1.3271506115794183, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6682912162131839, + "weighted_f1": 0.8331248932928772, + "macro_precision": 0.6574074074074074, + "macro_recall": 0.696493144035918, + "per_class_f1": { + "FAVOR": 0.8549618320610687, + "AGAINST": 0.2857142857142857, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 66.8291216213184, + "weighted_f1_pct": 83.31248932928771, + "macro_precision_pct": 65.74074074074075, + "macro_recall_pct": 69.6493144035918, + "per_class_f1_pct": { + "FAVOR": 85.49618320610686, + "AGAINST": 28.57142857142857, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6766054383903354, + "weighted_f1": 0.6956732217564874, + "macro_precision": 0.6711917686523039, + "macro_recall": 0.7394500561167229, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7243243243243244, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.66054383903354, + "weighted_f1_pct": 69.56732217564874, + "macro_precision_pct": 67.11917686523039, + "macro_recall_pct": 73.94500561167229, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 72.43243243243244, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5781693054420326, + "weighted_f1": 0.552934901206348, + "macro_precision": 0.5761704401255593, + "macro_recall": 0.6303418803418804, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.37362637362637363, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 57.81693054420326, + "weighted_f1_pct": 55.2934901206348, + "macro_precision_pct": 57.61704401255593, + "macro_recall_pct": 63.03418803418804, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 37.362637362637365, + "NONE": 69.42148760330579 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6916519694297473, + "weighted_f1": 0.6999162257495591, + "macro_precision": 0.7061117136527856, + "macro_recall": 0.7028421377818969, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6428571428571429, + "NONE": 0.7654320987654321 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 69.16519694297473, + "weighted_f1_pct": 69.99162257495591, + "macro_precision_pct": 70.61117136527855, + "macro_recall_pct": 70.28421377818968, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 64.28571428571429, + "NONE": 76.5432098765432 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6346249709886074, + "weighted_f1": 0.6640069930069931, + "macro_precision": 0.6280663780663781, + "macro_recall": 0.6468218943947099, + "per_class_f1": { + "FAVOR": 0.7076923076923077, + "AGAINST": 0.4523809523809524, + "NONE": 0.743801652892562 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.46249709886074, + "weighted_f1_pct": 66.4006993006993, + "macro_precision_pct": 62.80663780663781, + "macro_recall_pct": 64.68218943947099, + "per_class_f1_pct": { + "FAVOR": 70.76923076923077, + "AGAINST": 45.23809523809524, + "NONE": 74.3801652892562 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7151625129542339, + "weighted_f1": 0.7617962167340081, + "macro_precision": 0.6942195059842118, + "macro_recall": 0.7543678160919539, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.8165680473372781, + "NONE": 0.71875 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 71.51625129542339, + "weighted_f1_pct": 76.17962167340082, + "macro_precision_pct": 69.42195059842118, + "macro_recall_pct": 75.43678160919539, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 81.65680473372781, + "NONE": 71.875 + } + } + ], + "accuracy": 0.6960486322188449, + "macro_f1": 0.6962752756321606, + "weighted_f1": 0.6925823536077086, + "macro_precision": 0.6999464416790427, + "macro_recall": 0.705270579940067, + "per_class_f1": { + "FAVOR": 0.7050147492625368, + "AGAINST": 0.6412213740458015, + "NONE": 0.7425897035881436 + }, + "accuracy_pct": 69.6048632218845, + "macro_f1_pct": 69.62752756321606, + "weighted_f1_pct": 69.25823536077085, + "macro_precision_pct": 69.99464416790427, + "macro_recall_pct": 70.5270579940067, + "per_class_f1_pct": { + "FAVOR": 70.50147492625368, + "AGAINST": 64.12213740458014, + "NONE": 74.25897035881435 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.5939530234882558, + "losses": { + "current_ce": 0.5883172781178445, + "replay_ce": 0.223132795456684, + "dark_distillation": 0.4028841298637968, + "supervised_contrastive": 4.22716889814897, + "total": 1.4356090343359744, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6221139971139972, + "weighted_f1": 0.8268829253342528, + "macro_precision": 0.6233618233618233, + "macro_recall": 0.6298264773692512, + "per_class_f1": { + "FAVOR": 0.8484848484848485, + "AGAINST": 0.14285714285714285, + "NONE": 0.875 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 62.211399711399714, + "weighted_f1_pct": 82.68829253342528, + "macro_precision_pct": 62.33618233618233, + "macro_recall_pct": 62.982647736925124, + "per_class_f1_pct": { + "FAVOR": 84.84848484848484, + "AGAINST": 14.285714285714285, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.6239914344973965, + "weighted_f1": 0.6374270682692293, + "macro_precision": 0.6376633550546594, + "macro_recall": 0.7038720538720539, + "per_class_f1": { + "FAVOR": 0.5794392523364486, + "AGAINST": 0.6551724137931034, + "NONE": 0.6373626373626373 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.39914344973965, + "weighted_f1_pct": 63.742706826922934, + "macro_precision_pct": 63.766335505465946, + "macro_recall_pct": 70.38720538720538, + "per_class_f1_pct": { + "FAVOR": 57.943925233644855, + "AGAINST": 65.51724137931035, + "NONE": 63.73626373626373 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5653664259116197, + "weighted_f1": 0.5338596219433611, + "macro_precision": 0.5712820512820512, + "macro_recall": 0.6452991452991453, + "per_class_f1": { + "FAVOR": 0.6829268292682927, + "AGAINST": 0.32941176470588235, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 56.536642591161964, + "weighted_f1_pct": 53.38596219433611, + "macro_precision_pct": 57.128205128205124, + "macro_recall_pct": 64.52991452991454, + "per_class_f1_pct": { + "FAVOR": 68.29268292682927, + "AGAINST": 32.94117647058823, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6751771291648593, + "weighted_f1": 0.6845349197496436, + "macro_precision": 0.7251082991336069, + "macro_recall": 0.6966017917825146, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5858585858585859, + "NONE": 0.7730061349693251 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 67.51771291648593, + "weighted_f1_pct": 68.45349197496436, + "macro_precision_pct": 72.51082991336068, + "macro_recall_pct": 69.66017917825145, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 58.58585858585859, + "NONE": 77.30061349693251 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.69, + "macro_f1": 0.6413422794185045, + "weighted_f1": 0.6828094538854974, + "macro_precision": 0.6450165121677648, + "macro_recall": 0.6432536180108995, + "per_class_f1": { + "FAVOR": 0.7559808612440191, + "AGAINST": 0.4266666666666667, + "NONE": 0.7413793103448276 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 64.13422794185045, + "weighted_f1_pct": 68.28094538854974, + "macro_precision_pct": 64.50165121677648, + "macro_recall_pct": 64.32536180108995, + "per_class_f1_pct": { + "FAVOR": 75.5980861244019, + "AGAINST": 42.66666666666667, + "NONE": 74.13793103448276 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6906858340820605, + "weighted_f1": 0.7277935728439735, + "macro_precision": 0.6795174454524763, + "macro_recall": 0.7502098950524737, + "per_class_f1": { + "FAVOR": 0.5428571428571428, + "AGAINST": 0.7672955974842768, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.06858340820605, + "weighted_f1_pct": 72.77935728439735, + "macro_precision_pct": 67.95174454524762, + "macro_recall_pct": 75.02098950524737, + "per_class_f1_pct": { + "FAVOR": 54.285714285714285, + "AGAINST": 76.72955974842768, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5501464097955325, + "weighted_f1": 0.5667102365163307, + "macro_precision": 0.5405684754521963, + "macro_recall": 0.5682643427741466, + "per_class_f1": { + "FAVOR": 0.49122807017543857, + "AGAINST": 0.6137566137566137, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 55.014640979553256, + "weighted_f1_pct": 56.67102365163307, + "macro_precision_pct": 54.05684754521963, + "macro_recall_pct": 56.82643427741466, + "per_class_f1_pct": { + "FAVOR": 49.122807017543856, + "AGAINST": 61.37566137566137, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6610025488530161, + "macro_f1": 0.6629258206941381, + "weighted_f1": 0.6552159949421067, + "macro_precision": 0.6720840096396316, + "macro_recall": 0.6795214077335731, + "per_class_f1": { + "FAVOR": 0.6737089201877934, + "AGAINST": 0.5937106918238994, + "NONE": 0.7213578500707214 + }, + "accuracy_pct": 66.10025488530161, + "macro_f1_pct": 66.29258206941381, + "weighted_f1_pct": 65.52159949421068, + "macro_precision_pct": 67.20840096396316, + "macro_recall_pct": 67.9521407733573, + "per_class_f1_pct": { + "FAVOR": 67.37089201877934, + "AGAINST": 59.37106918238994, + "NONE": 72.13578500707214 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.6993172124101915, + "losses": { + "current_ce": 0.5525923003753026, + "replay_ce": 0.21369035108522935, + "dark_distillation": 0.37410288886590437, + "supervised_contrastive": 4.271607911948002, + "total": 1.3804948994607635, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6939311817360597, + "weighted_f1": 0.8302973617024987, + "macro_precision": 0.6820028011204481, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.36363636363636365, + "NONE": 0.8888888888888888 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.39311817360597, + "weighted_f1_pct": 83.02973617024986, + "macro_precision_pct": 68.20028011204481, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 36.36363636363637, + "NONE": 88.88888888888889 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7090851535295979, + "weighted_f1": 0.7306007413534295, + "macro_precision": 0.6909510930809105, + "macro_recall": 0.7513468013468013, + "per_class_f1": { + "FAVOR": 0.691358024691358, + "AGAINST": 0.7692307692307693, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.9085153529598, + "weighted_f1_pct": 73.06007413534294, + "macro_precision_pct": 69.09510930809105, + "macro_recall_pct": 75.13468013468014, + "per_class_f1_pct": { + "FAVOR": 69.1358024691358, + "AGAINST": 76.92307692307693, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6141077192937089, + "weighted_f1": 0.5815801679476806, + "macro_precision": 0.6265432098765432, + "macro_recall": 0.6717948717948717, + "per_class_f1": { + "FAVOR": 0.7428571428571429, + "AGAINST": 0.39080459770114945, + "NONE": 0.7086614173228346 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 61.41077192937089, + "weighted_f1_pct": 58.158016794768066, + "macro_precision_pct": 62.65432098765432, + "macro_recall_pct": 67.17948717948717, + "per_class_f1_pct": { + "FAVOR": 74.28571428571429, + "AGAINST": 39.08045977011494, + "NONE": 70.86614173228347 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.6993348371797096, + "weighted_f1": 0.7084004858580601, + "macro_precision": 0.7098180076628352, + "macro_recall": 0.7048038307074451, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6548672566371682, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 69.93348371797096, + "weighted_f1_pct": 70.84004858580602, + "macro_precision_pct": 70.98180076628353, + "macro_recall_pct": 70.48038307074451, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 65.48672566371681, + "NONE": 77.64705882352942 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6387944622931766, + "weighted_f1": 0.6673951038757339, + "macro_precision": 0.6322695035460993, + "macro_recall": 0.6562224183583408, + "per_class_f1": { + "FAVOR": 0.7127659574468085, + "AGAINST": 0.47191011235955055, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.87944622931766, + "weighted_f1_pct": 66.73951038757339, + "macro_precision_pct": 63.22695035460993, + "macro_recall_pct": 65.62224183583409, + "per_class_f1_pct": { + "FAVOR": 71.27659574468085, + "AGAINST": 47.19101123595505, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7034157240608853, + "weighted_f1": 0.7592832178868414, + "macro_precision": 0.6881465254959233, + "macro_recall": 0.7252373813093453, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8228571428571428, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 70.34157240608853, + "weighted_f1_pct": 75.92832178868414, + "macro_precision_pct": 68.81465254959232, + "macro_recall_pct": 72.52373813093453, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 82.28571428571428, + "NONE": 74.19354838709677 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5643281951693166, + "weighted_f1": 0.5861537871867434, + "macro_precision": 0.5560062737482845, + "macro_recall": 0.5784313725490197, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.6428571428571429, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 56.432819516931666, + "weighted_f1_pct": 58.615378718674336, + "macro_precision_pct": 55.60062737482845, + "macro_recall_pct": 57.84313725490197, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 64.28571428571429, + "NONE": 54.54545454545454 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7208121827411168, + "macro_f1": 0.7010562515352493, + "weighted_f1": 0.723390778154758, + "macro_precision": 0.6892546722651517, + "macro_recall": 0.726820521719012, + "per_class_f1": { + "FAVOR": 0.6376811594202898, + "AGAINST": 0.7536231884057971, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 72.08121827411168, + "macro_f1_pct": 70.10562515352493, + "weighted_f1_pct": 72.3390778154758, + "macro_precision_pct": 68.92546722651517, + "macro_recall_pct": 72.6820521719012, + "per_class_f1_pct": { + "FAVOR": 63.76811594202898, + "AGAINST": 75.36231884057972, + "NONE": 71.1864406779661 + } + } + ], + "accuracy": 0.6928675400291121, + "macro_f1": 0.6940431443519918, + "weighted_f1": 0.6915404324901548, + "macro_precision": 0.690797865313839, + "macro_recall": 0.7042599853888952, + "per_class_f1": { + "FAVOR": 0.6790123456790124, + "AGAINST": 0.6771217712177122, + "NONE": 0.7259953161592506 + }, + "accuracy_pct": 69.28675400291121, + "macro_f1_pct": 69.40431443519918, + "weighted_f1_pct": 69.15404324901549, + "macro_precision_pct": 69.0797865313839, + "macro_recall_pct": 70.42599853888952, + "per_class_f1_pct": { + "FAVOR": 67.90123456790124, + "AGAINST": 67.71217712177122, + "NONE": 72.59953161592506 + } + }, + { + "after_task": 9, + "best_epoch": 15, + "best_dev_macro_f1": 0.6514042025820047, + "losses": { + "current_ce": 0.5333497233334041, + "replay_ce": 0.21578899905795143, + "dark_distillation": 0.34363774250424095, + "supervised_contrastive": 4.232672637606424, + "total": 1.3442248700157045, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7121565500514967, + "weighted_f1": 0.843411432538111, + "macro_precision": 0.6959064327485379, + "macro_recall": 0.8198762286130323, + "per_class_f1": { + "FAVOR": 0.8503937007874016, + "AGAINST": 0.4, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.21565500514967, + "weighted_f1_pct": 84.3411432538111, + "macro_precision_pct": 69.59064327485379, + "macro_recall_pct": 81.98762286130324, + "per_class_f1_pct": { + "FAVOR": 85.03937007874016, + "AGAINST": 40.0, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6619776949221012, + "weighted_f1": 0.6782440221344732, + "macro_precision": 0.6533807462378891, + "macro_recall": 0.7231762065095397, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.7027027027027027, + "NONE": 0.6595744680851063 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.19776949221013, + "weighted_f1_pct": 67.82440221344731, + "macro_precision_pct": 65.33807462378891, + "macro_recall_pct": 72.31762065095397, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 70.27027027027027, + "NONE": 65.95744680851064 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5831333671446511, + "weighted_f1": 0.554765609987765, + "macro_precision": 0.5824684003513193, + "macro_recall": 0.6423076923076924, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.3595505617977528, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 58.313336714465116, + "weighted_f1_pct": 55.476560998776506, + "macro_precision_pct": 58.24684003513193, + "macro_recall_pct": 64.23076923076924, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 35.95505617977528, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7029181628644596, + "weighted_f1": 0.710612340035589, + "macro_precision": 0.7182772212598226, + "macro_recall": 0.7150602409638555, + "per_class_f1": { + "FAVOR": 0.6991869918699187, + "AGAINST": 0.6238532110091743, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.29181628644596, + "weighted_f1_pct": 71.06123400355891, + "macro_precision_pct": 71.82772212598226, + "macro_recall_pct": 71.50602409638554, + "per_class_f1_pct": { + "FAVOR": 69.91869918699187, + "AGAINST": 62.38532110091744, + "NONE": 78.57142857142857 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.705, + "macro_f1": 0.6672176469118737, + "weighted_f1": 0.7033029045124866, + "macro_precision": 0.663494231936855, + "macro_recall": 0.6727118620322504, + "per_class_f1": { + "FAVOR": 0.7623762376237624, + "AGAINST": 0.4634146341463415, + "NONE": 0.7758620689655172 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 66.72176469118737, + "weighted_f1_pct": 70.33029045124866, + "macro_precision_pct": 66.3494231936855, + "macro_recall_pct": 67.27118620322504, + "per_class_f1_pct": { + "FAVOR": 76.23762376237624, + "AGAINST": 46.34146341463415, + "NONE": 77.58620689655173 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.730108162550359, + "weighted_f1": 0.7792938240989353, + "macro_precision": 0.7105820105820105, + "macro_recall": 0.759720139930035, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8323699421965318, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.0108162550359, + "weighted_f1_pct": 77.92938240989353, + "macro_precision_pct": 71.05820105820105, + "macro_recall_pct": 75.9720139930035, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 83.23699421965318, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5808845297396096, + "weighted_f1": 0.5983629479537357, + "macro_precision": 0.5712678169542386, + "macro_recall": 0.5994916485112564, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.6382978723404256, + "NONE": 0.5526315789473685 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.088452973960955, + "weighted_f1_pct": 59.83629479537357, + "macro_precision_pct": 57.12678169542386, + "macro_recall_pct": 59.94916485112564, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 63.829787234042556, + "NONE": 55.26315789473685 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.672404474610357, + "weighted_f1": 0.6826960401497577, + "macro_precision": 0.6572688914461066, + "macro_recall": 0.7318263320345569, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.6875, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 67.2404474610357, + "weighted_f1_pct": 68.26960401497577, + "macro_precision_pct": 65.72688914461065, + "macro_recall_pct": 73.1826332034557, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 68.75, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6359586160347753, + "weighted_f1": 0.6240807606872479, + "macro_precision": 0.6326291079812206, + "macro_recall": 0.6426045599487333, + "per_class_f1": { + "FAVOR": 0.6052631578947368, + "AGAINST": 0.6049382716049383, + "NONE": 0.6976744186046512 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 63.59586160347753, + "weighted_f1_pct": 62.408076068724796, + "macro_precision_pct": 63.262910798122064, + "macro_recall_pct": 64.26045599487334, + "per_class_f1_pct": { + "FAVOR": 60.526315789473685, + "AGAINST": 60.49382716049383, + "NONE": 69.76744186046511 + } + } + ], + "accuracy": 0.6804320203303685, + "macro_f1": 0.6844792798659297, + "weighted_f1": 0.6778736543002519, + "macro_precision": 0.6810885152102513, + "macro_recall": 0.6971439048163566, + "per_class_f1": { + "FAVOR": 0.6764132553606238, + "AGAINST": 0.6416666666666667, + "NONE": 0.735357917570499 + }, + "accuracy_pct": 68.04320203303685, + "macro_f1_pct": 68.44792798659297, + "weighted_f1_pct": 67.78736543002519, + "macro_precision_pct": 68.10885152102513, + "macro_recall_pct": 69.71439048163566, + "per_class_f1_pct": { + "FAVOR": 67.64132553606238, + "AGAINST": 64.16666666666667, + "NONE": 73.53579175704989 + } + }, + { + "after_task": 10, + "best_epoch": 14, + "best_dev_macro_f1": 0.7270304950789348, + "losses": { + "current_ce": 0.4160077131042878, + "replay_ce": 0.20025991598765056, + "dark_distillation": 0.37701009716838596, + "supervised_contrastive": 4.04466478129228, + "total": 1.2092391716937223, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6974358974358975, + "weighted_f1": 0.8511912865895167, + "macro_precision": 0.6885346885346886, + "macro_recall": 0.7631598107025845, + "per_class_f1": { + "FAVOR": 0.8615384615384616, + "AGAINST": 0.3333333333333333, + "NONE": 0.8974358974358975 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 69.74358974358975, + "weighted_f1_pct": 85.11912865895167, + "macro_precision_pct": 68.85346885346885, + "macro_recall_pct": 76.31598107025846, + "per_class_f1_pct": { + "FAVOR": 86.15384615384616, + "AGAINST": 33.33333333333333, + "NONE": 89.74358974358975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6572943546479982, + "weighted_f1": 0.6771634071036368, + "macro_precision": 0.6626916274451045, + "macro_recall": 0.7271043771043771, + "per_class_f1": { + "FAVOR": 0.594059405940594, + "AGAINST": 0.7039106145251397, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.72943546479982, + "weighted_f1_pct": 67.71634071036368, + "macro_precision_pct": 66.26916274451045, + "macro_recall_pct": 72.71043771043772, + "per_class_f1_pct": { + "FAVOR": 59.4059405940594, + "AGAINST": 70.39106145251397, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5719683420700997, + "weighted_f1": 0.541319331344217, + "macro_precision": 0.5603174603174602, + "macro_recall": 0.6269230769230769, + "per_class_f1": { + "FAVOR": 0.6933333333333334, + "AGAINST": 0.3617021276595745, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.19683420700997, + "weighted_f1_pct": 54.13193313442169, + "macro_precision_pct": 56.031746031746025, + "macro_recall_pct": 62.69230769230769, + "per_class_f1_pct": { + "FAVOR": 69.33333333333334, + "AGAINST": 36.17021276595745, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.665, + "macro_f1": 0.6545612640583898, + "weighted_f1": 0.6624863972315785, + "macro_precision": 0.6692592592592592, + "macro_recall": 0.6678869323447637, + "per_class_f1": { + "FAVOR": 0.6456692913385826, + "AGAINST": 0.5818181818181818, + "NONE": 0.7361963190184049 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 65.45612640583897, + "weighted_f1_pct": 66.24863972315785, + "macro_precision_pct": 66.92592592592592, + "macro_recall_pct": 66.78869323447637, + "per_class_f1_pct": { + "FAVOR": 64.56692913385827, + "AGAINST": 58.18181818181818, + "NONE": 73.61963190184049 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6284076461785983, + "weighted_f1": 0.6666116211563775, + "macro_precision": 0.6325354384177914, + "macro_recall": 0.6251754717774135, + "per_class_f1": { + "FAVOR": 0.7307692307692307, + "AGAINST": 0.4186046511627907, + "NONE": 0.7358490566037735 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 62.84076461785983, + "weighted_f1_pct": 66.66116211563775, + "macro_precision_pct": 63.253543841779134, + "macro_recall_pct": 62.51754717774135, + "per_class_f1_pct": { + "FAVOR": 73.07692307692307, + "AGAINST": 41.86046511627907, + "NONE": 73.58490566037736 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7339459760512392, + "weighted_f1": 0.7800231171536147, + "macro_precision": 0.7123618105651589, + "macro_recall": 0.7694302848575711, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.8304093567251462, + "NONE": 0.8 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.39459760512392, + "weighted_f1_pct": 78.00231171536147, + "macro_precision_pct": 71.23618105651589, + "macro_recall_pct": 76.94302848575711, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 83.04093567251462, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.5886601661249549, + "weighted_f1": 0.6139932713691053, + "macro_precision": 0.5828113828113829, + "macro_recall": 0.599128540305011, + "per_class_f1": { + "FAVOR": 0.5641025641025641, + "AGAINST": 0.6666666666666666, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 58.866016612495486, + "weighted_f1_pct": 61.39932713691053, + "macro_precision_pct": 58.28113828113829, + "macro_recall_pct": 59.912854030501094, + "per_class_f1_pct": { + "FAVOR": 56.41025641025641, + "AGAINST": 66.66666666666666, + "NONE": 53.52112676056338 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6629735389301352, + "weighted_f1": 0.6716446294648164, + "macro_precision": 0.6455059277639923, + "macro_recall": 0.7079540853616856, + "per_class_f1": { + "FAVOR": 0.6588235294117647, + "AGAINST": 0.6893203883495146, + "NONE": 0.6407766990291263 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.29735389301352, + "weighted_f1_pct": 67.16446294648163, + "macro_precision_pct": 64.55059277639923, + "macro_recall_pct": 70.79540853616855, + "per_class_f1_pct": { + "FAVOR": 65.88235294117646, + "AGAINST": 68.93203883495146, + "NONE": 64.07766990291263 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.645, + "macro_f1": 0.6537829304639944, + "weighted_f1": 0.6460569523376525, + "macro_precision": 0.6687000726216413, + "macro_recall": 0.6433226013171813, + "per_class_f1": { + "FAVOR": 0.6265060240963856, + "AGAINST": 0.6415094339622641, + "NONE": 0.6933333333333334 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 65.37829304639943, + "weighted_f1_pct": 64.60569523376525, + "macro_precision_pct": 66.87000726216414, + "macro_recall_pct": 64.33226013171813, + "per_class_f1_pct": { + "FAVOR": 62.65060240963856, + "AGAINST": 64.15094339622641, + "NONE": 69.33333333333334 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7800320850253172, + "weighted_f1": 0.8132930515867048, + "macro_precision": 0.7630920859076199, + "macro_recall": 0.8046865456160809, + "per_class_f1": { + "FAVOR": 0.7, + "AGAINST": 0.7868852459016393, + "NONE": 0.8532110091743119 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 78.00320850253172, + "weighted_f1_pct": 81.32930515867048, + "macro_precision_pct": 76.30920859076198, + "macro_recall_pct": 80.46865456160809, + "per_class_f1_pct": { + "FAVOR": 70.0, + "AGAINST": 78.68852459016394, + "NONE": 85.3211009174312 + } + } + ], + "accuracy": 0.6860202931228861, + "macro_f1": 0.6885414110582025, + "weighted_f1": 0.6855558200102561, + "macro_precision": 0.6884620373937734, + "macro_recall": 0.6963016369121294, + "per_class_f1": { + "FAVOR": 0.6737777777777778, + "AGAINST": 0.6551982049364248, + "NONE": 0.7366482504604052 + }, + "accuracy_pct": 68.60202931228862, + "macro_f1_pct": 68.85414110582026, + "weighted_f1_pct": 68.55558200102561, + "macro_precision_pct": 68.84620373937734, + "macro_recall_pct": 69.63016369121294, + "per_class_f1_pct": { + "FAVOR": 67.37777777777778, + "AGAINST": 65.51982049364248, + "NONE": 73.66482504604052 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6629062200986624, + "average_forgetting": 0.03184929955844933, + "backward_transfer": -0.013263273288024013, + "final_mean_target_macro_f1_pct": 66.29062200986624, + "average_forgetting_pct": 3.1849299558449333, + "backward_transfer_pct": -1.3263273288024013, + "final_pooled_accuracy": 0.6860202931228861, + "final_pooled_accuracy_pct": 68.60202931228862, + "final_pooled_macro_f1": 0.6885414110582025, + "final_pooled_macro_f1_pct": 68.85414110582026, + "final_pooled_weighted_f1": 0.6855558200102561, + "final_pooled_weighted_f1_pct": 68.55558200102561, + "final_pooled_macro_precision": 0.6884620373937734, + "final_pooled_macro_precision_pct": 68.84620373937734, + "final_pooled_macro_recall": 0.6963016369121294, + "final_pooled_macro_recall_pct": 69.63016369121294, + "final_pooled_per_class_f1": { + "FAVOR": 0.6737777777777778, + "AGAINST": 0.6551982049364248, + "NONE": 0.7366482504604052 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 67.37777777777778, + "AGAINST": 65.51982049364248, + "NONE": 73.66482504604052 + } + }, + "efficiency": { + "training_wall_time_seconds": 989.7757089319639, + "training_wall_time_minutes": 16.49626181553273, + "peak_gpu_memory_allocated_gb": 10.41092300415039, + "peak_gpu_memory_reserved_gb": 10.630859375 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..bde7b9ff81d08cd61d58a902116aad77feb8a21f --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s2__seed13/metrics.partial.json @@ -0,0 +1,2178 @@ +{ + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.6902344189866794, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.734841290396846, + 0.624808935920047, + 0.5588476221387614, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.743662833768158, + 0.6551389465153578, + 0.5764754932617199, + 0.7159593594019823, + null, + null, + null, + null, + null, + null + ], + [ + 0.7453606091030789, + 0.6692171717171718, + 0.5918684000040498, + 0.7285974948896298, + 0.6323761472037615, + null, + null, + null, + null, + null + ], + [ + 0.6682912162131839, + 0.6766054383903354, + 0.5781693054420326, + 0.6916519694297473, + 0.6346249709886074, + 0.7151625129542339, + null, + null, + null, + null + ], + [ + 0.6221139971139972, + 0.6239914344973965, + 0.5653664259116197, + 0.6751771291648593, + 0.6413422794185045, + 0.6906858340820605, + 0.5501464097955325, + null, + null, + null + ], + [ + 0.6939311817360597, + 0.7090851535295979, + 0.6141077192937089, + 0.6993348371797096, + 0.6387944622931766, + 0.7034157240608853, + 0.5643281951693166, + 0.7010562515352493, + null, + null + ], + [ + 0.7121565500514967, + 0.6619776949221012, + 0.5831333671446511, + 0.7029181628644596, + 0.6672176469118737, + 0.730108162550359, + 0.5808845297396096, + 0.672404474610357, + 0.6359586160347753, + null + ], + [ + 0.6974358974358975, + 0.6572943546479982, + 0.5719683420700997, + 0.6545612640583898, + 0.6284076461785983, + 0.7339459760512392, + 0.5886601661249549, + 0.6629735389301352, + 0.6537829304639944, + 0.7800320850253172 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.5997024266560196, + "losses": { + "current_ce": 0.7054417070114252, + "replay_ce": 0.13542797391613323, + "dark_distillation": 0.35343743282737156, + "supervised_contrastive": 4.253260811892423, + "total": 1.4429144855701561, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6902344189866794, + "weighted_f1": 0.7223228267969563, + "macro_precision": 0.6751058433047626, + "macro_recall": 0.7257575757575757, + "per_class_f1": { + "FAVOR": 0.6075949367088608, + "AGAINST": 0.7692307692307693, + "NONE": 0.6938775510204082 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 69.02344189866794, + "weighted_f1_pct": 72.23228267969563, + "macro_precision_pct": 67.51058433047626, + "macro_recall_pct": 72.57575757575758, + "per_class_f1_pct": { + "FAVOR": 60.75949367088608, + "AGAINST": 76.92307692307693, + "NONE": 69.38775510204081 + } + } + ], + "accuracy": 0.7658862876254181, + "macro_f1": 0.7660669685015048, + "weighted_f1": 0.7651397808062407, + "macro_precision": 0.7672422847295511, + "macro_recall": 0.7708636893736709, + "per_class_f1": { + "FAVOR": 0.7772511848341233, + "AGAINST": 0.75, + "NONE": 0.770949720670391 + }, + "accuracy_pct": 76.58862876254182, + "macro_f1_pct": 76.60669685015048, + "weighted_f1_pct": 76.51397808062407, + "macro_precision_pct": 76.72422847295512, + "macro_recall_pct": 77.08636893736708, + "per_class_f1_pct": { + "FAVOR": 77.72511848341233, + "AGAINST": 75.0, + "NONE": 77.09497206703911 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.5960432482171613, + "losses": { + "current_ce": 0.7449606120586395, + "replay_ce": 0.24123859715958437, + "dark_distillation": 0.3530231888095538, + "supervised_contrastive": 4.430700246493021, + "total": 1.6057808419068655, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8495575221238938, + "macro_f1": 0.734841290396846, + "weighted_f1": 0.8550316721408165, + "macro_precision": 0.7141025641025641, + "macro_recall": 0.7731100594588035, + "per_class_f1": { + "FAVOR": 0.8787878787878788, + "AGAINST": 0.46153846153846156, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 84.95575221238938, + "macro_f1_pct": 73.4841290396846, + "weighted_f1_pct": 85.50316721408166, + "macro_precision_pct": 71.41025641025641, + "macro_recall_pct": 77.31100594588035, + "per_class_f1_pct": { + "FAVOR": 87.87878787878788, + "AGAINST": 46.15384615384615, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6236559139784946, + "macro_f1": 0.624808935920047, + "weighted_f1": 0.6244751664106503, + "macro_precision": 0.6466049382716049, + "macro_recall": 0.7267115600448935, + "per_class_f1": { + "FAVOR": 0.5904761904761905, + "AGAINST": 0.6172839506172839, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.365591397849464, + "macro_f1_pct": 62.480893592004705, + "weighted_f1_pct": 62.44751664106503, + "macro_precision_pct": 64.6604938271605, + "macro_recall_pct": 72.67115600448935, + "per_class_f1_pct": { + "FAVOR": 59.04761904761905, + "AGAINST": 61.72839506172839, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5588476221387614, + "weighted_f1": 0.5317774873845476, + "macro_precision": 0.5532148343208251, + "macro_recall": 0.6213675213675214, + "per_class_f1": { + "FAVOR": 0.6582278481012658, + "AGAINST": 0.3516483516483517, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.88476221387614, + "weighted_f1_pct": 53.17774873845475, + "macro_precision_pct": 55.32148343208251, + "macro_recall_pct": 62.136752136752136, + "per_class_f1_pct": { + "FAVOR": 65.82278481012658, + "AGAINST": 35.16483516483517, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6621315192743764, + "macro_f1": 0.6555483645823419, + "weighted_f1": 0.643383521233118, + "macro_precision": 0.6756272401433692, + "macro_recall": 0.688689336967799, + "per_class_f1": { + "FAVOR": 0.7278481012658228, + "AGAINST": 0.518796992481203, + "NONE": 0.72 + }, + "accuracy_pct": 66.21315192743764, + "macro_f1_pct": 65.55483645823419, + "weighted_f1_pct": 64.33835212331181, + "macro_precision_pct": 67.56272401433692, + "macro_recall_pct": 68.8689336967799, + "per_class_f1_pct": { + "FAVOR": 72.78481012658227, + "AGAINST": 51.8796992481203, + "NONE": 72.0 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.7218628612811903, + "losses": { + "current_ce": 0.6569286040637804, + "replay_ce": 0.23603768616482831, + "dark_distillation": 0.3524670530056608, + "supervised_contrastive": 4.381954227668651, + "total": 1.5073952502098635, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.743662833768158, + "weighted_f1": 0.8504113517872276, + "macro_precision": 0.716332735247209, + "macro_recall": 0.8266715204465478, + "per_class_f1": { + "FAVOR": 0.8702290076335878, + "AGAINST": 0.5, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 74.3662833768158, + "weighted_f1_pct": 85.04113517872275, + "macro_precision_pct": 71.63327352472089, + "macro_recall_pct": 82.66715204465478, + "per_class_f1_pct": { + "FAVOR": 87.02290076335878, + "AGAINST": 50.0, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6551389465153578, + "weighted_f1": 0.6759328854553298, + "macro_precision": 0.6565185518843345, + "macro_recall": 0.722783389450056, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7071823204419889, + "NONE": 0.6521739130434783 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.51389465153578, + "weighted_f1_pct": 67.59328854553299, + "macro_precision_pct": 65.65185518843344, + "macro_recall_pct": 72.2783389450056, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 70.71823204419888, + "NONE": 65.21739130434783 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5764754932617199, + "weighted_f1": 0.5511683202996074, + "macro_precision": 0.5884615384615385, + "macro_recall": 0.6525641025641026, + "per_class_f1": { + "FAVOR": 0.6585365853658537, + "AGAINST": 0.35294117647058826, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 57.64754932617199, + "weighted_f1_pct": 55.11683202996074, + "macro_precision_pct": 58.84615384615385, + "macro_recall_pct": 65.25641025641026, + "per_class_f1_pct": { + "FAVOR": 65.85365853658537, + "AGAINST": 35.294117647058826, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7159593594019823, + "weighted_f1": 0.7197524128876589, + "macro_precision": 0.7284469343292873, + "macro_recall": 0.7289928946555452, + "per_class_f1": { + "FAVOR": 0.7377049180327869, + "AGAINST": 0.6363636363636364, + "NONE": 0.7738095238095238 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 71.59593594019823, + "weighted_f1_pct": 71.97524128876589, + "macro_precision_pct": 72.84469343292874, + "macro_recall_pct": 72.89928946555452, + "per_class_f1_pct": { + "FAVOR": 73.77049180327869, + "AGAINST": 63.63636363636363, + "NONE": 77.38095238095238 + } + } + ], + "accuracy": 0.6989079563182528, + "macro_f1": 0.694919903445249, + "weighted_f1": 0.690156868947569, + "macro_precision": 0.7082819072180775, + "macro_recall": 0.7143934496875674, + "per_class_f1": { + "FAVOR": 0.7327188940092166, + "AGAINST": 0.6020408163265306, + "NONE": 0.75 + }, + "accuracy_pct": 69.89079563182527, + "macro_f1_pct": 69.49199034452491, + "weighted_f1_pct": 69.0156868947569, + "macro_precision_pct": 70.82819072180774, + "macro_recall_pct": 71.43934496875673, + "per_class_f1_pct": { + "FAVOR": 73.27188940092167, + "AGAINST": 60.204081632653065, + "NONE": 75.0 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6396795692570341, + "losses": { + "current_ce": 0.6728339387820317, + "replay_ce": 0.24794298800138326, + "dark_distillation": 0.38301720665051386, + "supervised_contrastive": 4.286247200843615, + "total": 1.5409102629392575, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7453606091030789, + "weighted_f1": 0.839142521413086, + "macro_precision": 0.7116575591985428, + "macro_recall": 0.8248513529911419, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.5333333333333333, + "NONE": 0.8433734939759037 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 74.5360609103079, + "weighted_f1_pct": 83.9142521413086, + "macro_precision_pct": 71.16575591985428, + "macro_recall_pct": 82.48513529911419, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 53.333333333333336, + "NONE": 84.33734939759037 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6692171717171718, + "weighted_f1": 0.674993890518084, + "macro_precision": 0.6641994171405936, + "macro_recall": 0.7501683501683502, + "per_class_f1": { + "FAVOR": 0.6458333333333334, + "AGAINST": 0.6818181818181818, + "NONE": 0.68 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.92171717171718, + "weighted_f1_pct": 67.49938905180841, + "macro_precision_pct": 66.41994171405936, + "macro_recall_pct": 75.01683501683502, + "per_class_f1_pct": { + "FAVOR": 64.58333333333334, + "AGAINST": 68.18181818181817, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5918684000040498, + "weighted_f1": 0.5701170849780516, + "macro_precision": 0.5996950696426622, + "macro_recall": 0.6495726495726496, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.38636363636363635, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.186840000404985, + "weighted_f1_pct": 57.01170849780516, + "macro_precision_pct": 59.969506964266216, + "macro_recall_pct": 64.95726495726495, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 38.63636363636363, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.745, + "macro_f1": 0.7285974948896298, + "weighted_f1": 0.7375361155698235, + "macro_precision": 0.7548582995951417, + "macro_recall": 0.7367624343527958, + "per_class_f1": { + "FAVOR": 0.717948717948718, + "AGAINST": 0.6476190476190476, + "NONE": 0.8202247191011236 + }, + "accuracy_pct": 74.5, + "macro_f1_pct": 72.85974948896298, + "weighted_f1_pct": 73.75361155698235, + "macro_precision_pct": 75.48582995951418, + "macro_recall_pct": 73.67624343527957, + "per_class_f1_pct": { + "FAVOR": 71.7948717948718, + "AGAINST": 64.76190476190476, + "NONE": 82.02247191011236 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.675, + "macro_f1": 0.6323761472037615, + "weighted_f1": 0.6666825172928251, + "macro_precision": 0.636881875591553, + "macro_recall": 0.6420179604645625, + "per_class_f1": { + "FAVOR": 0.7227722772277227, + "AGAINST": 0.4383561643835616, + "NONE": 0.736 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 63.23761472037614, + "weighted_f1_pct": 66.66825172928252, + "macro_precision_pct": 63.688187559155296, + "macro_recall_pct": 64.20179604645625, + "per_class_f1_pct": { + "FAVOR": 72.27722772277228, + "AGAINST": 43.83561643835616, + "NONE": 73.6 + } + } + ], + "accuracy": 0.7003567181926278, + "macro_f1": 0.6896708240743553, + "weighted_f1": 0.6895544997484406, + "macro_precision": 0.7065236674023606, + "macro_recall": 0.7005112418727979, + "per_class_f1": { + "FAVOR": 0.7305194805194806, + "AGAINST": 0.5733041575492341, + "NONE": 0.7651888341543513 + }, + "accuracy_pct": 70.03567181926277, + "macro_f1_pct": 68.96708240743553, + "weighted_f1_pct": 68.95544997484406, + "macro_precision_pct": 70.65236674023606, + "macro_recall_pct": 70.05112418727978, + "per_class_f1_pct": { + "FAVOR": 73.05194805194806, + "AGAINST": 57.330415754923415, + "NONE": 76.51888341543514 + } + }, + { + "after_task": 6, + "best_epoch": 15, + "best_dev_macro_f1": 0.8457912457912458, + "losses": { + "current_ce": 0.4921357918668676, + "replay_ce": 0.2652038168024134, + "dark_distillation": 0.38115012123490927, + "supervised_contrastive": 3.7923592388629914, + "total": 1.3271506115794183, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6682912162131839, + "weighted_f1": 0.8331248932928772, + "macro_precision": 0.6574074074074074, + "macro_recall": 0.696493144035918, + "per_class_f1": { + "FAVOR": 0.8549618320610687, + "AGAINST": 0.2857142857142857, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 66.8291216213184, + "weighted_f1_pct": 83.31248932928771, + "macro_precision_pct": 65.74074074074075, + "macro_recall_pct": 69.6493144035918, + "per_class_f1_pct": { + "FAVOR": 85.49618320610686, + "AGAINST": 28.57142857142857, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6766054383903354, + "weighted_f1": 0.6956732217564874, + "macro_precision": 0.6711917686523039, + "macro_recall": 0.7394500561167229, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7243243243243244, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.66054383903354, + "weighted_f1_pct": 69.56732217564874, + "macro_precision_pct": 67.11917686523039, + "macro_recall_pct": 73.94500561167229, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 72.43243243243244, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5781693054420326, + "weighted_f1": 0.552934901206348, + "macro_precision": 0.5761704401255593, + "macro_recall": 0.6303418803418804, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.37362637362637363, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 57.81693054420326, + "weighted_f1_pct": 55.2934901206348, + "macro_precision_pct": 57.61704401255593, + "macro_recall_pct": 63.03418803418804, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 37.362637362637365, + "NONE": 69.42148760330579 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6916519694297473, + "weighted_f1": 0.6999162257495591, + "macro_precision": 0.7061117136527856, + "macro_recall": 0.7028421377818969, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6428571428571429, + "NONE": 0.7654320987654321 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 69.16519694297473, + "weighted_f1_pct": 69.99162257495591, + "macro_precision_pct": 70.61117136527855, + "macro_recall_pct": 70.28421377818968, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 64.28571428571429, + "NONE": 76.5432098765432 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6346249709886074, + "weighted_f1": 0.6640069930069931, + "macro_precision": 0.6280663780663781, + "macro_recall": 0.6468218943947099, + "per_class_f1": { + "FAVOR": 0.7076923076923077, + "AGAINST": 0.4523809523809524, + "NONE": 0.743801652892562 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.46249709886074, + "weighted_f1_pct": 66.4006993006993, + "macro_precision_pct": 62.80663780663781, + "macro_recall_pct": 64.68218943947099, + "per_class_f1_pct": { + "FAVOR": 70.76923076923077, + "AGAINST": 45.23809523809524, + "NONE": 74.3801652892562 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7151625129542339, + "weighted_f1": 0.7617962167340081, + "macro_precision": 0.6942195059842118, + "macro_recall": 0.7543678160919539, + "per_class_f1": { + "FAVOR": 0.6101694915254238, + "AGAINST": 0.8165680473372781, + "NONE": 0.71875 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 71.51625129542339, + "weighted_f1_pct": 76.17962167340082, + "macro_precision_pct": 69.42195059842118, + "macro_recall_pct": 75.43678160919539, + "per_class_f1_pct": { + "FAVOR": 61.016949152542374, + "AGAINST": 81.65680473372781, + "NONE": 71.875 + } + } + ], + "accuracy": 0.6960486322188449, + "macro_f1": 0.6962752756321606, + "weighted_f1": 0.6925823536077086, + "macro_precision": 0.6999464416790427, + "macro_recall": 0.705270579940067, + "per_class_f1": { + "FAVOR": 0.7050147492625368, + "AGAINST": 0.6412213740458015, + "NONE": 0.7425897035881436 + }, + "accuracy_pct": 69.6048632218845, + "macro_f1_pct": 69.62752756321606, + "weighted_f1_pct": 69.25823536077085, + "macro_precision_pct": 69.99464416790427, + "macro_recall_pct": 70.5270579940067, + "per_class_f1_pct": { + "FAVOR": 70.50147492625368, + "AGAINST": 64.12213740458014, + "NONE": 74.25897035881435 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.5939530234882558, + "losses": { + "current_ce": 0.5883172781178445, + "replay_ce": 0.223132795456684, + "dark_distillation": 0.4028841298637968, + "supervised_contrastive": 4.22716889814897, + "total": 1.4356090343359744, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6221139971139972, + "weighted_f1": 0.8268829253342528, + "macro_precision": 0.6233618233618233, + "macro_recall": 0.6298264773692512, + "per_class_f1": { + "FAVOR": 0.8484848484848485, + "AGAINST": 0.14285714285714285, + "NONE": 0.875 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 62.211399711399714, + "weighted_f1_pct": 82.68829253342528, + "macro_precision_pct": 62.33618233618233, + "macro_recall_pct": 62.982647736925124, + "per_class_f1_pct": { + "FAVOR": 84.84848484848484, + "AGAINST": 14.285714285714285, + "NONE": 87.5 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6290322580645161, + "macro_f1": 0.6239914344973965, + "weighted_f1": 0.6374270682692293, + "macro_precision": 0.6376633550546594, + "macro_recall": 0.7038720538720539, + "per_class_f1": { + "FAVOR": 0.5794392523364486, + "AGAINST": 0.6551724137931034, + "NONE": 0.6373626373626373 + }, + "accuracy_pct": 62.903225806451616, + "macro_f1_pct": 62.39914344973965, + "weighted_f1_pct": 63.742706826922934, + "macro_precision_pct": 63.766335505465946, + "macro_recall_pct": 70.38720538720538, + "per_class_f1_pct": { + "FAVOR": 57.943925233644855, + "AGAINST": 65.51724137931035, + "NONE": 63.73626373626373 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5653664259116197, + "weighted_f1": 0.5338596219433611, + "macro_precision": 0.5712820512820512, + "macro_recall": 0.6452991452991453, + "per_class_f1": { + "FAVOR": 0.6829268292682927, + "AGAINST": 0.32941176470588235, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 56.536642591161964, + "weighted_f1_pct": 53.38596219433611, + "macro_precision_pct": 57.128205128205124, + "macro_recall_pct": 64.52991452991454, + "per_class_f1_pct": { + "FAVOR": 68.29268292682927, + "AGAINST": 32.94117647058823, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6751771291648593, + "weighted_f1": 0.6845349197496436, + "macro_precision": 0.7251082991336069, + "macro_recall": 0.6966017917825146, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.5858585858585859, + "NONE": 0.7730061349693251 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 67.51771291648593, + "weighted_f1_pct": 68.45349197496436, + "macro_precision_pct": 72.51082991336068, + "macro_recall_pct": 69.66017917825145, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 58.58585858585859, + "NONE": 77.30061349693251 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.69, + "macro_f1": 0.6413422794185045, + "weighted_f1": 0.6828094538854974, + "macro_precision": 0.6450165121677648, + "macro_recall": 0.6432536180108995, + "per_class_f1": { + "FAVOR": 0.7559808612440191, + "AGAINST": 0.4266666666666667, + "NONE": 0.7413793103448276 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 64.13422794185045, + "weighted_f1_pct": 68.28094538854974, + "macro_precision_pct": 64.50165121677648, + "macro_recall_pct": 64.32536180108995, + "per_class_f1_pct": { + "FAVOR": 75.5980861244019, + "AGAINST": 42.66666666666667, + "NONE": 74.13793103448276 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6906858340820605, + "weighted_f1": 0.7277935728439735, + "macro_precision": 0.6795174454524763, + "macro_recall": 0.7502098950524737, + "per_class_f1": { + "FAVOR": 0.5428571428571428, + "AGAINST": 0.7672955974842768, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.06858340820605, + "weighted_f1_pct": 72.77935728439735, + "macro_precision_pct": 67.95174454524762, + "macro_recall_pct": 75.02098950524737, + "per_class_f1_pct": { + "FAVOR": 54.285714285714285, + "AGAINST": 76.72955974842768, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5631578947368421, + "macro_f1": 0.5501464097955325, + "weighted_f1": 0.5667102365163307, + "macro_precision": 0.5405684754521963, + "macro_recall": 0.5682643427741466, + "per_class_f1": { + "FAVOR": 0.49122807017543857, + "AGAINST": 0.6137566137566137, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 56.315789473684205, + "macro_f1_pct": 55.014640979553256, + "weighted_f1_pct": 56.67102365163307, + "macro_precision_pct": 54.05684754521963, + "macro_recall_pct": 56.82643427741466, + "per_class_f1_pct": { + "FAVOR": 49.122807017543856, + "AGAINST": 61.37566137566137, + "NONE": 54.54545454545454 + } + } + ], + "accuracy": 0.6610025488530161, + "macro_f1": 0.6629258206941381, + "weighted_f1": 0.6552159949421067, + "macro_precision": 0.6720840096396316, + "macro_recall": 0.6795214077335731, + "per_class_f1": { + "FAVOR": 0.6737089201877934, + "AGAINST": 0.5937106918238994, + "NONE": 0.7213578500707214 + }, + "accuracy_pct": 66.10025488530161, + "macro_f1_pct": 66.29258206941381, + "weighted_f1_pct": 65.52159949421068, + "macro_precision_pct": 67.20840096396316, + "macro_recall_pct": 67.9521407733573, + "per_class_f1_pct": { + "FAVOR": 67.37089201877934, + "AGAINST": 59.37106918238994, + "NONE": 72.13578500707214 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.6993172124101915, + "losses": { + "current_ce": 0.5525923003753026, + "replay_ce": 0.21369035108522935, + "dark_distillation": 0.37410288886590437, + "supervised_contrastive": 4.271607911948002, + "total": 1.3804948994607635, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6939311817360597, + "weighted_f1": 0.8302973617024987, + "macro_precision": 0.6820028011204481, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8292682926829268, + "AGAINST": 0.36363636363636365, + "NONE": 0.8888888888888888 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.39311817360597, + "weighted_f1_pct": 83.02973617024986, + "macro_precision_pct": 68.20028011204481, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 82.92682926829268, + "AGAINST": 36.36363636363637, + "NONE": 88.88888888888889 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7090851535295979, + "weighted_f1": 0.7306007413534295, + "macro_precision": 0.6909510930809105, + "macro_recall": 0.7513468013468013, + "per_class_f1": { + "FAVOR": 0.691358024691358, + "AGAINST": 0.7692307692307693, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.9085153529598, + "weighted_f1_pct": 73.06007413534294, + "macro_precision_pct": 69.09510930809105, + "macro_recall_pct": 75.13468013468014, + "per_class_f1_pct": { + "FAVOR": 69.1358024691358, + "AGAINST": 76.92307692307693, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6141077192937089, + "weighted_f1": 0.5815801679476806, + "macro_precision": 0.6265432098765432, + "macro_recall": 0.6717948717948717, + "per_class_f1": { + "FAVOR": 0.7428571428571429, + "AGAINST": 0.39080459770114945, + "NONE": 0.7086614173228346 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 61.41077192937089, + "weighted_f1_pct": 58.158016794768066, + "macro_precision_pct": 62.65432098765432, + "macro_recall_pct": 67.17948717948717, + "per_class_f1_pct": { + "FAVOR": 74.28571428571429, + "AGAINST": 39.08045977011494, + "NONE": 70.86614173228347 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.6993348371797096, + "weighted_f1": 0.7084004858580601, + "macro_precision": 0.7098180076628352, + "macro_recall": 0.7048038307074451, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6548672566371682, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 69.93348371797096, + "weighted_f1_pct": 70.84004858580602, + "macro_precision_pct": 70.98180076628353, + "macro_recall_pct": 70.48038307074451, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 65.48672566371681, + "NONE": 77.64705882352942 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6387944622931766, + "weighted_f1": 0.6673951038757339, + "macro_precision": 0.6322695035460993, + "macro_recall": 0.6562224183583408, + "per_class_f1": { + "FAVOR": 0.7127659574468085, + "AGAINST": 0.47191011235955055, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 63.87944622931766, + "weighted_f1_pct": 66.73951038757339, + "macro_precision_pct": 63.22695035460993, + "macro_recall_pct": 65.62224183583409, + "per_class_f1_pct": { + "FAVOR": 71.27659574468085, + "AGAINST": 47.19101123595505, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7034157240608853, + "weighted_f1": 0.7592832178868414, + "macro_precision": 0.6881465254959233, + "macro_recall": 0.7252373813093453, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8228571428571428, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 70.34157240608853, + "weighted_f1_pct": 75.92832178868414, + "macro_precision_pct": 68.81465254959232, + "macro_recall_pct": 72.52373813093453, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 82.28571428571428, + "NONE": 74.19354838709677 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5643281951693166, + "weighted_f1": 0.5861537871867434, + "macro_precision": 0.5560062737482845, + "macro_recall": 0.5784313725490197, + "per_class_f1": { + "FAVOR": 0.5046728971962616, + "AGAINST": 0.6428571428571429, + "NONE": 0.5454545454545454 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 56.432819516931666, + "weighted_f1_pct": 58.615378718674336, + "macro_precision_pct": 55.60062737482845, + "macro_recall_pct": 57.84313725490197, + "per_class_f1_pct": { + "FAVOR": 50.467289719626166, + "AGAINST": 64.28571428571429, + "NONE": 54.54545454545454 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7208121827411168, + "macro_f1": 0.7010562515352493, + "weighted_f1": 0.723390778154758, + "macro_precision": 0.6892546722651517, + "macro_recall": 0.726820521719012, + "per_class_f1": { + "FAVOR": 0.6376811594202898, + "AGAINST": 0.7536231884057971, + "NONE": 0.711864406779661 + }, + "accuracy_pct": 72.08121827411168, + "macro_f1_pct": 70.10562515352493, + "weighted_f1_pct": 72.3390778154758, + "macro_precision_pct": 68.92546722651517, + "macro_recall_pct": 72.6820521719012, + "per_class_f1_pct": { + "FAVOR": 63.76811594202898, + "AGAINST": 75.36231884057972, + "NONE": 71.1864406779661 + } + } + ], + "accuracy": 0.6928675400291121, + "macro_f1": 0.6940431443519918, + "weighted_f1": 0.6915404324901548, + "macro_precision": 0.690797865313839, + "macro_recall": 0.7042599853888952, + "per_class_f1": { + "FAVOR": 0.6790123456790124, + "AGAINST": 0.6771217712177122, + "NONE": 0.7259953161592506 + }, + "accuracy_pct": 69.28675400291121, + "macro_f1_pct": 69.40431443519918, + "weighted_f1_pct": 69.15404324901549, + "macro_precision_pct": 69.0797865313839, + "macro_recall_pct": 70.42599853888952, + "per_class_f1_pct": { + "FAVOR": 67.90123456790124, + "AGAINST": 67.71217712177122, + "NONE": 72.59953161592506 + } + }, + { + "after_task": 9, + "best_epoch": 15, + "best_dev_macro_f1": 0.6514042025820047, + "losses": { + "current_ce": 0.5333497233334041, + "replay_ce": 0.21578899905795143, + "dark_distillation": 0.34363774250424095, + "supervised_contrastive": 4.232672637606424, + "total": 1.3442248700157045, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7121565500514967, + "weighted_f1": 0.843411432538111, + "macro_precision": 0.6959064327485379, + "macro_recall": 0.8198762286130323, + "per_class_f1": { + "FAVOR": 0.8503937007874016, + "AGAINST": 0.4, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.21565500514967, + "weighted_f1_pct": 84.3411432538111, + "macro_precision_pct": 69.59064327485379, + "macro_recall_pct": 81.98762286130324, + "per_class_f1_pct": { + "FAVOR": 85.03937007874016, + "AGAINST": 40.0, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6619776949221012, + "weighted_f1": 0.6782440221344732, + "macro_precision": 0.6533807462378891, + "macro_recall": 0.7231762065095397, + "per_class_f1": { + "FAVOR": 0.6236559139784946, + "AGAINST": 0.7027027027027027, + "NONE": 0.6595744680851063 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.19776949221013, + "weighted_f1_pct": 67.82440221344731, + "macro_precision_pct": 65.33807462378891, + "macro_recall_pct": 72.31762065095397, + "per_class_f1_pct": { + "FAVOR": 62.365591397849464, + "AGAINST": 70.27027027027027, + "NONE": 65.95744680851064 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5915492957746479, + "macro_f1": 0.5831333671446511, + "weighted_f1": 0.554765609987765, + "macro_precision": 0.5824684003513193, + "macro_recall": 0.6423076923076924, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.3595505617977528, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 59.154929577464785, + "macro_f1_pct": 58.313336714465116, + "weighted_f1_pct": 55.476560998776506, + "macro_precision_pct": 58.24684003513193, + "macro_recall_pct": 64.23076923076924, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 35.95505617977528, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.715, + "macro_f1": 0.7029181628644596, + "weighted_f1": 0.710612340035589, + "macro_precision": 0.7182772212598226, + "macro_recall": 0.7150602409638555, + "per_class_f1": { + "FAVOR": 0.6991869918699187, + "AGAINST": 0.6238532110091743, + "NONE": 0.7857142857142857 + }, + "accuracy_pct": 71.5, + "macro_f1_pct": 70.29181628644596, + "weighted_f1_pct": 71.06123400355891, + "macro_precision_pct": 71.82772212598226, + "macro_recall_pct": 71.50602409638554, + "per_class_f1_pct": { + "FAVOR": 69.91869918699187, + "AGAINST": 62.38532110091744, + "NONE": 78.57142857142857 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.705, + "macro_f1": 0.6672176469118737, + "weighted_f1": 0.7033029045124866, + "macro_precision": 0.663494231936855, + "macro_recall": 0.6727118620322504, + "per_class_f1": { + "FAVOR": 0.7623762376237624, + "AGAINST": 0.4634146341463415, + "NONE": 0.7758620689655172 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 66.72176469118737, + "weighted_f1_pct": 70.33029045124866, + "macro_precision_pct": 66.3494231936855, + "macro_recall_pct": 67.27118620322504, + "per_class_f1_pct": { + "FAVOR": 76.23762376237624, + "AGAINST": 46.34146341463415, + "NONE": 77.58620689655173 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.730108162550359, + "weighted_f1": 0.7792938240989353, + "macro_precision": 0.7105820105820105, + "macro_recall": 0.759720139930035, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8323699421965318, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.0108162550359, + "weighted_f1_pct": 77.92938240989353, + "macro_precision_pct": 71.05820105820105, + "macro_recall_pct": 75.9720139930035, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 83.23699421965318, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5808845297396096, + "weighted_f1": 0.5983629479537357, + "macro_precision": 0.5712678169542386, + "macro_recall": 0.5994916485112564, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.6382978723404256, + "NONE": 0.5526315789473685 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.088452973960955, + "weighted_f1_pct": 59.83629479537357, + "macro_precision_pct": 57.12678169542386, + "macro_recall_pct": 59.94916485112564, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 63.829787234042556, + "NONE": 55.26315789473685 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.672404474610357, + "weighted_f1": 0.6826960401497577, + "macro_precision": 0.6572688914461066, + "macro_recall": 0.7318263320345569, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.6875, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 67.2404474610357, + "weighted_f1_pct": 68.26960401497577, + "macro_precision_pct": 65.72688914461065, + "macro_recall_pct": 73.1826332034557, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 68.75, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6359586160347753, + "weighted_f1": 0.6240807606872479, + "macro_precision": 0.6326291079812206, + "macro_recall": 0.6426045599487333, + "per_class_f1": { + "FAVOR": 0.6052631578947368, + "AGAINST": 0.6049382716049383, + "NONE": 0.6976744186046512 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 63.59586160347753, + "weighted_f1_pct": 62.408076068724796, + "macro_precision_pct": 63.262910798122064, + "macro_recall_pct": 64.26045599487334, + "per_class_f1_pct": { + "FAVOR": 60.526315789473685, + "AGAINST": 60.49382716049383, + "NONE": 69.76744186046511 + } + } + ], + "accuracy": 0.6804320203303685, + "macro_f1": 0.6844792798659297, + "weighted_f1": 0.6778736543002519, + "macro_precision": 0.6810885152102513, + "macro_recall": 0.6971439048163566, + "per_class_f1": { + "FAVOR": 0.6764132553606238, + "AGAINST": 0.6416666666666667, + "NONE": 0.735357917570499 + }, + "accuracy_pct": 68.04320203303685, + "macro_f1_pct": 68.44792798659297, + "weighted_f1_pct": 67.78736543002519, + "macro_precision_pct": 68.10885152102513, + "macro_recall_pct": 69.71439048163566, + "per_class_f1_pct": { + "FAVOR": 67.64132553606238, + "AGAINST": 64.16666666666667, + "NONE": 73.53579175704989 + } + }, + { + "after_task": 10, + "best_epoch": 14, + "best_dev_macro_f1": 0.7270304950789348, + "losses": { + "current_ce": 0.4160077131042878, + "replay_ce": 0.20025991598765056, + "dark_distillation": 0.37701009716838596, + "supervised_contrastive": 4.04466478129228, + "total": 1.2092391716937223, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6974358974358975, + "weighted_f1": 0.8511912865895167, + "macro_precision": 0.6885346885346886, + "macro_recall": 0.7631598107025845, + "per_class_f1": { + "FAVOR": 0.8615384615384616, + "AGAINST": 0.3333333333333333, + "NONE": 0.8974358974358975 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 69.74358974358975, + "weighted_f1_pct": 85.11912865895167, + "macro_precision_pct": 68.85346885346885, + "macro_recall_pct": 76.31598107025846, + "per_class_f1_pct": { + "FAVOR": 86.15384615384616, + "AGAINST": 33.33333333333333, + "NONE": 89.74358974358975 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6666666666666666, + "macro_f1": 0.6572943546479982, + "weighted_f1": 0.6771634071036368, + "macro_precision": 0.6626916274451045, + "macro_recall": 0.7271043771043771, + "per_class_f1": { + "FAVOR": 0.594059405940594, + "AGAINST": 0.7039106145251397, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.72943546479982, + "weighted_f1_pct": 67.71634071036368, + "macro_precision_pct": 66.26916274451045, + "macro_recall_pct": 72.71043771043772, + "per_class_f1_pct": { + "FAVOR": 59.4059405940594, + "AGAINST": 70.39106145251397, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5704225352112676, + "macro_f1": 0.5719683420700997, + "weighted_f1": 0.541319331344217, + "macro_precision": 0.5603174603174602, + "macro_recall": 0.6269230769230769, + "per_class_f1": { + "FAVOR": 0.6933333333333334, + "AGAINST": 0.3617021276595745, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 57.04225352112676, + "macro_f1_pct": 57.19683420700997, + "weighted_f1_pct": 54.13193313442169, + "macro_precision_pct": 56.031746031746025, + "macro_recall_pct": 62.69230769230769, + "per_class_f1_pct": { + "FAVOR": 69.33333333333334, + "AGAINST": 36.17021276595745, + "NONE": 66.08695652173913 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.665, + "macro_f1": 0.6545612640583898, + "weighted_f1": 0.6624863972315785, + "macro_precision": 0.6692592592592592, + "macro_recall": 0.6678869323447637, + "per_class_f1": { + "FAVOR": 0.6456692913385826, + "AGAINST": 0.5818181818181818, + "NONE": 0.7361963190184049 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 65.45612640583897, + "weighted_f1_pct": 66.24863972315785, + "macro_precision_pct": 66.92592592592592, + "macro_recall_pct": 66.78869323447637, + "per_class_f1_pct": { + "FAVOR": 64.56692913385827, + "AGAINST": 58.18181818181818, + "NONE": 73.61963190184049 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6284076461785983, + "weighted_f1": 0.6666116211563775, + "macro_precision": 0.6325354384177914, + "macro_recall": 0.6251754717774135, + "per_class_f1": { + "FAVOR": 0.7307692307692307, + "AGAINST": 0.4186046511627907, + "NONE": 0.7358490566037735 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 62.84076461785983, + "weighted_f1_pct": 66.66116211563775, + "macro_precision_pct": 63.253543841779134, + "macro_recall_pct": 62.51754717774135, + "per_class_f1_pct": { + "FAVOR": 73.07692307692307, + "AGAINST": 41.86046511627907, + "NONE": 73.58490566037736 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.7339459760512392, + "weighted_f1": 0.7800231171536147, + "macro_precision": 0.7123618105651589, + "macro_recall": 0.7694302848575711, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.8304093567251462, + "NONE": 0.8 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.39459760512392, + "weighted_f1_pct": 78.00231171536147, + "macro_precision_pct": 71.23618105651589, + "macro_recall_pct": 76.94302848575711, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 83.04093567251462, + "NONE": 80.0 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.5886601661249549, + "weighted_f1": 0.6139932713691053, + "macro_precision": 0.5828113828113829, + "macro_recall": 0.599128540305011, + "per_class_f1": { + "FAVOR": 0.5641025641025641, + "AGAINST": 0.6666666666666666, + "NONE": 0.5352112676056338 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 58.866016612495486, + "weighted_f1_pct": 61.39932713691053, + "macro_precision_pct": 58.28113828113829, + "macro_recall_pct": 59.912854030501094, + "per_class_f1_pct": { + "FAVOR": 56.41025641025641, + "AGAINST": 66.66666666666666, + "NONE": 53.52112676056338 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6629735389301352, + "weighted_f1": 0.6716446294648164, + "macro_precision": 0.6455059277639923, + "macro_recall": 0.7079540853616856, + "per_class_f1": { + "FAVOR": 0.6588235294117647, + "AGAINST": 0.6893203883495146, + "NONE": 0.6407766990291263 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.29735389301352, + "weighted_f1_pct": 67.16446294648163, + "macro_precision_pct": 64.55059277639923, + "macro_recall_pct": 70.79540853616855, + "per_class_f1_pct": { + "FAVOR": 65.88235294117646, + "AGAINST": 68.93203883495146, + "NONE": 64.07766990291263 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.645, + "macro_f1": 0.6537829304639944, + "weighted_f1": 0.6460569523376525, + "macro_precision": 0.6687000726216413, + "macro_recall": 0.6433226013171813, + "per_class_f1": { + "FAVOR": 0.6265060240963856, + "AGAINST": 0.6415094339622641, + "NONE": 0.6933333333333334 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 65.37829304639943, + "weighted_f1_pct": 64.60569523376525, + "macro_precision_pct": 66.87000726216414, + "macro_recall_pct": 64.33226013171813, + "per_class_f1_pct": { + "FAVOR": 62.65060240963856, + "AGAINST": 64.15094339622641, + "NONE": 69.33333333333334 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7800320850253172, + "weighted_f1": 0.8132930515867048, + "macro_precision": 0.7630920859076199, + "macro_recall": 0.8046865456160809, + "per_class_f1": { + "FAVOR": 0.7, + "AGAINST": 0.7868852459016393, + "NONE": 0.8532110091743119 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 78.00320850253172, + "weighted_f1_pct": 81.32930515867048, + "macro_precision_pct": 76.30920859076198, + "macro_recall_pct": 80.46865456160809, + "per_class_f1_pct": { + "FAVOR": 70.0, + "AGAINST": 78.68852459016394, + "NONE": 85.3211009174312 + } + } + ], + "accuracy": 0.6860202931228861, + "macro_f1": 0.6885414110582025, + "weighted_f1": 0.6855558200102561, + "macro_precision": 0.6884620373937734, + "macro_recall": 0.6963016369121294, + "per_class_f1": { + "FAVOR": 0.6737777777777778, + "AGAINST": 0.6551982049364248, + "NONE": 0.7366482504604052 + }, + "accuracy_pct": 68.60202931228862, + "macro_f1_pct": 68.85414110582026, + "weighted_f1_pct": 68.55558200102561, + "macro_precision_pct": 68.84620373937734, + "macro_recall_pct": 69.63016369121294, + "per_class_f1_pct": { + "FAVOR": 67.37777777777778, + "AGAINST": 65.51982049364248, + "NONE": 73.66482504604052 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s2__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__x__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ec56c67613490a91b09946c54cb45fb56b5340cb --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__x__s2__seed13/run.log b/outputs/trace_s7_contrastive_replay__x__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..6267c4be4c5424b5d7b7a7ad3d1937ca7287d9c3 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 08:17:49,881 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__x__s2__seed13 device=cuda +2026-09-02 08:34:20,998 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6629062200986624, 'average_forgetting': 0.03184929955844933, 'backward_transfer': -0.013263273288024013, 'final_mean_target_macro_f1_pct': 66.29062200986624, 'average_forgetting_pct': 3.1849299558449333, 'backward_transfer_pct': -1.3263273288024013, 'final_pooled_accuracy': 0.6860202931228861, 'final_pooled_accuracy_pct': 68.60202931228862, 'final_pooled_macro_f1': 0.6885414110582025, 'final_pooled_macro_f1_pct': 68.85414110582026, 'final_pooled_weighted_f1': 0.6855558200102561, 'final_pooled_weighted_f1_pct': 68.55558200102561, 'final_pooled_macro_precision': 0.6884620373937734, 'final_pooled_macro_precision_pct': 68.84620373937734, 'final_pooled_macro_recall': 0.6963016369121294, 'final_pooled_macro_recall_pct': 69.63016369121294, 'final_pooled_per_class_f1': {'FAVOR': 0.6737777777777778, 'AGAINST': 0.6551982049364248, 'NONE': 0.7366482504604052}, 'final_pooled_per_class_f1_pct': {'FAVOR': 67.37777777777778, 'AGAINST': 65.51982049364248, 'NONE': 73.66482504604052}} +2026-09-02 08:34:20,999 [INFO] __main__: efficiency={'training_wall_time_seconds': 989.7757089319639, 'training_wall_time_minutes': 16.49626181553273, 'peak_gpu_memory_allocated_gb': 10.41092300415039, 'peak_gpu_memory_reserved_gb': 10.630859375} diff --git a/outputs/trace_s7_contrastive_replay__x__s3__seed13/environment.json b/outputs/trace_s7_contrastive_replay__x__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..c4f1b4e88d733470f10a6ac2bb5a494eb9d0a16a --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T08:34:29.020267+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.json b/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c5d020aac25cf9add7f540bfd2703d0671bf0cd2 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.json @@ -0,0 +1,2294 @@ +{ + "experiment": "trace_s7_contrastive_replay", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 7, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": false + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.5944026733500417, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7226722654769194, + 0.6199312215441247, + 0.6931462667311723, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6986629924474382, + 0.5844642560646262, + 0.688300069942582, + 0.6859853447419066, + null, + null, + null, + null, + null, + null + ], + [ + 0.6842565018860433, + 0.590117950593686, + 0.6978720475127194, + 0.6502264155021669, + 0.7957184628237259, + null, + null, + null, + null, + null + ], + [ + 0.6732414082732555, + 0.5913612363232242, + 0.6934350618561145, + 0.6327989776709387, + 0.7464593790724104, + 0.6973035697195904, + null, + null, + null, + null + ], + [ + 0.6718766613503456, + 0.6381502372638552, + 0.6688262049459159, + 0.678809178278275, + 0.7634259091037551, + 0.6984683876540748, + 0.6833985748812758, + null, + null, + null + ], + [ + 0.6984978454781756, + 0.5977448210922788, + 0.6762298558100085, + 0.6474902878388042, + 0.7830809611380851, + 0.7313869220527717, + 0.7105960663835428, + 0.6186811779769527, + null, + null + ], + [ + 0.7341011262907472, + 0.6013308678314905, + 0.6607497561623683, + 0.6964585523913165, + 0.7775192128133305, + 0.6922716627634661, + 0.7108116916102287, + 0.5934814814814814, + 0.7072935819980737, + null + ], + [ + 0.699632386729161, + 0.5808080808080809, + 0.696092682742745, + 0.6782023131501855, + 0.8049448647579488, + 0.7223614592035644, + 0.7052154195011338, + 0.5555284542297529, + 0.6891740938799763, + 0.6435605096349986 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6030075187969924, + "losses": { + "current_ce": 0.7280336425159917, + "replay_ce": 0.21442859954003132, + "dark_distillation": 0.3549410031600432, + "supervised_contrastive": 4.358035736372977, + "total": 1.5557363213914814, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5944026733500417, + "weighted_f1": 0.6198575385833003, + "macro_precision": 0.5900673400673401, + "macro_recall": 0.6053013798111838, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.6736842105263158, + "NONE": 0.5428571428571428 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.44026733500417, + "weighted_f1_pct": 61.98575385833003, + "macro_precision_pct": 59.00673400673401, + "macro_recall_pct": 60.53013798111838, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 67.36842105263158, + "NONE": 54.285714285714285 + } + } + ], + "accuracy": 0.6696428571428571, + "macro_f1": 0.6525327812284334, + "weighted_f1": 0.6769069801833777, + "macro_precision": 0.6408913308913309, + "macro_recall": 0.6810085845059093, + "per_class_f1": { + "FAVOR": 0.5652173913043478, + "AGAINST": 0.7257142857142858, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.96428571428571, + "macro_f1_pct": 65.25327812284334, + "weighted_f1_pct": 67.69069801833777, + "macro_precision_pct": 64.08913308913309, + "macro_recall_pct": 68.10085845059093, + "per_class_f1_pct": { + "FAVOR": 56.52173913043478, + "AGAINST": 72.57142857142857, + "NONE": 66.66666666666666 + } + }, + { + "after_task": 3, + "best_epoch": 13, + "best_dev_macro_f1": 0.7665425971877585, + "losses": { + "current_ce": 0.614460780584451, + "replay_ce": 0.20738965733484788, + "dark_distillation": 0.3309342642625173, + "supervised_contrastive": 4.328750132069443, + "total": 1.4201925938779658, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7226722654769194, + "weighted_f1": 0.7626167966704335, + "macro_precision": 0.6997239475500345, + "macro_recall": 0.7858520739630185, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.8074534161490683, + "NONE": 0.7605633802816901 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 72.26722654769195, + "weighted_f1_pct": 76.26167966704335, + "macro_precision_pct": 69.97239475500345, + "macro_recall_pct": 78.58520739630185, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 80.74534161490683, + "NONE": 76.05633802816901 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6199312215441247, + "weighted_f1": 0.6369930711696417, + "macro_precision": 0.6123015873015872, + "macro_recall": 0.6383442265795207, + "per_class_f1": { + "FAVOR": 0.55, + "AGAINST": 0.6881720430107527, + "NONE": 0.6216216216216216 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 61.993122154412475, + "weighted_f1_pct": 63.699307116964164, + "macro_precision_pct": 61.23015873015872, + "macro_recall_pct": 63.834422657952075, + "per_class_f1_pct": { + "FAVOR": 55.00000000000001, + "AGAINST": 68.81720430107528, + "NONE": 62.16216216216216 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.6931462667311723, + "weighted_f1": 0.708068045353746, + "macro_precision": 0.6887626262626263, + "macro_recall": 0.7072126785817572, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.7358490566037735, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 69.31462667311723, + "weighted_f1_pct": 70.8068045353746, + "macro_precision_pct": 68.87626262626263, + "macro_recall_pct": 70.72126785817572, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 73.58490566037736, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6923076923076923, + "macro_f1": 0.67197196976812, + "weighted_f1": 0.6971106287414733, + "macro_precision": 0.6572173677436836, + "macro_recall": 0.7022515232162926, + "per_class_f1": { + "FAVOR": 0.5959183673469388, + "AGAINST": 0.740608228980322, + "NONE": 0.6793893129770993 + }, + "accuracy_pct": 69.23076923076923, + "macro_f1_pct": 67.197196976812, + "weighted_f1_pct": 69.71106287414733, + "macro_precision_pct": 65.72173677436835, + "macro_recall_pct": 70.22515232162925, + "per_class_f1_pct": { + "FAVOR": 59.591836734693885, + "AGAINST": 74.0608228980322, + "NONE": 67.93893129770993 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.6943710920764463, + "losses": { + "current_ce": 0.6485477943269033, + "replay_ce": 0.19973071742625464, + "dark_distillation": 0.3263352944264336, + "supervised_contrastive": 4.321683437862093, + "total": 1.443614515425667, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6986629924474382, + "weighted_f1": 0.7245216384668943, + "macro_precision": 0.6865942028985508, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.75, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.86629924474383, + "weighted_f1_pct": 72.45216384668943, + "macro_precision_pct": 68.65942028985508, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 75.0, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5844642560646262, + "weighted_f1": 0.5868918721053684, + "macro_precision": 0.5986855699012332, + "macro_recall": 0.6292665214233841, + "per_class_f1": { + "FAVOR": 0.5673758865248227, + "AGAINST": 0.5962732919254659, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.44642560646262, + "weighted_f1_pct": 58.68918721053684, + "macro_precision_pct": 59.868556990123324, + "macro_recall_pct": 62.92665214233841, + "per_class_f1_pct": { + "FAVOR": 56.73758865248227, + "AGAINST": 59.62732919254658, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.688300069942582, + "weighted_f1": 0.7047982956170995, + "macro_precision": 0.6712418300653594, + "macro_recall": 0.736453551653968, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7272727272727273, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.8300069942582, + "weighted_f1_pct": 70.47982956170995, + "macro_precision_pct": 67.12418300653594, + "macro_recall_pct": 73.6453551653968, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 72.72727272727273, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6859853447419066, + "weighted_f1": 0.6755900141854225, + "macro_precision": 0.6784871304934247, + "macro_recall": 0.7111697897605757, + "per_class_f1": { + "FAVOR": 0.6871165644171779, + "AGAINST": 0.6285714285714286, + "NONE": 0.7422680412371134 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.59853447419067, + "weighted_f1_pct": 67.55900141854225, + "macro_precision_pct": 67.84871304934246, + "macro_recall_pct": 71.11697897605757, + "per_class_f1_pct": { + "FAVOR": 68.71165644171779, + "AGAINST": 62.857142857142854, + "NONE": 74.22680412371135 + } + } + ], + "accuracy": 0.6671214188267395, + "macro_f1": 0.6679041104723651, + "weighted_f1": 0.6690338775724239, + "macro_precision": 0.662799360288442, + "macro_recall": 0.7075601650564044, + "per_class_f1": { + "FAVOR": 0.6228070175438597, + "AGAINST": 0.6789554531490015, + "NONE": 0.7019498607242339 + }, + "accuracy_pct": 66.71214188267395, + "macro_f1_pct": 66.7904110472365, + "weighted_f1_pct": 66.9033877572424, + "macro_precision_pct": 66.2799360288442, + "macro_recall_pct": 70.75601650564045, + "per_class_f1_pct": { + "FAVOR": 62.28070175438597, + "AGAINST": 67.89554531490015, + "NONE": 70.19498607242339 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.6754541426672573, + "losses": { + "current_ce": 0.45398612692952156, + "replay_ce": 0.2076686985635509, + "dark_distillation": 0.3438919122641285, + "supervised_contrastive": 4.084514963626861, + "total": 1.2420522913336753, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6842565018860433, + "weighted_f1": 0.708297774550518, + "macro_precision": 0.6721525600835946, + "macro_recall": 0.7762868565717141, + "per_class_f1": { + "FAVOR": 0.5797101449275363, + "AGAINST": 0.7333333333333333, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 68.42565018860432, + "weighted_f1_pct": 70.8297774550518, + "macro_precision_pct": 67.21525600835946, + "macro_recall_pct": 77.62868565717142, + "per_class_f1_pct": { + "FAVOR": 57.971014492753625, + "AGAINST": 73.33333333333333, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.590117950593686, + "weighted_f1": 0.5924608909978035, + "macro_precision": 0.5920318364762809, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.562962962962963, + "AGAINST": 0.6046511627906976, + "NONE": 0.6027397260273972 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 59.0117950593686, + "weighted_f1_pct": 59.24608909978035, + "macro_precision_pct": 59.20318364762809, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 56.2962962962963, + "AGAINST": 60.46511627906976, + "NONE": 60.273972602739725 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6978720475127194, + "weighted_f1": 0.7184770996151844, + "macro_precision": 0.6847496001868789, + "macro_recall": 0.7166564131686464, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.7570093457943925, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.78720475127193, + "weighted_f1_pct": 71.84770996151843, + "macro_precision_pct": 68.4749600186879, + "macro_recall_pct": 71.66564131686464, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 75.70093457943925, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6502264155021669, + "weighted_f1": 0.6383397853645884, + "macro_precision": 0.6425110184919038, + "macro_recall": 0.6626633930156965, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6308724832214765, + "NONE": 0.7111111111111111 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 65.02264155021669, + "weighted_f1_pct": 63.83397853645884, + "macro_precision_pct": 64.25110184919038, + "macro_recall_pct": 66.26633930156966, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 63.08724832214765, + "NONE": 71.11111111111111 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.83, + "macro_f1": 0.7957184628237259, + "weighted_f1": 0.831108395989975, + "macro_precision": 0.785589424232492, + "macro_recall": 0.8114849982416198, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7678571428571429, + "NONE": 0.8859649122807017 + }, + "accuracy_pct": 83.0, + "macro_f1_pct": 79.57184628237259, + "weighted_f1_pct": 83.1108395989975, + "macro_precision_pct": 78.5589424232492, + "macro_recall_pct": 81.14849982416197, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 76.78571428571429, + "NONE": 88.59649122807018 + } + } + ], + "accuracy": 0.6956055734190782, + "macro_f1": 0.6918196828199387, + "weighted_f1": 0.6969075533865741, + "macro_precision": 0.6863768359149843, + "macro_recall": 0.710133818749319, + "per_class_f1": { + "FAVOR": 0.6169354838709677, + "AGAINST": 0.6976160602258469, + "NONE": 0.7609075043630017 + }, + "accuracy_pct": 69.56055734190782, + "macro_f1_pct": 69.18196828199386, + "weighted_f1_pct": 69.69075533865741, + "macro_precision_pct": 68.63768359149843, + "macro_recall_pct": 71.01338187493191, + "per_class_f1_pct": { + "FAVOR": 61.693548387096776, + "AGAINST": 69.76160602258469, + "NONE": 76.09075043630017 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.6946969076666939, + "losses": { + "current_ce": 0.4007276706752323, + "replay_ce": 0.20182080204997743, + "dark_distillation": 0.3912020226319631, + "supervised_contrastive": 3.8774871190388995, + "total": 1.185898204644521, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6732414082732555, + "weighted_f1": 0.7103214820255219, + "macro_precision": 0.6558161350844278, + "macro_recall": 0.7411244377811094, + "per_class_f1": { + "FAVOR": 0.5538461538461539, + "AGAINST": 0.7515923566878981, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 67.32414082732555, + "weighted_f1_pct": 71.0321482025522, + "macro_precision_pct": 65.58161350844279, + "macro_recall_pct": 74.11244377811094, + "per_class_f1_pct": { + "FAVOR": 55.38461538461539, + "AGAINST": 75.15923566878982, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5913612363232242, + "weighted_f1": 0.6020968356843309, + "macro_precision": 0.5953473735173082, + "macro_recall": 0.6023965141612201, + "per_class_f1": { + "FAVOR": 0.5116279069767442, + "AGAINST": 0.644808743169399, + "NONE": 0.6176470588235294 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 59.13612363232242, + "weighted_f1_pct": 60.20968356843309, + "macro_precision_pct": 59.53473735173082, + "macro_recall_pct": 60.23965141612201, + "per_class_f1_pct": { + "FAVOR": 51.162790697674424, + "AGAINST": 64.48087431693989, + "NONE": 61.76470588235294 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6934350618561145, + "weighted_f1": 0.7147422116698098, + "macro_precision": 0.6775575880758807, + "macro_recall": 0.7208577182549072, + "per_class_f1": { + "FAVOR": 0.6756756756756757, + "AGAINST": 0.7559808612440191, + "NONE": 0.6486486486486487 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.34350618561145, + "weighted_f1_pct": 71.47422116698098, + "macro_precision_pct": 67.75575880758808, + "macro_recall_pct": 72.08577182549072, + "per_class_f1_pct": { + "FAVOR": 67.56756756756756, + "AGAINST": 75.5980861244019, + "NONE": 64.86486486486487 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.62, + "macro_f1": 0.6327989776709387, + "weighted_f1": 0.619801399144967, + "macro_precision": 0.6313665758110203, + "macro_recall": 0.6343162007118647, + "per_class_f1": { + "FAVOR": 0.5802469135802469, + "AGAINST": 0.6193548387096774, + "NONE": 0.6987951807228916 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 63.27989776709387, + "weighted_f1_pct": 61.980139914496704, + "macro_precision_pct": 63.13665758110203, + "macro_recall_pct": 63.431620071186465, + "per_class_f1_pct": { + "FAVOR": 58.0246913580247, + "AGAINST": 61.935483870967744, + "NONE": 69.87951807228916 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7464593790724104, + "weighted_f1": 0.7856754737464854, + "macro_precision": 0.728974358974359, + "macro_recall": 0.775900938419679, + "per_class_f1": { + "FAVOR": 0.6774193548387096, + "AGAINST": 0.7154471544715447, + "NONE": 0.8465116279069768 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 74.64593790724105, + "weighted_f1_pct": 78.56754737464854, + "macro_precision_pct": 72.8974358974359, + "macro_recall_pct": 77.5900938419679, + "per_class_f1_pct": { + "FAVOR": 67.74193548387096, + "AGAINST": 71.54471544715447, + "NONE": 84.65116279069768 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6973035697195904, + "weighted_f1": 0.8357423160280402, + "macro_precision": 0.6782746823069403, + "macro_recall": 0.7581846863244751, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.375, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 69.73035697195904, + "weighted_f1_pct": 83.57423160280402, + "macro_precision_pct": 67.82746823069404, + "macro_recall_pct": 75.8184686324475, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 37.5, + "NONE": 86.41975308641975 + } + } + ], + "accuracy": 0.6959847036328872, + "macro_f1": 0.696147593238596, + "weighted_f1": 0.6969720091097147, + "macro_precision": 0.6942276529877592, + "macro_recall": 0.7024077943466619, + "per_class_f1": { + "FAVOR": 0.6409017713365539, + "AGAINST": 0.6927639383155397, + "NONE": 0.7547770700636943 + }, + "accuracy_pct": 69.59847036328873, + "macro_f1_pct": 69.6147593238596, + "weighted_f1_pct": 69.69720091097147, + "macro_precision_pct": 69.42276529877593, + "macro_recall_pct": 70.2407794346662, + "per_class_f1_pct": { + "FAVOR": 64.09017713365539, + "AGAINST": 69.27639383155397, + "NONE": 75.47770700636943 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.6513556618819777, + "losses": { + "current_ce": 0.589374444882075, + "replay_ce": 0.20022676194256003, + "dark_distillation": 0.386932716496063, + "supervised_contrastive": 4.155052943663164, + "total": 1.3985728632320058, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6917808219178082, + "macro_f1": 0.6718766613503456, + "weighted_f1": 0.7018592558606979, + "macro_precision": 0.6610149777213106, + "macro_recall": 0.7611144427786107, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.7333333333333333, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 69.17808219178082, + "macro_f1_pct": 67.18766613503456, + "weighted_f1_pct": 70.18592558606979, + "macro_precision_pct": 66.10149777213105, + "macro_recall_pct": 76.11144427786107, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 73.33333333333333, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6421052631578947, + "macro_f1": 0.6381502372638552, + "weighted_f1": 0.646197887951143, + "macro_precision": 0.6426752227388496, + "macro_recall": 0.6710239651416122, + "per_class_f1": { + "FAVOR": 0.6176470588235294, + "AGAINST": 0.6666666666666666, + "NONE": 0.6301369863013698 + }, + "accuracy_pct": 64.21052631578948, + "macro_f1_pct": 63.81502372638552, + "weighted_f1_pct": 64.6197887951143, + "macro_precision_pct": 64.26752227388496, + "macro_recall_pct": 67.10239651416123, + "per_class_f1_pct": { + "FAVOR": 61.76470588235294, + "AGAINST": 66.66666666666666, + "NONE": 63.013698630136986 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6688262049459159, + "weighted_f1": 0.6924580204035187, + "macro_precision": 0.6553125050727749, + "macro_recall": 0.6911278322001904, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7407407407407407, + "NONE": 0.6078431372549019 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 66.88262049459159, + "weighted_f1_pct": 69.24580204035186, + "macro_precision_pct": 65.53125050727749, + "macro_recall_pct": 69.11278322001904, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 74.07407407407408, + "NONE": 60.78431372549019 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.675, + "macro_f1": 0.678809178278275, + "weighted_f1": 0.6747409509065639, + "macro_precision": 0.6773809523809525, + "macro_recall": 0.6805333580401332, + "per_class_f1": { + "FAVOR": 0.6540880503144654, + "AGAINST": 0.6835443037974683, + "NONE": 0.6987951807228916 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 67.8809178278275, + "weighted_f1_pct": 67.47409509065639, + "macro_precision_pct": 67.73809523809526, + "macro_recall_pct": 68.05333580401332, + "per_class_f1_pct": { + "FAVOR": 65.40880503144653, + "AGAINST": 68.35443037974683, + "NONE": 69.87951807228916 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.795, + "macro_f1": 0.7634259091037551, + "weighted_f1": 0.7991457674745345, + "macro_precision": 0.7444157475917912, + "macro_recall": 0.7977936957447204, + "per_class_f1": { + "FAVOR": 0.7076923076923077, + "AGAINST": 0.7241379310344828, + "NONE": 0.8584474885844748 + }, + "accuracy_pct": 79.5, + "macro_f1_pct": 76.3425909103755, + "weighted_f1_pct": 79.91457674745345, + "macro_precision_pct": 74.44157475917912, + "macro_recall_pct": 79.77936957447204, + "per_class_f1_pct": { + "FAVOR": 70.76923076923077, + "AGAINST": 72.41379310344827, + "NONE": 85.84474885844749 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6984683876540748, + "weighted_f1": 0.8248884874604072, + "macro_precision": 0.6815789473684211, + "macro_recall": 0.80677102293411, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.4, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.84683876540748, + "weighted_f1_pct": 82.48884874604072, + "macro_precision_pct": 68.15789473684211, + "macro_recall_pct": 80.677102293411, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 40.0, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6833985748812758, + "weighted_f1": 0.7129749068014406, + "macro_precision": 0.6705040372651996, + "macro_recall": 0.7265993265993266, + "per_class_f1": { + "FAVOR": 0.5952380952380952, + "AGAINST": 0.7539267015706806, + "NONE": 0.7010309278350515 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 68.33985748812758, + "weighted_f1_pct": 71.29749068014407, + "macro_precision_pct": 67.05040372651996, + "macro_recall_pct": 72.65993265993266, + "per_class_f1_pct": { + "FAVOR": 59.523809523809526, + "AGAINST": 75.39267015706807, + "NONE": 70.10309278350515 + } + } + ], + "accuracy": 0.7102272727272727, + "macro_f1": 0.7096338811894595, + "weighted_f1": 0.7112995762548214, + "macro_precision": 0.7062057764865362, + "macro_recall": 0.7217600638590252, + "per_class_f1": { + "FAVOR": 0.667601683029453, + "AGAINST": 0.7123287671232876, + "NONE": 0.7489711934156379 + }, + "accuracy_pct": 71.02272727272727, + "macro_f1_pct": 70.96338811894594, + "weighted_f1_pct": 71.12995762548215, + "macro_precision_pct": 70.62057764865361, + "macro_recall_pct": 72.17600638590253, + "per_class_f1_pct": { + "FAVOR": 66.7601683029453, + "AGAINST": 71.23287671232876, + "NONE": 74.8971193415638 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.6494871794871795, + "losses": { + "current_ce": 0.6413807926078637, + "replay_ce": 0.19874954614788293, + "dark_distillation": 0.39132465521494547, + "supervised_contrastive": 4.285132644573848, + "total": 1.4643059392770132, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6984978454781756, + "weighted_f1": 0.7199296553171288, + "macro_precision": 0.6825217469285265, + "macro_recall": 0.7914042978510745, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7417218543046358, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.84978454781756, + "weighted_f1_pct": 71.99296553171288, + "macro_precision_pct": 68.25217469285265, + "macro_recall_pct": 79.14042978510744, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 74.17218543046357, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5977448210922788, + "weighted_f1": 0.6106135890573892, + "macro_precision": 0.5943572841133817, + "macro_recall": 0.6241830065359477, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.6440677966101694, + "NONE": 0.5866666666666667 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 59.774482109227876, + "weighted_f1_pct": 61.06135890573891, + "macro_precision_pct": 59.435728411338175, + "macro_recall_pct": 62.41830065359477, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 64.40677966101694, + "NONE": 58.666666666666664 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6751269035532995, + "macro_f1": 0.6762298558100085, + "weighted_f1": 0.6733324937678408, + "macro_precision": 0.6694655521783183, + "macro_recall": 0.7467859227359487, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.6666666666666666, + "NONE": 0.6870229007633588 + }, + "accuracy_pct": 67.51269035532995, + "macro_f1_pct": 67.62298558100085, + "weighted_f1_pct": 67.33324937678408, + "macro_precision_pct": 66.94655521783183, + "macro_recall_pct": 74.67859227359487, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 66.66666666666666, + "NONE": 68.70229007633588 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6474902878388042, + "weighted_f1": 0.6382081903608513, + "macro_precision": 0.6429263565891473, + "macro_recall": 0.658173704244165, + "per_class_f1": { + "FAVOR": 0.6347305389221557, + "AGAINST": 0.6111111111111112, + "NONE": 0.6966292134831461 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 64.74902878388042, + "weighted_f1_pct": 63.82081903608513, + "macro_precision_pct": 64.29263565891473, + "macro_recall_pct": 65.8173704244165, + "per_class_f1_pct": { + "FAVOR": 63.47305389221557, + "AGAINST": 61.111111111111114, + "NONE": 69.66292134831461 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7830809611380851, + "weighted_f1": 0.8151887331333656, + "macro_precision": 0.7728831883630645, + "macro_recall": 0.8036907472189831, + "per_class_f1": { + "FAVOR": 0.7419354838709677, + "AGAINST": 0.7339449541284404, + "NONE": 0.8733624454148472 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 78.30809611380852, + "weighted_f1_pct": 81.51887331333656, + "macro_precision_pct": 77.28831883630644, + "macro_recall_pct": 80.36907472189831, + "per_class_f1_pct": { + "FAVOR": 74.19354838709677, + "AGAINST": 73.39449541284404, + "NONE": 87.33624454148472 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7313869220527717, + "weighted_f1": 0.8439218136451565, + "macro_precision": 0.7033242258652095, + "macro_recall": 0.8248513529911419, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.47058823529411764, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 73.13869220527717, + "weighted_f1_pct": 84.39218136451565, + "macro_precision_pct": 70.33242258652095, + "macro_recall_pct": 82.48513529911419, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 47.05882352941176, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7105960663835428, + "weighted_f1": 0.7328653625862416, + "macro_precision": 0.6962636606599455, + "macro_recall": 0.7659371492704826, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.7634408602150538, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 71.05960663835428, + "weighted_f1_pct": 73.28653625862415, + "macro_precision_pct": 69.62636606599455, + "macro_recall_pct": 76.59371492704827, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 76.34408602150538, + "NONE": 72.54901960784314 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6267605633802817, + "macro_f1": 0.6186811779769527, + "weighted_f1": 0.5949003624821915, + "macro_precision": 0.6349219607655959, + "macro_recall": 0.6717948717948717, + "per_class_f1": { + "FAVOR": 0.704225352112676, + "AGAINST": 0.4318181818181818, + "NONE": 0.72 + }, + "accuracy_pct": 62.676056338028175, + "macro_f1_pct": 61.86811779769526, + "weighted_f1_pct": 59.49003624821915, + "macro_precision_pct": 63.492196076559594, + "macro_recall_pct": 67.17948717948717, + "per_class_f1_pct": { + "FAVOR": 70.4225352112676, + "AGAINST": 43.18181818181818, + "NONE": 72.0 + } + } + ], + "accuracy": 0.6994177583697234, + "macro_f1": 0.7001525744528895, + "weighted_f1": 0.6961992283600937, + "macro_precision": 0.699757665238883, + "macro_recall": 0.7240382819794585, + "per_class_f1": { + "FAVOR": 0.6759847522236341, + "AGAINST": 0.6672985781990521, + "NONE": 0.7571743929359823 + }, + "accuracy_pct": 69.94177583697234, + "macro_f1_pct": 70.01525744528895, + "weighted_f1_pct": 69.61992283600938, + "macro_precision_pct": 69.97576652388831, + "macro_recall_pct": 72.40382819794586, + "per_class_f1_pct": { + "FAVOR": 67.5984752223634, + "AGAINST": 66.72985781990522, + "NONE": 75.71743929359823 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.700836320191159, + "losses": { + "current_ce": 0.578133224317993, + "replay_ce": 0.19593904035678808, + "dark_distillation": 0.3500406695448834, + "supervised_contrastive": 4.266744485108749, + "total": 1.3757670606392016, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7341011262907472, + "weighted_f1": 0.7602863626785094, + "macro_precision": 0.7105413105413105, + "macro_recall": 0.8131434282858572, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.7898089171974523, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 73.41011262907472, + "weighted_f1_pct": 76.02863626785094, + "macro_precision_pct": 71.05413105413105, + "macro_recall_pct": 81.31434282858572, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 78.98089171974523, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.6013308678314905, + "weighted_f1": 0.620572375124384, + "macro_precision": 0.594673154374647, + "macro_recall": 0.6183732752360204, + "per_class_f1": { + "FAVOR": 0.5619834710743802, + "AGAINST": 0.6666666666666666, + "NONE": 0.5753424657534246 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 60.13308678314905, + "weighted_f1_pct": 62.0572375124384, + "macro_precision_pct": 59.4673154374647, + "macro_recall_pct": 61.837327523602035, + "per_class_f1_pct": { + "FAVOR": 56.19834710743802, + "AGAINST": 66.66666666666666, + "NONE": 57.534246575342465 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6607497561623683, + "weighted_f1": 0.6721925618966471, + "macro_precision": 0.6472222222222223, + "macro_recall": 0.7116243118325367, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6839378238341969, + "NONE": 0.6829268292682927 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.07497561623683, + "weighted_f1_pct": 67.2192561896647, + "macro_precision_pct": 64.72222222222223, + "macro_recall_pct": 71.16243118325367, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 68.39378238341969, + "NONE": 68.29268292682927 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.69, + "macro_f1": 0.6964585523913165, + "weighted_f1": 0.6881858690137428, + "macro_precision": 0.6917582417582416, + "macro_recall": 0.7131115898053567, + "per_class_f1": { + "FAVOR": 0.6712328767123288, + "AGAINST": 0.6790123456790124, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 69.64585523913165, + "weighted_f1_pct": 68.81858690137427, + "macro_precision_pct": 69.17582417582416, + "macro_recall_pct": 71.31115898053567, + "per_class_f1_pct": { + "FAVOR": 67.12328767123287, + "AGAINST": 67.90123456790124, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7775192128133305, + "weighted_f1": 0.8077976729153201, + "macro_precision": 0.7614015190569474, + "macro_recall": 0.7998408203305755, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7394957983193278, + "NONE": 0.8597285067873304 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 77.75192128133305, + "weighted_f1_pct": 80.77976729153201, + "macro_precision_pct": 76.14015190569474, + "macro_recall_pct": 79.98408203305755, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 73.94957983193278, + "NONE": 85.97285067873304 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6922716627634661, + "weighted_f1": 0.8146981409711715, + "macro_precision": 0.6709889593610524, + "macro_recall": 0.8081058124014077, + "per_class_f1": { + "FAVOR": 0.819672131147541, + "AGAINST": 0.4, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 69.22716627634661, + "weighted_f1_pct": 81.46981409711715, + "macro_precision_pct": 67.09889593610524, + "macro_recall_pct": 80.81058124014076, + "per_class_f1_pct": { + "FAVOR": 81.9672131147541, + "AGAINST": 40.0, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.7108116916102287, + "weighted_f1": 0.7372639846511895, + "macro_precision": 0.6940379022422366, + "macro_recall": 0.7506734006734007, + "per_class_f1": { + "FAVOR": 0.6493506493506493, + "AGAINST": 0.7772020725388601, + "NONE": 0.7058823529411765 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 71.08116916102287, + "weighted_f1_pct": 73.72639846511895, + "macro_precision_pct": 69.40379022422366, + "macro_recall_pct": 75.06734006734007, + "per_class_f1_pct": { + "FAVOR": 64.93506493506493, + "AGAINST": 77.720207253886, + "NONE": 70.58823529411765 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5934814814814814, + "weighted_f1": 0.584075117370892, + "macro_precision": 0.5980798501346446, + "macro_recall": 0.617094017094017, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.4444444444444444, + "NONE": 0.736 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.34814814814814, + "weighted_f1_pct": 58.4075117370892, + "macro_precision_pct": 59.80798501346446, + "macro_recall_pct": 61.7094017094017, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 44.44444444444444, + "NONE": 73.6 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7072935819980737, + "weighted_f1": 0.7117507223535593, + "macro_precision": 0.7218203737191079, + "macro_recall": 0.7026722273710225, + "per_class_f1": { + "FAVOR": 0.7021276595744681, + "AGAINST": 0.6666666666666666, + "NONE": 0.7530864197530864 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.72935819980736, + "weighted_f1_pct": 71.17507223535593, + "macro_precision_pct": 72.1820373719108, + "macro_recall_pct": 70.26722273710226, + "per_class_f1_pct": { + "FAVOR": 70.2127659574468, + "AGAINST": 66.66666666666666, + "NONE": 75.30864197530865 + } + } + ], + "accuracy": 0.7064803049555273, + "macro_f1": 0.705123581940157, + "weighted_f1": 0.7049715978948906, + "macro_precision": 0.7018439923975673, + "macro_recall": 0.7132928997999377, + "per_class_f1": { + "FAVOR": 0.6715506715506715, + "AGAINST": 0.6881382560879812, + "NONE": 0.7556818181818182 + }, + "accuracy_pct": 70.64803049555273, + "macro_f1_pct": 70.5123581940157, + "weighted_f1_pct": 70.49715978948906, + "macro_precision_pct": 70.18439923975673, + "macro_recall_pct": 71.32928997999377, + "per_class_f1_pct": { + "FAVOR": 67.15506715506714, + "AGAINST": 68.81382560879811, + "NONE": 75.56818181818183 + } + }, + { + "after_task": 10, + "best_epoch": 5, + "best_dev_macro_f1": 0.637968622305098, + "losses": { + "current_ce": 0.6123842968390538, + "replay_ce": 0.19816253085931143, + "dark_distillation": 0.35369483759770026, + "supervised_contrastive": 4.1878715148338905, + "total": 1.4061814057521331, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.699632386729161, + "weighted_f1": 0.7285463664827341, + "macro_precision": 0.6803863664328781, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7612903225806451, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.96323867291609, + "weighted_f1_pct": 72.85463664827341, + "macro_precision_pct": 68.0386366432878, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.12903225806451, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5808080808080809, + "weighted_f1": 0.5888357256778309, + "macro_precision": 0.5812999234051865, + "macro_recall": 0.6045751633986928, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.6136363636363636, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.08080808080809, + "weighted_f1_pct": 58.88357256778309, + "macro_precision_pct": 58.129992340518655, + "macro_recall_pct": 60.45751633986928, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 61.36363636363637, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.696092682742745, + "weighted_f1": 0.7071399188897655, + "macro_precision": 0.680623973727422, + "macro_recall": 0.7358593745892028, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.72, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 69.6092682742745, + "weighted_f1_pct": 70.71399188897655, + "macro_precision_pct": 68.0623973727422, + "macro_recall_pct": 73.58593745892028, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 72.0, + "NONE": 71.07438016528926 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6782023131501855, + "weighted_f1": 0.662413623765953, + "macro_precision": 0.6739229564424355, + "macro_recall": 0.6883623252186938, + "per_class_f1": { + "FAVOR": 0.6174496644295302, + "AGAINST": 0.6585365853658537, + "NONE": 0.7586206896551724 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.82023131501856, + "weighted_f1_pct": 66.2413623765953, + "macro_precision_pct": 67.39229564424355, + "macro_recall_pct": 68.83623252186938, + "per_class_f1_pct": { + "FAVOR": 61.74496644295302, + "AGAINST": 65.85365853658537, + "NONE": 75.86206896551724 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8049448647579488, + "weighted_f1": 0.839672748850319, + "macro_precision": 0.7871855813032284, + "macro_recall": 0.8343235789512651, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7936507936507936, + "NONE": 0.8878504672897196 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 80.49448647579489, + "weighted_f1_pct": 83.9672748850319, + "macro_precision_pct": 78.71855813032283, + "macro_recall_pct": 83.4323578951265, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 79.36507936507937, + "NONE": 88.78504672897196 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7223614592035644, + "weighted_f1": 0.849365301158501, + "macro_precision": 0.7003228410008071, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.42105263157894735, + "NONE": 0.8888888888888888 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.23614592035645, + "weighted_f1_pct": 84.93653011585009, + "macro_precision_pct": 70.03228410008072, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 42.10526315789473, + "NONE": 88.88888888888889 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7052154195011338, + "weighted_f1": 0.7315119596225587, + "macro_precision": 0.6880185502827012, + "macro_recall": 0.741638608305275, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7755102040816326, + "NONE": 0.673469387755102 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.52154195011337, + "weighted_f1_pct": 73.15119596225587, + "macro_precision_pct": 68.80185502827013, + "macro_recall_pct": 74.1638608305275, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 77.55102040816327, + "NONE": 67.3469387755102 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5555284542297529, + "weighted_f1": 0.5415700085537292, + "macro_precision": 0.5505175983436853, + "macro_recall": 0.5858974358974359, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.3877551020408163, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.552845422975295, + "weighted_f1_pct": 54.157000855372914, + "macro_precision_pct": 55.05175983436853, + "macro_recall_pct": 58.58974358974359, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 38.775510204081634, + "NONE": 69.42148760330579 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6891740938799763, + "weighted_f1": 0.7019982174688058, + "macro_precision": 0.7227888642522787, + "macro_recall": 0.6852332406549273, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.68, + "NONE": 0.7757575757575758 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 68.91740938799764, + "weighted_f1_pct": 70.19982174688057, + "macro_precision_pct": 72.27888642522787, + "macro_recall_pct": 68.52332406549273, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 68.0, + "NONE": 77.57575757575758 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6435605096349986, + "weighted_f1": 0.6701077285109232, + "macro_precision": 0.6423148148148149, + "macro_recall": 0.6477493380406002, + "per_class_f1": { + "FAVOR": 0.7035175879396985, + "AGAINST": 0.45652173913043476, + "NONE": 0.7706422018348624 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.35605096349985, + "weighted_f1_pct": 67.01077285109231, + "macro_precision_pct": 64.2314814814815, + "macro_recall_pct": 64.77493380406003, + "per_class_f1_pct": { + "FAVOR": 70.35175879396985, + "AGAINST": 45.65217391304348, + "NONE": 77.06422018348624 + } + } + ], + "accuracy": 0.6978579481397971, + "macro_f1": 0.6988304093567251, + "weighted_f1": 0.6966027947086456, + "macro_precision": 0.6958727212046442, + "macro_recall": 0.7040497050480207, + "per_class_f1": { + "FAVOR": 0.6647286821705426, + "AGAINST": 0.6686046511627907, + "NONE": 0.7631578947368421 + }, + "accuracy_pct": 69.7857948139797, + "macro_f1_pct": 69.88304093567251, + "weighted_f1_pct": 69.66027947086457, + "macro_precision_pct": 69.58727212046442, + "macro_recall_pct": 70.40497050480207, + "per_class_f1_pct": { + "FAVOR": 66.47286821705426, + "AGAINST": 66.86046511627907, + "NONE": 76.31578947368422 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6775520264637548, + "average_forgetting": 0.02205711610198248, + "backward_transfer": -0.006583798700073319, + "final_mean_target_macro_f1_pct": 67.75520264637548, + "average_forgetting_pct": 2.205711610198248, + "backward_transfer_pct": -0.6583798700073319, + "final_pooled_accuracy": 0.6978579481397971, + "final_pooled_accuracy_pct": 69.7857948139797, + "final_pooled_macro_f1": 0.6988304093567251, + "final_pooled_macro_f1_pct": 69.88304093567251, + "final_pooled_weighted_f1": 0.6966027947086456, + "final_pooled_weighted_f1_pct": 69.66027947086457, + "final_pooled_macro_precision": 0.6958727212046442, + "final_pooled_macro_precision_pct": 69.58727212046442, + "final_pooled_macro_recall": 0.7040497050480207, + "final_pooled_macro_recall_pct": 70.40497050480207, + "final_pooled_per_class_f1": { + "FAVOR": 0.6647286821705426, + "AGAINST": 0.6686046511627907, + "NONE": 0.7631578947368421 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 66.47286821705426, + "AGAINST": 66.86046511627907, + "NONE": 76.31578947368422 + } + }, + "efficiency": { + "training_wall_time_seconds": 972.2049490110949, + "training_wall_time_minutes": 16.20341581685158, + "peak_gpu_memory_allocated_gb": 10.40872573852539, + "peak_gpu_memory_reserved_gb": 10.587890625 + } +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.partial.json b/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..76788c1f5a35c809dd53f31d14579cee3fe27262 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s3__seed13/metrics.partial.json @@ -0,0 +1,2178 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.5944026733500417, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7226722654769194, + 0.6199312215441247, + 0.6931462667311723, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6986629924474382, + 0.5844642560646262, + 0.688300069942582, + 0.6859853447419066, + null, + null, + null, + null, + null, + null + ], + [ + 0.6842565018860433, + 0.590117950593686, + 0.6978720475127194, + 0.6502264155021669, + 0.7957184628237259, + null, + null, + null, + null, + null + ], + [ + 0.6732414082732555, + 0.5913612363232242, + 0.6934350618561145, + 0.6327989776709387, + 0.7464593790724104, + 0.6973035697195904, + null, + null, + null, + null + ], + [ + 0.6718766613503456, + 0.6381502372638552, + 0.6688262049459159, + 0.678809178278275, + 0.7634259091037551, + 0.6984683876540748, + 0.6833985748812758, + null, + null, + null + ], + [ + 0.6984978454781756, + 0.5977448210922788, + 0.6762298558100085, + 0.6474902878388042, + 0.7830809611380851, + 0.7313869220527717, + 0.7105960663835428, + 0.6186811779769527, + null, + null + ], + [ + 0.7341011262907472, + 0.6013308678314905, + 0.6607497561623683, + 0.6964585523913165, + 0.7775192128133305, + 0.6922716627634661, + 0.7108116916102287, + 0.5934814814814814, + 0.7072935819980737, + null + ], + [ + 0.699632386729161, + 0.5808080808080809, + 0.696092682742745, + 0.6782023131501855, + 0.8049448647579488, + 0.7223614592035644, + 0.7052154195011338, + 0.5555284542297529, + 0.6891740938799763, + 0.6435605096349986 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6030075187969924, + "losses": { + "current_ce": 0.7280336425159917, + "replay_ce": 0.21442859954003132, + "dark_distillation": 0.3549410031600432, + "supervised_contrastive": 4.358035736372977, + "total": 1.5557363213914814, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5944026733500417, + "weighted_f1": 0.6198575385833003, + "macro_precision": 0.5900673400673401, + "macro_recall": 0.6053013798111838, + "per_class_f1": { + "FAVOR": 0.5666666666666667, + "AGAINST": 0.6736842105263158, + "NONE": 0.5428571428571428 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.44026733500417, + "weighted_f1_pct": 61.98575385833003, + "macro_precision_pct": 59.00673400673401, + "macro_recall_pct": 60.53013798111838, + "per_class_f1_pct": { + "FAVOR": 56.666666666666664, + "AGAINST": 67.36842105263158, + "NONE": 54.285714285714285 + } + } + ], + "accuracy": 0.6696428571428571, + "macro_f1": 0.6525327812284334, + "weighted_f1": 0.6769069801833777, + "macro_precision": 0.6408913308913309, + "macro_recall": 0.6810085845059093, + "per_class_f1": { + "FAVOR": 0.5652173913043478, + "AGAINST": 0.7257142857142858, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.96428571428571, + "macro_f1_pct": 65.25327812284334, + "weighted_f1_pct": 67.69069801833777, + "macro_precision_pct": 64.08913308913309, + "macro_recall_pct": 68.10085845059093, + "per_class_f1_pct": { + "FAVOR": 56.52173913043478, + "AGAINST": 72.57142857142857, + "NONE": 66.66666666666666 + } + }, + { + "after_task": 3, + "best_epoch": 13, + "best_dev_macro_f1": 0.7665425971877585, + "losses": { + "current_ce": 0.614460780584451, + "replay_ce": 0.20738965733484788, + "dark_distillation": 0.3309342642625173, + "supervised_contrastive": 4.328750132069443, + "total": 1.4201925938779658, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7226722654769194, + "weighted_f1": 0.7626167966704335, + "macro_precision": 0.6997239475500345, + "macro_recall": 0.7858520739630185, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.8074534161490683, + "NONE": 0.7605633802816901 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 72.26722654769195, + "weighted_f1_pct": 76.26167966704335, + "macro_precision_pct": 69.97239475500345, + "macro_recall_pct": 78.58520739630185, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 80.74534161490683, + "NONE": 76.05633802816901 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6199312215441247, + "weighted_f1": 0.6369930711696417, + "macro_precision": 0.6123015873015872, + "macro_recall": 0.6383442265795207, + "per_class_f1": { + "FAVOR": 0.55, + "AGAINST": 0.6881720430107527, + "NONE": 0.6216216216216216 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 61.993122154412475, + "weighted_f1_pct": 63.699307116964164, + "macro_precision_pct": 61.23015873015872, + "macro_recall_pct": 63.834422657952075, + "per_class_f1_pct": { + "FAVOR": 55.00000000000001, + "AGAINST": 68.81720430107528, + "NONE": 62.16216216216216 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.6931462667311723, + "weighted_f1": 0.708068045353746, + "macro_precision": 0.6887626262626263, + "macro_recall": 0.7072126785817572, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.7358490566037735, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 69.31462667311723, + "weighted_f1_pct": 70.8068045353746, + "macro_precision_pct": 68.87626262626263, + "macro_recall_pct": 70.72126785817572, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 73.58490566037736, + "NONE": 66.66666666666666 + } + } + ], + "accuracy": 0.6923076923076923, + "macro_f1": 0.67197196976812, + "weighted_f1": 0.6971106287414733, + "macro_precision": 0.6572173677436836, + "macro_recall": 0.7022515232162926, + "per_class_f1": { + "FAVOR": 0.5959183673469388, + "AGAINST": 0.740608228980322, + "NONE": 0.6793893129770993 + }, + "accuracy_pct": 69.23076923076923, + "macro_f1_pct": 67.197196976812, + "weighted_f1_pct": 69.71106287414733, + "macro_precision_pct": 65.72173677436835, + "macro_recall_pct": 70.22515232162925, + "per_class_f1_pct": { + "FAVOR": 59.591836734693885, + "AGAINST": 74.0608228980322, + "NONE": 67.93893129770993 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.6943710920764463, + "losses": { + "current_ce": 0.6485477943269033, + "replay_ce": 0.19973071742625464, + "dark_distillation": 0.3263352944264336, + "supervised_contrastive": 4.321683437862093, + "total": 1.443614515425667, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6986629924474382, + "weighted_f1": 0.7245216384668943, + "macro_precision": 0.6865942028985508, + "macro_recall": 0.7835332333833084, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.75, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.86629924474383, + "weighted_f1_pct": 72.45216384668943, + "macro_precision_pct": 68.65942028985508, + "macro_recall_pct": 78.35332333833084, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 75.0, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5844642560646262, + "weighted_f1": 0.5868918721053684, + "macro_precision": 0.5986855699012332, + "macro_recall": 0.6292665214233841, + "per_class_f1": { + "FAVOR": 0.5673758865248227, + "AGAINST": 0.5962732919254659, + "NONE": 0.5897435897435898 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.44642560646262, + "weighted_f1_pct": 58.68918721053684, + "macro_precision_pct": 59.868556990123324, + "macro_recall_pct": 62.92665214233841, + "per_class_f1_pct": { + "FAVOR": 56.73758865248227, + "AGAINST": 59.62732919254658, + "NONE": 58.97435897435898 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.688300069942582, + "weighted_f1": 0.7047982956170995, + "macro_precision": 0.6712418300653594, + "macro_recall": 0.736453551653968, + "per_class_f1": { + "FAVOR": 0.6419753086419753, + "AGAINST": 0.7272727272727273, + "NONE": 0.6956521739130435 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.8300069942582, + "weighted_f1_pct": 70.47982956170995, + "macro_precision_pct": 67.12418300653594, + "macro_recall_pct": 73.6453551653968, + "per_class_f1_pct": { + "FAVOR": 64.19753086419753, + "AGAINST": 72.72727272727273, + "NONE": 69.56521739130434 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6859853447419066, + "weighted_f1": 0.6755900141854225, + "macro_precision": 0.6784871304934247, + "macro_recall": 0.7111697897605757, + "per_class_f1": { + "FAVOR": 0.6871165644171779, + "AGAINST": 0.6285714285714286, + "NONE": 0.7422680412371134 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.59853447419067, + "weighted_f1_pct": 67.55900141854225, + "macro_precision_pct": 67.84871304934246, + "macro_recall_pct": 71.11697897605757, + "per_class_f1_pct": { + "FAVOR": 68.71165644171779, + "AGAINST": 62.857142857142854, + "NONE": 74.22680412371135 + } + } + ], + "accuracy": 0.6671214188267395, + "macro_f1": 0.6679041104723651, + "weighted_f1": 0.6690338775724239, + "macro_precision": 0.662799360288442, + "macro_recall": 0.7075601650564044, + "per_class_f1": { + "FAVOR": 0.6228070175438597, + "AGAINST": 0.6789554531490015, + "NONE": 0.7019498607242339 + }, + "accuracy_pct": 66.71214188267395, + "macro_f1_pct": 66.7904110472365, + "weighted_f1_pct": 66.9033877572424, + "macro_precision_pct": 66.2799360288442, + "macro_recall_pct": 70.75601650564045, + "per_class_f1_pct": { + "FAVOR": 62.28070175438597, + "AGAINST": 67.89554531490015, + "NONE": 70.19498607242339 + } + }, + { + "after_task": 5, + "best_epoch": 12, + "best_dev_macro_f1": 0.6754541426672573, + "losses": { + "current_ce": 0.45398612692952156, + "replay_ce": 0.2076686985635509, + "dark_distillation": 0.3438919122641285, + "supervised_contrastive": 4.084514963626861, + "total": 1.2420522913336753, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6842565018860433, + "weighted_f1": 0.708297774550518, + "macro_precision": 0.6721525600835946, + "macro_recall": 0.7762868565717141, + "per_class_f1": { + "FAVOR": 0.5797101449275363, + "AGAINST": 0.7333333333333333, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 68.42565018860432, + "weighted_f1_pct": 70.8297774550518, + "macro_precision_pct": 67.21525600835946, + "macro_recall_pct": 77.62868565717142, + "per_class_f1_pct": { + "FAVOR": 57.971014492753625, + "AGAINST": 73.33333333333333, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.590117950593686, + "weighted_f1": 0.5924608909978035, + "macro_precision": 0.5920318364762809, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.562962962962963, + "AGAINST": 0.6046511627906976, + "NONE": 0.6027397260273972 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 59.0117950593686, + "weighted_f1_pct": 59.24608909978035, + "macro_precision_pct": 59.20318364762809, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 56.2962962962963, + "AGAINST": 60.46511627906976, + "NONE": 60.273972602739725 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.6978720475127194, + "weighted_f1": 0.7184770996151844, + "macro_precision": 0.6847496001868789, + "macro_recall": 0.7166564131686464, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.7570093457943925, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.78720475127193, + "weighted_f1_pct": 71.84770996151843, + "macro_precision_pct": 68.4749600186879, + "macro_recall_pct": 71.66564131686464, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 75.70093457943925, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6502264155021669, + "weighted_f1": 0.6383397853645884, + "macro_precision": 0.6425110184919038, + "macro_recall": 0.6626633930156965, + "per_class_f1": { + "FAVOR": 0.6086956521739131, + "AGAINST": 0.6308724832214765, + "NONE": 0.7111111111111111 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 65.02264155021669, + "weighted_f1_pct": 63.83397853645884, + "macro_precision_pct": 64.25110184919038, + "macro_recall_pct": 66.26633930156966, + "per_class_f1_pct": { + "FAVOR": 60.86956521739131, + "AGAINST": 63.08724832214765, + "NONE": 71.11111111111111 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.83, + "macro_f1": 0.7957184628237259, + "weighted_f1": 0.831108395989975, + "macro_precision": 0.785589424232492, + "macro_recall": 0.8114849982416198, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7678571428571429, + "NONE": 0.8859649122807017 + }, + "accuracy_pct": 83.0, + "macro_f1_pct": 79.57184628237259, + "weighted_f1_pct": 83.1108395989975, + "macro_precision_pct": 78.5589424232492, + "macro_recall_pct": 81.14849982416197, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 76.78571428571429, + "NONE": 88.59649122807018 + } + } + ], + "accuracy": 0.6956055734190782, + "macro_f1": 0.6918196828199387, + "weighted_f1": 0.6969075533865741, + "macro_precision": 0.6863768359149843, + "macro_recall": 0.710133818749319, + "per_class_f1": { + "FAVOR": 0.6169354838709677, + "AGAINST": 0.6976160602258469, + "NONE": 0.7609075043630017 + }, + "accuracy_pct": 69.56055734190782, + "macro_f1_pct": 69.18196828199386, + "weighted_f1_pct": 69.69075533865741, + "macro_precision_pct": 68.63768359149843, + "macro_recall_pct": 71.01338187493191, + "per_class_f1_pct": { + "FAVOR": 61.693548387096776, + "AGAINST": 69.76160602258469, + "NONE": 76.09075043630017 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.6946969076666939, + "losses": { + "current_ce": 0.4007276706752323, + "replay_ce": 0.20182080204997743, + "dark_distillation": 0.3912020226319631, + "supervised_contrastive": 3.8774871190388995, + "total": 1.185898204644521, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6986301369863014, + "macro_f1": 0.6732414082732555, + "weighted_f1": 0.7103214820255219, + "macro_precision": 0.6558161350844278, + "macro_recall": 0.7411244377811094, + "per_class_f1": { + "FAVOR": 0.5538461538461539, + "AGAINST": 0.7515923566878981, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 69.86301369863014, + "macro_f1_pct": 67.32414082732555, + "weighted_f1_pct": 71.0321482025522, + "macro_precision_pct": 65.58161350844279, + "macro_recall_pct": 74.11244377811094, + "per_class_f1_pct": { + "FAVOR": 55.38461538461539, + "AGAINST": 75.15923566878982, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5913612363232242, + "weighted_f1": 0.6020968356843309, + "macro_precision": 0.5953473735173082, + "macro_recall": 0.6023965141612201, + "per_class_f1": { + "FAVOR": 0.5116279069767442, + "AGAINST": 0.644808743169399, + "NONE": 0.6176470588235294 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 59.13612363232242, + "weighted_f1_pct": 60.20968356843309, + "macro_precision_pct": 59.53473735173082, + "macro_recall_pct": 60.23965141612201, + "per_class_f1_pct": { + "FAVOR": 51.162790697674424, + "AGAINST": 64.48087431693989, + "NONE": 61.76470588235294 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6934350618561145, + "weighted_f1": 0.7147422116698098, + "macro_precision": 0.6775575880758807, + "macro_recall": 0.7208577182549072, + "per_class_f1": { + "FAVOR": 0.6756756756756757, + "AGAINST": 0.7559808612440191, + "NONE": 0.6486486486486487 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.34350618561145, + "weighted_f1_pct": 71.47422116698098, + "macro_precision_pct": 67.75575880758808, + "macro_recall_pct": 72.08577182549072, + "per_class_f1_pct": { + "FAVOR": 67.56756756756756, + "AGAINST": 75.5980861244019, + "NONE": 64.86486486486487 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.62, + "macro_f1": 0.6327989776709387, + "weighted_f1": 0.619801399144967, + "macro_precision": 0.6313665758110203, + "macro_recall": 0.6343162007118647, + "per_class_f1": { + "FAVOR": 0.5802469135802469, + "AGAINST": 0.6193548387096774, + "NONE": 0.6987951807228916 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 63.27989776709387, + "weighted_f1_pct": 61.980139914496704, + "macro_precision_pct": 63.13665758110203, + "macro_recall_pct": 63.431620071186465, + "per_class_f1_pct": { + "FAVOR": 58.0246913580247, + "AGAINST": 61.935483870967744, + "NONE": 69.87951807228916 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7464593790724104, + "weighted_f1": 0.7856754737464854, + "macro_precision": 0.728974358974359, + "macro_recall": 0.775900938419679, + "per_class_f1": { + "FAVOR": 0.6774193548387096, + "AGAINST": 0.7154471544715447, + "NONE": 0.8465116279069768 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 74.64593790724105, + "weighted_f1_pct": 78.56754737464854, + "macro_precision_pct": 72.8974358974359, + "macro_recall_pct": 77.5900938419679, + "per_class_f1_pct": { + "FAVOR": 67.74193548387096, + "AGAINST": 71.54471544715447, + "NONE": 84.65116279069768 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6973035697195904, + "weighted_f1": 0.8357423160280402, + "macro_precision": 0.6782746823069403, + "macro_recall": 0.7581846863244751, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.375, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 69.73035697195904, + "weighted_f1_pct": 83.57423160280402, + "macro_precision_pct": 67.82746823069404, + "macro_recall_pct": 75.8184686324475, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 37.5, + "NONE": 86.41975308641975 + } + } + ], + "accuracy": 0.6959847036328872, + "macro_f1": 0.696147593238596, + "weighted_f1": 0.6969720091097147, + "macro_precision": 0.6942276529877592, + "macro_recall": 0.7024077943466619, + "per_class_f1": { + "FAVOR": 0.6409017713365539, + "AGAINST": 0.6927639383155397, + "NONE": 0.7547770700636943 + }, + "accuracy_pct": 69.59847036328873, + "macro_f1_pct": 69.6147593238596, + "weighted_f1_pct": 69.69720091097147, + "macro_precision_pct": 69.42276529877593, + "macro_recall_pct": 70.2407794346662, + "per_class_f1_pct": { + "FAVOR": 64.09017713365539, + "AGAINST": 69.27639383155397, + "NONE": 75.47770700636943 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.6513556618819777, + "losses": { + "current_ce": 0.589374444882075, + "replay_ce": 0.20022676194256003, + "dark_distillation": 0.386932716496063, + "supervised_contrastive": 4.155052943663164, + "total": 1.3985728632320058, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6917808219178082, + "macro_f1": 0.6718766613503456, + "weighted_f1": 0.7018592558606979, + "macro_precision": 0.6610149777213106, + "macro_recall": 0.7611144427786107, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.7333333333333333, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 69.17808219178082, + "macro_f1_pct": 67.18766613503456, + "weighted_f1_pct": 70.18592558606979, + "macro_precision_pct": 66.10149777213105, + "macro_recall_pct": 76.11144427786107, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 73.33333333333333, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6421052631578947, + "macro_f1": 0.6381502372638552, + "weighted_f1": 0.646197887951143, + "macro_precision": 0.6426752227388496, + "macro_recall": 0.6710239651416122, + "per_class_f1": { + "FAVOR": 0.6176470588235294, + "AGAINST": 0.6666666666666666, + "NONE": 0.6301369863013698 + }, + "accuracy_pct": 64.21052631578948, + "macro_f1_pct": 63.81502372638552, + "weighted_f1_pct": 64.6197887951143, + "macro_precision_pct": 64.26752227388496, + "macro_recall_pct": 67.10239651416123, + "per_class_f1_pct": { + "FAVOR": 61.76470588235294, + "AGAINST": 66.66666666666666, + "NONE": 63.013698630136986 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6688262049459159, + "weighted_f1": 0.6924580204035187, + "macro_precision": 0.6553125050727749, + "macro_recall": 0.6911278322001904, + "per_class_f1": { + "FAVOR": 0.6578947368421053, + "AGAINST": 0.7407407407407407, + "NONE": 0.6078431372549019 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 66.88262049459159, + "weighted_f1_pct": 69.24580204035186, + "macro_precision_pct": 65.53125050727749, + "macro_recall_pct": 69.11278322001904, + "per_class_f1_pct": { + "FAVOR": 65.78947368421053, + "AGAINST": 74.07407407407408, + "NONE": 60.78431372549019 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.675, + "macro_f1": 0.678809178278275, + "weighted_f1": 0.6747409509065639, + "macro_precision": 0.6773809523809525, + "macro_recall": 0.6805333580401332, + "per_class_f1": { + "FAVOR": 0.6540880503144654, + "AGAINST": 0.6835443037974683, + "NONE": 0.6987951807228916 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 67.8809178278275, + "weighted_f1_pct": 67.47409509065639, + "macro_precision_pct": 67.73809523809526, + "macro_recall_pct": 68.05333580401332, + "per_class_f1_pct": { + "FAVOR": 65.40880503144653, + "AGAINST": 68.35443037974683, + "NONE": 69.87951807228916 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.795, + "macro_f1": 0.7634259091037551, + "weighted_f1": 0.7991457674745345, + "macro_precision": 0.7444157475917912, + "macro_recall": 0.7977936957447204, + "per_class_f1": { + "FAVOR": 0.7076923076923077, + "AGAINST": 0.7241379310344828, + "NONE": 0.8584474885844748 + }, + "accuracy_pct": 79.5, + "macro_f1_pct": 76.3425909103755, + "weighted_f1_pct": 79.91457674745345, + "macro_precision_pct": 74.44157475917912, + "macro_recall_pct": 79.77936957447204, + "per_class_f1_pct": { + "FAVOR": 70.76923076923077, + "AGAINST": 72.41379310344827, + "NONE": 85.84474885844749 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6984683876540748, + "weighted_f1": 0.8248884874604072, + "macro_precision": 0.6815789473684211, + "macro_recall": 0.80677102293411, + "per_class_f1": { + "FAVOR": 0.8346456692913385, + "AGAINST": 0.4, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.84683876540748, + "weighted_f1_pct": 82.48884874604072, + "macro_precision_pct": 68.15789473684211, + "macro_recall_pct": 80.677102293411, + "per_class_f1_pct": { + "FAVOR": 83.46456692913385, + "AGAINST": 40.0, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6833985748812758, + "weighted_f1": 0.7129749068014406, + "macro_precision": 0.6705040372651996, + "macro_recall": 0.7265993265993266, + "per_class_f1": { + "FAVOR": 0.5952380952380952, + "AGAINST": 0.7539267015706806, + "NONE": 0.7010309278350515 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 68.33985748812758, + "weighted_f1_pct": 71.29749068014407, + "macro_precision_pct": 67.05040372651996, + "macro_recall_pct": 72.65993265993266, + "per_class_f1_pct": { + "FAVOR": 59.523809523809526, + "AGAINST": 75.39267015706807, + "NONE": 70.10309278350515 + } + } + ], + "accuracy": 0.7102272727272727, + "macro_f1": 0.7096338811894595, + "weighted_f1": 0.7112995762548214, + "macro_precision": 0.7062057764865362, + "macro_recall": 0.7217600638590252, + "per_class_f1": { + "FAVOR": 0.667601683029453, + "AGAINST": 0.7123287671232876, + "NONE": 0.7489711934156379 + }, + "accuracy_pct": 71.02272727272727, + "macro_f1_pct": 70.96338811894594, + "weighted_f1_pct": 71.12995762548215, + "macro_precision_pct": 70.62057764865361, + "macro_recall_pct": 72.17600638590253, + "per_class_f1_pct": { + "FAVOR": 66.7601683029453, + "AGAINST": 71.23287671232876, + "NONE": 74.8971193415638 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.6494871794871795, + "losses": { + "current_ce": 0.6413807926078637, + "replay_ce": 0.19874954614788293, + "dark_distillation": 0.39132465521494547, + "supervised_contrastive": 4.285132644573848, + "total": 1.4643059392770132, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6984978454781756, + "weighted_f1": 0.7199296553171288, + "macro_precision": 0.6825217469285265, + "macro_recall": 0.7914042978510745, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7417218543046358, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.84978454781756, + "weighted_f1_pct": 71.99296553171288, + "macro_precision_pct": 68.25217469285265, + "macro_recall_pct": 79.14042978510744, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 74.17218543046357, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5977448210922788, + "weighted_f1": 0.6106135890573892, + "macro_precision": 0.5943572841133817, + "macro_recall": 0.6241830065359477, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.6440677966101694, + "NONE": 0.5866666666666667 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 59.774482109227876, + "weighted_f1_pct": 61.06135890573891, + "macro_precision_pct": 59.435728411338175, + "macro_recall_pct": 62.41830065359477, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 64.40677966101694, + "NONE": 58.666666666666664 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6751269035532995, + "macro_f1": 0.6762298558100085, + "weighted_f1": 0.6733324937678408, + "macro_precision": 0.6694655521783183, + "macro_recall": 0.7467859227359487, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.6666666666666666, + "NONE": 0.6870229007633588 + }, + "accuracy_pct": 67.51269035532995, + "macro_f1_pct": 67.62298558100085, + "weighted_f1_pct": 67.33324937678408, + "macro_precision_pct": 66.94655521783183, + "macro_recall_pct": 74.67859227359487, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 66.66666666666666, + "NONE": 68.70229007633588 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6474902878388042, + "weighted_f1": 0.6382081903608513, + "macro_precision": 0.6429263565891473, + "macro_recall": 0.658173704244165, + "per_class_f1": { + "FAVOR": 0.6347305389221557, + "AGAINST": 0.6111111111111112, + "NONE": 0.6966292134831461 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 64.74902878388042, + "weighted_f1_pct": 63.82081903608513, + "macro_precision_pct": 64.29263565891473, + "macro_recall_pct": 65.8173704244165, + "per_class_f1_pct": { + "FAVOR": 63.47305389221557, + "AGAINST": 61.111111111111114, + "NONE": 69.66292134831461 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.815, + "macro_f1": 0.7830809611380851, + "weighted_f1": 0.8151887331333656, + "macro_precision": 0.7728831883630645, + "macro_recall": 0.8036907472189831, + "per_class_f1": { + "FAVOR": 0.7419354838709677, + "AGAINST": 0.7339449541284404, + "NONE": 0.8733624454148472 + }, + "accuracy_pct": 81.5, + "macro_f1_pct": 78.30809611380852, + "weighted_f1_pct": 81.51887331333656, + "macro_precision_pct": 77.28831883630644, + "macro_recall_pct": 80.36907472189831, + "per_class_f1_pct": { + "FAVOR": 74.19354838709677, + "AGAINST": 73.39449541284404, + "NONE": 87.33624454148472 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7313869220527717, + "weighted_f1": 0.8439218136451565, + "macro_precision": 0.7033242258652095, + "macro_recall": 0.8248513529911419, + "per_class_f1": { + "FAVOR": 0.859375, + "AGAINST": 0.47058823529411764, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 73.13869220527717, + "weighted_f1_pct": 84.39218136451565, + "macro_precision_pct": 70.33242258652095, + "macro_recall_pct": 82.48513529911419, + "per_class_f1_pct": { + "FAVOR": 85.9375, + "AGAINST": 47.05882352941176, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7105960663835428, + "weighted_f1": 0.7328653625862416, + "macro_precision": 0.6962636606599455, + "macro_recall": 0.7659371492704826, + "per_class_f1": { + "FAVOR": 0.6428571428571429, + "AGAINST": 0.7634408602150538, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 71.05960663835428, + "weighted_f1_pct": 73.28653625862415, + "macro_precision_pct": 69.62636606599455, + "macro_recall_pct": 76.59371492704827, + "per_class_f1_pct": { + "FAVOR": 64.28571428571429, + "AGAINST": 76.34408602150538, + "NONE": 72.54901960784314 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6267605633802817, + "macro_f1": 0.6186811779769527, + "weighted_f1": 0.5949003624821915, + "macro_precision": 0.6349219607655959, + "macro_recall": 0.6717948717948717, + "per_class_f1": { + "FAVOR": 0.704225352112676, + "AGAINST": 0.4318181818181818, + "NONE": 0.72 + }, + "accuracy_pct": 62.676056338028175, + "macro_f1_pct": 61.86811779769526, + "weighted_f1_pct": 59.49003624821915, + "macro_precision_pct": 63.492196076559594, + "macro_recall_pct": 67.17948717948717, + "per_class_f1_pct": { + "FAVOR": 70.4225352112676, + "AGAINST": 43.18181818181818, + "NONE": 72.0 + } + } + ], + "accuracy": 0.6994177583697234, + "macro_f1": 0.7001525744528895, + "weighted_f1": 0.6961992283600937, + "macro_precision": 0.699757665238883, + "macro_recall": 0.7240382819794585, + "per_class_f1": { + "FAVOR": 0.6759847522236341, + "AGAINST": 0.6672985781990521, + "NONE": 0.7571743929359823 + }, + "accuracy_pct": 69.94177583697234, + "macro_f1_pct": 70.01525744528895, + "weighted_f1_pct": 69.61992283600938, + "macro_precision_pct": 69.97576652388831, + "macro_recall_pct": 72.40382819794586, + "per_class_f1_pct": { + "FAVOR": 67.5984752223634, + "AGAINST": 66.72985781990522, + "NONE": 75.71743929359823 + } + }, + { + "after_task": 9, + "best_epoch": 9, + "best_dev_macro_f1": 0.700836320191159, + "losses": { + "current_ce": 0.578133224317993, + "replay_ce": 0.19593904035678808, + "dark_distillation": 0.3500406695448834, + "supervised_contrastive": 4.266744485108749, + "total": 1.3757670606392016, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7341011262907472, + "weighted_f1": 0.7602863626785094, + "macro_precision": 0.7105413105413105, + "macro_recall": 0.8131434282858572, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.7898089171974523, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 73.41011262907472, + "weighted_f1_pct": 76.02863626785094, + "macro_precision_pct": 71.05413105413105, + "macro_recall_pct": 81.31434282858572, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 78.98089171974523, + "NONE": 75.67567567567568 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.6013308678314905, + "weighted_f1": 0.620572375124384, + "macro_precision": 0.594673154374647, + "macro_recall": 0.6183732752360204, + "per_class_f1": { + "FAVOR": 0.5619834710743802, + "AGAINST": 0.6666666666666666, + "NONE": 0.5753424657534246 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 60.13308678314905, + "weighted_f1_pct": 62.0572375124384, + "macro_precision_pct": 59.4673154374647, + "macro_recall_pct": 61.837327523602035, + "per_class_f1_pct": { + "FAVOR": 56.19834710743802, + "AGAINST": 66.66666666666666, + "NONE": 57.534246575342465 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6700507614213198, + "macro_f1": 0.6607497561623683, + "weighted_f1": 0.6721925618966471, + "macro_precision": 0.6472222222222223, + "macro_recall": 0.7116243118325367, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6839378238341969, + "NONE": 0.6829268292682927 + }, + "accuracy_pct": 67.00507614213198, + "macro_f1_pct": 66.07497561623683, + "weighted_f1_pct": 67.2192561896647, + "macro_precision_pct": 64.72222222222223, + "macro_recall_pct": 71.16243118325367, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 68.39378238341969, + "NONE": 68.29268292682927 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.69, + "macro_f1": 0.6964585523913165, + "weighted_f1": 0.6881858690137428, + "macro_precision": 0.6917582417582416, + "macro_recall": 0.7131115898053567, + "per_class_f1": { + "FAVOR": 0.6712328767123288, + "AGAINST": 0.6790123456790124, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 69.64585523913165, + "weighted_f1_pct": 68.81858690137427, + "macro_precision_pct": 69.17582417582416, + "macro_recall_pct": 71.31115898053567, + "per_class_f1_pct": { + "FAVOR": 67.12328767123287, + "AGAINST": 67.90123456790124, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7775192128133305, + "weighted_f1": 0.8077976729153201, + "macro_precision": 0.7614015190569474, + "macro_recall": 0.7998408203305755, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7394957983193278, + "NONE": 0.8597285067873304 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 77.75192128133305, + "weighted_f1_pct": 80.77976729153201, + "macro_precision_pct": 76.14015190569474, + "macro_recall_pct": 79.98408203305755, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 73.94957983193278, + "NONE": 85.97285067873304 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6922716627634661, + "weighted_f1": 0.8146981409711715, + "macro_precision": 0.6709889593610524, + "macro_recall": 0.8081058124014077, + "per_class_f1": { + "FAVOR": 0.819672131147541, + "AGAINST": 0.4, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 69.22716627634661, + "weighted_f1_pct": 81.46981409711715, + "macro_precision_pct": 67.09889593610524, + "macro_recall_pct": 80.81058124014076, + "per_class_f1_pct": { + "FAVOR": 81.9672131147541, + "AGAINST": 40.0, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.7108116916102287, + "weighted_f1": 0.7372639846511895, + "macro_precision": 0.6940379022422366, + "macro_recall": 0.7506734006734007, + "per_class_f1": { + "FAVOR": 0.6493506493506493, + "AGAINST": 0.7772020725388601, + "NONE": 0.7058823529411765 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 71.08116916102287, + "weighted_f1_pct": 73.72639846511895, + "macro_precision_pct": 69.40379022422366, + "macro_recall_pct": 75.06734006734007, + "per_class_f1_pct": { + "FAVOR": 64.93506493506493, + "AGAINST": 77.720207253886, + "NONE": 70.58823529411765 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6056338028169014, + "macro_f1": 0.5934814814814814, + "weighted_f1": 0.584075117370892, + "macro_precision": 0.5980798501346446, + "macro_recall": 0.617094017094017, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.4444444444444444, + "NONE": 0.736 + }, + "accuracy_pct": 60.56338028169014, + "macro_f1_pct": 59.34814814814814, + "weighted_f1_pct": 58.4075117370892, + "macro_precision_pct": 59.80798501346446, + "macro_recall_pct": 61.7094017094017, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 44.44444444444444, + "NONE": 73.6 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.71, + "macro_f1": 0.7072935819980737, + "weighted_f1": 0.7117507223535593, + "macro_precision": 0.7218203737191079, + "macro_recall": 0.7026722273710225, + "per_class_f1": { + "FAVOR": 0.7021276595744681, + "AGAINST": 0.6666666666666666, + "NONE": 0.7530864197530864 + }, + "accuracy_pct": 71.0, + "macro_f1_pct": 70.72935819980736, + "weighted_f1_pct": 71.17507223535593, + "macro_precision_pct": 72.1820373719108, + "macro_recall_pct": 70.26722273710226, + "per_class_f1_pct": { + "FAVOR": 70.2127659574468, + "AGAINST": 66.66666666666666, + "NONE": 75.30864197530865 + } + } + ], + "accuracy": 0.7064803049555273, + "macro_f1": 0.705123581940157, + "weighted_f1": 0.7049715978948906, + "macro_precision": 0.7018439923975673, + "macro_recall": 0.7132928997999377, + "per_class_f1": { + "FAVOR": 0.6715506715506715, + "AGAINST": 0.6881382560879812, + "NONE": 0.7556818181818182 + }, + "accuracy_pct": 70.64803049555273, + "macro_f1_pct": 70.5123581940157, + "weighted_f1_pct": 70.49715978948906, + "macro_precision_pct": 70.18439923975673, + "macro_recall_pct": 71.32928997999377, + "per_class_f1_pct": { + "FAVOR": 67.15506715506714, + "AGAINST": 68.81382560879811, + "NONE": 75.56818181818183 + } + }, + { + "after_task": 10, + "best_epoch": 5, + "best_dev_macro_f1": 0.637968622305098, + "losses": { + "current_ce": 0.6123842968390538, + "replay_ce": 0.19816253085931143, + "dark_distillation": 0.35369483759770026, + "supervised_contrastive": 4.1878715148338905, + "total": 1.4061814057521331, + "replay_multiplier": 1.0 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.699632386729161, + "weighted_f1": 0.7285463664827341, + "macro_precision": 0.6803863664328781, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7612903225806451, + "NONE": 0.7222222222222222 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.96323867291609, + "weighted_f1_pct": 72.85463664827341, + "macro_precision_pct": 68.0386366432878, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.12903225806451, + "NONE": 72.22222222222221 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5808080808080809, + "weighted_f1": 0.5888357256778309, + "macro_precision": 0.5812999234051865, + "macro_recall": 0.6045751633986928, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.6136363636363636, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 58.08080808080809, + "weighted_f1_pct": 58.88357256778309, + "macro_precision_pct": 58.129992340518655, + "macro_recall_pct": 60.45751633986928, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 61.36363636363637, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7055837563451777, + "macro_f1": 0.696092682742745, + "weighted_f1": 0.7071399188897655, + "macro_precision": 0.680623973727422, + "macro_recall": 0.7358593745892028, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.72, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 70.55837563451777, + "macro_f1_pct": 69.6092682742745, + "weighted_f1_pct": 70.71399188897655, + "macro_precision_pct": 68.0623973727422, + "macro_recall_pct": 73.58593745892028, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 72.0, + "NONE": 71.07438016528926 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.665, + "macro_f1": 0.6782023131501855, + "weighted_f1": 0.662413623765953, + "macro_precision": 0.6739229564424355, + "macro_recall": 0.6883623252186938, + "per_class_f1": { + "FAVOR": 0.6174496644295302, + "AGAINST": 0.6585365853658537, + "NONE": 0.7586206896551724 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 67.82023131501856, + "weighted_f1_pct": 66.2413623765953, + "macro_precision_pct": 67.39229564424355, + "macro_recall_pct": 68.83623252186938, + "per_class_f1_pct": { + "FAVOR": 61.74496644295302, + "AGAINST": 65.85365853658537, + "NONE": 75.86206896551724 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8049448647579488, + "weighted_f1": 0.839672748850319, + "macro_precision": 0.7871855813032284, + "macro_recall": 0.8343235789512651, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.7936507936507936, + "NONE": 0.8878504672897196 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 80.49448647579489, + "weighted_f1_pct": 83.9672748850319, + "macro_precision_pct": 78.71855813032283, + "macro_recall_pct": 83.4323578951265, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 79.36507936507937, + "NONE": 88.78504672897196 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7223614592035644, + "weighted_f1": 0.849365301158501, + "macro_precision": 0.7003228410008071, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.8571428571428571, + "AGAINST": 0.42105263157894735, + "NONE": 0.8888888888888888 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.23614592035645, + "weighted_f1_pct": 84.93653011585009, + "macro_precision_pct": 70.03228410008072, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 85.71428571428571, + "AGAINST": 42.10526315789473, + "NONE": 88.88888888888889 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7052154195011338, + "weighted_f1": 0.7315119596225587, + "macro_precision": 0.6880185502827012, + "macro_recall": 0.741638608305275, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7755102040816326, + "NONE": 0.673469387755102 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.52154195011337, + "weighted_f1_pct": 73.15119596225587, + "macro_precision_pct": 68.80185502827013, + "macro_recall_pct": 74.1638608305275, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 77.55102040816327, + "NONE": 67.3469387755102 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.5555284542297529, + "weighted_f1": 0.5415700085537292, + "macro_precision": 0.5505175983436853, + "macro_recall": 0.5858974358974359, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.3877551020408163, + "NONE": 0.6942148760330579 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 55.552845422975295, + "weighted_f1_pct": 54.157000855372914, + "macro_precision_pct": 55.05175983436853, + "macro_recall_pct": 58.58974358974359, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 38.775510204081634, + "NONE": 69.42148760330579 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6891740938799763, + "weighted_f1": 0.7019982174688058, + "macro_precision": 0.7227888642522787, + "macro_recall": 0.6852332406549273, + "per_class_f1": { + "FAVOR": 0.611764705882353, + "AGAINST": 0.68, + "NONE": 0.7757575757575758 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 68.91740938799764, + "weighted_f1_pct": 70.19982174688057, + "macro_precision_pct": 72.27888642522787, + "macro_recall_pct": 68.52332406549273, + "per_class_f1_pct": { + "FAVOR": 61.1764705882353, + "AGAINST": 68.0, + "NONE": 77.57575757575758 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6435605096349986, + "weighted_f1": 0.6701077285109232, + "macro_precision": 0.6423148148148149, + "macro_recall": 0.6477493380406002, + "per_class_f1": { + "FAVOR": 0.7035175879396985, + "AGAINST": 0.45652173913043476, + "NONE": 0.7706422018348624 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.35605096349985, + "weighted_f1_pct": 67.01077285109231, + "macro_precision_pct": 64.2314814814815, + "macro_recall_pct": 64.77493380406003, + "per_class_f1_pct": { + "FAVOR": 70.35175879396985, + "AGAINST": 45.65217391304348, + "NONE": 77.06422018348624 + } + } + ], + "accuracy": 0.6978579481397971, + "macro_f1": 0.6988304093567251, + "weighted_f1": 0.6966027947086456, + "macro_precision": 0.6958727212046442, + "macro_recall": 0.7040497050480207, + "per_class_f1": { + "FAVOR": 0.6647286821705426, + "AGAINST": 0.6686046511627907, + "NONE": 0.7631578947368421 + }, + "accuracy_pct": 69.7857948139797, + "macro_f1_pct": 69.88304093567251, + "weighted_f1_pct": 69.66027947086457, + "macro_precision_pct": 69.58727212046442, + "macro_recall_pct": 70.40497050480207, + "per_class_f1_pct": { + "FAVOR": 66.47286821705426, + "AGAINST": 66.86046511627907, + "NONE": 76.31578947368422 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s7_contrastive_replay__x__s3__seed13/resolved_config.yaml b/outputs/trace_s7_contrastive_replay__x__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d4aedc80d839f2b7e2001336f8b3086a75572b21 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s7_contrastive_replay + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 7 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s7_contrastive_replay__x__s3__seed13/run.log b/outputs/trace_s7_contrastive_replay__x__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..e936b5a359754fbc787ede4d889d1ad0d0923f60 --- /dev/null +++ b/outputs/trace_s7_contrastive_replay__x__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 08:34:29,073 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s7_contrastive_replay__x__s3__seed13 device=cuda +2026-09-02 08:50:42,622 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6775520264637548, 'average_forgetting': 0.02205711610198248, 'backward_transfer': -0.006583798700073319, 'final_mean_target_macro_f1_pct': 67.75520264637548, 'average_forgetting_pct': 2.205711610198248, 'backward_transfer_pct': -0.6583798700073319, 'final_pooled_accuracy': 0.6978579481397971, 'final_pooled_accuracy_pct': 69.7857948139797, 'final_pooled_macro_f1': 0.6988304093567251, 'final_pooled_macro_f1_pct': 69.88304093567251, 'final_pooled_weighted_f1': 0.6966027947086456, 'final_pooled_weighted_f1_pct': 69.66027947086457, 'final_pooled_macro_precision': 0.6958727212046442, 'final_pooled_macro_precision_pct': 69.58727212046442, 'final_pooled_macro_recall': 0.7040497050480207, 'final_pooled_macro_recall_pct': 70.40497050480207, 'final_pooled_per_class_f1': {'FAVOR': 0.6647286821705426, 'AGAINST': 0.6686046511627907, 'NONE': 0.7631578947368421}, 'final_pooled_per_class_f1_pct': {'FAVOR': 66.47286821705426, 'AGAINST': 66.86046511627907, 'NONE': 76.31578947368422}} +2026-09-02 08:50:42,622 [INFO] __main__: efficiency={'training_wall_time_seconds': 972.2049490110949, 'training_wall_time_minutes': 16.20341581685158, 'peak_gpu_memory_allocated_gb': 10.40872573852539, 'peak_gpu_memory_reserved_gb': 10.587890625} diff --git a/outputs/trace_s8_full__doc__s1__seed13/environment.json b/outputs/trace_s8_full__doc__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..62df995258a23b84867ad1c52efe77fcc071eb6c --- /dev/null +++ b/outputs/trace_s8_full__doc__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T08:50:50.715668+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s1__seed13/metrics.json b/outputs/trace_s8_full__doc__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..10a105f1022fb06b659ee5609e08505cc84e2f36 --- /dev/null +++ b/outputs/trace_s8_full__doc__s1__seed13/metrics.json @@ -0,0 +1,1649 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5521828566854664, + 0.7024807388842477, + null, + null, + null, + null, + null, + null + ], + [ + 0.5514262676047573, + 0.6728619435802518, + 0.5879023146125258, + null, + null, + null, + null, + null + ], + [ + 0.5731378418059848, + 0.6793344692780678, + 0.5592688514128775, + 0.6052984649157862, + null, + null, + null, + null + ], + [ + 0.516494119769407, + 0.6670962669946734, + 0.5398826032838026, + 0.5614970603629533, + 0.6886960123959924, + null, + null, + null + ], + [ + 0.5399580446255724, + 0.6905418574933396, + 0.5519461699316053, + 0.5785336807655703, + 0.6855251191869449, + 0.6953117199096991, + null, + null + ], + [ + 0.5116545173797081, + 0.690278917145201, + 0.5481803324729765, + 0.5660650757412758, + 0.6800826920403474, + 0.6681143928738864, + 0.672418883322757, + null + ], + [ + 0.5036648473182558, + 0.6469653114594304, + 0.5155360266172196, + 0.5641442947958578, + 0.6550370871815985, + 0.6408548403304598, + 0.6368391557424404, + 0.6946931848988314 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 0.6157282389728079, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09377898486040985, + "ema_update_magnitude": 0.06227830254384934, + "final_multiplier": 0.6640965738384851 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7252063433385922, + "losses": { + "current_ce": 0.5511746357452302, + "replay_ce": 0.30564706095627375, + "dark_distillation": 0.43235585470994314, + "supervised_contrastive": 4.234550108682542, + "total": 1.2801156005405245, + "replay_multiplier": 0.5757239149154829, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10423027288050819, + "ema_update_magnitude": 0.05826846665653003, + "final_multiplier": 0.559035921582305 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6073697585768743, + "macro_f1": 0.5521828566854664, + "weighted_f1": 0.6352519527559964, + "macro_precision": 0.5587512624549662, + "macro_recall": 0.5930638997087798, + "per_class_f1": { + "FAVOR": 0.45382585751978893, + "AGAINST": 0.45576407506702415, + "NONE": 0.7469586374695864 + }, + "accuracy_pct": 60.736975857687426, + "macro_f1_pct": 55.21828566854664, + "weighted_f1_pct": 63.52519527559964, + "macro_precision_pct": 55.87512624549662, + "macro_recall_pct": 59.30638997087798, + "per_class_f1_pct": { + "FAVOR": 45.382585751978894, + "AGAINST": 45.57640750670242, + "NONE": 74.69586374695864 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7175697865353038, + "macro_f1": 0.7024807388842477, + "weighted_f1": 0.7256805350358198, + "macro_precision": 0.7028499374815308, + "macro_recall": 0.7184522968356403, + "per_class_f1": { + "FAVOR": 0.6130952380952381, + "AGAINST": 0.7017543859649122, + "NONE": 0.7925925925925926 + }, + "accuracy_pct": 71.75697865353038, + "macro_f1_pct": 70.24807388842477, + "weighted_f1_pct": 72.56805350358198, + "macro_precision_pct": 70.28499374815308, + "macro_recall_pct": 71.84522968356403, + "per_class_f1_pct": { + "FAVOR": 61.30952380952381, + "AGAINST": 70.17543859649122, + "NONE": 79.25925925925927 + } + } + ], + "accuracy": 0.6554441260744985, + "macro_f1": 0.6223830990350815, + "weighted_f1": 0.6722686414057607, + "macro_precision": 0.6240299225553477, + "macro_recall": 0.6530553327464843, + "per_class_f1": { + "FAVOR": 0.5286713286713287, + "AGAINST": 0.5734265734265734, + "NONE": 0.7650513950073421 + }, + "accuracy_pct": 65.54441260744986, + "macro_f1_pct": 62.238309903508146, + "weighted_f1_pct": 67.22686414057607, + "macro_precision_pct": 62.40299225553477, + "macro_recall_pct": 65.30553327464843, + "per_class_f1_pct": { + "FAVOR": 52.867132867132874, + "AGAINST": 57.34265734265735, + "NONE": 76.5051395007342 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.7119542251121199, + "losses": { + "current_ce": 0.5329047585482753, + "replay_ce": 0.20085617160380242, + "dark_distillation": 0.35809316140791725, + "supervised_contrastive": 4.138732387953334, + "total": 1.1647706680354617, + "replay_multiplier": 0.570103481893771, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0990981362310118, + "ema_update_magnitude": 0.05680334383553882, + "final_multiplier": 0.5732029480667747 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6086404066073697, + "macro_f1": 0.5514262676047573, + "weighted_f1": 0.6345650475743899, + "macro_precision": 0.5574386499035119, + "macro_recall": 0.5868993126618398, + "per_class_f1": { + "FAVOR": 0.42948717948717946, + "AGAINST": 0.4805491990846682, + "NONE": 0.7442424242424243 + }, + "accuracy_pct": 60.864040660736975, + "macro_f1_pct": 55.142626760475736, + "weighted_f1_pct": 63.45650475743899, + "macro_precision_pct": 55.74386499035119, + "macro_recall_pct": 58.68993126618398, + "per_class_f1_pct": { + "FAVOR": 42.94871794871795, + "AGAINST": 48.05491990846682, + "NONE": 74.42424242424242 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6830870279146142, + "macro_f1": 0.6728619435802518, + "weighted_f1": 0.6921832245378512, + "macro_precision": 0.6809552301499981, + "macro_recall": 0.6981548338713065, + "per_class_f1": { + "FAVOR": 0.608955223880597, + "AGAINST": 0.6596306068601583, + "NONE": 0.75 + }, + "accuracy_pct": 68.30870279146141, + "macro_f1_pct": 67.28619435802517, + "weighted_f1_pct": 69.21832245378512, + "macro_precision_pct": 68.09552301499981, + "macro_recall_pct": 69.81548338713066, + "per_class_f1_pct": { + "FAVOR": 60.895522388059696, + "AGAINST": 65.96306068601582, + "NONE": 75.0 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6210670314637483, + "macro_f1": 0.5879023146125258, + "weighted_f1": 0.6333945587670041, + "macro_precision": 0.5912416175008213, + "macro_recall": 0.6198250781444621, + "per_class_f1": { + "FAVOR": 0.46825396825396826, + "AGAINST": 0.5988909426987061, + "NONE": 0.6965620328849028 + }, + "accuracy_pct": 62.106703146374834, + "macro_f1_pct": 58.79023146125258, + "weighted_f1_pct": 63.33945587670041, + "macro_precision_pct": 59.12416175008212, + "macro_recall_pct": 61.98250781444621, + "per_class_f1_pct": { + "FAVOR": 46.82539682539682, + "AGAINST": 59.88909426987061, + "NONE": 69.65620328849027 + } + } + ], + "accuracy": 0.6342266102491773, + "macro_f1": 0.6049016704539135, + "weighted_f1": 0.6495023403189326, + "macro_precision": 0.6091309322643472, + "macro_recall": 0.6365013345404389, + "per_class_f1": { + "FAVOR": 0.5072302558398221, + "AGAINST": 0.5777450257921887, + "NONE": 0.7297297297297297 + }, + "accuracy_pct": 63.42266102491772, + "macro_f1_pct": 60.49016704539135, + "weighted_f1_pct": 64.95023403189326, + "macro_precision_pct": 60.913093226434725, + "macro_recall_pct": 63.65013345404389, + "per_class_f1_pct": { + "FAVOR": 50.723025583982206, + "AGAINST": 57.77450257921887, + "NONE": 72.97297297297297 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.6932320373404837, + "losses": { + "current_ce": 0.4968336216713253, + "replay_ce": 0.17727527264178844, + "dark_distillation": 0.33596014069360597, + "supervised_contrastive": 4.153137310345968, + "total": 1.111669364205578, + "replay_multiplier": 0.5743553565917187, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09877235114585066, + "ema_update_magnitude": 0.055328508784162816, + "final_multiplier": 0.5601619091000763 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6467598475222364, + "macro_f1": 0.5731378418059848, + "weighted_f1": 0.6678431169438671, + "macro_precision": 0.5688569069886892, + "macro_recall": 0.5976203575332116, + "per_class_f1": { + "FAVOR": 0.4430769230769231, + "AGAINST": 0.4827586206896552, + "NONE": 0.7935779816513762 + }, + "accuracy_pct": 64.67598475222364, + "macro_f1_pct": 57.31378418059848, + "weighted_f1_pct": 66.78431169438672, + "macro_precision_pct": 56.88569069886892, + "macro_recall_pct": 59.762035753321165, + "per_class_f1_pct": { + "FAVOR": 44.30769230769231, + "AGAINST": 48.275862068965516, + "NONE": 79.35779816513761 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6929392446633826, + "macro_f1": 0.6793344692780678, + "weighted_f1": 0.7026812161103249, + "macro_precision": 0.6833724340555638, + "macro_recall": 0.697889398823263, + "per_class_f1": { + "FAVOR": 0.591304347826087, + "AGAINST": 0.6763005780346821, + "NONE": 0.7703984819734345 + }, + "accuracy_pct": 69.29392446633827, + "macro_f1_pct": 67.93344692780678, + "weighted_f1_pct": 70.26812161103248, + "macro_precision_pct": 68.33724340555638, + "macro_recall_pct": 69.78893988232629, + "per_class_f1_pct": { + "FAVOR": 59.130434782608695, + "AGAINST": 67.63005780346822, + "NONE": 77.03984819734345 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.600547195622435, + "macro_f1": 0.5592688514128775, + "weighted_f1": 0.6204540199113022, + "macro_precision": 0.5658791449540711, + "macro_recall": 0.5905092099434248, + "per_class_f1": { + "FAVOR": 0.40425531914893614, + "AGAINST": 0.5634920634920635, + "NONE": 0.7100591715976331 + }, + "accuracy_pct": 60.0547195622435, + "macro_f1_pct": 55.926885141287755, + "weighted_f1_pct": 62.04540199113022, + "macro_precision_pct": 56.58791449540711, + "macro_recall_pct": 59.05092099434248, + "per_class_f1_pct": { + "FAVOR": 40.42553191489361, + "AGAINST": 56.34920634920635, + "NONE": 71.00591715976331 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6502242152466368, + "macro_f1": 0.6052984649157862, + "weighted_f1": 0.6644933114380682, + "macro_precision": 0.6109275146996885, + "macro_recall": 0.629994465770615, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.4740740740740741, + "NONE": 0.770392749244713 + }, + "accuracy_pct": 65.02242152466367, + "macro_f1_pct": 60.529846491578624, + "weighted_f1_pct": 66.44933114380682, + "macro_precision_pct": 61.09275146996885, + "macro_recall_pct": 62.999446577061505, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 47.40740740740741, + "NONE": 77.0392749244713 + } + } + ], + "accuracy": 0.6455650929899857, + "macro_f1": 0.6090103108232493, + "weighted_f1": 0.6619635931114013, + "macro_precision": 0.6093609953506981, + "macro_recall": 0.6355226780990644, + "per_class_f1": { + "FAVOR": 0.5110456553755522, + "AGAINST": 0.5531062124248497, + "NONE": 0.762879064669346 + }, + "accuracy_pct": 64.55650929899856, + "macro_f1_pct": 60.90103108232493, + "weighted_f1_pct": 66.19635931114013, + "macro_precision_pct": 60.93609953506981, + "macro_recall_pct": 63.55226780990644, + "per_class_f1_pct": { + "FAVOR": 51.104565537555224, + "AGAINST": 55.31062124248497, + "NONE": 76.2879064669346 + } + }, + { + "after_task": 5, + "best_epoch": 15, + "best_dev_macro_f1": 0.6779180228096636, + "losses": { + "current_ce": 0.4747798019221851, + "replay_ce": 0.17091749629804065, + "dark_distillation": 0.3337059034832886, + "supervised_contrastive": 4.133141693614778, + "total": 1.085216860544114, + "replay_multiplier": 0.5802333703632864, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09865730625482667, + "ema_update_magnitude": 0.055509740591312445, + "final_multiplier": 0.5626521004732653 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5552731893265566, + "macro_f1": 0.516494119769407, + "weighted_f1": 0.5900406413809998, + "macro_precision": 0.5459615440097827, + "macro_recall": 0.5685993926190006, + "per_class_f1": { + "FAVOR": 0.40616966580976865, + "AGAINST": 0.45647058823529413, + "NONE": 0.6868421052631579 + }, + "accuracy_pct": 55.52731893265565, + "macro_f1_pct": 51.6494119769407, + "weighted_f1_pct": 59.00406413809998, + "macro_precision_pct": 54.59615440097827, + "macro_recall_pct": 56.85993926190006, + "per_class_f1_pct": { + "FAVOR": 40.616966580976865, + "AGAINST": 45.64705882352941, + "NONE": 68.6842105263158 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6732348111658456, + "macro_f1": 0.6670962669946734, + "weighted_f1": 0.6817502198094194, + "macro_precision": 0.6846646392738109, + "macro_recall": 0.7002367232005934, + "per_class_f1": { + "FAVOR": 0.6062322946175638, + "AGAINST": 0.6719160104986877, + "NONE": 0.7231404958677686 + }, + "accuracy_pct": 67.32348111658456, + "macro_f1_pct": 66.70962669946734, + "weighted_f1_pct": 68.17502198094193, + "macro_precision_pct": 68.4664639273811, + "macro_recall_pct": 70.02367232005933, + "per_class_f1_pct": { + "FAVOR": 60.62322946175638, + "AGAINST": 67.19160104986877, + "NONE": 72.31404958677686 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5398826032838026, + "weighted_f1": 0.5899766647634495, + "macro_precision": 0.5582801783264747, + "macro_recall": 0.5837164469153087, + "per_class_f1": { + "FAVOR": 0.40924092409240925, + "AGAINST": 0.55, + "NONE": 0.6604068857589984 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 53.98826032838026, + "weighted_f1_pct": 58.997666476344946, + "macro_precision_pct": 55.82801783264747, + "macro_recall_pct": 58.37164469153087, + "per_class_f1_pct": { + "FAVOR": 40.92409240924093, + "AGAINST": 55.00000000000001, + "NONE": 66.04068857589984 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5614970603629533, + "weighted_f1": 0.6225208250766868, + "macro_precision": 0.5798539658181671, + "macro_recall": 0.5911978727879239, + "per_class_f1": { + "FAVOR": 0.5302325581395348, + "AGAINST": 0.4229390681003584, + "NONE": 0.7313195548489666 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.149706036295335, + "weighted_f1_pct": 62.25208250766868, + "macro_precision_pct": 57.98539658181671, + "macro_recall_pct": 59.11978727879239, + "per_class_f1_pct": { + "FAVOR": 53.02325581395348, + "AGAINST": 42.29390681003584, + "NONE": 73.13195548489666 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.6886960123959924, + "weighted_f1": 0.7176042995928901, + "macro_precision": 0.6863356382133206, + "macro_recall": 0.6964119561761418, + "per_class_f1": { + "FAVOR": 0.6335877862595419, + "AGAINST": 0.6341463414634146, + "NONE": 0.7983539094650206 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 68.86960123959925, + "weighted_f1_pct": 71.76042995928901, + "macro_precision_pct": 68.63356382133206, + "macro_recall_pct": 69.64119561761419, + "per_class_f1_pct": { + "FAVOR": 63.358778625954194, + "AGAINST": 63.41463414634146, + "NONE": 79.83539094650206 + } + } + ], + "accuracy": 0.613726085636198, + "macro_f1": 0.5908349932138371, + "weighted_f1": 0.6319100016090985, + "macro_precision": 0.6022653055325016, + "macro_recall": 0.6270816123640404, + "per_class_f1": { + "FAVOR": 0.51237766263673, + "AGAINST": 0.5456509994597515, + "NONE": 0.71447631754503 + }, + "accuracy_pct": 61.3726085636198, + "macro_f1_pct": 59.08349932138371, + "weighted_f1_pct": 63.19100016090985, + "macro_precision_pct": 60.22653055325016, + "macro_recall_pct": 62.70816123640404, + "per_class_f1_pct": { + "FAVOR": 51.237766263673, + "AGAINST": 54.56509994597515, + "NONE": 71.447631754503 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.7190146750524109, + "losses": { + "current_ce": 0.37990011626056264, + "replay_ce": 0.13867790674169858, + "dark_distillation": 0.31864081465062644, + "supervised_contrastive": 4.059075568403516, + "total": 0.9565982834214256, + "replay_multiplier": 0.5683018831519859, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09782326756857436, + "ema_update_magnitude": 0.05448669148246428, + "final_multiplier": 0.5569911211999637 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5399580446255724, + "weighted_f1": 0.627200774318466, + "macro_precision": 0.5546492715516013, + "macro_recall": 0.5744733097129612, + "per_class_f1": { + "FAVOR": 0.4028268551236749, + "AGAINST": 0.47558386411889597, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 53.99580446255724, + "weighted_f1_pct": 62.7200774318466, + "macro_precision_pct": 55.464927155160126, + "macro_recall_pct": 57.44733097129612, + "per_class_f1_pct": { + "FAVOR": 40.282685512367486, + "AGAINST": 47.5583864118896, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7027914614121511, + "macro_f1": 0.6905418574933396, + "weighted_f1": 0.7094660071410845, + "macro_precision": 0.6928889957008199, + "macro_recall": 0.712324205358498, + "per_class_f1": { + "FAVOR": 0.6198083067092651, + "AGAINST": 0.6873385012919897, + "NONE": 0.7644787644787645 + }, + "accuracy_pct": 70.27914614121511, + "macro_f1_pct": 69.05418574933397, + "weighted_f1_pct": 70.94660071410846, + "macro_precision_pct": 69.28889957008198, + "macro_recall_pct": 71.23242053584981, + "per_class_f1_pct": { + "FAVOR": 61.980830670926515, + "AGAINST": 68.73385012919897, + "NONE": 76.44787644787645 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5937072503419972, + "macro_f1": 0.5519461699316053, + "weighted_f1": 0.6102337171233613, + "macro_precision": 0.5658123696994135, + "macro_recall": 0.5802495437580807, + "per_class_f1": { + "FAVOR": 0.39215686274509803, + "AGAINST": 0.5776173285198556, + "NONE": 0.6860643185298622 + }, + "accuracy_pct": 59.37072503419972, + "macro_f1_pct": 55.19461699316053, + "weighted_f1_pct": 61.02337171233613, + "macro_precision_pct": 56.58123696994135, + "macro_recall_pct": 58.02495437580807, + "per_class_f1_pct": { + "FAVOR": 39.21568627450981, + "AGAINST": 57.761732851985556, + "NONE": 68.60643185298622 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6263079222720478, + "macro_f1": 0.5785336807655703, + "weighted_f1": 0.6468238508983174, + "macro_precision": 0.5818109325530109, + "macro_recall": 0.5981456219075585, + "per_class_f1": { + "FAVOR": 0.47802197802197804, + "AGAINST": 0.48231511254019294, + "NONE": 0.77526395173454 + }, + "accuracy_pct": 62.63079222720478, + "macro_f1_pct": 57.85336807655703, + "weighted_f1_pct": 64.68238508983174, + "macro_precision_pct": 58.181093255301086, + "macro_recall_pct": 59.81456219075585, + "per_class_f1_pct": { + "FAVOR": 47.8021978021978, + "AGAINST": 48.231511254019296, + "NONE": 77.526395173454 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.710261569416499, + "macro_f1": 0.6855251191869449, + "weighted_f1": 0.7145949401055764, + "macro_precision": 0.6814432946990623, + "macro_recall": 0.6938493968707675, + "per_class_f1": { + "FAVOR": 0.6307053941908713, + "AGAINST": 0.63003663003663, + "NONE": 0.7958333333333333 + }, + "accuracy_pct": 71.0261569416499, + "macro_f1_pct": 68.55251191869449, + "weighted_f1_pct": 71.45949401055765, + "macro_precision_pct": 68.14432946990622, + "macro_recall_pct": 69.38493968707675, + "per_class_f1_pct": { + "FAVOR": 63.07053941908713, + "AGAINST": 63.003663003663, + "NONE": 79.58333333333333 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7263581488933601, + "macro_f1": 0.6953117199096991, + "weighted_f1": 0.7387450302477794, + "macro_precision": 0.6913463778243653, + "macro_recall": 0.7182734458596528, + "per_class_f1": { + "FAVOR": 0.6015037593984962, + "AGAINST": 0.6504065040650406, + "NONE": 0.8340248962655602 + }, + "accuracy_pct": 72.63581488933602, + "macro_f1_pct": 69.53117199096991, + "weighted_f1_pct": 73.87450302477794, + "macro_precision_pct": 69.13463778243653, + "macro_recall_pct": 71.82734458596528, + "per_class_f1_pct": { + "FAVOR": 60.150375939849624, + "AGAINST": 65.04065040650406, + "NONE": 83.40248962655602 + } + } + ], + "accuracy": 0.6511873350923483, + "macro_f1": 0.6186479330214674, + "weighted_f1": 0.6658328688524984, + "macro_precision": 0.6205444032655976, + "macro_recall": 0.6431994185322474, + "per_class_f1": { + "FAVOR": 0.519163763066202, + "AGAINST": 0.576271186440678, + "NONE": 0.7605088495575221 + }, + "accuracy_pct": 65.11873350923483, + "macro_f1_pct": 61.86479330214674, + "weighted_f1_pct": 66.58328688524983, + "macro_precision_pct": 62.05444032655976, + "macro_recall_pct": 64.31994185322473, + "per_class_f1_pct": { + "FAVOR": 51.91637630662021, + "AGAINST": 57.6271186440678, + "NONE": 76.05088495575221 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.7455270655270656, + "losses": { + "current_ce": 0.3234014348013372, + "replay_ce": 0.13548697528800344, + "dark_distillation": 0.3017914269755526, + "supervised_contrastive": 3.9907342930150227, + "total": 0.8831816029257891, + "replay_multiplier": 0.5576599656599283, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09962185720878244, + "ema_update_magnitude": 0.051961950418673256, + "final_multiplier": 0.5215918662284525 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5116545173797081, + "weighted_f1": 0.600083762518169, + "macro_precision": 0.545293547779049, + "macro_recall": 0.5517218297719387, + "per_class_f1": { + "FAVOR": 0.35877862595419846, + "AGAINST": 0.46153846153846156, + "NONE": 0.7146464646464646 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 51.16545173797081, + "weighted_f1_pct": 60.0083762518169, + "macro_precision_pct": 54.529354777904906, + "macro_recall_pct": 55.17218297719387, + "per_class_f1_pct": { + "FAVOR": 35.87786259541985, + "AGAINST": 46.15384615384615, + "NONE": 71.46464646464646 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.690278917145201, + "weighted_f1": 0.7117972684816621, + "macro_precision": 0.6895242148612936, + "macro_recall": 0.7084438066526247, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.6684782608695652, + "NONE": 0.7773584905660378 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.0278917145201, + "weighted_f1_pct": 71.17972684816621, + "macro_precision_pct": 68.95242148612935, + "macro_recall_pct": 70.84438066526248, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 66.84782608695652, + "NONE": 77.73584905660378 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5772913816689467, + "macro_f1": 0.5481803324729765, + "weighted_f1": 0.5933111760236676, + "macro_precision": 0.5649628098741202, + "macro_recall": 0.5833398148429954, + "per_class_f1": { + "FAVOR": 0.4351145038167939, + "AGAINST": 0.5490196078431373, + "NONE": 0.6604068857589984 + }, + "accuracy_pct": 57.729138166894664, + "macro_f1_pct": 54.81803324729765, + "weighted_f1_pct": 59.33111760236676, + "macro_precision_pct": 56.49628098741202, + "macro_recall_pct": 58.333981484299535, + "per_class_f1_pct": { + "FAVOR": 43.51145038167939, + "AGAINST": 54.90196078431373, + "NONE": 66.04068857589984 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6068759342301944, + "macro_f1": 0.5660650757412758, + "weighted_f1": 0.6304498894179335, + "macro_precision": 0.575518417455123, + "macro_recall": 0.5911585104345486, + "per_class_f1": { + "FAVOR": 0.49025069637883006, + "AGAINST": 0.4583333333333333, + "NONE": 0.749611197511664 + }, + "accuracy_pct": 60.687593423019436, + "macro_f1_pct": 56.60650757412758, + "weighted_f1_pct": 63.04498894179334, + "macro_precision_pct": 57.5518417455123, + "macro_recall_pct": 59.11585104345486, + "per_class_f1_pct": { + "FAVOR": 49.02506963788301, + "AGAINST": 45.83333333333333, + "NONE": 74.96111975116641 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6800826920403474, + "weighted_f1": 0.7108859018019771, + "macro_precision": 0.6754504682657735, + "macro_recall": 0.6894551721375598, + "per_class_f1": { + "FAVOR": 0.6303501945525292, + "AGAINST": 0.6124031007751938, + "NONE": 0.7974947807933194 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 68.00826920403475, + "weighted_f1_pct": 71.08859018019771, + "macro_precision_pct": 67.54504682657735, + "macro_recall_pct": 68.94551721375598, + "per_class_f1_pct": { + "FAVOR": 63.035019455252915, + "AGAINST": 61.240310077519375, + "NONE": 79.74947807933194 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7002012072434608, + "macro_f1": 0.6681143928738864, + "weighted_f1": 0.7153377876258961, + "macro_precision": 0.6680788295261423, + "macro_recall": 0.6918723321022172, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.608, + "NONE": 0.8185654008438819 + }, + "accuracy_pct": 70.02012072434609, + "macro_f1_pct": 66.81143928738864, + "weighted_f1_pct": 71.5337787625896, + "macro_precision_pct": 66.80788295261422, + "macro_recall_pct": 69.18723321022172, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 60.8, + "NONE": 81.85654008438819 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7196652719665272, + "macro_f1": 0.672418883322757, + "weighted_f1": 0.7296119580386282, + "macro_precision": 0.6640642135862489, + "macro_recall": 0.7008576575989928, + "per_class_f1": { + "FAVOR": 0.5482625482625483, + "AGAINST": 0.6585365853658537, + "NONE": 0.8104575163398693 + }, + "accuracy_pct": 71.96652719665272, + "macro_f1_pct": 67.2418883322757, + "weighted_f1_pct": 72.96119580386282, + "macro_precision_pct": 66.4064213586249, + "macro_recall_pct": 70.08576575989927, + "per_class_f1_pct": { + "FAVOR": 54.826254826254825, + "AGAINST": 65.85365853658537, + "NONE": 81.04575163398692 + } + } + ], + "accuracy": 0.6483248280452629, + "macro_f1": 0.6157472043872111, + "weighted_f1": 0.6638956685746421, + "macro_precision": 0.6188087106906376, + "macro_recall": 0.6422836668788158, + "per_class_f1": { + "FAVOR": 0.5248868778280543, + "AGAINST": 0.5652978172401036, + "NONE": 0.7570569180934752 + }, + "accuracy_pct": 64.8324828045263, + "macro_f1_pct": 61.57472043872111, + "weighted_f1_pct": 66.38956685746422, + "macro_precision_pct": 61.88087106906376, + "macro_recall_pct": 64.22836668788158, + "per_class_f1_pct": { + "FAVOR": 52.488687782805435, + "AGAINST": 56.52978172401036, + "NONE": 75.70569180934751 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.7801174049850865, + "losses": { + "current_ce": 0.30776806522672084, + "replay_ce": 0.12483645702401797, + "dark_distillation": 0.29410668415766134, + "supervised_contrastive": 4.0093502114800845, + "total": 0.8577304455579496, + "replay_multiplier": 0.5454794998550193, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10364761917169144, + "ema_update_magnitude": 0.05460331420835193, + "final_multiplier": 0.5268168689712205 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5501905972045743, + "macro_f1": 0.5036648473182558, + "weighted_f1": 0.5845183803725745, + "macro_precision": 0.5374770198079385, + "macro_recall": 0.5465656932432532, + "per_class_f1": { + "FAVOR": 0.38311688311688313, + "AGAINST": 0.43687374749499, + "NONE": 0.6910039113428944 + }, + "accuracy_pct": 55.01905972045743, + "macro_f1_pct": 50.36648473182558, + "weighted_f1_pct": 58.45183803725745, + "macro_precision_pct": 53.74770198079385, + "macro_recall_pct": 54.65656932432532, + "per_class_f1_pct": { + "FAVOR": 38.311688311688314, + "AGAINST": 43.687374749499, + "NONE": 69.10039113428944 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6518883415435139, + "macro_f1": 0.6469653114594304, + "weighted_f1": 0.6610862460956662, + "macro_precision": 0.6691896502425573, + "macro_recall": 0.6812547525958426, + "per_class_f1": { + "FAVOR": 0.5907859078590786, + "AGAINST": 0.6486486486486487, + "NONE": 0.7014613778705637 + }, + "accuracy_pct": 65.1888341543514, + "macro_f1_pct": 64.69653114594304, + "weighted_f1_pct": 66.10862460956662, + "macro_precision_pct": 66.91896502425573, + "macro_recall_pct": 68.12547525958425, + "per_class_f1_pct": { + "FAVOR": 59.078590785907856, + "AGAINST": 64.86486486486487, + "NONE": 70.14613778705638 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.533515731874145, + "macro_f1": 0.5155360266172196, + "weighted_f1": 0.5579125621463014, + "macro_precision": 0.5543455652949042, + "macro_recall": 0.5804493433385298, + "per_class_f1": { + "FAVOR": 0.41456582633053224, + "AGAINST": 0.5070422535211268, + "NONE": 0.625 + }, + "accuracy_pct": 53.3515731874145, + "macro_f1_pct": 51.55360266172197, + "weighted_f1_pct": 55.79125621463014, + "macro_precision_pct": 55.43455652949042, + "macro_recall_pct": 58.04493433385298, + "per_class_f1_pct": { + "FAVOR": 41.45658263305322, + "AGAINST": 50.70422535211267, + "NONE": 62.5 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5964125560538116, + "macro_f1": 0.5641442947958578, + "weighted_f1": 0.6161736909232732, + "macro_precision": 0.5843596840223722, + "macro_recall": 0.6011007358997588, + "per_class_f1": { + "FAVOR": 0.5467289719626168, + "AGAINST": 0.4377104377104377, + "NONE": 0.7079934747145188 + }, + "accuracy_pct": 59.64125560538116, + "macro_f1_pct": 56.414429479585785, + "weighted_f1_pct": 61.61736909232732, + "macro_precision_pct": 58.435968402237215, + "macro_recall_pct": 60.110073589975876, + "per_class_f1_pct": { + "FAVOR": 54.67289719626168, + "AGAINST": 43.77104377104377, + "NONE": 70.79934747145188 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6550370871815985, + "weighted_f1": 0.6896607333298811, + "macro_precision": 0.6537439909896304, + "macro_recall": 0.6668256414019126, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.5905511811023622, + "NONE": 0.7863247863247863 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 65.50370871815984, + "weighted_f1_pct": 68.9660733329881, + "macro_precision_pct": 65.37439909896304, + "macro_recall_pct": 66.68256414019126, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 59.055118110236215, + "NONE": 78.63247863247864 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.670020120724346, + "macro_f1": 0.6408548403304598, + "weighted_f1": 0.6880991232704631, + "macro_precision": 0.6529306066569692, + "macro_recall": 0.6638644391517955, + "per_class_f1": { + "FAVOR": 0.5496688741721855, + "AGAINST": 0.5814977973568282, + "NONE": 0.7913978494623656 + }, + "accuracy_pct": 67.0020120724346, + "macro_f1_pct": 64.08548403304599, + "weighted_f1_pct": 68.8099123270463, + "macro_precision_pct": 65.29306066569693, + "macro_recall_pct": 66.38644391517956, + "per_class_f1_pct": { + "FAVOR": 54.966887417218544, + "AGAINST": 58.14977973568281, + "NONE": 79.13978494623656 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6638772663877266, + "macro_f1": 0.6368391557424404, + "weighted_f1": 0.679497119275485, + "macro_precision": 0.6377108770902448, + "macro_recall": 0.6947101166153864, + "per_class_f1": { + "FAVOR": 0.5266457680250783, + "AGAINST": 0.6474820143884892, + "NONE": 0.7363896848137536 + }, + "accuracy_pct": 66.38772663877266, + "macro_f1_pct": 63.68391557424404, + "weighted_f1_pct": 67.94971192754849, + "macro_precision_pct": 63.77108770902448, + "macro_recall_pct": 69.47101166153864, + "per_class_f1_pct": { + "FAVOR": 52.66457680250783, + "AGAINST": 64.74820143884892, + "NONE": 73.63896848137536 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7259136212624585, + "macro_f1": 0.6946931848988314, + "weighted_f1": 0.7346643845237473, + "macro_precision": 0.6827680686114421, + "macro_recall": 0.7328767726916422, + "per_class_f1": { + "FAVOR": 0.6545454545454545, + "AGAINST": 0.6225165562913907, + "NONE": 0.8070175438596491 + }, + "accuracy_pct": 72.59136212624585, + "macro_f1_pct": 69.46931848988314, + "weighted_f1_pct": 73.46643845237473, + "macro_precision_pct": 68.27680686114421, + "macro_recall_pct": 73.28767726916422, + "per_class_f1_pct": { + "FAVOR": 65.45454545454545, + "AGAINST": 62.251655629139066, + "NONE": 80.7017543859649 + } + } + ], + "accuracy": 0.6271285965942455, + "macro_f1": 0.6027806676545912, + "weighted_f1": 0.6453364209217395, + "macro_precision": 0.6120526566603584, + "macro_recall": 0.6421302193459932, + "per_class_f1": { + "FAVOR": 0.529277566539924, + "AGAINST": 0.5518686692280824, + "NONE": 0.7271957671957672 + }, + "accuracy_pct": 62.712859659424545, + "macro_f1_pct": 60.278066765459116, + "weighted_f1_pct": 64.53364209217395, + "macro_precision_pct": 61.205265666035835, + "macro_recall_pct": 64.21302193459933, + "per_class_f1_pct": { + "FAVOR": 52.9277566539924, + "AGAINST": 55.18686692280824, + "NONE": 72.71957671957672 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6072168435430118, + "average_forgetting": 0.05425147172332497, + "backward_transfer": -0.05425147172332497, + "final_mean_target_macro_f1_pct": 60.721684354301175, + "average_forgetting_pct": 5.4251471723324975, + "backward_transfer_pct": -5.4251471723324975, + "final_pooled_accuracy": 0.6271285965942455, + "final_pooled_accuracy_pct": 62.712859659424545, + "final_pooled_macro_f1": 0.6027806676545912, + "final_pooled_macro_f1_pct": 60.278066765459116, + "final_pooled_weighted_f1": 0.6453364209217395, + "final_pooled_weighted_f1_pct": 64.53364209217395, + "final_pooled_macro_precision": 0.6120526566603584, + "final_pooled_macro_precision_pct": 61.205265666035835, + "final_pooled_macro_recall": 0.6421302193459932, + "final_pooled_macro_recall_pct": 64.21302193459933, + "final_pooled_per_class_f1": { + "FAVOR": 0.529277566539924, + "AGAINST": 0.5518686692280824, + "NONE": 0.7271957671957672 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 52.9277566539924, + "AGAINST": 55.18686692280824, + "NONE": 72.71957671957672 + } + }, + "efficiency": { + "training_wall_time_seconds": 1838.6547880358994, + "training_wall_time_minutes": 30.64424646726499, + "peak_gpu_memory_allocated_gb": 10.430231094360352, + "peak_gpu_memory_reserved_gb": 10.6015625 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s1__seed13/metrics.partial.json b/outputs/trace_s8_full__doc__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..d2abcd058a41d1a76a4f68c7aa72dba0376b8046 --- /dev/null +++ b/outputs/trace_s8_full__doc__s1__seed13/metrics.partial.json @@ -0,0 +1,1543 @@ +{ + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5521828566854664, + 0.7024807388842477, + null, + null, + null, + null, + null, + null + ], + [ + 0.5514262676047573, + 0.6728619435802518, + 0.5879023146125258, + null, + null, + null, + null, + null + ], + [ + 0.5731378418059848, + 0.6793344692780678, + 0.5592688514128775, + 0.6052984649157862, + null, + null, + null, + null + ], + [ + 0.516494119769407, + 0.6670962669946734, + 0.5398826032838026, + 0.5614970603629533, + 0.6886960123959924, + null, + null, + null + ], + [ + 0.5399580446255724, + 0.6905418574933396, + 0.5519461699316053, + 0.5785336807655703, + 0.6855251191869449, + 0.6953117199096991, + null, + null + ], + [ + 0.5116545173797081, + 0.690278917145201, + 0.5481803324729765, + 0.5660650757412758, + 0.6800826920403474, + 0.6681143928738864, + 0.672418883322757, + null + ], + [ + 0.5036648473182558, + 0.6469653114594304, + 0.5155360266172196, + 0.5641442947958578, + 0.6550370871815985, + 0.6408548403304598, + 0.6368391557424404, + 0.6946931848988314 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 0.6157282389728079, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09377898486040985, + "ema_update_magnitude": 0.06227830254384934, + "final_multiplier": 0.6640965738384851 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.7252063433385922, + "losses": { + "current_ce": 0.5511746357452302, + "replay_ce": 0.30564706095627375, + "dark_distillation": 0.43235585470994314, + "supervised_contrastive": 4.234550108682542, + "total": 1.2801156005405245, + "replay_multiplier": 0.5757239149154829, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10423027288050819, + "ema_update_magnitude": 0.05826846665653003, + "final_multiplier": 0.559035921582305 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6073697585768743, + "macro_f1": 0.5521828566854664, + "weighted_f1": 0.6352519527559964, + "macro_precision": 0.5587512624549662, + "macro_recall": 0.5930638997087798, + "per_class_f1": { + "FAVOR": 0.45382585751978893, + "AGAINST": 0.45576407506702415, + "NONE": 0.7469586374695864 + }, + "accuracy_pct": 60.736975857687426, + "macro_f1_pct": 55.21828566854664, + "weighted_f1_pct": 63.52519527559964, + "macro_precision_pct": 55.87512624549662, + "macro_recall_pct": 59.30638997087798, + "per_class_f1_pct": { + "FAVOR": 45.382585751978894, + "AGAINST": 45.57640750670242, + "NONE": 74.69586374695864 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7175697865353038, + "macro_f1": 0.7024807388842477, + "weighted_f1": 0.7256805350358198, + "macro_precision": 0.7028499374815308, + "macro_recall": 0.7184522968356403, + "per_class_f1": { + "FAVOR": 0.6130952380952381, + "AGAINST": 0.7017543859649122, + "NONE": 0.7925925925925926 + }, + "accuracy_pct": 71.75697865353038, + "macro_f1_pct": 70.24807388842477, + "weighted_f1_pct": 72.56805350358198, + "macro_precision_pct": 70.28499374815308, + "macro_recall_pct": 71.84522968356403, + "per_class_f1_pct": { + "FAVOR": 61.30952380952381, + "AGAINST": 70.17543859649122, + "NONE": 79.25925925925927 + } + } + ], + "accuracy": 0.6554441260744985, + "macro_f1": 0.6223830990350815, + "weighted_f1": 0.6722686414057607, + "macro_precision": 0.6240299225553477, + "macro_recall": 0.6530553327464843, + "per_class_f1": { + "FAVOR": 0.5286713286713287, + "AGAINST": 0.5734265734265734, + "NONE": 0.7650513950073421 + }, + "accuracy_pct": 65.54441260744986, + "macro_f1_pct": 62.238309903508146, + "weighted_f1_pct": 67.22686414057607, + "macro_precision_pct": 62.40299225553477, + "macro_recall_pct": 65.30553327464843, + "per_class_f1_pct": { + "FAVOR": 52.867132867132874, + "AGAINST": 57.34265734265735, + "NONE": 76.5051395007342 + } + }, + { + "after_task": 3, + "best_epoch": 10, + "best_dev_macro_f1": 0.7119542251121199, + "losses": { + "current_ce": 0.5329047585482753, + "replay_ce": 0.20085617160380242, + "dark_distillation": 0.35809316140791725, + "supervised_contrastive": 4.138732387953334, + "total": 1.1647706680354617, + "replay_multiplier": 0.570103481893771, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0990981362310118, + "ema_update_magnitude": 0.05680334383553882, + "final_multiplier": 0.5732029480667747 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6086404066073697, + "macro_f1": 0.5514262676047573, + "weighted_f1": 0.6345650475743899, + "macro_precision": 0.5574386499035119, + "macro_recall": 0.5868993126618398, + "per_class_f1": { + "FAVOR": 0.42948717948717946, + "AGAINST": 0.4805491990846682, + "NONE": 0.7442424242424243 + }, + "accuracy_pct": 60.864040660736975, + "macro_f1_pct": 55.142626760475736, + "weighted_f1_pct": 63.45650475743899, + "macro_precision_pct": 55.74386499035119, + "macro_recall_pct": 58.68993126618398, + "per_class_f1_pct": { + "FAVOR": 42.94871794871795, + "AGAINST": 48.05491990846682, + "NONE": 74.42424242424242 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6830870279146142, + "macro_f1": 0.6728619435802518, + "weighted_f1": 0.6921832245378512, + "macro_precision": 0.6809552301499981, + "macro_recall": 0.6981548338713065, + "per_class_f1": { + "FAVOR": 0.608955223880597, + "AGAINST": 0.6596306068601583, + "NONE": 0.75 + }, + "accuracy_pct": 68.30870279146141, + "macro_f1_pct": 67.28619435802517, + "weighted_f1_pct": 69.21832245378512, + "macro_precision_pct": 68.09552301499981, + "macro_recall_pct": 69.81548338713066, + "per_class_f1_pct": { + "FAVOR": 60.895522388059696, + "AGAINST": 65.96306068601582, + "NONE": 75.0 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.6210670314637483, + "macro_f1": 0.5879023146125258, + "weighted_f1": 0.6333945587670041, + "macro_precision": 0.5912416175008213, + "macro_recall": 0.6198250781444621, + "per_class_f1": { + "FAVOR": 0.46825396825396826, + "AGAINST": 0.5988909426987061, + "NONE": 0.6965620328849028 + }, + "accuracy_pct": 62.106703146374834, + "macro_f1_pct": 58.79023146125258, + "weighted_f1_pct": 63.33945587670041, + "macro_precision_pct": 59.12416175008212, + "macro_recall_pct": 61.98250781444621, + "per_class_f1_pct": { + "FAVOR": 46.82539682539682, + "AGAINST": 59.88909426987061, + "NONE": 69.65620328849027 + } + } + ], + "accuracy": 0.6342266102491773, + "macro_f1": 0.6049016704539135, + "weighted_f1": 0.6495023403189326, + "macro_precision": 0.6091309322643472, + "macro_recall": 0.6365013345404389, + "per_class_f1": { + "FAVOR": 0.5072302558398221, + "AGAINST": 0.5777450257921887, + "NONE": 0.7297297297297297 + }, + "accuracy_pct": 63.42266102491772, + "macro_f1_pct": 60.49016704539135, + "weighted_f1_pct": 64.95023403189326, + "macro_precision_pct": 60.913093226434725, + "macro_recall_pct": 63.65013345404389, + "per_class_f1_pct": { + "FAVOR": 50.723025583982206, + "AGAINST": 57.77450257921887, + "NONE": 72.97297297297297 + } + }, + { + "after_task": 4, + "best_epoch": 12, + "best_dev_macro_f1": 0.6932320373404837, + "losses": { + "current_ce": 0.4968336216713253, + "replay_ce": 0.17727527264178844, + "dark_distillation": 0.33596014069360597, + "supervised_contrastive": 4.153137310345968, + "total": 1.111669364205578, + "replay_multiplier": 0.5743553565917187, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09877235114585066, + "ema_update_magnitude": 0.055328508784162816, + "final_multiplier": 0.5601619091000763 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6467598475222364, + "macro_f1": 0.5731378418059848, + "weighted_f1": 0.6678431169438671, + "macro_precision": 0.5688569069886892, + "macro_recall": 0.5976203575332116, + "per_class_f1": { + "FAVOR": 0.4430769230769231, + "AGAINST": 0.4827586206896552, + "NONE": 0.7935779816513762 + }, + "accuracy_pct": 64.67598475222364, + "macro_f1_pct": 57.31378418059848, + "weighted_f1_pct": 66.78431169438672, + "macro_precision_pct": 56.88569069886892, + "macro_recall_pct": 59.762035753321165, + "per_class_f1_pct": { + "FAVOR": 44.30769230769231, + "AGAINST": 48.275862068965516, + "NONE": 79.35779816513761 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6929392446633826, + "macro_f1": 0.6793344692780678, + "weighted_f1": 0.7026812161103249, + "macro_precision": 0.6833724340555638, + "macro_recall": 0.697889398823263, + "per_class_f1": { + "FAVOR": 0.591304347826087, + "AGAINST": 0.6763005780346821, + "NONE": 0.7703984819734345 + }, + "accuracy_pct": 69.29392446633827, + "macro_f1_pct": 67.93344692780678, + "weighted_f1_pct": 70.26812161103248, + "macro_precision_pct": 68.33724340555638, + "macro_recall_pct": 69.78893988232629, + "per_class_f1_pct": { + "FAVOR": 59.130434782608695, + "AGAINST": 67.63005780346822, + "NONE": 77.03984819734345 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.600547195622435, + "macro_f1": 0.5592688514128775, + "weighted_f1": 0.6204540199113022, + "macro_precision": 0.5658791449540711, + "macro_recall": 0.5905092099434248, + "per_class_f1": { + "FAVOR": 0.40425531914893614, + "AGAINST": 0.5634920634920635, + "NONE": 0.7100591715976331 + }, + "accuracy_pct": 60.0547195622435, + "macro_f1_pct": 55.926885141287755, + "weighted_f1_pct": 62.04540199113022, + "macro_precision_pct": 56.58791449540711, + "macro_recall_pct": 59.05092099434248, + "per_class_f1_pct": { + "FAVOR": 40.42553191489361, + "AGAINST": 56.34920634920635, + "NONE": 71.00591715976331 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6502242152466368, + "macro_f1": 0.6052984649157862, + "weighted_f1": 0.6644933114380682, + "macro_precision": 0.6109275146996885, + "macro_recall": 0.629994465770615, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.4740740740740741, + "NONE": 0.770392749244713 + }, + "accuracy_pct": 65.02242152466367, + "macro_f1_pct": 60.529846491578624, + "weighted_f1_pct": 66.44933114380682, + "macro_precision_pct": 61.09275146996885, + "macro_recall_pct": 62.999446577061505, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 47.40740740740741, + "NONE": 77.0392749244713 + } + } + ], + "accuracy": 0.6455650929899857, + "macro_f1": 0.6090103108232493, + "weighted_f1": 0.6619635931114013, + "macro_precision": 0.6093609953506981, + "macro_recall": 0.6355226780990644, + "per_class_f1": { + "FAVOR": 0.5110456553755522, + "AGAINST": 0.5531062124248497, + "NONE": 0.762879064669346 + }, + "accuracy_pct": 64.55650929899856, + "macro_f1_pct": 60.90103108232493, + "weighted_f1_pct": 66.19635931114013, + "macro_precision_pct": 60.93609953506981, + "macro_recall_pct": 63.55226780990644, + "per_class_f1_pct": { + "FAVOR": 51.104565537555224, + "AGAINST": 55.31062124248497, + "NONE": 76.2879064669346 + } + }, + { + "after_task": 5, + "best_epoch": 15, + "best_dev_macro_f1": 0.6779180228096636, + "losses": { + "current_ce": 0.4747798019221851, + "replay_ce": 0.17091749629804065, + "dark_distillation": 0.3337059034832886, + "supervised_contrastive": 4.133141693614778, + "total": 1.085216860544114, + "replay_multiplier": 0.5802333703632864, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09865730625482667, + "ema_update_magnitude": 0.055509740591312445, + "final_multiplier": 0.5626521004732653 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5552731893265566, + "macro_f1": 0.516494119769407, + "weighted_f1": 0.5900406413809998, + "macro_precision": 0.5459615440097827, + "macro_recall": 0.5685993926190006, + "per_class_f1": { + "FAVOR": 0.40616966580976865, + "AGAINST": 0.45647058823529413, + "NONE": 0.6868421052631579 + }, + "accuracy_pct": 55.52731893265565, + "macro_f1_pct": 51.6494119769407, + "weighted_f1_pct": 59.00406413809998, + "macro_precision_pct": 54.59615440097827, + "macro_recall_pct": 56.85993926190006, + "per_class_f1_pct": { + "FAVOR": 40.616966580976865, + "AGAINST": 45.64705882352941, + "NONE": 68.6842105263158 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6732348111658456, + "macro_f1": 0.6670962669946734, + "weighted_f1": 0.6817502198094194, + "macro_precision": 0.6846646392738109, + "macro_recall": 0.7002367232005934, + "per_class_f1": { + "FAVOR": 0.6062322946175638, + "AGAINST": 0.6719160104986877, + "NONE": 0.7231404958677686 + }, + "accuracy_pct": 67.32348111658456, + "macro_f1_pct": 66.70962669946734, + "weighted_f1_pct": 68.17502198094193, + "macro_precision_pct": 68.4664639273811, + "macro_recall_pct": 70.02367232005933, + "per_class_f1_pct": { + "FAVOR": 60.62322946175638, + "AGAINST": 67.19160104986877, + "NONE": 72.31404958677686 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5690834473324213, + "macro_f1": 0.5398826032838026, + "weighted_f1": 0.5899766647634495, + "macro_precision": 0.5582801783264747, + "macro_recall": 0.5837164469153087, + "per_class_f1": { + "FAVOR": 0.40924092409240925, + "AGAINST": 0.55, + "NONE": 0.6604068857589984 + }, + "accuracy_pct": 56.90834473324213, + "macro_f1_pct": 53.98826032838026, + "weighted_f1_pct": 58.997666476344946, + "macro_precision_pct": 55.82801783264747, + "macro_recall_pct": 58.37164469153087, + "per_class_f1_pct": { + "FAVOR": 40.92409240924093, + "AGAINST": 55.00000000000001, + "NONE": 66.04068857589984 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6023916292974589, + "macro_f1": 0.5614970603629533, + "weighted_f1": 0.6225208250766868, + "macro_precision": 0.5798539658181671, + "macro_recall": 0.5911978727879239, + "per_class_f1": { + "FAVOR": 0.5302325581395348, + "AGAINST": 0.4229390681003584, + "NONE": 0.7313195548489666 + }, + "accuracy_pct": 60.239162929745895, + "macro_f1_pct": 56.149706036295335, + "weighted_f1_pct": 62.25208250766868, + "macro_precision_pct": 57.98539658181671, + "macro_recall_pct": 59.11978727879239, + "per_class_f1_pct": { + "FAVOR": 53.02325581395348, + "AGAINST": 42.29390681003584, + "NONE": 73.13195548489666 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7142857142857143, + "macro_f1": 0.6886960123959924, + "weighted_f1": 0.7176042995928901, + "macro_precision": 0.6863356382133206, + "macro_recall": 0.6964119561761418, + "per_class_f1": { + "FAVOR": 0.6335877862595419, + "AGAINST": 0.6341463414634146, + "NONE": 0.7983539094650206 + }, + "accuracy_pct": 71.42857142857143, + "macro_f1_pct": 68.86960123959925, + "weighted_f1_pct": 71.76042995928901, + "macro_precision_pct": 68.63356382133206, + "macro_recall_pct": 69.64119561761419, + "per_class_f1_pct": { + "FAVOR": 63.358778625954194, + "AGAINST": 63.41463414634146, + "NONE": 79.83539094650206 + } + } + ], + "accuracy": 0.613726085636198, + "macro_f1": 0.5908349932138371, + "weighted_f1": 0.6319100016090985, + "macro_precision": 0.6022653055325016, + "macro_recall": 0.6270816123640404, + "per_class_f1": { + "FAVOR": 0.51237766263673, + "AGAINST": 0.5456509994597515, + "NONE": 0.71447631754503 + }, + "accuracy_pct": 61.3726085636198, + "macro_f1_pct": 59.08349932138371, + "weighted_f1_pct": 63.19100016090985, + "macro_precision_pct": 60.22653055325016, + "macro_recall_pct": 62.70816123640404, + "per_class_f1_pct": { + "FAVOR": 51.237766263673, + "AGAINST": 54.56509994597515, + "NONE": 71.447631754503 + } + }, + { + "after_task": 6, + "best_epoch": 8, + "best_dev_macro_f1": 0.7190146750524109, + "losses": { + "current_ce": 0.37990011626056264, + "replay_ce": 0.13867790674169858, + "dark_distillation": 0.31864081465062644, + "supervised_contrastive": 4.059075568403516, + "total": 0.9565982834214256, + "replay_multiplier": 0.5683018831519859, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09782326756857436, + "ema_update_magnitude": 0.05448669148246428, + "final_multiplier": 0.5569911211999637 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5399580446255724, + "weighted_f1": 0.627200774318466, + "macro_precision": 0.5546492715516013, + "macro_recall": 0.5744733097129612, + "per_class_f1": { + "FAVOR": 0.4028268551236749, + "AGAINST": 0.47558386411889597, + "NONE": 0.7414634146341463 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 53.99580446255724, + "weighted_f1_pct": 62.7200774318466, + "macro_precision_pct": 55.464927155160126, + "macro_recall_pct": 57.44733097129612, + "per_class_f1_pct": { + "FAVOR": 40.282685512367486, + "AGAINST": 47.5583864118896, + "NONE": 74.14634146341463 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7027914614121511, + "macro_f1": 0.6905418574933396, + "weighted_f1": 0.7094660071410845, + "macro_precision": 0.6928889957008199, + "macro_recall": 0.712324205358498, + "per_class_f1": { + "FAVOR": 0.6198083067092651, + "AGAINST": 0.6873385012919897, + "NONE": 0.7644787644787645 + }, + "accuracy_pct": 70.27914614121511, + "macro_f1_pct": 69.05418574933397, + "weighted_f1_pct": 70.94660071410846, + "macro_precision_pct": 69.28889957008198, + "macro_recall_pct": 71.23242053584981, + "per_class_f1_pct": { + "FAVOR": 61.980830670926515, + "AGAINST": 68.73385012919897, + "NONE": 76.44787644787645 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5937072503419972, + "macro_f1": 0.5519461699316053, + "weighted_f1": 0.6102337171233613, + "macro_precision": 0.5658123696994135, + "macro_recall": 0.5802495437580807, + "per_class_f1": { + "FAVOR": 0.39215686274509803, + "AGAINST": 0.5776173285198556, + "NONE": 0.6860643185298622 + }, + "accuracy_pct": 59.37072503419972, + "macro_f1_pct": 55.19461699316053, + "weighted_f1_pct": 61.02337171233613, + "macro_precision_pct": 56.58123696994135, + "macro_recall_pct": 58.02495437580807, + "per_class_f1_pct": { + "FAVOR": 39.21568627450981, + "AGAINST": 57.761732851985556, + "NONE": 68.60643185298622 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6263079222720478, + "macro_f1": 0.5785336807655703, + "weighted_f1": 0.6468238508983174, + "macro_precision": 0.5818109325530109, + "macro_recall": 0.5981456219075585, + "per_class_f1": { + "FAVOR": 0.47802197802197804, + "AGAINST": 0.48231511254019294, + "NONE": 0.77526395173454 + }, + "accuracy_pct": 62.63079222720478, + "macro_f1_pct": 57.85336807655703, + "weighted_f1_pct": 64.68238508983174, + "macro_precision_pct": 58.181093255301086, + "macro_recall_pct": 59.81456219075585, + "per_class_f1_pct": { + "FAVOR": 47.8021978021978, + "AGAINST": 48.231511254019296, + "NONE": 77.526395173454 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.710261569416499, + "macro_f1": 0.6855251191869449, + "weighted_f1": 0.7145949401055764, + "macro_precision": 0.6814432946990623, + "macro_recall": 0.6938493968707675, + "per_class_f1": { + "FAVOR": 0.6307053941908713, + "AGAINST": 0.63003663003663, + "NONE": 0.7958333333333333 + }, + "accuracy_pct": 71.0261569416499, + "macro_f1_pct": 68.55251191869449, + "weighted_f1_pct": 71.45949401055765, + "macro_precision_pct": 68.14432946990622, + "macro_recall_pct": 69.38493968707675, + "per_class_f1_pct": { + "FAVOR": 63.07053941908713, + "AGAINST": 63.003663003663, + "NONE": 79.58333333333333 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7263581488933601, + "macro_f1": 0.6953117199096991, + "weighted_f1": 0.7387450302477794, + "macro_precision": 0.6913463778243653, + "macro_recall": 0.7182734458596528, + "per_class_f1": { + "FAVOR": 0.6015037593984962, + "AGAINST": 0.6504065040650406, + "NONE": 0.8340248962655602 + }, + "accuracy_pct": 72.63581488933602, + "macro_f1_pct": 69.53117199096991, + "weighted_f1_pct": 73.87450302477794, + "macro_precision_pct": 69.13463778243653, + "macro_recall_pct": 71.82734458596528, + "per_class_f1_pct": { + "FAVOR": 60.150375939849624, + "AGAINST": 65.04065040650406, + "NONE": 83.40248962655602 + } + } + ], + "accuracy": 0.6511873350923483, + "macro_f1": 0.6186479330214674, + "weighted_f1": 0.6658328688524984, + "macro_precision": 0.6205444032655976, + "macro_recall": 0.6431994185322474, + "per_class_f1": { + "FAVOR": 0.519163763066202, + "AGAINST": 0.576271186440678, + "NONE": 0.7605088495575221 + }, + "accuracy_pct": 65.11873350923483, + "macro_f1_pct": 61.86479330214674, + "weighted_f1_pct": 66.58328688524983, + "macro_precision_pct": 62.05444032655976, + "macro_recall_pct": 64.31994185322473, + "per_class_f1_pct": { + "FAVOR": 51.91637630662021, + "AGAINST": 57.6271186440678, + "NONE": 76.05088495575221 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.7455270655270656, + "losses": { + "current_ce": 0.3234014348013372, + "replay_ce": 0.13548697528800344, + "dark_distillation": 0.3017914269755526, + "supervised_contrastive": 3.9907342930150227, + "total": 0.8831816029257891, + "replay_multiplier": 0.5576599656599283, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09962185720878244, + "ema_update_magnitude": 0.051961950418673256, + "final_multiplier": 0.5215918662284525 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5717916137229987, + "macro_f1": 0.5116545173797081, + "weighted_f1": 0.600083762518169, + "macro_precision": 0.545293547779049, + "macro_recall": 0.5517218297719387, + "per_class_f1": { + "FAVOR": 0.35877862595419846, + "AGAINST": 0.46153846153846156, + "NONE": 0.7146464646464646 + }, + "accuracy_pct": 57.179161372299866, + "macro_f1_pct": 51.16545173797081, + "weighted_f1_pct": 60.0083762518169, + "macro_precision_pct": 54.529354777904906, + "macro_recall_pct": 55.17218297719387, + "per_class_f1_pct": { + "FAVOR": 35.87786259541985, + "AGAINST": 46.15384615384615, + "NONE": 71.46464646464646 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.7044334975369458, + "macro_f1": 0.690278917145201, + "weighted_f1": 0.7117972684816621, + "macro_precision": 0.6895242148612936, + "macro_recall": 0.7084438066526247, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.6684782608695652, + "NONE": 0.7773584905660378 + }, + "accuracy_pct": 70.44334975369459, + "macro_f1_pct": 69.0278917145201, + "weighted_f1_pct": 71.17972684816621, + "macro_precision_pct": 68.95242148612935, + "macro_recall_pct": 70.84438066526248, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 66.84782608695652, + "NONE": 77.73584905660378 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.5772913816689467, + "macro_f1": 0.5481803324729765, + "weighted_f1": 0.5933111760236676, + "macro_precision": 0.5649628098741202, + "macro_recall": 0.5833398148429954, + "per_class_f1": { + "FAVOR": 0.4351145038167939, + "AGAINST": 0.5490196078431373, + "NONE": 0.6604068857589984 + }, + "accuracy_pct": 57.729138166894664, + "macro_f1_pct": 54.81803324729765, + "weighted_f1_pct": 59.33111760236676, + "macro_precision_pct": 56.49628098741202, + "macro_recall_pct": 58.333981484299535, + "per_class_f1_pct": { + "FAVOR": 43.51145038167939, + "AGAINST": 54.90196078431373, + "NONE": 66.04068857589984 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.6068759342301944, + "macro_f1": 0.5660650757412758, + "weighted_f1": 0.6304498894179335, + "macro_precision": 0.575518417455123, + "macro_recall": 0.5911585104345486, + "per_class_f1": { + "FAVOR": 0.49025069637883006, + "AGAINST": 0.4583333333333333, + "NONE": 0.749611197511664 + }, + "accuracy_pct": 60.687593423019436, + "macro_f1_pct": 56.60650757412758, + "weighted_f1_pct": 63.04498894179334, + "macro_precision_pct": 57.5518417455123, + "macro_recall_pct": 59.11585104345486, + "per_class_f1_pct": { + "FAVOR": 49.02506963788301, + "AGAINST": 45.83333333333333, + "NONE": 74.96111975116641 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6800826920403474, + "weighted_f1": 0.7108859018019771, + "macro_precision": 0.6754504682657735, + "macro_recall": 0.6894551721375598, + "per_class_f1": { + "FAVOR": 0.6303501945525292, + "AGAINST": 0.6124031007751938, + "NONE": 0.7974947807933194 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 68.00826920403475, + "weighted_f1_pct": 71.08859018019771, + "macro_precision_pct": 67.54504682657735, + "macro_recall_pct": 68.94551721375598, + "per_class_f1_pct": { + "FAVOR": 63.035019455252915, + "AGAINST": 61.240310077519375, + "NONE": 79.74947807933194 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.7002012072434608, + "macro_f1": 0.6681143928738864, + "weighted_f1": 0.7153377876258961, + "macro_precision": 0.6680788295261423, + "macro_recall": 0.6918723321022172, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.608, + "NONE": 0.8185654008438819 + }, + "accuracy_pct": 70.02012072434609, + "macro_f1_pct": 66.81143928738864, + "weighted_f1_pct": 71.5337787625896, + "macro_precision_pct": 66.80788295261422, + "macro_recall_pct": 69.18723321022172, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 60.8, + "NONE": 81.85654008438819 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.7196652719665272, + "macro_f1": 0.672418883322757, + "weighted_f1": 0.7296119580386282, + "macro_precision": 0.6640642135862489, + "macro_recall": 0.7008576575989928, + "per_class_f1": { + "FAVOR": 0.5482625482625483, + "AGAINST": 0.6585365853658537, + "NONE": 0.8104575163398693 + }, + "accuracy_pct": 71.96652719665272, + "macro_f1_pct": 67.2418883322757, + "weighted_f1_pct": 72.96119580386282, + "macro_precision_pct": 66.4064213586249, + "macro_recall_pct": 70.08576575989927, + "per_class_f1_pct": { + "FAVOR": 54.826254826254825, + "AGAINST": 65.85365853658537, + "NONE": 81.04575163398692 + } + } + ], + "accuracy": 0.6483248280452629, + "macro_f1": 0.6157472043872111, + "weighted_f1": 0.6638956685746421, + "macro_precision": 0.6188087106906376, + "macro_recall": 0.6422836668788158, + "per_class_f1": { + "FAVOR": 0.5248868778280543, + "AGAINST": 0.5652978172401036, + "NONE": 0.7570569180934752 + }, + "accuracy_pct": 64.8324828045263, + "macro_f1_pct": 61.57472043872111, + "weighted_f1_pct": 66.38956685746422, + "macro_precision_pct": 61.88087106906376, + "macro_recall_pct": 64.22836668788158, + "per_class_f1_pct": { + "FAVOR": 52.488687782805435, + "AGAINST": 56.52978172401036, + "NONE": 75.70569180934751 + } + }, + { + "after_task": 8, + "best_epoch": 15, + "best_dev_macro_f1": 0.7801174049850865, + "losses": { + "current_ce": 0.30776806522672084, + "replay_ce": 0.12483645702401797, + "dark_distillation": 0.29410668415766134, + "supervised_contrastive": 4.0093502114800845, + "total": 0.8577304455579496, + "replay_multiplier": 0.5454794998550193, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10364761917169144, + "ema_update_magnitude": 0.05460331420835193, + "final_multiplier": 0.5268168689712205 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5501905972045743, + "macro_f1": 0.5036648473182558, + "weighted_f1": 0.5845183803725745, + "macro_precision": 0.5374770198079385, + "macro_recall": 0.5465656932432532, + "per_class_f1": { + "FAVOR": 0.38311688311688313, + "AGAINST": 0.43687374749499, + "NONE": 0.6910039113428944 + }, + "accuracy_pct": 55.01905972045743, + "macro_f1_pct": 50.36648473182558, + "weighted_f1_pct": 58.45183803725745, + "macro_precision_pct": 53.74770198079385, + "macro_recall_pct": 54.65656932432532, + "per_class_f1_pct": { + "FAVOR": 38.311688311688314, + "AGAINST": 43.687374749499, + "NONE": 69.10039113428944 + } + }, + { + "task_index": 1, + "target": "cloning", + "accuracy": 0.6518883415435139, + "macro_f1": 0.6469653114594304, + "weighted_f1": 0.6610862460956662, + "macro_precision": 0.6691896502425573, + "macro_recall": 0.6812547525958426, + "per_class_f1": { + "FAVOR": 0.5907859078590786, + "AGAINST": 0.6486486486486487, + "NONE": 0.7014613778705637 + }, + "accuracy_pct": 65.1888341543514, + "macro_f1_pct": 64.69653114594304, + "weighted_f1_pct": 66.10862460956662, + "macro_precision_pct": 66.91896502425573, + "macro_recall_pct": 68.12547525958425, + "per_class_f1_pct": { + "FAVOR": 59.078590785907856, + "AGAINST": 64.86486486486487, + "NONE": 70.14613778705638 + } + }, + { + "task_index": 2, + "target": "death penalty", + "accuracy": 0.533515731874145, + "macro_f1": 0.5155360266172196, + "weighted_f1": 0.5579125621463014, + "macro_precision": 0.5543455652949042, + "macro_recall": 0.5804493433385298, + "per_class_f1": { + "FAVOR": 0.41456582633053224, + "AGAINST": 0.5070422535211268, + "NONE": 0.625 + }, + "accuracy_pct": 53.3515731874145, + "macro_f1_pct": 51.55360266172197, + "weighted_f1_pct": 55.79125621463014, + "macro_precision_pct": 55.43455652949042, + "macro_recall_pct": 58.04493433385298, + "per_class_f1_pct": { + "FAVOR": 41.45658263305322, + "AGAINST": 50.70422535211267, + "NONE": 62.5 + } + }, + { + "task_index": 3, + "target": "gun control", + "accuracy": 0.5964125560538116, + "macro_f1": 0.5641442947958578, + "weighted_f1": 0.6161736909232732, + "macro_precision": 0.5843596840223722, + "macro_recall": 0.6011007358997588, + "per_class_f1": { + "FAVOR": 0.5467289719626168, + "AGAINST": 0.4377104377104377, + "NONE": 0.7079934747145188 + }, + "accuracy_pct": 59.64125560538116, + "macro_f1_pct": 56.414429479585785, + "weighted_f1_pct": 61.61736909232732, + "macro_precision_pct": 58.435968402237215, + "macro_recall_pct": 60.110073589975876, + "per_class_f1_pct": { + "FAVOR": 54.67289719626168, + "AGAINST": 43.77104377104377, + "NONE": 70.79934747145188 + } + }, + { + "task_index": 4, + "target": "marijuana legalization", + "accuracy": 0.682092555331992, + "macro_f1": 0.6550370871815985, + "weighted_f1": 0.6896607333298811, + "macro_precision": 0.6537439909896304, + "macro_recall": 0.6668256414019126, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.5905511811023622, + "NONE": 0.7863247863247863 + }, + "accuracy_pct": 68.2092555331992, + "macro_f1_pct": 65.50370871815984, + "weighted_f1_pct": 68.9660733329881, + "macro_precision_pct": 65.37439909896304, + "macro_recall_pct": 66.68256414019126, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 59.055118110236215, + "NONE": 78.63247863247864 + } + }, + { + "task_index": 5, + "target": "minimum wage", + "accuracy": 0.670020120724346, + "macro_f1": 0.6408548403304598, + "weighted_f1": 0.6880991232704631, + "macro_precision": 0.6529306066569692, + "macro_recall": 0.6638644391517955, + "per_class_f1": { + "FAVOR": 0.5496688741721855, + "AGAINST": 0.5814977973568282, + "NONE": 0.7913978494623656 + }, + "accuracy_pct": 67.0020120724346, + "macro_f1_pct": 64.08548403304599, + "weighted_f1_pct": 68.8099123270463, + "macro_precision_pct": 65.29306066569693, + "macro_recall_pct": 66.38644391517956, + "per_class_f1_pct": { + "FAVOR": 54.966887417218544, + "AGAINST": 58.14977973568281, + "NONE": 79.13978494623656 + } + }, + { + "task_index": 6, + "target": "nuclear energy", + "accuracy": 0.6638772663877266, + "macro_f1": 0.6368391557424404, + "weighted_f1": 0.679497119275485, + "macro_precision": 0.6377108770902448, + "macro_recall": 0.6947101166153864, + "per_class_f1": { + "FAVOR": 0.5266457680250783, + "AGAINST": 0.6474820143884892, + "NONE": 0.7363896848137536 + }, + "accuracy_pct": 66.38772663877266, + "macro_f1_pct": 63.68391557424404, + "weighted_f1_pct": 67.94971192754849, + "macro_precision_pct": 63.77108770902448, + "macro_recall_pct": 69.47101166153864, + "per_class_f1_pct": { + "FAVOR": 52.66457680250783, + "AGAINST": 64.74820143884892, + "NONE": 73.63896848137536 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7259136212624585, + "macro_f1": 0.6946931848988314, + "weighted_f1": 0.7346643845237473, + "macro_precision": 0.6827680686114421, + "macro_recall": 0.7328767726916422, + "per_class_f1": { + "FAVOR": 0.6545454545454545, + "AGAINST": 0.6225165562913907, + "NONE": 0.8070175438596491 + }, + "accuracy_pct": 72.59136212624585, + "macro_f1_pct": 69.46931848988314, + "weighted_f1_pct": 73.46643845237473, + "macro_precision_pct": 68.27680686114421, + "macro_recall_pct": 73.28767726916422, + "per_class_f1_pct": { + "FAVOR": 65.45454545454545, + "AGAINST": 62.251655629139066, + "NONE": 80.7017543859649 + } + } + ], + "accuracy": 0.6271285965942455, + "macro_f1": 0.6027806676545912, + "weighted_f1": 0.6453364209217395, + "macro_precision": 0.6120526566603584, + "macro_recall": 0.6421302193459932, + "per_class_f1": { + "FAVOR": 0.529277566539924, + "AGAINST": 0.5518686692280824, + "NONE": 0.7271957671957672 + }, + "accuracy_pct": 62.712859659424545, + "macro_f1_pct": 60.278066765459116, + "weighted_f1_pct": 64.53364209217395, + "macro_precision_pct": 61.205265666035835, + "macro_recall_pct": 64.21302193459933, + "per_class_f1_pct": { + "FAVOR": 52.9277566539924, + "AGAINST": 55.18686692280824, + "NONE": 72.71957671957672 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s1__seed13/resolved_config.yaml b/outputs/trace_s8_full__doc__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4b92464e9a0d467443f1e57096937d6f001d5c79 --- /dev/null +++ b/outputs/trace_s8_full__doc__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__doc__s1__seed13/run.log b/outputs/trace_s8_full__doc__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..8bf20dfcee63070ee3e332d043eab9e1bec65c86 --- /dev/null +++ b/outputs/trace_s8_full__doc__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 08:50:50,765 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__doc__s1__seed13 device=cuda +2026-09-02 09:21:30,860 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6072168435430118, 'average_forgetting': 0.05425147172332497, 'backward_transfer': -0.05425147172332497, 'final_mean_target_macro_f1_pct': 60.721684354301175, 'average_forgetting_pct': 5.4251471723324975, 'backward_transfer_pct': -5.4251471723324975, 'final_pooled_accuracy': 0.6271285965942455, 'final_pooled_accuracy_pct': 62.712859659424545, 'final_pooled_macro_f1': 0.6027806676545912, 'final_pooled_macro_f1_pct': 60.278066765459116, 'final_pooled_weighted_f1': 0.6453364209217395, 'final_pooled_weighted_f1_pct': 64.53364209217395, 'final_pooled_macro_precision': 0.6120526566603584, 'final_pooled_macro_precision_pct': 61.205265666035835, 'final_pooled_macro_recall': 0.6421302193459932, 'final_pooled_macro_recall_pct': 64.21302193459933, 'final_pooled_per_class_f1': {'FAVOR': 0.529277566539924, 'AGAINST': 0.5518686692280824, 'NONE': 0.7271957671957672}, 'final_pooled_per_class_f1_pct': {'FAVOR': 52.9277566539924, 'AGAINST': 55.18686692280824, 'NONE': 72.71957671957672}} +2026-09-02 09:21:30,860 [INFO] __main__: efficiency={'training_wall_time_seconds': 1838.6547880358994, 'training_wall_time_minutes': 30.64424646726499, 'peak_gpu_memory_allocated_gb': 10.430231094360352, 'peak_gpu_memory_reserved_gb': 10.6015625} diff --git a/outputs/trace_s8_full__doc__s2__seed13/environment.json b/outputs/trace_s8_full__doc__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..73c56e997931cb80fd44505ccc76167ed2bab2d8 --- /dev/null +++ b/outputs/trace_s8_full__doc__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T09:21:39.011431+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s2__seed13/metrics.json b/outputs/trace_s8_full__doc__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..acdd4985339d91229c067c3c7276aeba40b017c5 --- /dev/null +++ b/outputs/trace_s8_full__doc__s2__seed13/metrics.json @@ -0,0 +1,1649 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.6531689645051564, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6611865866234831, + 0.5583363924488084, + null, + null, + null, + null, + null, + null + ], + [ + 0.6493811709080796, + 0.5709253584576212, + 0.7009701964489053, + null, + null, + null, + null, + null + ], + [ + 0.6239002390810516, + 0.5334669805702301, + 0.6652278977588794, + 0.6864626888785647, + null, + null, + null, + null + ], + [ + 0.624536020241719, + 0.5354867455692319, + 0.6519081170955975, + 0.6938552575303544, + 0.6396981015974215, + null, + null, + null + ], + [ + 0.6440466649833313, + 0.5251482314455121, + 0.6247995961837218, + 0.6785214092815774, + 0.6039032434053807, + 0.6190672475272646, + null, + null + ], + [ + 0.6573455732386064, + 0.5212829618023841, + 0.6768189239318553, + 0.6881520504757727, + 0.6017427254942028, + 0.5953878164667785, + 0.6657501795143738, + null + ], + [ + 0.6673926045981432, + 0.5427200767163765, + 0.6722041785562971, + 0.7065971163370396, + 0.5977050948445176, + 0.5930950622079046, + 0.662554648120005, + 0.6816434266939136 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6959209496522929, + "losses": { + "current_ce": 0.5369516231332506, + "total": 0.5369516231332506, + "replay_multiplier": 0.6309892731588317, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09500305559777947, + "ema_update_magnitude": 0.054923990581633164, + "final_multiplier": 0.5781286742414727 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6617405582922824, + "macro_f1": 0.6531689645051564, + "weighted_f1": 0.6717170205511607, + "macro_precision": 0.6637560337842204, + "macro_recall": 0.6783035634842127, + "per_class_f1": { + "FAVOR": 0.5793871866295265, + "AGAINST": 0.655367231638418, + "NONE": 0.7247524752475247 + }, + "accuracy_pct": 66.17405582922824, + "macro_f1_pct": 65.31689645051564, + "weighted_f1_pct": 67.17170205511607, + "macro_precision_pct": 66.37560337842204, + "macro_recall_pct": 67.83035634842128, + "per_class_f1_pct": { + "FAVOR": 57.938718662952645, + "AGAINST": 65.5367231638418, + "NONE": 72.47524752475248 + } + } + ], + "accuracy": 0.6617405582922824, + "macro_f1": 0.6531689645051564, + "weighted_f1": 0.6717170205511607, + "macro_precision": 0.6637560337842204, + "macro_recall": 0.6783035634842127, + "per_class_f1": { + "FAVOR": 0.5793871866295265, + "AGAINST": 0.655367231638418, + "NONE": 0.7247524752475247 + }, + "accuracy_pct": 66.17405582922824, + "macro_f1_pct": 65.31689645051564, + "weighted_f1_pct": 67.17170205511607, + "macro_precision_pct": 66.37560337842204, + "macro_recall_pct": 67.83035634842128, + "per_class_f1_pct": { + "FAVOR": 57.938718662952645, + "AGAINST": 65.5367231638418, + "NONE": 72.47524752475248 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6808517283222878, + "losses": { + "current_ce": 0.5742792376823592, + "replay_ce": 0.11091023103186959, + "dark_distillation": 0.2696209774205559, + "supervised_contrastive": 4.2479217525114095, + "total": 1.1378239785370075, + "replay_multiplier": 0.5501218868882206, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10218153654180324, + "ema_update_magnitude": 0.05620731856191398, + "final_multiplier": 0.5500731390833914 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6650246305418719, + "macro_f1": 0.6611865866234831, + "weighted_f1": 0.669300222256354, + "macro_precision": 0.6773180410375531, + "macro_recall": 0.6986457952471424, + "per_class_f1": { + "FAVOR": 0.6109510086455331, + "AGAINST": 0.6836734693877551, + "NONE": 0.6889352818371608 + }, + "accuracy_pct": 66.50246305418719, + "macro_f1_pct": 66.11865866234831, + "weighted_f1_pct": 66.9300222256354, + "macro_precision_pct": 67.73180410375531, + "macro_recall_pct": 69.86457952471424, + "per_class_f1_pct": { + "FAVOR": 61.09510086455331, + "AGAINST": 68.36734693877551, + "NONE": 68.89352818371609 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5583363924488084, + "weighted_f1": 0.6233545281498355, + "macro_precision": 0.5712719862719863, + "macro_recall": 0.5830533906261199, + "per_class_f1": { + "FAVOR": 0.5083798882681564, + "AGAINST": 0.42565597667638483, + "NONE": 0.7409733124018838 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 55.83363924488084, + "weighted_f1_pct": 62.33545281498355, + "macro_precision_pct": 57.12719862719863, + "macro_recall_pct": 58.30533906261199, + "per_class_f1_pct": { + "FAVOR": 50.83798882681564, + "AGAINST": 42.565597667638485, + "NONE": 74.09733124018838 + } + } + ], + "accuracy": 0.6298904538341158, + "macro_f1": 0.6135895157269802, + "weighted_f1": 0.644538617181452, + "macro_precision": 0.6256059713224498, + "macro_recall": 0.6455644541718534, + "per_class_f1": { + "FAVOR": 0.5588652482269504, + "AGAINST": 0.563265306122449, + "NONE": 0.7186379928315412 + }, + "accuracy_pct": 62.98904538341158, + "macro_f1_pct": 61.35895157269802, + "weighted_f1_pct": 64.4538617181452, + "macro_precision_pct": 62.56059713224498, + "macro_recall_pct": 64.55644541718534, + "per_class_f1_pct": { + "FAVOR": 55.88652482269504, + "AGAINST": 56.326530612244895, + "NONE": 71.86379928315412 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.7293284493284493, + "losses": { + "current_ce": 0.46401691392418887, + "replay_ce": 0.16284544592102368, + "dark_distillation": 0.31413190109389166, + "supervised_contrastive": 4.16954949753625, + "total": 1.0630513106073651, + "replay_multiplier": 0.5620751171367094, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10033671587530903, + "ema_update_magnitude": 0.05317476582602582, + "final_multiplier": 0.5299631880727236 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6493811709080796, + "weighted_f1": 0.6555182209836735, + "macro_precision": 0.6845213809718094, + "macro_recall": 0.6890035867378183, + "per_class_f1": { + "FAVOR": 0.5784313725490197, + "AGAINST": 0.7058823529411765, + "NONE": 0.6638297872340425 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.93811709080795, + "weighted_f1_pct": 65.55182209836735, + "macro_precision_pct": 68.45213809718094, + "macro_recall_pct": 68.90035867378182, + "per_class_f1_pct": { + "FAVOR": 57.84313725490197, + "AGAINST": 70.58823529411765, + "NONE": 66.38297872340425 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5709253584576212, + "weighted_f1": 0.6205569555690633, + "macro_precision": 0.593700616225371, + "macro_recall": 0.6096317916215762, + "per_class_f1": { + "FAVOR": 0.5288461538461539, + "AGAINST": 0.47318611987381703, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 57.092535845762114, + "weighted_f1_pct": 62.05569555690633, + "macro_precision_pct": 59.370061622537094, + "macro_recall_pct": 60.96317916215762, + "per_class_f1_pct": { + "FAVOR": 52.88461538461539, + "AGAINST": 47.3186119873817, + "NONE": 71.07438016528926 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7344064386317908, + "macro_f1": 0.7009701964489053, + "weighted_f1": 0.7434722404056137, + "macro_precision": 0.6941846296703992, + "macro_recall": 0.7182947315131223, + "per_class_f1": { + "FAVOR": 0.5992217898832685, + "AGAINST": 0.6666666666666666, + "NONE": 0.8370221327967807 + }, + "accuracy_pct": 73.44064386317908, + "macro_f1_pct": 70.09701964489054, + "weighted_f1_pct": 74.34722404056137, + "macro_precision_pct": 69.41846296703991, + "macro_recall_pct": 71.82947315131223, + "per_class_f1_pct": { + "FAVOR": 59.92217898832685, + "AGAINST": 66.66666666666666, + "NONE": 83.70221327967808 + } + } + ], + "accuracy": 0.6529577464788733, + "macro_f1": 0.6380295014264127, + "weighted_f1": 0.6675857388652716, + "macro_precision": 0.6506856419399013, + "macro_recall": 0.6706844133788317, + "per_class_f1": { + "FAVOR": 0.5642923219241444, + "AGAINST": 0.6131549609810479, + "NONE": 0.7366412213740458 + }, + "accuracy_pct": 65.29577464788733, + "macro_f1_pct": 63.802950142641265, + "weighted_f1_pct": 66.75857388652716, + "macro_precision_pct": 65.06856419399013, + "macro_recall_pct": 67.06844133788317, + "per_class_f1_pct": { + "FAVOR": 56.42923219241444, + "AGAINST": 61.315496098104795, + "NONE": 73.66412213740458 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7942840256670044, + "losses": { + "current_ce": 0.3819822433678543, + "replay_ce": 0.16660696886041584, + "dark_distillation": 0.31244127271222133, + "supervised_contrastive": 4.1073419893489165, + "total": 0.9827905578940522, + "replay_multiplier": 0.5840755558292984, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09852493640596918, + "ema_update_magnitude": 0.05240707358582121, + "final_multiplier": 0.5319168476280904 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6239737274220033, + "macro_f1": 0.6239002390810516, + "weighted_f1": 0.6330359958775136, + "macro_precision": 0.6583823812459172, + "macro_recall": 0.6620172540258272, + "per_class_f1": { + "FAVOR": 0.5607940446650124, + "AGAINST": 0.6570605187319885, + "NONE": 0.6538461538461539 + }, + "accuracy_pct": 62.39737274220033, + "macro_f1_pct": 62.390023908105164, + "weighted_f1_pct": 63.30359958775136, + "macro_precision_pct": 65.83823812459173, + "macro_recall_pct": 66.20172540258272, + "per_class_f1_pct": { + "FAVOR": 56.07940446650124, + "AGAINST": 65.70605187319885, + "NONE": 65.38461538461539 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5560538116591929, + "macro_f1": 0.5334669805702301, + "weighted_f1": 0.5805728534160934, + "macro_precision": 0.5670220068674671, + "macro_recall": 0.5751556575451756, + "per_class_f1": { + "FAVOR": 0.5221674876847291, + "AGAINST": 0.414985590778098, + "NONE": 0.6632478632478632 + }, + "accuracy_pct": 55.60538116591929, + "macro_f1_pct": 53.34669805702301, + "weighted_f1_pct": 58.057285341609344, + "macro_precision_pct": 56.70220068674671, + "macro_recall_pct": 57.51556575451756, + "per_class_f1_pct": { + "FAVOR": 52.21674876847291, + "AGAINST": 41.4985590778098, + "NONE": 66.32478632478632 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6881287726358148, + "macro_f1": 0.6652278977588794, + "weighted_f1": 0.7031151304807804, + "macro_precision": 0.6711509453548227, + "macro_recall": 0.6940400170285228, + "per_class_f1": { + "FAVOR": 0.5850340136054422, + "AGAINST": 0.6244725738396625, + "NONE": 0.7861771058315334 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 66.52278977588793, + "weighted_f1_pct": 70.31151304807804, + "macro_precision_pct": 67.11509453548227, + "macro_recall_pct": 69.40400170285228, + "per_class_f1_pct": { + "FAVOR": 58.50340136054422, + "AGAINST": 62.447257383966246, + "NONE": 78.61771058315334 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7308970099667774, + "macro_f1": 0.6864626888785647, + "weighted_f1": 0.7357691128022409, + "macro_precision": 0.6773249475890984, + "macro_recall": 0.6993263020241965, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6229508196721312, + "NONE": 0.8210526315789474 + }, + "accuracy_pct": 73.08970099667775, + "macro_f1_pct": 68.64626888785646, + "weighted_f1_pct": 73.5769112802241, + "macro_precision_pct": 67.73249475890985, + "macro_recall_pct": 69.93263020241966, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 62.295081967213115, + "NONE": 82.10526315789474 + } + } + ], + "accuracy": 0.6453512831299958, + "macro_f1": 0.624608418137187, + "weighted_f1": 0.6597685398111296, + "macro_precision": 0.6306852994978426, + "macro_recall": 0.6580265238778477, + "per_class_f1": { + "FAVOR": 0.5639491398653702, + "AGAINST": 0.5744336569579288, + "NONE": 0.7354424575882622 + }, + "accuracy_pct": 64.53512831299957, + "macro_f1_pct": 62.4608418137187, + "weighted_f1_pct": 65.97685398111295, + "macro_precision_pct": 63.068529949784256, + "macro_recall_pct": 65.80265238778476, + "per_class_f1_pct": { + "FAVOR": 56.39491398653702, + "AGAINST": 57.44336569579288, + "NONE": 73.54424575882622 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6809780922236702, + "losses": { + "current_ce": 0.5113522084774794, + "replay_ce": 0.13679645480933011, + "dark_distillation": 0.3153090239012683, + "supervised_contrastive": 4.152721416685316, + "total": 1.0968406403506243, + "replay_multiplier": 0.5754165599666786, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09590627963932682, + "ema_update_magnitude": 0.053931072266029265, + "final_multiplier": 0.5623309805035391 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6239737274220033, + "macro_f1": 0.624536020241719, + "weighted_f1": 0.6365885483149184, + "macro_precision": 0.6698946886446887, + "macro_recall": 0.6608625181037918, + "per_class_f1": { + "FAVOR": 0.5545023696682464, + "AGAINST": 0.6524390243902439, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.39737274220033, + "macro_f1_pct": 62.45360202417189, + "weighted_f1_pct": 63.65885483149184, + "macro_precision_pct": 66.98946886446888, + "macro_recall_pct": 66.08625181037918, + "per_class_f1_pct": { + "FAVOR": 55.45023696682464, + "AGAINST": 65.2439024390244, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5575485799701047, + "macro_f1": 0.5354867455692319, + "weighted_f1": 0.5816408313751352, + "macro_precision": 0.5691597454872214, + "macro_recall": 0.5780966715898983, + "per_class_f1": { + "FAVOR": 0.5409429280397022, + "AGAINST": 0.4045584045584046, + "NONE": 0.660958904109589 + }, + "accuracy_pct": 55.75485799701047, + "macro_f1_pct": 53.548674556923196, + "weighted_f1_pct": 58.16408313751352, + "macro_precision_pct": 56.91597454872214, + "macro_recall_pct": 57.80966715898983, + "per_class_f1_pct": { + "FAVOR": 54.09429280397022, + "AGAINST": 40.45584045584046, + "NONE": 66.0958904109589 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6519081170955975, + "weighted_f1": 0.6880735589461767, + "macro_precision": 0.6624086021505377, + "macro_recall": 0.6823950387168779, + "per_class_f1": { + "FAVOR": 0.5695364238410596, + "AGAINST": 0.6186440677966102, + "NONE": 0.7675438596491229 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 65.19081170955975, + "weighted_f1_pct": 68.80735589461767, + "macro_precision_pct": 66.24086021505377, + "macro_recall_pct": 68.23950387168779, + "per_class_f1_pct": { + "FAVOR": 56.95364238410596, + "AGAINST": 61.86440677966102, + "NONE": 76.75438596491229 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.729235880398671, + "macro_f1": 0.6938552575303544, + "weighted_f1": 0.7370647709391056, + "macro_precision": 0.6808383288832248, + "macro_recall": 0.7198880703280613, + "per_class_f1": { + "FAVOR": 0.6454183266932271, + "AGAINST": 0.6217948717948718, + "NONE": 0.8143525741029641 + }, + "accuracy_pct": 72.9235880398671, + "macro_f1_pct": 69.38552575303544, + "weighted_f1_pct": 73.70647709391056, + "macro_precision_pct": 68.08383288832248, + "macro_recall_pct": 71.98880703280614, + "per_class_f1_pct": { + "FAVOR": 64.5418326693227, + "AGAINST": 62.17948717948718, + "NONE": 81.43525741029642 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6975857687420585, + "macro_f1": 0.6396981015974215, + "weighted_f1": 0.7097043020558996, + "macro_precision": 0.6352202201545095, + "macro_recall": 0.6810239284640156, + "per_class_f1": { + "FAVOR": 0.5844504021447721, + "AGAINST": 0.5283018867924528, + "NONE": 0.8063420158550396 + }, + "accuracy_pct": 69.75857687420584, + "macro_f1_pct": 63.969810159742146, + "weighted_f1_pct": 70.97043020558996, + "macro_precision_pct": 63.52202201545095, + "macro_recall_pct": 68.10239284640156, + "per_class_f1_pct": { + "FAVOR": 58.4450402144772, + "AGAINST": 52.83018867924528, + "NONE": 80.63420158550396 + } + } + ], + "accuracy": 0.6558154235145386, + "macro_f1": 0.6284161532020652, + "weighted_f1": 0.6709935471100493, + "macro_precision": 0.6324799190020716, + "macro_recall": 0.6659923266831896, + "per_class_f1": { + "FAVOR": 0.5733866362078812, + "AGAINST": 0.5592233009708738, + "NONE": 0.7526385224274407 + }, + "accuracy_pct": 65.58154235145386, + "macro_f1_pct": 62.84161532020652, + "weighted_f1_pct": 67.09935471100493, + "macro_precision_pct": 63.247991900207154, + "macro_recall_pct": 66.59923266831896, + "per_class_f1_pct": { + "FAVOR": 57.338663620788125, + "AGAINST": 55.92233009708738, + "NONE": 75.26385224274406 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.7147138374986476, + "losses": { + "current_ce": 0.46405755235325724, + "replay_ce": 0.1386613651356172, + "dark_distillation": 0.307944294881253, + "supervised_contrastive": 4.076478501633992, + "total": 1.044828422154699, + "replay_multiplier": 0.5879203419005052, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09738311319483217, + "ema_update_magnitude": 0.05319258874510215, + "final_multiplier": 0.5462198424349092 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.6440466649833313, + "weighted_f1": 0.6529778387946281, + "macro_precision": 0.6825933430475161, + "macro_recall": 0.6846720419133158, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6819484240687679, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 64.40466649833313, + "weighted_f1_pct": 65.29778387946281, + "macro_precision_pct": 68.2593343047516, + "macro_recall_pct": 68.46720419133158, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 68.19484240687679, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5440956651718983, + "macro_f1": 0.5251482314455121, + "weighted_f1": 0.5726906349143771, + "macro_precision": 0.5659715801222339, + "macro_recall": 0.5654924935404616, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.417910447761194, + "NONE": 0.6575342465753424 + }, + "accuracy_pct": 54.40956651718983, + "macro_f1_pct": 52.514823144551215, + "weighted_f1_pct": 57.26906349143771, + "macro_precision_pct": 56.59715801222339, + "macro_recall_pct": 56.54924935404616, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 41.7910447761194, + "NONE": 65.75342465753424 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6519114688128773, + "macro_f1": 0.6247995961837218, + "weighted_f1": 0.6691668649543532, + "macro_precision": 0.6308779186704551, + "macro_recall": 0.6526676101388745, + "per_class_f1": { + "FAVOR": 0.5166051660516605, + "AGAINST": 0.5909090909090909, + "NONE": 0.7668845315904139 + }, + "accuracy_pct": 65.19114688128774, + "macro_f1_pct": 62.479959618372185, + "weighted_f1_pct": 66.91668649543531, + "macro_precision_pct": 63.0877918670455, + "macro_recall_pct": 65.26676101388745, + "per_class_f1_pct": { + "FAVOR": 51.66051660516605, + "AGAINST": 59.09090909090909, + "NONE": 76.68845315904139 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7159468438538206, + "macro_f1": 0.6785214092815774, + "weighted_f1": 0.7245442681851216, + "macro_precision": 0.6680131696311146, + "macro_recall": 0.7007131994412329, + "per_class_f1": { + "FAVOR": 0.6134453781512605, + "AGAINST": 0.617737003058104, + "NONE": 0.8043818466353677 + }, + "accuracy_pct": 71.59468438538205, + "macro_f1_pct": 67.85214092815775, + "weighted_f1_pct": 72.45442681851216, + "macro_precision_pct": 66.80131696311146, + "macro_recall_pct": 70.07131994412329, + "per_class_f1_pct": { + "FAVOR": 61.34453781512605, + "AGAINST": 61.773700305810394, + "NONE": 80.43818466353677 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6607369758576874, + "macro_f1": 0.6039032434053807, + "weighted_f1": 0.6772899237148085, + "macro_precision": 0.6033333333333334, + "macro_recall": 0.6487359066335101, + "per_class_f1": { + "FAVOR": 0.5412371134020618, + "AGAINST": 0.49230769230769234, + "NONE": 0.778164924506388 + }, + "accuracy_pct": 66.07369758576874, + "macro_f1_pct": 60.39032434053807, + "weighted_f1_pct": 67.72899237148084, + "macro_precision_pct": 60.333333333333336, + "macro_recall_pct": 64.87359066335101, + "per_class_f1_pct": { + "FAVOR": 54.123711340206185, + "AGAINST": 49.23076923076923, + "NONE": 77.81649245063879 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6607387140902873, + "macro_f1": 0.6190672475272646, + "weighted_f1": 0.6719824376788095, + "macro_precision": 0.619628517984702, + "macro_recall": 0.6453971700037987, + "per_class_f1": { + "FAVOR": 0.475, + "AGAINST": 0.6405959031657356, + "NONE": 0.7416058394160584 + }, + "accuracy_pct": 66.07387140902873, + "macro_f1_pct": 61.90672475272646, + "weighted_f1_pct": 67.19824376788095, + "macro_precision_pct": 61.9628517984702, + "macro_recall_pct": 64.53971700037987, + "per_class_f1_pct": { + "FAVOR": 47.5, + "AGAINST": 64.05959031657356, + "NONE": 74.16058394160584 + } + } + ], + "accuracy": 0.6454428754813863, + "macro_f1": 0.6186685165840272, + "weighted_f1": 0.6613049794245358, + "macro_precision": 0.6208758166315432, + "macro_recall": 0.6536697336002464, + "per_class_f1": { + "FAVOR": 0.5385427666314678, + "AGAINST": 0.5753176043557169, + "NONE": 0.742145178764897 + }, + "accuracy_pct": 64.54428754813864, + "macro_f1_pct": 61.86685165840272, + "weighted_f1_pct": 66.13049794245359, + "macro_precision_pct": 62.087581663154324, + "macro_recall_pct": 65.36697336002464, + "per_class_f1_pct": { + "FAVOR": 53.854276663146784, + "AGAINST": 57.53176043557169, + "NONE": 74.21451787648971 + } + }, + { + "after_task": 7, + "best_epoch": 4, + "best_dev_macro_f1": 0.6313638358569235, + "losses": { + "current_ce": 0.4475596746695893, + "replay_ce": 0.1338053308959518, + "dark_distillation": 0.3153361161904676, + "supervised_contrastive": 4.11189615215574, + "total": 1.0235067956504367, + "replay_multiplier": 0.5621885679273254, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10042728456554599, + "ema_update_magnitude": 0.05473592847182507, + "final_multiplier": 0.545030453712014 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6600985221674877, + "macro_f1": 0.6573455732386064, + "weighted_f1": 0.6682489055469358, + "macro_precision": 0.6829236325627535, + "macro_recall": 0.6954570230749043, + "per_class_f1": { + "FAVOR": 0.6041666666666666, + "AGAINST": 0.6703910614525139, + "NONE": 0.6974789915966386 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.73455732386064, + "weighted_f1_pct": 66.82489055469358, + "macro_precision_pct": 68.29236325627535, + "macro_recall_pct": 69.54570230749043, + "per_class_f1_pct": { + "FAVOR": 60.416666666666664, + "AGAINST": 67.0391061452514, + "NONE": 69.74789915966386 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5426008968609866, + "macro_f1": 0.5212829618023841, + "weighted_f1": 0.5700261111727465, + "macro_precision": 0.5584132924290884, + "macro_recall": 0.560248488074382, + "per_class_f1": { + "FAVOR": 0.5129533678756477, + "AGAINST": 0.3956639566395664, + "NONE": 0.6552315608919382 + }, + "accuracy_pct": 54.26008968609865, + "macro_f1_pct": 52.12829618023841, + "weighted_f1_pct": 57.00261111727465, + "macro_precision_pct": 55.84132924290884, + "macro_recall_pct": 56.0248488074382, + "per_class_f1_pct": { + "FAVOR": 51.29533678756477, + "AGAINST": 39.56639566395664, + "NONE": 65.52315608919382 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7002012072434608, + "macro_f1": 0.6768189239318553, + "weighted_f1": 0.7144968070912432, + "macro_precision": 0.6803691551317333, + "macro_recall": 0.7033452993223107, + "per_class_f1": { + "FAVOR": 0.5862068965517241, + "AGAINST": 0.6468085106382979, + "NONE": 0.7974413646055437 + }, + "accuracy_pct": 70.02012072434609, + "macro_f1_pct": 67.68189239318552, + "weighted_f1_pct": 71.44968070912432, + "macro_precision_pct": 68.03691551317334, + "macro_recall_pct": 70.33452993223108, + "per_class_f1_pct": { + "FAVOR": 58.620689655172406, + "AGAINST": 64.68085106382979, + "NONE": 79.74413646055437 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7275747508305648, + "macro_f1": 0.6881520504757727, + "weighted_f1": 0.7350099547642708, + "macro_precision": 0.6783994186310577, + "macro_recall": 0.7066625040428036, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6276923076923077, + "NONE": 0.8160741885625966 + }, + "accuracy_pct": 72.75747508305648, + "macro_f1_pct": 68.81520504757727, + "weighted_f1_pct": 73.50099547642708, + "macro_precision_pct": 67.83994186310576, + "macro_recall_pct": 70.66625040428036, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 62.76923076923077, + "NONE": 81.60741885625966 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6594663278271918, + "macro_f1": 0.6017427254942028, + "weighted_f1": 0.6744028565550348, + "macro_precision": 0.6088662415515119, + "macro_recall": 0.652830593516868, + "per_class_f1": { + "FAVOR": 0.551219512195122, + "AGAINST": 0.4786885245901639, + "NONE": 0.7753201396973225 + }, + "accuracy_pct": 65.94663278271918, + "macro_f1_pct": 60.17427254942028, + "weighted_f1_pct": 67.44028565550349, + "macro_precision_pct": 60.88662415515119, + "macro_recall_pct": 65.2830593516868, + "per_class_f1_pct": { + "FAVOR": 55.1219512195122, + "AGAINST": 47.868852459016395, + "NONE": 77.53201396973225 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6306429548563611, + "macro_f1": 0.5953878164667785, + "weighted_f1": 0.6431986071060065, + "macro_precision": 0.6003618113912231, + "macro_recall": 0.6257029619935546, + "per_class_f1": { + "FAVOR": 0.46963562753036436, + "AGAINST": 0.6069469835466179, + "NONE": 0.7095808383233533 + }, + "accuracy_pct": 63.064295485636116, + "macro_f1_pct": 59.538781646677855, + "weighted_f1_pct": 64.31986071060065, + "macro_precision_pct": 60.03618113912231, + "macro_recall_pct": 62.570296199355454, + "per_class_f1_pct": { + "FAVOR": 46.963562753036435, + "AGAINST": 60.69469835466179, + "NONE": 70.95808383233533 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6657501795143738, + "weighted_f1": 0.7018253949290945, + "macro_precision": 0.6619201559934319, + "macro_recall": 0.674254741682892, + "per_class_f1": { + "FAVOR": 0.5691056910569106, + "AGAINST": 0.6273062730627307, + "NONE": 0.80083857442348 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 66.57501795143737, + "weighted_f1_pct": 70.18253949290944, + "macro_precision_pct": 66.1920155993432, + "macro_recall_pct": 67.4254741682892, + "per_class_f1_pct": { + "FAVOR": 56.91056910569105, + "AGAINST": 62.73062730627307, + "NONE": 80.083857442348 + } + } + ], + "accuracy": 0.6550546448087432, + "macro_f1": 0.6287343235231383, + "weighted_f1": 0.6694091859305455, + "macro_precision": 0.6287969262086331, + "macro_recall": 0.6612931301248806, + "per_class_f1": { + "FAVOR": 0.5585421412300683, + "AGAINST": 0.5767634854771784, + "NONE": 0.750897343862168 + }, + "accuracy_pct": 65.50546448087432, + "macro_f1_pct": 62.87343235231383, + "weighted_f1_pct": 66.94091859305456, + "macro_precision_pct": 62.87969262086332, + "macro_recall_pct": 66.12931301248805, + "per_class_f1_pct": { + "FAVOR": 55.85421412300683, + "AGAINST": 57.676348547717836, + "NONE": 75.0897343862168 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7442821851459724, + "losses": { + "current_ce": 0.3323441031502514, + "replay_ce": 0.15160475447773933, + "dark_distillation": 0.29525542545124767, + "supervised_contrastive": 4.010183724349107, + "total": 0.9115058193361856, + "replay_multiplier": 0.591498833193117, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09775717520264009, + "ema_update_magnitude": 0.05400273177333391, + "final_multiplier": 0.552417064644023 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6732348111658456, + "macro_f1": 0.6673926045981432, + "weighted_f1": 0.6838248674318488, + "macro_precision": 0.6875326007001238, + "macro_recall": 0.697421242430428, + "per_class_f1": { + "FAVOR": 0.6025974025974026, + "AGAINST": 0.6686567164179105, + "NONE": 0.7309236947791165 + }, + "accuracy_pct": 67.32348111658456, + "macro_f1_pct": 66.73926045981432, + "weighted_f1_pct": 68.38248674318488, + "macro_precision_pct": 68.75326007001237, + "macro_recall_pct": 69.74212424304281, + "per_class_f1_pct": { + "FAVOR": 60.259740259740255, + "AGAINST": 66.86567164179105, + "NONE": 73.09236947791165 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5665171898355755, + "macro_f1": 0.5427200767163765, + "weighted_f1": 0.5946877056244732, + "macro_precision": 0.5731881320698293, + "macro_recall": 0.5783998204606688, + "per_class_f1": { + "FAVOR": 0.532967032967033, + "AGAINST": 0.4095744680851064, + "NONE": 0.68561872909699 + }, + "accuracy_pct": 56.65171898355755, + "macro_f1_pct": 54.272007671637645, + "weighted_f1_pct": 59.46877056244732, + "macro_precision_pct": 57.31881320698293, + "macro_recall_pct": 57.839982046066886, + "per_class_f1_pct": { + "FAVOR": 53.2967032967033, + "AGAINST": 40.95744680851064, + "NONE": 68.56187290969899 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6722041785562971, + "weighted_f1": 0.7103459000845467, + "macro_precision": 0.6798470532390516, + "macro_recall": 0.6969509739624682, + "per_class_f1": { + "FAVOR": 0.5886287625418061, + "AGAINST": 0.6339285714285714, + "NONE": 0.7940552016985138 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 67.2204178556297, + "weighted_f1_pct": 71.03459000845467, + "macro_precision_pct": 67.98470532390516, + "macro_recall_pct": 69.69509739624682, + "per_class_f1_pct": { + "FAVOR": 58.862876254180605, + "AGAINST": 63.39285714285714, + "NONE": 79.40552016985139 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7065971163370396, + "weighted_f1": 0.7513374421211609, + "macro_precision": 0.6942564173685456, + "macro_recall": 0.7286260793938407, + "per_class_f1": { + "FAVOR": 0.6307053941908713, + "AGAINST": 0.6624605678233438, + "NONE": 0.826625386996904 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.65971163370396, + "weighted_f1_pct": 75.1337442121161, + "macro_precision_pct": 69.42564173685456, + "macro_recall_pct": 72.86260793938408, + "per_class_f1_pct": { + "FAVOR": 63.07053941908713, + "AGAINST": 66.24605678233438, + "NONE": 82.6625386996904 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6709021601016518, + "macro_f1": 0.5977050948445176, + "weighted_f1": 0.6778934502695632, + "macro_precision": 0.6141120481795921, + "macro_recall": 0.6412263521195983, + "per_class_f1": { + "FAVOR": 0.5572139303482587, + "AGAINST": 0.44525547445255476, + "NONE": 0.7906458797327395 + }, + "accuracy_pct": 67.09021601016518, + "macro_f1_pct": 59.770509484451765, + "weighted_f1_pct": 67.78934502695631, + "macro_precision_pct": 61.41120481795921, + "macro_recall_pct": 64.12263521195983, + "per_class_f1_pct": { + "FAVOR": 55.72139303482587, + "AGAINST": 44.52554744525548, + "NONE": 79.06458797327394 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6347469220246238, + "macro_f1": 0.5930950622079046, + "weighted_f1": 0.6484044091822824, + "macro_precision": 0.5924318345224268, + "macro_recall": 0.6198834117314197, + "per_class_f1": { + "FAVOR": 0.4505928853754941, + "AGAINST": 0.6011560693641619, + "NONE": 0.7275362318840579 + }, + "accuracy_pct": 63.47469220246238, + "macro_f1_pct": 59.30950622079047, + "weighted_f1_pct": 64.84044091822824, + "macro_precision_pct": 59.243183452242675, + "macro_recall_pct": 61.98834117314197, + "per_class_f1_pct": { + "FAVOR": 45.059288537549406, + "AGAINST": 60.115606936416185, + "NONE": 72.7536231884058 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6881287726358148, + "macro_f1": 0.662554648120005, + "weighted_f1": 0.6943057532302708, + "macro_precision": 0.6587951849145878, + "macro_recall": 0.6728961014148928, + "per_class_f1": { + "FAVOR": 0.5900383141762452, + "AGAINST": 0.6153846153846154, + "NONE": 0.7822410147991543 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 66.2554648120005, + "weighted_f1_pct": 69.43057532302707, + "macro_precision_pct": 65.87951849145878, + "macro_recall_pct": 67.28961014148928, + "per_class_f1_pct": { + "FAVOR": 59.00383141762452, + "AGAINST": 61.53846153846154, + "NONE": 78.22410147991543 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.7168758716875872, + "macro_f1": 0.6816434266939136, + "weighted_f1": 0.7284552526580028, + "macro_precision": 0.6686391382218254, + "macro_recall": 0.721497558981483, + "per_class_f1": { + "FAVOR": 0.5830508474576271, + "AGAINST": 0.6666666666666666, + "NONE": 0.7952127659574468 + }, + "accuracy_pct": 71.68758716875871, + "macro_f1_pct": 68.16434266939136, + "weighted_f1_pct": 72.84552526580028, + "macro_precision_pct": 66.86391382218254, + "macro_recall_pct": 72.1497558981483, + "per_class_f1_pct": { + "FAVOR": 58.30508474576271, + "AGAINST": 66.66666666666666, + "NONE": 79.52127659574468 + } + } + ], + "accuracy": 0.6715599921706792, + "macro_f1": 0.6410904865378585, + "weighted_f1": 0.6845734092733181, + "macro_precision": 0.6372897542646264, + "macro_recall": 0.6715986139076676, + "per_class_f1": { + "FAVOR": 0.5672, + "AGAINST": 0.587667161961367, + "NONE": 0.7684042976522085 + }, + "accuracy_pct": 67.15599921706792, + "macro_f1_pct": 64.10904865378585, + "weighted_f1_pct": 68.45734092733181, + "macro_precision_pct": 63.72897542646264, + "macro_recall_pct": 67.15986139076676, + "per_class_f1_pct": { + "FAVOR": 56.720000000000006, + "AGAINST": 58.766716196136706, + "NONE": 76.84042976522085 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6404890260092746, + "average_forgetting": 0.0155977351884486, + "backward_transfer": -0.011597855648601576, + "final_mean_target_macro_f1_pct": 64.04890260092746, + "average_forgetting_pct": 1.55977351884486, + "backward_transfer_pct": -1.1597855648601576, + "final_pooled_accuracy": 0.6715599921706792, + "final_pooled_accuracy_pct": 67.15599921706792, + "final_pooled_macro_f1": 0.6410904865378585, + "final_pooled_macro_f1_pct": 64.10904865378585, + "final_pooled_weighted_f1": 0.6845734092733181, + "final_pooled_weighted_f1_pct": 68.45734092733181, + "final_pooled_macro_precision": 0.6372897542646264, + "final_pooled_macro_precision_pct": 63.72897542646264, + "final_pooled_macro_recall": 0.6715986139076676, + "final_pooled_macro_recall_pct": 67.15986139076676, + "final_pooled_per_class_f1": { + "FAVOR": 0.5672, + "AGAINST": 0.587667161961367, + "NONE": 0.7684042976522085 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 56.720000000000006, + "AGAINST": 58.766716196136706, + "NONE": 76.84042976522085 + } + }, + "efficiency": { + "training_wall_time_seconds": 1857.3751796099823, + "training_wall_time_minutes": 30.956252993499707, + "peak_gpu_memory_allocated_gb": 10.430231094360352, + "peak_gpu_memory_reserved_gb": 10.6015625 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s2__seed13/metrics.partial.json b/outputs/trace_s8_full__doc__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..eaabb016d9fd5e0f4cd87c0de15630d38d47021f --- /dev/null +++ b/outputs/trace_s8_full__doc__s2__seed13/metrics.partial.json @@ -0,0 +1,1543 @@ +{ + "matrix": [ + [ + 0.6531689645051564, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6611865866234831, + 0.5583363924488084, + null, + null, + null, + null, + null, + null + ], + [ + 0.6493811709080796, + 0.5709253584576212, + 0.7009701964489053, + null, + null, + null, + null, + null + ], + [ + 0.6239002390810516, + 0.5334669805702301, + 0.6652278977588794, + 0.6864626888785647, + null, + null, + null, + null + ], + [ + 0.624536020241719, + 0.5354867455692319, + 0.6519081170955975, + 0.6938552575303544, + 0.6396981015974215, + null, + null, + null + ], + [ + 0.6440466649833313, + 0.5251482314455121, + 0.6247995961837218, + 0.6785214092815774, + 0.6039032434053807, + 0.6190672475272646, + null, + null + ], + [ + 0.6573455732386064, + 0.5212829618023841, + 0.6768189239318553, + 0.6881520504757727, + 0.6017427254942028, + 0.5953878164667785, + 0.6657501795143738, + null + ], + [ + 0.6673926045981432, + 0.5427200767163765, + 0.6722041785562971, + 0.7065971163370396, + 0.5977050948445176, + 0.5930950622079046, + 0.662554648120005, + 0.6816434266939136 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 9, + "best_dev_macro_f1": 0.6959209496522929, + "losses": { + "current_ce": 0.5369516231332506, + "total": 0.5369516231332506, + "replay_multiplier": 0.6309892731588317, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09500305559777947, + "ema_update_magnitude": 0.054923990581633164, + "final_multiplier": 0.5781286742414727 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6617405582922824, + "macro_f1": 0.6531689645051564, + "weighted_f1": 0.6717170205511607, + "macro_precision": 0.6637560337842204, + "macro_recall": 0.6783035634842127, + "per_class_f1": { + "FAVOR": 0.5793871866295265, + "AGAINST": 0.655367231638418, + "NONE": 0.7247524752475247 + }, + "accuracy_pct": 66.17405582922824, + "macro_f1_pct": 65.31689645051564, + "weighted_f1_pct": 67.17170205511607, + "macro_precision_pct": 66.37560337842204, + "macro_recall_pct": 67.83035634842128, + "per_class_f1_pct": { + "FAVOR": 57.938718662952645, + "AGAINST": 65.5367231638418, + "NONE": 72.47524752475248 + } + } + ], + "accuracy": 0.6617405582922824, + "macro_f1": 0.6531689645051564, + "weighted_f1": 0.6717170205511607, + "macro_precision": 0.6637560337842204, + "macro_recall": 0.6783035634842127, + "per_class_f1": { + "FAVOR": 0.5793871866295265, + "AGAINST": 0.655367231638418, + "NONE": 0.7247524752475247 + }, + "accuracy_pct": 66.17405582922824, + "macro_f1_pct": 65.31689645051564, + "weighted_f1_pct": 67.17170205511607, + "macro_precision_pct": 66.37560337842204, + "macro_recall_pct": 67.83035634842128, + "per_class_f1_pct": { + "FAVOR": 57.938718662952645, + "AGAINST": 65.5367231638418, + "NONE": 72.47524752475248 + } + }, + { + "after_task": 2, + "best_epoch": 10, + "best_dev_macro_f1": 0.6808517283222878, + "losses": { + "current_ce": 0.5742792376823592, + "replay_ce": 0.11091023103186959, + "dark_distillation": 0.2696209774205559, + "supervised_contrastive": 4.2479217525114095, + "total": 1.1378239785370075, + "replay_multiplier": 0.5501218868882206, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10218153654180324, + "ema_update_magnitude": 0.05620731856191398, + "final_multiplier": 0.5500731390833914 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6650246305418719, + "macro_f1": 0.6611865866234831, + "weighted_f1": 0.669300222256354, + "macro_precision": 0.6773180410375531, + "macro_recall": 0.6986457952471424, + "per_class_f1": { + "FAVOR": 0.6109510086455331, + "AGAINST": 0.6836734693877551, + "NONE": 0.6889352818371608 + }, + "accuracy_pct": 66.50246305418719, + "macro_f1_pct": 66.11865866234831, + "weighted_f1_pct": 66.9300222256354, + "macro_precision_pct": 67.73180410375531, + "macro_recall_pct": 69.86457952471424, + "per_class_f1_pct": { + "FAVOR": 61.09510086455331, + "AGAINST": 68.36734693877551, + "NONE": 68.89352818371609 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5583363924488084, + "weighted_f1": 0.6233545281498355, + "macro_precision": 0.5712719862719863, + "macro_recall": 0.5830533906261199, + "per_class_f1": { + "FAVOR": 0.5083798882681564, + "AGAINST": 0.42565597667638483, + "NONE": 0.7409733124018838 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 55.83363924488084, + "weighted_f1_pct": 62.33545281498355, + "macro_precision_pct": 57.12719862719863, + "macro_recall_pct": 58.30533906261199, + "per_class_f1_pct": { + "FAVOR": 50.83798882681564, + "AGAINST": 42.565597667638485, + "NONE": 74.09733124018838 + } + } + ], + "accuracy": 0.6298904538341158, + "macro_f1": 0.6135895157269802, + "weighted_f1": 0.644538617181452, + "macro_precision": 0.6256059713224498, + "macro_recall": 0.6455644541718534, + "per_class_f1": { + "FAVOR": 0.5588652482269504, + "AGAINST": 0.563265306122449, + "NONE": 0.7186379928315412 + }, + "accuracy_pct": 62.98904538341158, + "macro_f1_pct": 61.35895157269802, + "weighted_f1_pct": 64.4538617181452, + "macro_precision_pct": 62.56059713224498, + "macro_recall_pct": 64.55644541718534, + "per_class_f1_pct": { + "FAVOR": 55.88652482269504, + "AGAINST": 56.326530612244895, + "NONE": 71.86379928315412 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.7293284493284493, + "losses": { + "current_ce": 0.46401691392418887, + "replay_ce": 0.16284544592102368, + "dark_distillation": 0.31413190109389166, + "supervised_contrastive": 4.16954949753625, + "total": 1.0630513106073651, + "replay_multiplier": 0.5620751171367094, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10033671587530903, + "ema_update_magnitude": 0.05317476582602582, + "final_multiplier": 0.5299631880727236 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6469622331691297, + "macro_f1": 0.6493811709080796, + "weighted_f1": 0.6555182209836735, + "macro_precision": 0.6845213809718094, + "macro_recall": 0.6890035867378183, + "per_class_f1": { + "FAVOR": 0.5784313725490197, + "AGAINST": 0.7058823529411765, + "NONE": 0.6638297872340425 + }, + "accuracy_pct": 64.69622331691298, + "macro_f1_pct": 64.93811709080795, + "weighted_f1_pct": 65.55182209836735, + "macro_precision_pct": 68.45213809718094, + "macro_recall_pct": 68.90035867378182, + "per_class_f1_pct": { + "FAVOR": 57.84313725490197, + "AGAINST": 70.58823529411765, + "NONE": 66.38297872340425 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5979073243647235, + "macro_f1": 0.5709253584576212, + "weighted_f1": 0.6205569555690633, + "macro_precision": 0.593700616225371, + "macro_recall": 0.6096317916215762, + "per_class_f1": { + "FAVOR": 0.5288461538461539, + "AGAINST": 0.47318611987381703, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 59.79073243647235, + "macro_f1_pct": 57.092535845762114, + "weighted_f1_pct": 62.05569555690633, + "macro_precision_pct": 59.370061622537094, + "macro_recall_pct": 60.96317916215762, + "per_class_f1_pct": { + "FAVOR": 52.88461538461539, + "AGAINST": 47.3186119873817, + "NONE": 71.07438016528926 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7344064386317908, + "macro_f1": 0.7009701964489053, + "weighted_f1": 0.7434722404056137, + "macro_precision": 0.6941846296703992, + "macro_recall": 0.7182947315131223, + "per_class_f1": { + "FAVOR": 0.5992217898832685, + "AGAINST": 0.6666666666666666, + "NONE": 0.8370221327967807 + }, + "accuracy_pct": 73.44064386317908, + "macro_f1_pct": 70.09701964489054, + "weighted_f1_pct": 74.34722404056137, + "macro_precision_pct": 69.41846296703991, + "macro_recall_pct": 71.82947315131223, + "per_class_f1_pct": { + "FAVOR": 59.92217898832685, + "AGAINST": 66.66666666666666, + "NONE": 83.70221327967808 + } + } + ], + "accuracy": 0.6529577464788733, + "macro_f1": 0.6380295014264127, + "weighted_f1": 0.6675857388652716, + "macro_precision": 0.6506856419399013, + "macro_recall": 0.6706844133788317, + "per_class_f1": { + "FAVOR": 0.5642923219241444, + "AGAINST": 0.6131549609810479, + "NONE": 0.7366412213740458 + }, + "accuracy_pct": 65.29577464788733, + "macro_f1_pct": 63.802950142641265, + "weighted_f1_pct": 66.75857388652716, + "macro_precision_pct": 65.06856419399013, + "macro_recall_pct": 67.06844133788317, + "per_class_f1_pct": { + "FAVOR": 56.42923219241444, + "AGAINST": 61.315496098104795, + "NONE": 73.66412213740458 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7942840256670044, + "losses": { + "current_ce": 0.3819822433678543, + "replay_ce": 0.16660696886041584, + "dark_distillation": 0.31244127271222133, + "supervised_contrastive": 4.1073419893489165, + "total": 0.9827905578940522, + "replay_multiplier": 0.5840755558292984, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09852493640596918, + "ema_update_magnitude": 0.05240707358582121, + "final_multiplier": 0.5319168476280904 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6239737274220033, + "macro_f1": 0.6239002390810516, + "weighted_f1": 0.6330359958775136, + "macro_precision": 0.6583823812459172, + "macro_recall": 0.6620172540258272, + "per_class_f1": { + "FAVOR": 0.5607940446650124, + "AGAINST": 0.6570605187319885, + "NONE": 0.6538461538461539 + }, + "accuracy_pct": 62.39737274220033, + "macro_f1_pct": 62.390023908105164, + "weighted_f1_pct": 63.30359958775136, + "macro_precision_pct": 65.83823812459173, + "macro_recall_pct": 66.20172540258272, + "per_class_f1_pct": { + "FAVOR": 56.07940446650124, + "AGAINST": 65.70605187319885, + "NONE": 65.38461538461539 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5560538116591929, + "macro_f1": 0.5334669805702301, + "weighted_f1": 0.5805728534160934, + "macro_precision": 0.5670220068674671, + "macro_recall": 0.5751556575451756, + "per_class_f1": { + "FAVOR": 0.5221674876847291, + "AGAINST": 0.414985590778098, + "NONE": 0.6632478632478632 + }, + "accuracy_pct": 55.60538116591929, + "macro_f1_pct": 53.34669805702301, + "weighted_f1_pct": 58.057285341609344, + "macro_precision_pct": 56.70220068674671, + "macro_recall_pct": 57.51556575451756, + "per_class_f1_pct": { + "FAVOR": 52.21674876847291, + "AGAINST": 41.4985590778098, + "NONE": 66.32478632478632 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6881287726358148, + "macro_f1": 0.6652278977588794, + "weighted_f1": 0.7031151304807804, + "macro_precision": 0.6711509453548227, + "macro_recall": 0.6940400170285228, + "per_class_f1": { + "FAVOR": 0.5850340136054422, + "AGAINST": 0.6244725738396625, + "NONE": 0.7861771058315334 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 66.52278977588793, + "weighted_f1_pct": 70.31151304807804, + "macro_precision_pct": 67.11509453548227, + "macro_recall_pct": 69.40400170285228, + "per_class_f1_pct": { + "FAVOR": 58.50340136054422, + "AGAINST": 62.447257383966246, + "NONE": 78.61771058315334 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7308970099667774, + "macro_f1": 0.6864626888785647, + "weighted_f1": 0.7357691128022409, + "macro_precision": 0.6773249475890984, + "macro_recall": 0.6993263020241965, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.6229508196721312, + "NONE": 0.8210526315789474 + }, + "accuracy_pct": 73.08970099667775, + "macro_f1_pct": 68.64626888785646, + "weighted_f1_pct": 73.5769112802241, + "macro_precision_pct": 67.73249475890985, + "macro_recall_pct": 69.93263020241966, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 62.295081967213115, + "NONE": 82.10526315789474 + } + } + ], + "accuracy": 0.6453512831299958, + "macro_f1": 0.624608418137187, + "weighted_f1": 0.6597685398111296, + "macro_precision": 0.6306852994978426, + "macro_recall": 0.6580265238778477, + "per_class_f1": { + "FAVOR": 0.5639491398653702, + "AGAINST": 0.5744336569579288, + "NONE": 0.7354424575882622 + }, + "accuracy_pct": 64.53512831299957, + "macro_f1_pct": 62.4608418137187, + "weighted_f1_pct": 65.97685398111295, + "macro_precision_pct": 63.068529949784256, + "macro_recall_pct": 65.80265238778476, + "per_class_f1_pct": { + "FAVOR": 56.39491398653702, + "AGAINST": 57.44336569579288, + "NONE": 73.54424575882622 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6809780922236702, + "losses": { + "current_ce": 0.5113522084774794, + "replay_ce": 0.13679645480933011, + "dark_distillation": 0.3153090239012683, + "supervised_contrastive": 4.152721416685316, + "total": 1.0968406403506243, + "replay_multiplier": 0.5754165599666786, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09590627963932682, + "ema_update_magnitude": 0.053931072266029265, + "final_multiplier": 0.5623309805035391 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6239737274220033, + "macro_f1": 0.624536020241719, + "weighted_f1": 0.6365885483149184, + "macro_precision": 0.6698946886446887, + "macro_recall": 0.6608625181037918, + "per_class_f1": { + "FAVOR": 0.5545023696682464, + "AGAINST": 0.6524390243902439, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 62.39737274220033, + "macro_f1_pct": 62.45360202417189, + "weighted_f1_pct": 63.65885483149184, + "macro_precision_pct": 66.98946886446888, + "macro_recall_pct": 66.08625181037918, + "per_class_f1_pct": { + "FAVOR": 55.45023696682464, + "AGAINST": 65.2439024390244, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5575485799701047, + "macro_f1": 0.5354867455692319, + "weighted_f1": 0.5816408313751352, + "macro_precision": 0.5691597454872214, + "macro_recall": 0.5780966715898983, + "per_class_f1": { + "FAVOR": 0.5409429280397022, + "AGAINST": 0.4045584045584046, + "NONE": 0.660958904109589 + }, + "accuracy_pct": 55.75485799701047, + "macro_f1_pct": 53.548674556923196, + "weighted_f1_pct": 58.16408313751352, + "macro_precision_pct": 56.91597454872214, + "macro_recall_pct": 57.80966715898983, + "per_class_f1_pct": { + "FAVOR": 54.09429280397022, + "AGAINST": 40.45584045584046, + "NONE": 66.0958904109589 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6720321931589537, + "macro_f1": 0.6519081170955975, + "weighted_f1": 0.6880735589461767, + "macro_precision": 0.6624086021505377, + "macro_recall": 0.6823950387168779, + "per_class_f1": { + "FAVOR": 0.5695364238410596, + "AGAINST": 0.6186440677966102, + "NONE": 0.7675438596491229 + }, + "accuracy_pct": 67.20321931589537, + "macro_f1_pct": 65.19081170955975, + "weighted_f1_pct": 68.80735589461767, + "macro_precision_pct": 66.24086021505377, + "macro_recall_pct": 68.23950387168779, + "per_class_f1_pct": { + "FAVOR": 56.95364238410596, + "AGAINST": 61.86440677966102, + "NONE": 76.75438596491229 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.729235880398671, + "macro_f1": 0.6938552575303544, + "weighted_f1": 0.7370647709391056, + "macro_precision": 0.6808383288832248, + "macro_recall": 0.7198880703280613, + "per_class_f1": { + "FAVOR": 0.6454183266932271, + "AGAINST": 0.6217948717948718, + "NONE": 0.8143525741029641 + }, + "accuracy_pct": 72.9235880398671, + "macro_f1_pct": 69.38552575303544, + "weighted_f1_pct": 73.70647709391056, + "macro_precision_pct": 68.08383288832248, + "macro_recall_pct": 71.98880703280614, + "per_class_f1_pct": { + "FAVOR": 64.5418326693227, + "AGAINST": 62.17948717948718, + "NONE": 81.43525741029642 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6975857687420585, + "macro_f1": 0.6396981015974215, + "weighted_f1": 0.7097043020558996, + "macro_precision": 0.6352202201545095, + "macro_recall": 0.6810239284640156, + "per_class_f1": { + "FAVOR": 0.5844504021447721, + "AGAINST": 0.5283018867924528, + "NONE": 0.8063420158550396 + }, + "accuracy_pct": 69.75857687420584, + "macro_f1_pct": 63.969810159742146, + "weighted_f1_pct": 70.97043020558996, + "macro_precision_pct": 63.52202201545095, + "macro_recall_pct": 68.10239284640156, + "per_class_f1_pct": { + "FAVOR": 58.4450402144772, + "AGAINST": 52.83018867924528, + "NONE": 80.63420158550396 + } + } + ], + "accuracy": 0.6558154235145386, + "macro_f1": 0.6284161532020652, + "weighted_f1": 0.6709935471100493, + "macro_precision": 0.6324799190020716, + "macro_recall": 0.6659923266831896, + "per_class_f1": { + "FAVOR": 0.5733866362078812, + "AGAINST": 0.5592233009708738, + "NONE": 0.7526385224274407 + }, + "accuracy_pct": 65.58154235145386, + "macro_f1_pct": 62.84161532020652, + "weighted_f1_pct": 67.09935471100493, + "macro_precision_pct": 63.247991900207154, + "macro_recall_pct": 66.59923266831896, + "per_class_f1_pct": { + "FAVOR": 57.338663620788125, + "AGAINST": 55.92233009708738, + "NONE": 75.26385224274406 + } + }, + { + "after_task": 6, + "best_epoch": 10, + "best_dev_macro_f1": 0.7147138374986476, + "losses": { + "current_ce": 0.46405755235325724, + "replay_ce": 0.1386613651356172, + "dark_distillation": 0.307944294881253, + "supervised_contrastive": 4.076478501633992, + "total": 1.044828422154699, + "replay_multiplier": 0.5879203419005052, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09738311319483217, + "ema_update_magnitude": 0.05319258874510215, + "final_multiplier": 0.5462198424349092 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6436781609195402, + "macro_f1": 0.6440466649833313, + "weighted_f1": 0.6529778387946281, + "macro_precision": 0.6825933430475161, + "macro_recall": 0.6846720419133158, + "per_class_f1": { + "FAVOR": 0.5777777777777777, + "AGAINST": 0.6819484240687679, + "NONE": 0.6724137931034483 + }, + "accuracy_pct": 64.36781609195403, + "macro_f1_pct": 64.40466649833313, + "weighted_f1_pct": 65.29778387946281, + "macro_precision_pct": 68.2593343047516, + "macro_recall_pct": 68.46720419133158, + "per_class_f1_pct": { + "FAVOR": 57.77777777777777, + "AGAINST": 68.19484240687679, + "NONE": 67.24137931034483 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5440956651718983, + "macro_f1": 0.5251482314455121, + "weighted_f1": 0.5726906349143771, + "macro_precision": 0.5659715801222339, + "macro_recall": 0.5654924935404616, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.417910447761194, + "NONE": 0.6575342465753424 + }, + "accuracy_pct": 54.40956651718983, + "macro_f1_pct": 52.514823144551215, + "weighted_f1_pct": 57.26906349143771, + "macro_precision_pct": 56.59715801222339, + "macro_recall_pct": 56.54924935404616, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 41.7910447761194, + "NONE": 65.75342465753424 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6519114688128773, + "macro_f1": 0.6247995961837218, + "weighted_f1": 0.6691668649543532, + "macro_precision": 0.6308779186704551, + "macro_recall": 0.6526676101388745, + "per_class_f1": { + "FAVOR": 0.5166051660516605, + "AGAINST": 0.5909090909090909, + "NONE": 0.7668845315904139 + }, + "accuracy_pct": 65.19114688128774, + "macro_f1_pct": 62.479959618372185, + "weighted_f1_pct": 66.91668649543531, + "macro_precision_pct": 63.0877918670455, + "macro_recall_pct": 65.26676101388745, + "per_class_f1_pct": { + "FAVOR": 51.66051660516605, + "AGAINST": 59.09090909090909, + "NONE": 76.68845315904139 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7159468438538206, + "macro_f1": 0.6785214092815774, + "weighted_f1": 0.7245442681851216, + "macro_precision": 0.6680131696311146, + "macro_recall": 0.7007131994412329, + "per_class_f1": { + "FAVOR": 0.6134453781512605, + "AGAINST": 0.617737003058104, + "NONE": 0.8043818466353677 + }, + "accuracy_pct": 71.59468438538205, + "macro_f1_pct": 67.85214092815775, + "weighted_f1_pct": 72.45442681851216, + "macro_precision_pct": 66.80131696311146, + "macro_recall_pct": 70.07131994412329, + "per_class_f1_pct": { + "FAVOR": 61.34453781512605, + "AGAINST": 61.773700305810394, + "NONE": 80.43818466353677 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6607369758576874, + "macro_f1": 0.6039032434053807, + "weighted_f1": 0.6772899237148085, + "macro_precision": 0.6033333333333334, + "macro_recall": 0.6487359066335101, + "per_class_f1": { + "FAVOR": 0.5412371134020618, + "AGAINST": 0.49230769230769234, + "NONE": 0.778164924506388 + }, + "accuracy_pct": 66.07369758576874, + "macro_f1_pct": 60.39032434053807, + "weighted_f1_pct": 67.72899237148084, + "macro_precision_pct": 60.333333333333336, + "macro_recall_pct": 64.87359066335101, + "per_class_f1_pct": { + "FAVOR": 54.123711340206185, + "AGAINST": 49.23076923076923, + "NONE": 77.81649245063879 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6607387140902873, + "macro_f1": 0.6190672475272646, + "weighted_f1": 0.6719824376788095, + "macro_precision": 0.619628517984702, + "macro_recall": 0.6453971700037987, + "per_class_f1": { + "FAVOR": 0.475, + "AGAINST": 0.6405959031657356, + "NONE": 0.7416058394160584 + }, + "accuracy_pct": 66.07387140902873, + "macro_f1_pct": 61.90672475272646, + "weighted_f1_pct": 67.19824376788095, + "macro_precision_pct": 61.9628517984702, + "macro_recall_pct": 64.53971700037987, + "per_class_f1_pct": { + "FAVOR": 47.5, + "AGAINST": 64.05959031657356, + "NONE": 74.16058394160584 + } + } + ], + "accuracy": 0.6454428754813863, + "macro_f1": 0.6186685165840272, + "weighted_f1": 0.6613049794245358, + "macro_precision": 0.6208758166315432, + "macro_recall": 0.6536697336002464, + "per_class_f1": { + "FAVOR": 0.5385427666314678, + "AGAINST": 0.5753176043557169, + "NONE": 0.742145178764897 + }, + "accuracy_pct": 64.54428754813864, + "macro_f1_pct": 61.86685165840272, + "weighted_f1_pct": 66.13049794245359, + "macro_precision_pct": 62.087581663154324, + "macro_recall_pct": 65.36697336002464, + "per_class_f1_pct": { + "FAVOR": 53.854276663146784, + "AGAINST": 57.53176043557169, + "NONE": 74.21451787648971 + } + }, + { + "after_task": 7, + "best_epoch": 4, + "best_dev_macro_f1": 0.6313638358569235, + "losses": { + "current_ce": 0.4475596746695893, + "replay_ce": 0.1338053308959518, + "dark_distillation": 0.3153361161904676, + "supervised_contrastive": 4.11189615215574, + "total": 1.0235067956504367, + "replay_multiplier": 0.5621885679273254, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10042728456554599, + "ema_update_magnitude": 0.05473592847182507, + "final_multiplier": 0.545030453712014 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6600985221674877, + "macro_f1": 0.6573455732386064, + "weighted_f1": 0.6682489055469358, + "macro_precision": 0.6829236325627535, + "macro_recall": 0.6954570230749043, + "per_class_f1": { + "FAVOR": 0.6041666666666666, + "AGAINST": 0.6703910614525139, + "NONE": 0.6974789915966386 + }, + "accuracy_pct": 66.00985221674877, + "macro_f1_pct": 65.73455732386064, + "weighted_f1_pct": 66.82489055469358, + "macro_precision_pct": 68.29236325627535, + "macro_recall_pct": 69.54570230749043, + "per_class_f1_pct": { + "FAVOR": 60.416666666666664, + "AGAINST": 67.0391061452514, + "NONE": 69.74789915966386 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5426008968609866, + "macro_f1": 0.5212829618023841, + "weighted_f1": 0.5700261111727465, + "macro_precision": 0.5584132924290884, + "macro_recall": 0.560248488074382, + "per_class_f1": { + "FAVOR": 0.5129533678756477, + "AGAINST": 0.3956639566395664, + "NONE": 0.6552315608919382 + }, + "accuracy_pct": 54.26008968609865, + "macro_f1_pct": 52.12829618023841, + "weighted_f1_pct": 57.00261111727465, + "macro_precision_pct": 55.84132924290884, + "macro_recall_pct": 56.0248488074382, + "per_class_f1_pct": { + "FAVOR": 51.29533678756477, + "AGAINST": 39.56639566395664, + "NONE": 65.52315608919382 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.7002012072434608, + "macro_f1": 0.6768189239318553, + "weighted_f1": 0.7144968070912432, + "macro_precision": 0.6803691551317333, + "macro_recall": 0.7033452993223107, + "per_class_f1": { + "FAVOR": 0.5862068965517241, + "AGAINST": 0.6468085106382979, + "NONE": 0.7974413646055437 + }, + "accuracy_pct": 70.02012072434609, + "macro_f1_pct": 67.68189239318552, + "weighted_f1_pct": 71.44968070912432, + "macro_precision_pct": 68.03691551317334, + "macro_recall_pct": 70.33452993223108, + "per_class_f1_pct": { + "FAVOR": 58.620689655172406, + "AGAINST": 64.68085106382979, + "NONE": 79.74413646055437 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7275747508305648, + "macro_f1": 0.6881520504757727, + "weighted_f1": 0.7350099547642708, + "macro_precision": 0.6783994186310577, + "macro_recall": 0.7066625040428036, + "per_class_f1": { + "FAVOR": 0.6206896551724138, + "AGAINST": 0.6276923076923077, + "NONE": 0.8160741885625966 + }, + "accuracy_pct": 72.75747508305648, + "macro_f1_pct": 68.81520504757727, + "weighted_f1_pct": 73.50099547642708, + "macro_precision_pct": 67.83994186310576, + "macro_recall_pct": 70.66625040428036, + "per_class_f1_pct": { + "FAVOR": 62.06896551724138, + "AGAINST": 62.76923076923077, + "NONE": 81.60741885625966 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6594663278271918, + "macro_f1": 0.6017427254942028, + "weighted_f1": 0.6744028565550348, + "macro_precision": 0.6088662415515119, + "macro_recall": 0.652830593516868, + "per_class_f1": { + "FAVOR": 0.551219512195122, + "AGAINST": 0.4786885245901639, + "NONE": 0.7753201396973225 + }, + "accuracy_pct": 65.94663278271918, + "macro_f1_pct": 60.17427254942028, + "weighted_f1_pct": 67.44028565550349, + "macro_precision_pct": 60.88662415515119, + "macro_recall_pct": 65.2830593516868, + "per_class_f1_pct": { + "FAVOR": 55.1219512195122, + "AGAINST": 47.868852459016395, + "NONE": 77.53201396973225 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6306429548563611, + "macro_f1": 0.5953878164667785, + "weighted_f1": 0.6431986071060065, + "macro_precision": 0.6003618113912231, + "macro_recall": 0.6257029619935546, + "per_class_f1": { + "FAVOR": 0.46963562753036436, + "AGAINST": 0.6069469835466179, + "NONE": 0.7095808383233533 + }, + "accuracy_pct": 63.064295485636116, + "macro_f1_pct": 59.538781646677855, + "weighted_f1_pct": 64.31986071060065, + "macro_precision_pct": 60.03618113912231, + "macro_recall_pct": 62.570296199355454, + "per_class_f1_pct": { + "FAVOR": 46.963562753036435, + "AGAINST": 60.69469835466179, + "NONE": 70.95808383233533 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6657501795143738, + "weighted_f1": 0.7018253949290945, + "macro_precision": 0.6619201559934319, + "macro_recall": 0.674254741682892, + "per_class_f1": { + "FAVOR": 0.5691056910569106, + "AGAINST": 0.6273062730627307, + "NONE": 0.80083857442348 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 66.57501795143737, + "weighted_f1_pct": 70.18253949290944, + "macro_precision_pct": 66.1920155993432, + "macro_recall_pct": 67.4254741682892, + "per_class_f1_pct": { + "FAVOR": 56.91056910569105, + "AGAINST": 62.73062730627307, + "NONE": 80.083857442348 + } + } + ], + "accuracy": 0.6550546448087432, + "macro_f1": 0.6287343235231383, + "weighted_f1": 0.6694091859305455, + "macro_precision": 0.6287969262086331, + "macro_recall": 0.6612931301248806, + "per_class_f1": { + "FAVOR": 0.5585421412300683, + "AGAINST": 0.5767634854771784, + "NONE": 0.750897343862168 + }, + "accuracy_pct": 65.50546448087432, + "macro_f1_pct": 62.87343235231383, + "weighted_f1_pct": 66.94091859305456, + "macro_precision_pct": 62.87969262086332, + "macro_recall_pct": 66.12931301248805, + "per_class_f1_pct": { + "FAVOR": 55.85421412300683, + "AGAINST": 57.676348547717836, + "NONE": 75.0897343862168 + } + }, + { + "after_task": 8, + "best_epoch": 3, + "best_dev_macro_f1": 0.7442821851459724, + "losses": { + "current_ce": 0.3323441031502514, + "replay_ce": 0.15160475447773933, + "dark_distillation": 0.29525542545124767, + "supervised_contrastive": 4.010183724349107, + "total": 0.9115058193361856, + "replay_multiplier": 0.591498833193117, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09775717520264009, + "ema_update_magnitude": 0.05400273177333391, + "final_multiplier": 0.552417064644023 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "cloning", + "accuracy": 0.6732348111658456, + "macro_f1": 0.6673926045981432, + "weighted_f1": 0.6838248674318488, + "macro_precision": 0.6875326007001238, + "macro_recall": 0.697421242430428, + "per_class_f1": { + "FAVOR": 0.6025974025974026, + "AGAINST": 0.6686567164179105, + "NONE": 0.7309236947791165 + }, + "accuracy_pct": 67.32348111658456, + "macro_f1_pct": 66.73926045981432, + "weighted_f1_pct": 68.38248674318488, + "macro_precision_pct": 68.75326007001237, + "macro_recall_pct": 69.74212424304281, + "per_class_f1_pct": { + "FAVOR": 60.259740259740255, + "AGAINST": 66.86567164179105, + "NONE": 73.09236947791165 + } + }, + { + "task_index": 1, + "target": "gun control", + "accuracy": 0.5665171898355755, + "macro_f1": 0.5427200767163765, + "weighted_f1": 0.5946877056244732, + "macro_precision": 0.5731881320698293, + "macro_recall": 0.5783998204606688, + "per_class_f1": { + "FAVOR": 0.532967032967033, + "AGAINST": 0.4095744680851064, + "NONE": 0.68561872909699 + }, + "accuracy_pct": 56.65171898355755, + "macro_f1_pct": 54.272007671637645, + "weighted_f1_pct": 59.46877056244732, + "macro_precision_pct": 57.31881320698293, + "macro_recall_pct": 57.839982046066886, + "per_class_f1_pct": { + "FAVOR": 53.2967032967033, + "AGAINST": 40.95744680851064, + "NONE": 68.56187290969899 + } + }, + { + "task_index": 2, + "target": "minimum wage", + "accuracy": 0.6961770623742455, + "macro_f1": 0.6722041785562971, + "weighted_f1": 0.7103459000845467, + "macro_precision": 0.6798470532390516, + "macro_recall": 0.6969509739624682, + "per_class_f1": { + "FAVOR": 0.5886287625418061, + "AGAINST": 0.6339285714285714, + "NONE": 0.7940552016985138 + }, + "accuracy_pct": 69.61770623742454, + "macro_f1_pct": 67.2204178556297, + "weighted_f1_pct": 71.03459000845467, + "macro_precision_pct": 67.98470532390516, + "macro_recall_pct": 69.69509739624682, + "per_class_f1_pct": { + "FAVOR": 58.862876254180605, + "AGAINST": 63.39285714285714, + "NONE": 79.40552016985139 + } + }, + { + "task_index": 3, + "target": "school uniforms", + "accuracy": 0.7441860465116279, + "macro_f1": 0.7065971163370396, + "weighted_f1": 0.7513374421211609, + "macro_precision": 0.6942564173685456, + "macro_recall": 0.7286260793938407, + "per_class_f1": { + "FAVOR": 0.6307053941908713, + "AGAINST": 0.6624605678233438, + "NONE": 0.826625386996904 + }, + "accuracy_pct": 74.4186046511628, + "macro_f1_pct": 70.65971163370396, + "weighted_f1_pct": 75.1337442121161, + "macro_precision_pct": 69.42564173685456, + "macro_recall_pct": 72.86260793938408, + "per_class_f1_pct": { + "FAVOR": 63.07053941908713, + "AGAINST": 66.24605678233438, + "NONE": 82.6625386996904 + } + }, + { + "task_index": 4, + "target": "abortion", + "accuracy": 0.6709021601016518, + "macro_f1": 0.5977050948445176, + "weighted_f1": 0.6778934502695632, + "macro_precision": 0.6141120481795921, + "macro_recall": 0.6412263521195983, + "per_class_f1": { + "FAVOR": 0.5572139303482587, + "AGAINST": 0.44525547445255476, + "NONE": 0.7906458797327395 + }, + "accuracy_pct": 67.09021601016518, + "macro_f1_pct": 59.770509484451765, + "weighted_f1_pct": 67.78934502695631, + "macro_precision_pct": 61.41120481795921, + "macro_recall_pct": 64.12263521195983, + "per_class_f1_pct": { + "FAVOR": 55.72139303482587, + "AGAINST": 44.52554744525548, + "NONE": 79.06458797327394 + } + }, + { + "task_index": 5, + "target": "death penalty", + "accuracy": 0.6347469220246238, + "macro_f1": 0.5930950622079046, + "weighted_f1": 0.6484044091822824, + "macro_precision": 0.5924318345224268, + "macro_recall": 0.6198834117314197, + "per_class_f1": { + "FAVOR": 0.4505928853754941, + "AGAINST": 0.6011560693641619, + "NONE": 0.7275362318840579 + }, + "accuracy_pct": 63.47469220246238, + "macro_f1_pct": 59.30950622079047, + "weighted_f1_pct": 64.84044091822824, + "macro_precision_pct": 59.243183452242675, + "macro_recall_pct": 61.98834117314197, + "per_class_f1_pct": { + "FAVOR": 45.059288537549406, + "AGAINST": 60.115606936416185, + "NONE": 72.7536231884058 + } + }, + { + "task_index": 6, + "target": "marijuana legalization", + "accuracy": 0.6881287726358148, + "macro_f1": 0.662554648120005, + "weighted_f1": 0.6943057532302708, + "macro_precision": 0.6587951849145878, + "macro_recall": 0.6728961014148928, + "per_class_f1": { + "FAVOR": 0.5900383141762452, + "AGAINST": 0.6153846153846154, + "NONE": 0.7822410147991543 + }, + "accuracy_pct": 68.81287726358148, + "macro_f1_pct": 66.2554648120005, + "weighted_f1_pct": 69.43057532302707, + "macro_precision_pct": 65.87951849145878, + "macro_recall_pct": 67.28961014148928, + "per_class_f1_pct": { + "FAVOR": 59.00383141762452, + "AGAINST": 61.53846153846154, + "NONE": 78.22410147991543 + } + }, + { + "task_index": 7, + "target": "nuclear energy", + "accuracy": 0.7168758716875872, + "macro_f1": 0.6816434266939136, + "weighted_f1": 0.7284552526580028, + "macro_precision": 0.6686391382218254, + "macro_recall": 0.721497558981483, + "per_class_f1": { + "FAVOR": 0.5830508474576271, + "AGAINST": 0.6666666666666666, + "NONE": 0.7952127659574468 + }, + "accuracy_pct": 71.68758716875871, + "macro_f1_pct": 68.16434266939136, + "weighted_f1_pct": 72.84552526580028, + "macro_precision_pct": 66.86391382218254, + "macro_recall_pct": 72.1497558981483, + "per_class_f1_pct": { + "FAVOR": 58.30508474576271, + "AGAINST": 66.66666666666666, + "NONE": 79.52127659574468 + } + } + ], + "accuracy": 0.6715599921706792, + "macro_f1": 0.6410904865378585, + "weighted_f1": 0.6845734092733181, + "macro_precision": 0.6372897542646264, + "macro_recall": 0.6715986139076676, + "per_class_f1": { + "FAVOR": 0.5672, + "AGAINST": 0.587667161961367, + "NONE": 0.7684042976522085 + }, + "accuracy_pct": 67.15599921706792, + "macro_f1_pct": 64.10904865378585, + "weighted_f1_pct": 68.45734092733181, + "macro_precision_pct": 63.72897542646264, + "macro_recall_pct": 67.15986139076676, + "per_class_f1_pct": { + "FAVOR": 56.720000000000006, + "AGAINST": 58.766716196136706, + "NONE": 76.84042976522085 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s2__seed13/resolved_config.yaml b/outputs/trace_s8_full__doc__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1bcc64acaaf0622c4971b3b19723d3564fe53fd5 --- /dev/null +++ b/outputs/trace_s8_full__doc__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__doc__s2__seed13/run.log b/outputs/trace_s8_full__doc__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..9958d91669e5814407565e8efb991954739574a3 --- /dev/null +++ b/outputs/trace_s8_full__doc__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 09:21:39,061 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__doc__s2__seed13 device=cuda +2026-09-02 09:52:37,805 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6404890260092746, 'average_forgetting': 0.0155977351884486, 'backward_transfer': -0.011597855648601576, 'final_mean_target_macro_f1_pct': 64.04890260092746, 'average_forgetting_pct': 1.55977351884486, 'backward_transfer_pct': -1.1597855648601576, 'final_pooled_accuracy': 0.6715599921706792, 'final_pooled_accuracy_pct': 67.15599921706792, 'final_pooled_macro_f1': 0.6410904865378585, 'final_pooled_macro_f1_pct': 64.10904865378585, 'final_pooled_weighted_f1': 0.6845734092733181, 'final_pooled_weighted_f1_pct': 68.45734092733181, 'final_pooled_macro_precision': 0.6372897542646264, 'final_pooled_macro_precision_pct': 63.72897542646264, 'final_pooled_macro_recall': 0.6715986139076676, 'final_pooled_macro_recall_pct': 67.15986139076676, 'final_pooled_per_class_f1': {'FAVOR': 0.5672, 'AGAINST': 0.587667161961367, 'NONE': 0.7684042976522085}, 'final_pooled_per_class_f1_pct': {'FAVOR': 56.720000000000006, 'AGAINST': 58.766716196136706, 'NONE': 76.84042976522085}} +2026-09-02 09:52:37,805 [INFO] __main__: efficiency={'training_wall_time_seconds': 1857.3751796099823, 'training_wall_time_minutes': 30.956252993499707, 'peak_gpu_memory_allocated_gb': 10.430231094360352, 'peak_gpu_memory_reserved_gb': 10.6015625} diff --git a/outputs/trace_s8_full__doc__s3__seed13/environment.json b/outputs/trace_s8_full__doc__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..e05ebc013f907dd957a1162fd91daa52ee2b925f --- /dev/null +++ b/outputs/trace_s8_full__doc__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T09:52:45.819180+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s3__seed13/metrics.json b/outputs/trace_s8_full__doc__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..0b4bed8082b2ecd628bf9ea8de301d15e8df6a24 --- /dev/null +++ b/outputs/trace_s8_full__doc__s3__seed13/metrics.json @@ -0,0 +1,1649 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "doc", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5292270687553361, + 0.5631093252583498, + null, + null, + null, + null, + null, + null + ], + [ + 0.5204063837975658, + 0.5478952936580055, + 0.6489022632267165, + null, + null, + null, + null, + null + ], + [ + 0.49977267795057373, + 0.5061909509825049, + 0.6342860580148716, + 0.6320522674541541, + null, + null, + null, + null + ], + [ + 0.5275531881258319, + 0.5272783864541832, + 0.6732493321930463, + 0.6258358901738652, + 0.7142560068932152, + null, + null, + null + ], + [ + 0.5387679675590371, + 0.5255729246382672, + 0.6493462016149997, + 0.624478397819645, + 0.6924104219697439, + 0.582177616621394, + null, + null + ], + [ + 0.557708009911873, + 0.5440573535811631, + 0.6760579671838519, + 0.6363153676548962, + 0.6932640677585229, + 0.5882016794591084, + 0.7097149898502301, + null + ], + [ + 0.5440200839137009, + 0.549122801488911, + 0.6700716758231308, + 0.6466862025461412, + 0.6768535934053176, + 0.5819464960208162, + 0.7062754692559148, + 0.7152910321439679 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 0.6157282389728079, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09377898486040985, + "ema_update_magnitude": 0.06227830254384934, + "final_multiplier": 0.6640965738384851 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 9, + "best_dev_macro_f1": 0.6623891911193713, + "losses": { + "current_ce": 0.5793638345681958, + "replay_ce": 0.30043413045978734, + "dark_distillation": 0.42657383261691956, + "supervised_contrastive": 4.21360826683896, + "total": 1.2879408689718397, + "replay_multiplier": 0.5511182736436044, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10717160958299814, + "ema_update_magnitude": 0.05686726362737617, + "final_multiplier": 0.5306187324109918 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.590851334180432, + "macro_f1": 0.5292270687553361, + "weighted_f1": 0.6159473441352522, + "macro_precision": 0.5563652936320073, + "macro_recall": 0.5695431071792728, + "per_class_f1": { + "FAVOR": 0.37547892720306514, + "AGAINST": 0.484251968503937, + "NONE": 0.7279503105590062 + }, + "accuracy_pct": 59.0851334180432, + "macro_f1_pct": 52.922706875533606, + "weighted_f1_pct": 61.594734413525224, + "macro_precision_pct": 55.63652936320073, + "macro_recall_pct": 56.95431071792728, + "per_class_f1_pct": { + "FAVOR": 37.547892720306514, + "AGAINST": 48.4251968503937, + "NONE": 72.79503105590062 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.6032831737346102, + "macro_f1": 0.5631093252583498, + "weighted_f1": 0.6173178876674911, + "macro_precision": 0.5643926992743689, + "macro_recall": 0.5903645031951025, + "per_class_f1": { + "FAVOR": 0.4186046511627907, + "AGAINST": 0.5796545105566219, + "NONE": 0.6910688140556369 + }, + "accuracy_pct": 60.32831737346102, + "macro_f1_pct": 56.31093252583498, + "weighted_f1_pct": 61.731788766749105, + "macro_precision_pct": 56.43926992743688, + "macro_recall_pct": 59.036450319510244, + "per_class_f1_pct": { + "FAVOR": 41.86046511627907, + "AGAINST": 57.965451055662186, + "NONE": 69.1068814055637 + } + } + ], + "accuracy": 0.5968379446640316, + "macro_f1": 0.5468315110777447, + "weighted_f1": 0.6148938408167903, + "macro_precision": 0.558112913461328, + "macro_recall": 0.5736373027149185, + "per_class_f1": { + "FAVOR": 0.3969171483622351, + "AGAINST": 0.532555879494655, + "NONE": 0.7110215053763441 + }, + "accuracy_pct": 59.683794466403164, + "macro_f1_pct": 54.68315110777448, + "weighted_f1_pct": 61.489384081679034, + "macro_precision_pct": 55.8112913461328, + "macro_recall_pct": 57.36373027149185, + "per_class_f1_pct": { + "FAVOR": 39.691714836223504, + "AGAINST": 53.2555879494655, + "NONE": 71.10215053763442 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.672886916030738, + "losses": { + "current_ce": 0.5586497469317345, + "replay_ce": 0.2243478038126514, + "dark_distillation": 0.3522329228974524, + "supervised_contrastive": 4.244264116741362, + "total": 1.2260460795391173, + "replay_multiplier": 0.6008182736393368, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09750293760391796, + "ema_update_magnitude": 0.0579264029093888, + "final_multiplier": 0.5940990531454629 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5819567979669632, + "macro_f1": 0.5204063837975658, + "weighted_f1": 0.6120312044125907, + "macro_precision": 0.5295236122366919, + "macro_recall": 0.5520922734975023, + "per_class_f1": { + "FAVOR": 0.3727810650887574, + "AGAINST": 0.4567307692307692, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 58.19567979669632, + "macro_f1_pct": 52.04063837975658, + "weighted_f1_pct": 61.203120441259074, + "macro_precision_pct": 52.952361223669186, + "macro_recall_pct": 55.209227349750236, + "per_class_f1_pct": { + "FAVOR": 37.278106508875744, + "AGAINST": 45.67307692307692, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5786593707250342, + "macro_f1": 0.5478952936580055, + "weighted_f1": 0.5971854768719987, + "macro_precision": 0.5621019817706779, + "macro_recall": 0.5877948664189039, + "per_class_f1": { + "FAVOR": 0.4195804195804196, + "AGAINST": 0.5574387947269304, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 57.865937072503414, + "macro_f1_pct": 54.78952936580055, + "weighted_f1_pct": 59.71854768719987, + "macro_precision_pct": 56.210198177067795, + "macro_recall_pct": 58.77948664189039, + "per_class_f1_pct": { + "FAVOR": 41.95804195804196, + "AGAINST": 55.74387947269304, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6489022632267165, + "weighted_f1": 0.6883144805546445, + "macro_precision": 0.645833745706372, + "macro_recall": 0.6538597187086502, + "per_class_f1": { + "FAVOR": 0.5469387755102041, + "AGAINST": 0.6030534351145038, + "NONE": 0.7967145790554415 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 64.89022632267165, + "weighted_f1_pct": 68.83144805546445, + "macro_precision_pct": 64.5833745706372, + "macro_recall_pct": 65.38597187086502, + "per_class_f1_pct": { + "FAVOR": 54.69387755102041, + "AGAINST": 60.30534351145038, + "NONE": 79.67145790554414 + } + } + ], + "accuracy": 0.605955334987593, + "macro_f1": 0.5654634301426401, + "weighted_f1": 0.6249135522468147, + "macro_precision": 0.5694301875431935, + "macro_recall": 0.5908537549896405, + "per_class_f1": { + "FAVOR": 0.4372842347525892, + "AGAINST": 0.532671629445823, + "NONE": 0.7264344262295082 + }, + "accuracy_pct": 60.5955334987593, + "macro_f1_pct": 56.54634301426401, + "weighted_f1_pct": 62.49135522468146, + "macro_precision_pct": 56.943018754319354, + "macro_recall_pct": 59.085375498964055, + "per_class_f1_pct": { + "FAVOR": 43.72842347525892, + "AGAINST": 53.2671629445823, + "NONE": 72.64344262295081 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.731393655889082, + "losses": { + "current_ce": 0.40162161117404455, + "replay_ce": 0.2393797692971501, + "dark_distillation": 0.3441159891403788, + "supervised_contrastive": 4.088513990340194, + "total": 1.0516372580838398, + "replay_multiplier": 0.5829096767213594, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0959712189937064, + "ema_update_magnitude": 0.055459863856492764, + "final_multiplier": 0.5778801648870346 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5667090216010165, + "macro_f1": 0.49977267795057373, + "weighted_f1": 0.5936835157219332, + "macro_precision": 0.5250545256270448, + "macro_recall": 0.5356770317554631, + "per_class_f1": { + "FAVOR": 0.3295880149812734, + "AGAINST": 0.4550898203592814, + "NONE": 0.7146401985111662 + }, + "accuracy_pct": 56.670902160101654, + "macro_f1_pct": 49.97726779505737, + "weighted_f1_pct": 59.368351572193326, + "macro_precision_pct": 52.50545256270448, + "macro_recall_pct": 53.56770317554631, + "per_class_f1_pct": { + "FAVOR": 32.958801498127336, + "AGAINST": 45.50898203592814, + "NONE": 71.46401985111662 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5348837209302325, + "macro_f1": 0.5061909509825049, + "weighted_f1": 0.5512548000810898, + "macro_precision": 0.5421494913754975, + "macro_recall": 0.5440686296059579, + "per_class_f1": { + "FAVOR": 0.3805970149253731, + "AGAINST": 0.5297113752122241, + "NONE": 0.6082644628099173 + }, + "accuracy_pct": 53.48837209302325, + "macro_f1_pct": 50.61909509825049, + "weighted_f1_pct": 55.12548000810899, + "macro_precision_pct": 54.21494913754975, + "macro_recall_pct": 54.406862960595795, + "per_class_f1_pct": { + "FAVOR": 38.059701492537314, + "AGAINST": 52.97113752122241, + "NONE": 60.82644628099173 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6659959758551308, + "macro_f1": 0.6342860580148716, + "weighted_f1": 0.675149999093661, + "macro_precision": 0.6368646850517029, + "macro_recall": 0.6452685255190782, + "per_class_f1": { + "FAVOR": 0.5338983050847458, + "AGAINST": 0.581081081081081, + "NONE": 0.7878787878787878 + }, + "accuracy_pct": 66.59959758551308, + "macro_f1_pct": 63.42860580148716, + "weighted_f1_pct": 67.5149999093661, + "macro_precision_pct": 63.68646850517029, + "macro_recall_pct": 64.52685255190782, + "per_class_f1_pct": { + "FAVOR": 53.38983050847458, + "AGAINST": 58.108108108108105, + "NONE": 78.78787878787878 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6806136680613668, + "macro_f1": 0.6320522674541541, + "weighted_f1": 0.6938634490967182, + "macro_precision": 0.6301708469647401, + "macro_recall": 0.6663856947933983, + "per_class_f1": { + "FAVOR": 0.48854961832061067, + "AGAINST": 0.628175519630485, + "NONE": 0.7794316644113667 + }, + "accuracy_pct": 68.06136680613668, + "macro_f1_pct": 63.20522674541541, + "weighted_f1_pct": 69.38634490967182, + "macro_precision_pct": 63.017084696474015, + "macro_recall_pct": 66.63856947933982, + "per_class_f1_pct": { + "FAVOR": 48.854961832061065, + "AGAINST": 62.8175519630485, + "NONE": 77.94316644113667 + } + } + ], + "accuracy": 0.6061493411420205, + "macro_f1": 0.5640196699469507, + "weighted_f1": 0.6243745272367086, + "macro_precision": 0.5775474165269592, + "macro_recall": 0.5922109179390277, + "per_class_f1": { + "FAVOR": 0.4298160696999032, + "AGAINST": 0.5409565695437053, + "NONE": 0.7212863705972435 + }, + "accuracy_pct": 60.61493411420204, + "macro_f1_pct": 56.40196699469507, + "weighted_f1_pct": 62.43745272367086, + "macro_precision_pct": 57.75474165269592, + "macro_recall_pct": 59.22109179390277, + "per_class_f1_pct": { + "FAVOR": 42.981606969990324, + "AGAINST": 54.09565695437053, + "NONE": 72.12863705972435 + } + }, + { + "after_task": 5, + "best_epoch": 11, + "best_dev_macro_f1": 0.7553833855992128, + "losses": { + "current_ce": 0.43858537752003895, + "replay_ce": 0.21158687879641852, + "dark_distillation": 0.3585774808270591, + "supervised_contrastive": 4.0848351614815845, + "total": 1.0772660090809776, + "replay_multiplier": 0.5870782915174904, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0960261785006973, + "ema_update_magnitude": 0.0515125538225056, + "final_multiplier": 0.5364428182688905 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.5275531881258319, + "weighted_f1": 0.621562506244403, + "macro_precision": 0.536990746681773, + "macro_recall": 0.5580520161674846, + "per_class_f1": { + "FAVOR": 0.38485804416403785, + "AGAINST": 0.45265588914549654, + "NONE": 0.7451456310679612 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 52.755318812583184, + "weighted_f1_pct": 62.156250624440304, + "macro_precision_pct": 53.6990746681773, + "macro_recall_pct": 55.805201616748455, + "per_class_f1_pct": { + "FAVOR": 38.48580441640379, + "AGAINST": 45.26558891454965, + "NONE": 74.51456310679612 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5581395348837209, + "macro_f1": 0.5272783864541832, + "weighted_f1": 0.584684688060344, + "macro_precision": 0.5512469723381058, + "macro_recall": 0.5704514737825752, + "per_class_f1": { + "FAVOR": 0.39375, + "AGAINST": 0.5099601593625498, + "NONE": 0.678125 + }, + "accuracy_pct": 55.81395348837209, + "macro_f1_pct": 52.72783864541832, + "weighted_f1_pct": 58.4684688060344, + "macro_precision_pct": 55.12469723381058, + "macro_recall_pct": 57.045147378257525, + "per_class_f1_pct": { + "FAVOR": 39.375, + "AGAINST": 50.99601593625498, + "NONE": 67.8125 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6732493321930463, + "weighted_f1": 0.7115282204774049, + "macro_precision": 0.6696268490193483, + "macro_recall": 0.6800524457119889, + "per_class_f1": { + "FAVOR": 0.584, + "AGAINST": 0.6183206106870229, + "NONE": 0.8174273858921162 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 67.32493321930463, + "weighted_f1_pct": 71.1528220477405, + "macro_precision_pct": 66.96268490193484, + "macro_recall_pct": 68.00524457119889, + "per_class_f1_pct": { + "FAVOR": 58.4, + "AGAINST": 61.832061068702295, + "NONE": 81.74273858921161 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6569037656903766, + "macro_f1": 0.6258358901738652, + "weighted_f1": 0.6718066814596391, + "macro_precision": 0.6288733099182338, + "macro_recall": 0.6755909152025681, + "per_class_f1": { + "FAVOR": 0.5263157894736842, + "AGAINST": 0.6143497757847534, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 65.69037656903765, + "macro_f1_pct": 62.58358901738652, + "weighted_f1_pct": 67.1806681459639, + "macro_precision_pct": 62.887330991823376, + "macro_recall_pct": 67.55909152025681, + "per_class_f1_pct": { + "FAVOR": 52.63157894736842, + "AGAINST": 61.43497757847533, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7274220032840722, + "macro_f1": 0.7142560068932152, + "weighted_f1": 0.7344172966771585, + "macro_precision": 0.7126322680415721, + "macro_recall": 0.732822240017586, + "per_class_f1": { + "FAVOR": 0.6363636363636364, + "AGAINST": 0.7138810198300283, + "NONE": 0.7925233644859813 + }, + "accuracy_pct": 72.74220032840722, + "macro_f1_pct": 71.42560068932153, + "weighted_f1_pct": 73.44172966771585, + "macro_precision_pct": 71.26322680415721, + "macro_recall_pct": 73.2822240017586, + "per_class_f1_pct": { + "FAVOR": 63.63636363636363, + "AGAINST": 71.38810198300283, + "NONE": 79.25233644859813 + } + } + ], + "accuracy": 0.6402274768033522, + "macro_f1": 0.6072944466820859, + "weighted_f1": 0.6570580139627015, + "macro_precision": 0.6098375035634389, + "macro_recall": 0.6366294334230376, + "per_class_f1": { + "FAVOR": 0.5019973368841545, + "AGAINST": 0.5711422845691383, + "NONE": 0.7487437185929648 + }, + "accuracy_pct": 64.02274768033523, + "macro_f1_pct": 60.729444668208586, + "weighted_f1_pct": 65.70580139627015, + "macro_precision_pct": 60.98375035634389, + "macro_recall_pct": 63.662943342303755, + "per_class_f1_pct": { + "FAVOR": 50.19973368841545, + "AGAINST": 57.11422845691383, + "NONE": 74.87437185929649 + } + }, + { + "after_task": 6, + "best_epoch": 9, + "best_dev_macro_f1": 0.6959419189304247, + "losses": { + "current_ce": 0.4662171551663625, + "replay_ce": 0.18734145224617238, + "dark_distillation": 0.3516244075538819, + "supervised_contrastive": 4.112886892703542, + "total": 1.0843591910705233, + "replay_multiplier": 0.5667504979604322, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10015700949464298, + "ema_update_magnitude": 0.054220822197950054, + "final_multiplier": 0.54135823814558 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5387679675590371, + "weighted_f1": 0.6281414736793608, + "macro_precision": 0.5479114452798663, + "macro_recall": 0.5730969829009044, + "per_class_f1": { + "FAVOR": 0.39087947882736157, + "AGAINST": 0.48089887640449436, + "NONE": 0.7445255474452555 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 53.87679675590371, + "weighted_f1_pct": 62.81414736793608, + "macro_precision_pct": 54.79114452798663, + "macro_recall_pct": 57.30969829009044, + "per_class_f1_pct": { + "FAVOR": 39.08794788273616, + "AGAINST": 48.08988764044943, + "NONE": 74.45255474452554 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5595075239398085, + "macro_f1": 0.5255729246382672, + "weighted_f1": 0.5831682987596867, + "macro_precision": 0.5451234700730975, + "macro_recall": 0.5616865000862322, + "per_class_f1": { + "FAVOR": 0.3838383838383838, + "AGAINST": 0.5220729366602687, + "NONE": 0.6708074534161491 + }, + "accuracy_pct": 55.95075239398085, + "macro_f1_pct": 52.55729246382672, + "weighted_f1_pct": 58.316829875968665, + "macro_precision_pct": 54.51234700730975, + "macro_recall_pct": 56.168650008623224, + "per_class_f1_pct": { + "FAVOR": 38.38383838383838, + "AGAINST": 52.207293666026864, + "NONE": 67.08074534161491 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6493462016149997, + "weighted_f1": 0.6944415126011254, + "macro_precision": 0.6487503741143158, + "macro_recall": 0.6532902807773846, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.5766423357664233, + "NONE": 0.819672131147541 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 64.93462016149996, + "weighted_f1_pct": 69.44415126011253, + "macro_precision_pct": 64.87503741143158, + "macro_recall_pct": 65.32902807773846, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 57.66423357664233, + "NONE": 81.9672131147541 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6527196652719666, + "macro_f1": 0.624478397819645, + "weighted_f1": 0.6678314128002614, + "macro_precision": 0.6251609971155005, + "macro_recall": 0.6771470381283655, + "per_class_f1": { + "FAVOR": 0.53156146179402, + "AGAINST": 0.6125290023201856, + "NONE": 0.7293447293447294 + }, + "accuracy_pct": 65.27196652719665, + "macro_f1_pct": 62.4478397819645, + "weighted_f1_pct": 66.78314128002614, + "macro_precision_pct": 62.516099711550055, + "macro_recall_pct": 67.71470381283655, + "per_class_f1_pct": { + "FAVOR": 53.156146179402, + "AGAINST": 61.25290023201856, + "NONE": 72.93447293447294 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6995073891625616, + "macro_f1": 0.6924104219697439, + "weighted_f1": 0.7075336331958191, + "macro_precision": 0.7020324510166406, + "macro_recall": 0.7228616029136544, + "per_class_f1": { + "FAVOR": 0.6384180790960452, + "AGAINST": 0.6868131868131868, + "NONE": 0.752 + }, + "accuracy_pct": 69.95073891625616, + "macro_f1_pct": 69.24104219697439, + "weighted_f1_pct": 70.75336331958191, + "macro_precision_pct": 70.20324510166405, + "macro_recall_pct": 72.28616029136545, + "per_class_f1_pct": { + "FAVOR": 63.84180790960452, + "AGAINST": 68.68131868131869, + "NONE": 75.2 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.617339312406577, + "macro_f1": 0.582177616621394, + "weighted_f1": 0.6370206755769637, + "macro_precision": 0.5901918202732708, + "macro_recall": 0.6133700401966004, + "per_class_f1": { + "FAVOR": 0.5633074935400517, + "AGAINST": 0.44936708860759494, + "NONE": 0.7338582677165354 + }, + "accuracy_pct": 61.733931240657704, + "macro_f1_pct": 58.2177616621394, + "weighted_f1_pct": 63.702067557696374, + "macro_precision_pct": 59.019182027327076, + "macro_recall_pct": 61.337004019660036, + "per_class_f1_pct": { + "FAVOR": 56.33074935400517, + "AGAINST": 44.936708860759495, + "NONE": 73.38582677165354 + } + } + ], + "accuracy": 0.6314214463840399, + "macro_f1": 0.6017990070201829, + "weighted_f1": 0.6486787591960806, + "macro_precision": 0.6061733116332967, + "macro_recall": 0.6329876154621205, + "per_class_f1": { + "FAVOR": 0.5143769968051118, + "AGAINST": 0.552956188855806, + "NONE": 0.7380638353996307 + }, + "accuracy_pct": 63.142144638403984, + "macro_f1_pct": 60.17990070201829, + "weighted_f1_pct": 64.86787591960807, + "macro_precision_pct": 60.617331163329666, + "macro_recall_pct": 63.29876154621205, + "per_class_f1_pct": { + "FAVOR": 51.43769968051119, + "AGAINST": 55.2956188855806, + "NONE": 73.80638353996306 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.7326662497394204, + "losses": { + "current_ce": 0.38933320352364154, + "replay_ce": 0.18236051325996716, + "dark_distillation": 0.3250864769731249, + "supervised_contrastive": 4.069064423583803, + "total": 0.998428778563227, + "replay_multiplier": 0.5838337234251252, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09645133756624359, + "ema_update_magnitude": 0.05368551393621005, + "final_multiplier": 0.5566072518106697 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.557708009911873, + "weighted_f1": 0.6552853074673407, + "macro_precision": 0.557199323445225, + "macro_recall": 0.5826685518951314, + "per_class_f1": { + "FAVOR": 0.40131578947368424, + "AGAINST": 0.4889975550122249, + "NONE": 0.7828106852497096 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 55.7708009911873, + "weighted_f1_pct": 65.52853074673408, + "macro_precision_pct": 55.719932344522505, + "macro_recall_pct": 58.266855189513144, + "per_class_f1_pct": { + "FAVOR": 40.131578947368425, + "AGAINST": 48.899755501222494, + "NONE": 78.28106852497096 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5800273597811217, + "macro_f1": 0.5440573535811631, + "weighted_f1": 0.6005443568077924, + "macro_precision": 0.5513765325638397, + "macro_recall": 0.5809299613702024, + "per_class_f1": { + "FAVOR": 0.4175084175084175, + "AGAINST": 0.5183673469387755, + "NONE": 0.6962962962962963 + }, + "accuracy_pct": 58.00273597811218, + "macro_f1_pct": 54.40573535811632, + "weighted_f1_pct": 60.054435680779235, + "macro_precision_pct": 55.13765325638397, + "macro_recall_pct": 58.092996137020236, + "per_class_f1_pct": { + "FAVOR": 41.75084175084175, + "AGAINST": 51.83673469387755, + "NONE": 69.62962962962963 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6760579671838519, + "weighted_f1": 0.7142228313676188, + "macro_precision": 0.6742482494181813, + "macro_recall": 0.6790518087496716, + "per_class_f1": { + "FAVOR": 0.6048387096774194, + "AGAINST": 0.6024096385542169, + "NONE": 0.8209255533199196 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.60579671838519, + "weighted_f1_pct": 71.42228313676188, + "macro_precision_pct": 67.42482494181813, + "macro_recall_pct": 67.90518087496716, + "per_class_f1_pct": { + "FAVOR": 60.483870967741936, + "AGAINST": 60.24096385542169, + "NONE": 82.09255533199196 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6764295676429568, + "macro_f1": 0.6363153676548962, + "weighted_f1": 0.6925431477277065, + "macro_precision": 0.6310926738149693, + "macro_recall": 0.6769117392581016, + "per_class_f1": { + "FAVOR": 0.5117845117845118, + "AGAINST": 0.625615763546798, + "NONE": 0.771545827633379 + }, + "accuracy_pct": 67.64295676429568, + "macro_f1_pct": 63.631536765489614, + "weighted_f1_pct": 69.25431477277066, + "macro_precision_pct": 63.10926738149693, + "macro_recall_pct": 67.69117392581016, + "per_class_f1_pct": { + "FAVOR": 51.178451178451176, + "AGAINST": 62.5615763546798, + "NONE": 77.1545827633379 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7027914614121511, + "macro_f1": 0.6932640677585229, + "weighted_f1": 0.7110026916481256, + "macro_precision": 0.6974469231994315, + "macro_recall": 0.713712505430203, + "per_class_f1": { + "FAVOR": 0.6136363636363636, + "AGAINST": 0.7062314540059347, + "NONE": 0.7599243856332704 + }, + "accuracy_pct": 70.27914614121511, + "macro_f1_pct": 69.32640677585229, + "weighted_f1_pct": 71.10026916481256, + "macro_precision_pct": 69.74469231994314, + "macro_recall_pct": 71.37125054302031, + "per_class_f1_pct": { + "FAVOR": 61.36363636363637, + "AGAINST": 70.62314540059347, + "NONE": 75.99243856332704 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6322869955156951, + "macro_f1": 0.5882016794591084, + "weighted_f1": 0.6502956969727246, + "macro_precision": 0.5994170805254248, + "macro_recall": 0.6161635922906187, + "per_class_f1": { + "FAVOR": 0.5576923076923077, + "AGAINST": 0.4460431654676259, + "NONE": 0.7608695652173914 + }, + "accuracy_pct": 63.22869955156951, + "macro_f1_pct": 58.82016794591084, + "weighted_f1_pct": 65.02956969727246, + "macro_precision_pct": 59.941708052542474, + "macro_recall_pct": 61.61635922906187, + "per_class_f1_pct": { + "FAVOR": 55.769230769230774, + "AGAINST": 44.60431654676259, + "NONE": 76.08695652173914 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7364185110663984, + "macro_f1": 0.7097149898502301, + "weighted_f1": 0.7460579842626739, + "macro_precision": 0.7143041427097154, + "macro_recall": 0.7266887577232405, + "per_class_f1": { + "FAVOR": 0.627177700348432, + "AGAINST": 0.676056338028169, + "NONE": 0.8259109311740891 + }, + "accuracy_pct": 73.64185110663985, + "macro_f1_pct": 70.971498985023, + "weighted_f1_pct": 74.60579842626738, + "macro_precision_pct": 71.43041427097154, + "macro_recall_pct": 72.66887577232404, + "per_class_f1_pct": { + "FAVOR": 62.7177700348432, + "AGAINST": 67.6056338028169, + "NONE": 82.5910931174089 + } + } + ], + "accuracy": 0.66074994453073, + "macro_f1": 0.6259542343009993, + "weighted_f1": 0.6752791658068897, + "macro_precision": 0.623995064976213, + "macro_recall": 0.6517280046874842, + "per_class_f1": { + "FAVOR": 0.53430258973194, + "AGAINST": 0.5726280436607892, + "NONE": 0.7709320695102686 + }, + "accuracy_pct": 66.074994453073, + "macro_f1_pct": 62.59542343009993, + "weighted_f1_pct": 67.52791658068898, + "macro_precision_pct": 62.3995064976213, + "macro_recall_pct": 65.17280046874842, + "per_class_f1_pct": { + "FAVOR": 53.430258973194, + "AGAINST": 57.262804366078925, + "NONE": 77.09320695102686 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.7734156910627498, + "losses": { + "current_ce": 0.3188671692183205, + "replay_ce": 0.16653647843529196, + "dark_distillation": 0.32476043297963986, + "supervised_contrastive": 4.014220697739545, + "total": 0.9070791248012992, + "replay_multiplier": 0.5651531295176306, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10054323849101035, + "ema_update_magnitude": 0.053728406064887546, + "final_multiplier": 0.5343810968421456 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6149936467598475, + "macro_f1": 0.5440200839137009, + "weighted_f1": 0.6406154897767269, + "macro_precision": 0.5495214315439035, + "macro_recall": 0.5699669168078754, + "per_class_f1": { + "FAVOR": 0.40540540540540543, + "AGAINST": 0.4583333333333333, + "NONE": 0.7683215130023641 + }, + "accuracy_pct": 61.499364675984744, + "macro_f1_pct": 54.402008391370096, + "weighted_f1_pct": 64.06154897767269, + "macro_precision_pct": 54.952143154390356, + "macro_recall_pct": 56.996691680787535, + "per_class_f1_pct": { + "FAVOR": 40.54054054054054, + "AGAINST": 45.83333333333333, + "NONE": 76.8321513002364 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.549122801488911, + "weighted_f1": 0.5981715214635795, + "macro_precision": 0.5631806108897742, + "macro_recall": 0.582280510720417, + "per_class_f1": { + "FAVOR": 0.42585551330798477, + "AGAINST": 0.5507246376811594, + "NONE": 0.6707882534775889 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.9122801488911, + "weighted_f1_pct": 59.817152146357955, + "macro_precision_pct": 56.31806108897742, + "macro_recall_pct": 58.228051072041694, + "per_class_f1_pct": { + "FAVOR": 42.585551330798474, + "AGAINST": 55.072463768115945, + "NONE": 67.07882534775888 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6700716758231308, + "weighted_f1": 0.7140202129619635, + "macro_precision": 0.678380221653879, + "macro_recall": 0.6706699234333723, + "per_class_f1": { + "FAVOR": 0.5523809523809524, + "AGAINST": 0.6231884057971014, + "NONE": 0.8346456692913385 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 67.00716758231307, + "weighted_f1_pct": 71.40202129619635, + "macro_precision_pct": 67.8380221653879, + "macro_recall_pct": 67.06699234333723, + "per_class_f1_pct": { + "FAVOR": 55.23809523809524, + "AGAINST": 62.31884057971014, + "NONE": 83.46456692913385 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6847977684797768, + "macro_f1": 0.6466862025461412, + "weighted_f1": 0.6977892409902622, + "macro_precision": 0.6461160691666847, + "macro_recall": 0.6881824948493515, + "per_class_f1": { + "FAVOR": 0.5338345864661654, + "AGAINST": 0.636568848758465, + "NONE": 0.7696551724137931 + }, + "accuracy_pct": 68.47977684797767, + "macro_f1_pct": 64.66862025461411, + "weighted_f1_pct": 69.77892409902621, + "macro_precision_pct": 64.61160691666848, + "macro_recall_pct": 68.81824948493515, + "per_class_f1_pct": { + "FAVOR": 53.383458646616546, + "AGAINST": 63.6568848758465, + "NONE": 76.96551724137932 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6896551724137931, + "macro_f1": 0.6768535934053176, + "weighted_f1": 0.6949662413158252, + "macro_precision": 0.6793693757392099, + "macro_recall": 0.6968793436950204, + "per_class_f1": { + "FAVOR": 0.6066666666666667, + "AGAINST": 0.6767676767676768, + "NONE": 0.7471264367816092 + }, + "accuracy_pct": 68.96551724137932, + "macro_f1_pct": 67.68535934053176, + "weighted_f1_pct": 69.49662413158252, + "macro_precision_pct": 67.93693757392099, + "macro_recall_pct": 69.68793436950203, + "per_class_f1_pct": { + "FAVOR": 60.66666666666667, + "AGAINST": 67.67676767676768, + "NONE": 74.71264367816092 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6158445440956651, + "macro_f1": 0.5819464960208162, + "weighted_f1": 0.6368768823908122, + "macro_precision": 0.5979635388632777, + "macro_recall": 0.6178344360667256, + "per_class_f1": { + "FAVOR": 0.5672371638141809, + "AGAINST": 0.44516129032258067, + "NONE": 0.7334410339256866 + }, + "accuracy_pct": 61.58445440956651, + "macro_f1_pct": 58.19464960208162, + "weighted_f1_pct": 63.68768823908122, + "macro_precision_pct": 59.79635388632777, + "macro_recall_pct": 61.78344360667256, + "per_class_f1_pct": { + "FAVOR": 56.723716381418086, + "AGAINST": 44.516129032258064, + "NONE": 73.34410339256866 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7283702213279678, + "macro_f1": 0.7062754692559148, + "weighted_f1": 0.7387901109222667, + "macro_precision": 0.706749032069545, + "macro_recall": 0.7273146709928319, + "per_class_f1": { + "FAVOR": 0.6175438596491228, + "AGAINST": 0.6905829596412556, + "NONE": 0.8106995884773662 + }, + "accuracy_pct": 72.83702213279678, + "macro_f1_pct": 70.62754692559147, + "weighted_f1_pct": 73.87901109222666, + "macro_precision_pct": 70.6749032069545, + "macro_recall_pct": 72.7314670992832, + "per_class_f1_pct": { + "FAVOR": 61.75438596491228, + "AGAINST": 69.05829596412556, + "NONE": 81.06995884773663 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7508305647840532, + "macro_f1": 0.7152910321439679, + "weighted_f1": 0.7578587159325979, + "macro_precision": 0.7023505547263792, + "macro_recall": 0.7382135148372551, + "per_class_f1": { + "FAVOR": 0.6611570247933884, + "AGAINST": 0.6518987341772152, + "NONE": 0.8328173374613003 + }, + "accuracy_pct": 75.08305647840533, + "macro_f1_pct": 71.5291032143968, + "weighted_f1_pct": 75.78587159325978, + "macro_precision_pct": 70.23505547263792, + "macro_recall_pct": 73.82135148372551, + "per_class_f1_pct": { + "FAVOR": 66.11570247933885, + "AGAINST": 65.18987341772153, + "NONE": 83.28173374613003 + } + } + ], + "accuracy": 0.6658837345860247, + "macro_f1": 0.6328978375046277, + "weighted_f1": 0.6796255736276681, + "macro_precision": 0.6307212593460513, + "macro_recall": 0.6595186280526214, + "per_class_f1": { + "FAVOR": 0.5460149713782475, + "AGAINST": 0.5841248303934871, + "NONE": 0.7685537107421484 + }, + "accuracy_pct": 66.58837345860246, + "macro_f1_pct": 63.28978375046277, + "weighted_f1_pct": 67.9625573627668, + "macro_precision_pct": 63.072125934605126, + "macro_recall_pct": 65.95186280526214, + "per_class_f1_pct": { + "FAVOR": 54.60149713782475, + "AGAINST": 58.41248303934871, + "NONE": 76.85537107421484 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6362834193247375, + "average_forgetting": 0.014767535044749718, + "backward_transfer": -0.009418554045379455, + "final_mean_target_macro_f1_pct": 63.62834193247375, + "average_forgetting_pct": 1.4767535044749718, + "backward_transfer_pct": -0.9418554045379455, + "final_pooled_accuracy": 0.6658837345860247, + "final_pooled_accuracy_pct": 66.58837345860246, + "final_pooled_macro_f1": 0.6328978375046277, + "final_pooled_macro_f1_pct": 63.28978375046277, + "final_pooled_weighted_f1": 0.6796255736276681, + "final_pooled_weighted_f1_pct": 67.9625573627668, + "final_pooled_macro_precision": 0.6307212593460513, + "final_pooled_macro_precision_pct": 63.072125934605126, + "final_pooled_macro_recall": 0.6595186280526214, + "final_pooled_macro_recall_pct": 65.95186280526214, + "final_pooled_per_class_f1": { + "FAVOR": 0.5460149713782475, + "AGAINST": 0.5841248303934871, + "NONE": 0.7685537107421484 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 54.60149713782475, + "AGAINST": 58.41248303934871, + "NONE": 76.85537107421484 + } + }, + "efficiency": { + "training_wall_time_seconds": 1836.928849024931, + "training_wall_time_minutes": 30.615480817082183, + "peak_gpu_memory_allocated_gb": 10.430231094360352, + "peak_gpu_memory_reserved_gb": 10.6015625 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s3__seed13/metrics.partial.json b/outputs/trace_s8_full__doc__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..92811f12c21fc7a3c0bd7396b225caa99c283140 --- /dev/null +++ b/outputs/trace_s8_full__doc__s3__seed13/metrics.partial.json @@ -0,0 +1,1543 @@ +{ + "matrix": [ + [ + 0.5906937314675289, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.5292270687553361, + 0.5631093252583498, + null, + null, + null, + null, + null, + null + ], + [ + 0.5204063837975658, + 0.5478952936580055, + 0.6489022632267165, + null, + null, + null, + null, + null + ], + [ + 0.49977267795057373, + 0.5061909509825049, + 0.6342860580148716, + 0.6320522674541541, + null, + null, + null, + null + ], + [ + 0.5275531881258319, + 0.5272783864541832, + 0.6732493321930463, + 0.6258358901738652, + 0.7142560068932152, + null, + null, + null + ], + [ + 0.5387679675590371, + 0.5255729246382672, + 0.6493462016149997, + 0.624478397819645, + 0.6924104219697439, + 0.582177616621394, + null, + null + ], + [ + 0.557708009911873, + 0.5440573535811631, + 0.6760579671838519, + 0.6363153676548962, + 0.6932640677585229, + 0.5882016794591084, + 0.7097149898502301, + null + ], + [ + 0.5440200839137009, + 0.549122801488911, + 0.6700716758231308, + 0.6466862025461412, + 0.6768535934053176, + 0.5819464960208162, + 0.7062754692559148, + 0.7152910321439679 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 7, + "best_dev_macro_f1": 0.6682621382082549, + "losses": { + "current_ce": 0.6031686561858213, + "total": 0.6031686561858213, + "replay_multiplier": 0.6157282389728079, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09377898486040985, + "ema_update_magnitude": 0.06227830254384934, + "final_multiplier": 0.6640965738384851 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + } + ], + "accuracy": 0.6454891994917408, + "macro_f1": 0.5906937314675289, + "weighted_f1": 0.6669670470265637, + "macro_precision": 0.6099031341569652, + "macro_recall": 0.6273193078203971, + "per_class_f1": { + "FAVOR": 0.4806201550387597, + "AGAINST": 0.5237113402061856, + "NONE": 0.7677496991576414 + }, + "accuracy_pct": 64.54891994917408, + "macro_f1_pct": 59.06937314675289, + "weighted_f1_pct": 66.69670470265638, + "macro_precision_pct": 60.99031341569652, + "macro_recall_pct": 62.73193078203971, + "per_class_f1_pct": { + "FAVOR": 48.06201550387597, + "AGAINST": 52.37113402061856, + "NONE": 76.77496991576413 + } + }, + { + "after_task": 2, + "best_epoch": 9, + "best_dev_macro_f1": 0.6623891911193713, + "losses": { + "current_ce": 0.5793638345681958, + "replay_ce": 0.30043413045978734, + "dark_distillation": 0.42657383261691956, + "supervised_contrastive": 4.21360826683896, + "total": 1.2879408689718397, + "replay_multiplier": 0.5511182736436044, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10717160958299814, + "ema_update_magnitude": 0.05686726362737617, + "final_multiplier": 0.5306187324109918 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.590851334180432, + "macro_f1": 0.5292270687553361, + "weighted_f1": 0.6159473441352522, + "macro_precision": 0.5563652936320073, + "macro_recall": 0.5695431071792728, + "per_class_f1": { + "FAVOR": 0.37547892720306514, + "AGAINST": 0.484251968503937, + "NONE": 0.7279503105590062 + }, + "accuracy_pct": 59.0851334180432, + "macro_f1_pct": 52.922706875533606, + "weighted_f1_pct": 61.594734413525224, + "macro_precision_pct": 55.63652936320073, + "macro_recall_pct": 56.95431071792728, + "per_class_f1_pct": { + "FAVOR": 37.547892720306514, + "AGAINST": 48.4251968503937, + "NONE": 72.79503105590062 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.6032831737346102, + "macro_f1": 0.5631093252583498, + "weighted_f1": 0.6173178876674911, + "macro_precision": 0.5643926992743689, + "macro_recall": 0.5903645031951025, + "per_class_f1": { + "FAVOR": 0.4186046511627907, + "AGAINST": 0.5796545105566219, + "NONE": 0.6910688140556369 + }, + "accuracy_pct": 60.32831737346102, + "macro_f1_pct": 56.31093252583498, + "weighted_f1_pct": 61.731788766749105, + "macro_precision_pct": 56.43926992743688, + "macro_recall_pct": 59.036450319510244, + "per_class_f1_pct": { + "FAVOR": 41.86046511627907, + "AGAINST": 57.965451055662186, + "NONE": 69.1068814055637 + } + } + ], + "accuracy": 0.5968379446640316, + "macro_f1": 0.5468315110777447, + "weighted_f1": 0.6148938408167903, + "macro_precision": 0.558112913461328, + "macro_recall": 0.5736373027149185, + "per_class_f1": { + "FAVOR": 0.3969171483622351, + "AGAINST": 0.532555879494655, + "NONE": 0.7110215053763441 + }, + "accuracy_pct": 59.683794466403164, + "macro_f1_pct": 54.68315110777448, + "weighted_f1_pct": 61.489384081679034, + "macro_precision_pct": 55.8112913461328, + "macro_recall_pct": 57.36373027149185, + "per_class_f1_pct": { + "FAVOR": 39.691714836223504, + "AGAINST": 53.2555879494655, + "NONE": 71.10215053763442 + } + }, + { + "after_task": 3, + "best_epoch": 5, + "best_dev_macro_f1": 0.672886916030738, + "losses": { + "current_ce": 0.5586497469317345, + "replay_ce": 0.2243478038126514, + "dark_distillation": 0.3522329228974524, + "supervised_contrastive": 4.244264116741362, + "total": 1.2260460795391173, + "replay_multiplier": 0.6008182736393368, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09750293760391796, + "ema_update_magnitude": 0.0579264029093888, + "final_multiplier": 0.5940990531454629 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5819567979669632, + "macro_f1": 0.5204063837975658, + "weighted_f1": 0.6120312044125907, + "macro_precision": 0.5295236122366919, + "macro_recall": 0.5520922734975023, + "per_class_f1": { + "FAVOR": 0.3727810650887574, + "AGAINST": 0.4567307692307692, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 58.19567979669632, + "macro_f1_pct": 52.04063837975658, + "weighted_f1_pct": 61.203120441259074, + "macro_precision_pct": 52.952361223669186, + "macro_recall_pct": 55.209227349750236, + "per_class_f1_pct": { + "FAVOR": 37.278106508875744, + "AGAINST": 45.67307692307692, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5786593707250342, + "macro_f1": 0.5478952936580055, + "weighted_f1": 0.5971854768719987, + "macro_precision": 0.5621019817706779, + "macro_recall": 0.5877948664189039, + "per_class_f1": { + "FAVOR": 0.4195804195804196, + "AGAINST": 0.5574387947269304, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 57.865937072503414, + "macro_f1_pct": 54.78952936580055, + "weighted_f1_pct": 59.71854768719987, + "macro_precision_pct": 56.210198177067795, + "macro_recall_pct": 58.77948664189039, + "per_class_f1_pct": { + "FAVOR": 41.95804195804196, + "AGAINST": 55.74387947269304, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6841046277665996, + "macro_f1": 0.6489022632267165, + "weighted_f1": 0.6883144805546445, + "macro_precision": 0.645833745706372, + "macro_recall": 0.6538597187086502, + "per_class_f1": { + "FAVOR": 0.5469387755102041, + "AGAINST": 0.6030534351145038, + "NONE": 0.7967145790554415 + }, + "accuracy_pct": 68.41046277665997, + "macro_f1_pct": 64.89022632267165, + "weighted_f1_pct": 68.83144805546445, + "macro_precision_pct": 64.5833745706372, + "macro_recall_pct": 65.38597187086502, + "per_class_f1_pct": { + "FAVOR": 54.69387755102041, + "AGAINST": 60.30534351145038, + "NONE": 79.67145790554414 + } + } + ], + "accuracy": 0.605955334987593, + "macro_f1": 0.5654634301426401, + "weighted_f1": 0.6249135522468147, + "macro_precision": 0.5694301875431935, + "macro_recall": 0.5908537549896405, + "per_class_f1": { + "FAVOR": 0.4372842347525892, + "AGAINST": 0.532671629445823, + "NONE": 0.7264344262295082 + }, + "accuracy_pct": 60.5955334987593, + "macro_f1_pct": 56.54634301426401, + "weighted_f1_pct": 62.49135522468146, + "macro_precision_pct": 56.943018754319354, + "macro_recall_pct": 59.085375498964055, + "per_class_f1_pct": { + "FAVOR": 43.72842347525892, + "AGAINST": 53.2671629445823, + "NONE": 72.64344262295081 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.731393655889082, + "losses": { + "current_ce": 0.40162161117404455, + "replay_ce": 0.2393797692971501, + "dark_distillation": 0.3441159891403788, + "supervised_contrastive": 4.088513990340194, + "total": 1.0516372580838398, + "replay_multiplier": 0.5829096767213594, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0959712189937064, + "ema_update_magnitude": 0.055459863856492764, + "final_multiplier": 0.5778801648870346 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5667090216010165, + "macro_f1": 0.49977267795057373, + "weighted_f1": 0.5936835157219332, + "macro_precision": 0.5250545256270448, + "macro_recall": 0.5356770317554631, + "per_class_f1": { + "FAVOR": 0.3295880149812734, + "AGAINST": 0.4550898203592814, + "NONE": 0.7146401985111662 + }, + "accuracy_pct": 56.670902160101654, + "macro_f1_pct": 49.97726779505737, + "weighted_f1_pct": 59.368351572193326, + "macro_precision_pct": 52.50545256270448, + "macro_recall_pct": 53.56770317554631, + "per_class_f1_pct": { + "FAVOR": 32.958801498127336, + "AGAINST": 45.50898203592814, + "NONE": 71.46401985111662 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5348837209302325, + "macro_f1": 0.5061909509825049, + "weighted_f1": 0.5512548000810898, + "macro_precision": 0.5421494913754975, + "macro_recall": 0.5440686296059579, + "per_class_f1": { + "FAVOR": 0.3805970149253731, + "AGAINST": 0.5297113752122241, + "NONE": 0.6082644628099173 + }, + "accuracy_pct": 53.48837209302325, + "macro_f1_pct": 50.61909509825049, + "weighted_f1_pct": 55.12548000810899, + "macro_precision_pct": 54.21494913754975, + "macro_recall_pct": 54.406862960595795, + "per_class_f1_pct": { + "FAVOR": 38.059701492537314, + "AGAINST": 52.97113752122241, + "NONE": 60.82644628099173 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6659959758551308, + "macro_f1": 0.6342860580148716, + "weighted_f1": 0.675149999093661, + "macro_precision": 0.6368646850517029, + "macro_recall": 0.6452685255190782, + "per_class_f1": { + "FAVOR": 0.5338983050847458, + "AGAINST": 0.581081081081081, + "NONE": 0.7878787878787878 + }, + "accuracy_pct": 66.59959758551308, + "macro_f1_pct": 63.42860580148716, + "weighted_f1_pct": 67.5149999093661, + "macro_precision_pct": 63.68646850517029, + "macro_recall_pct": 64.52685255190782, + "per_class_f1_pct": { + "FAVOR": 53.38983050847458, + "AGAINST": 58.108108108108105, + "NONE": 78.78787878787878 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6806136680613668, + "macro_f1": 0.6320522674541541, + "weighted_f1": 0.6938634490967182, + "macro_precision": 0.6301708469647401, + "macro_recall": 0.6663856947933983, + "per_class_f1": { + "FAVOR": 0.48854961832061067, + "AGAINST": 0.628175519630485, + "NONE": 0.7794316644113667 + }, + "accuracy_pct": 68.06136680613668, + "macro_f1_pct": 63.20522674541541, + "weighted_f1_pct": 69.38634490967182, + "macro_precision_pct": 63.017084696474015, + "macro_recall_pct": 66.63856947933982, + "per_class_f1_pct": { + "FAVOR": 48.854961832061065, + "AGAINST": 62.8175519630485, + "NONE": 77.94316644113667 + } + } + ], + "accuracy": 0.6061493411420205, + "macro_f1": 0.5640196699469507, + "weighted_f1": 0.6243745272367086, + "macro_precision": 0.5775474165269592, + "macro_recall": 0.5922109179390277, + "per_class_f1": { + "FAVOR": 0.4298160696999032, + "AGAINST": 0.5409565695437053, + "NONE": 0.7212863705972435 + }, + "accuracy_pct": 60.61493411420204, + "macro_f1_pct": 56.40196699469507, + "weighted_f1_pct": 62.43745272367086, + "macro_precision_pct": 57.75474165269592, + "macro_recall_pct": 59.22109179390277, + "per_class_f1_pct": { + "FAVOR": 42.981606969990324, + "AGAINST": 54.09565695437053, + "NONE": 72.12863705972435 + } + }, + { + "after_task": 5, + "best_epoch": 11, + "best_dev_macro_f1": 0.7553833855992128, + "losses": { + "current_ce": 0.43858537752003895, + "replay_ce": 0.21158687879641852, + "dark_distillation": 0.3585774808270591, + "supervised_contrastive": 4.0848351614815845, + "total": 1.0772660090809776, + "replay_multiplier": 0.5870782915174904, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0960261785006973, + "ema_update_magnitude": 0.0515125538225056, + "final_multiplier": 0.5364428182688905 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.5921219822109276, + "macro_f1": 0.5275531881258319, + "weighted_f1": 0.621562506244403, + "macro_precision": 0.536990746681773, + "macro_recall": 0.5580520161674846, + "per_class_f1": { + "FAVOR": 0.38485804416403785, + "AGAINST": 0.45265588914549654, + "NONE": 0.7451456310679612 + }, + "accuracy_pct": 59.21219822109276, + "macro_f1_pct": 52.755318812583184, + "weighted_f1_pct": 62.156250624440304, + "macro_precision_pct": 53.6990746681773, + "macro_recall_pct": 55.805201616748455, + "per_class_f1_pct": { + "FAVOR": 38.48580441640379, + "AGAINST": 45.26558891454965, + "NONE": 74.51456310679612 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5581395348837209, + "macro_f1": 0.5272783864541832, + "weighted_f1": 0.584684688060344, + "macro_precision": 0.5512469723381058, + "macro_recall": 0.5704514737825752, + "per_class_f1": { + "FAVOR": 0.39375, + "AGAINST": 0.5099601593625498, + "NONE": 0.678125 + }, + "accuracy_pct": 55.81395348837209, + "macro_f1_pct": 52.72783864541832, + "weighted_f1_pct": 58.4684688060344, + "macro_precision_pct": 55.12469723381058, + "macro_recall_pct": 57.045147378257525, + "per_class_f1_pct": { + "FAVOR": 39.375, + "AGAINST": 50.99601593625498, + "NONE": 67.8125 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7062374245472837, + "macro_f1": 0.6732493321930463, + "weighted_f1": 0.7115282204774049, + "macro_precision": 0.6696268490193483, + "macro_recall": 0.6800524457119889, + "per_class_f1": { + "FAVOR": 0.584, + "AGAINST": 0.6183206106870229, + "NONE": 0.8174273858921162 + }, + "accuracy_pct": 70.62374245472837, + "macro_f1_pct": 67.32493321930463, + "weighted_f1_pct": 71.1528220477405, + "macro_precision_pct": 66.96268490193484, + "macro_recall_pct": 68.00524457119889, + "per_class_f1_pct": { + "FAVOR": 58.4, + "AGAINST": 61.832061068702295, + "NONE": 81.74273858921161 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6569037656903766, + "macro_f1": 0.6258358901738652, + "weighted_f1": 0.6718066814596391, + "macro_precision": 0.6288733099182338, + "macro_recall": 0.6755909152025681, + "per_class_f1": { + "FAVOR": 0.5263157894736842, + "AGAINST": 0.6143497757847534, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 65.69037656903765, + "macro_f1_pct": 62.58358901738652, + "weighted_f1_pct": 67.1806681459639, + "macro_precision_pct": 62.887330991823376, + "macro_recall_pct": 67.55909152025681, + "per_class_f1_pct": { + "FAVOR": 52.63157894736842, + "AGAINST": 61.43497757847533, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7274220032840722, + "macro_f1": 0.7142560068932152, + "weighted_f1": 0.7344172966771585, + "macro_precision": 0.7126322680415721, + "macro_recall": 0.732822240017586, + "per_class_f1": { + "FAVOR": 0.6363636363636364, + "AGAINST": 0.7138810198300283, + "NONE": 0.7925233644859813 + }, + "accuracy_pct": 72.74220032840722, + "macro_f1_pct": 71.42560068932153, + "weighted_f1_pct": 73.44172966771585, + "macro_precision_pct": 71.26322680415721, + "macro_recall_pct": 73.2822240017586, + "per_class_f1_pct": { + "FAVOR": 63.63636363636363, + "AGAINST": 71.38810198300283, + "NONE": 79.25233644859813 + } + } + ], + "accuracy": 0.6402274768033522, + "macro_f1": 0.6072944466820859, + "weighted_f1": 0.6570580139627015, + "macro_precision": 0.6098375035634389, + "macro_recall": 0.6366294334230376, + "per_class_f1": { + "FAVOR": 0.5019973368841545, + "AGAINST": 0.5711422845691383, + "NONE": 0.7487437185929648 + }, + "accuracy_pct": 64.02274768033523, + "macro_f1_pct": 60.729444668208586, + "weighted_f1_pct": 65.70580139627015, + "macro_precision_pct": 60.98375035634389, + "macro_recall_pct": 63.662943342303755, + "per_class_f1_pct": { + "FAVOR": 50.19973368841545, + "AGAINST": 57.11422845691383, + "NONE": 74.87437185929649 + } + }, + { + "after_task": 6, + "best_epoch": 9, + "best_dev_macro_f1": 0.6959419189304247, + "losses": { + "current_ce": 0.4662171551663625, + "replay_ce": 0.18734145224617238, + "dark_distillation": 0.3516244075538819, + "supervised_contrastive": 4.112886892703542, + "total": 1.0843591910705233, + "replay_multiplier": 0.5667504979604322, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10015700949464298, + "ema_update_magnitude": 0.054220822197950054, + "final_multiplier": 0.54135823814558 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 132, + "1": 126, + "2": 126 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6010165184243964, + "macro_f1": 0.5387679675590371, + "weighted_f1": 0.6281414736793608, + "macro_precision": 0.5479114452798663, + "macro_recall": 0.5730969829009044, + "per_class_f1": { + "FAVOR": 0.39087947882736157, + "AGAINST": 0.48089887640449436, + "NONE": 0.7445255474452555 + }, + "accuracy_pct": 60.101651842439644, + "macro_f1_pct": 53.87679675590371, + "weighted_f1_pct": 62.81414736793608, + "macro_precision_pct": 54.79114452798663, + "macro_recall_pct": 57.30969829009044, + "per_class_f1_pct": { + "FAVOR": 39.08794788273616, + "AGAINST": 48.08988764044943, + "NONE": 74.45255474452554 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5595075239398085, + "macro_f1": 0.5255729246382672, + "weighted_f1": 0.5831682987596867, + "macro_precision": 0.5451234700730975, + "macro_recall": 0.5616865000862322, + "per_class_f1": { + "FAVOR": 0.3838383838383838, + "AGAINST": 0.5220729366602687, + "NONE": 0.6708074534161491 + }, + "accuracy_pct": 55.95075239398085, + "macro_f1_pct": 52.55729246382672, + "weighted_f1_pct": 58.316829875968665, + "macro_precision_pct": 54.51234700730975, + "macro_recall_pct": 56.168650008623224, + "per_class_f1_pct": { + "FAVOR": 38.38383838383838, + "AGAINST": 52.207293666026864, + "NONE": 67.08074534161491 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.6901408450704225, + "macro_f1": 0.6493462016149997, + "weighted_f1": 0.6944415126011254, + "macro_precision": 0.6487503741143158, + "macro_recall": 0.6532902807773846, + "per_class_f1": { + "FAVOR": 0.5517241379310345, + "AGAINST": 0.5766423357664233, + "NONE": 0.819672131147541 + }, + "accuracy_pct": 69.01408450704226, + "macro_f1_pct": 64.93462016149996, + "weighted_f1_pct": 69.44415126011253, + "macro_precision_pct": 64.87503741143158, + "macro_recall_pct": 65.32902807773846, + "per_class_f1_pct": { + "FAVOR": 55.172413793103445, + "AGAINST": 57.66423357664233, + "NONE": 81.9672131147541 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6527196652719666, + "macro_f1": 0.624478397819645, + "weighted_f1": 0.6678314128002614, + "macro_precision": 0.6251609971155005, + "macro_recall": 0.6771470381283655, + "per_class_f1": { + "FAVOR": 0.53156146179402, + "AGAINST": 0.6125290023201856, + "NONE": 0.7293447293447294 + }, + "accuracy_pct": 65.27196652719665, + "macro_f1_pct": 62.4478397819645, + "weighted_f1_pct": 66.78314128002614, + "macro_precision_pct": 62.516099711550055, + "macro_recall_pct": 67.71470381283655, + "per_class_f1_pct": { + "FAVOR": 53.156146179402, + "AGAINST": 61.25290023201856, + "NONE": 72.93447293447294 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6995073891625616, + "macro_f1": 0.6924104219697439, + "weighted_f1": 0.7075336331958191, + "macro_precision": 0.7020324510166406, + "macro_recall": 0.7228616029136544, + "per_class_f1": { + "FAVOR": 0.6384180790960452, + "AGAINST": 0.6868131868131868, + "NONE": 0.752 + }, + "accuracy_pct": 69.95073891625616, + "macro_f1_pct": 69.24104219697439, + "weighted_f1_pct": 70.75336331958191, + "macro_precision_pct": 70.20324510166405, + "macro_recall_pct": 72.28616029136545, + "per_class_f1_pct": { + "FAVOR": 63.84180790960452, + "AGAINST": 68.68131868131869, + "NONE": 75.2 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.617339312406577, + "macro_f1": 0.582177616621394, + "weighted_f1": 0.6370206755769637, + "macro_precision": 0.5901918202732708, + "macro_recall": 0.6133700401966004, + "per_class_f1": { + "FAVOR": 0.5633074935400517, + "AGAINST": 0.44936708860759494, + "NONE": 0.7338582677165354 + }, + "accuracy_pct": 61.733931240657704, + "macro_f1_pct": 58.2177616621394, + "weighted_f1_pct": 63.702067557696374, + "macro_precision_pct": 59.019182027327076, + "macro_recall_pct": 61.337004019660036, + "per_class_f1_pct": { + "FAVOR": 56.33074935400517, + "AGAINST": 44.936708860759495, + "NONE": 73.38582677165354 + } + } + ], + "accuracy": 0.6314214463840399, + "macro_f1": 0.6017990070201829, + "weighted_f1": 0.6486787591960806, + "macro_precision": 0.6061733116332967, + "macro_recall": 0.6329876154621205, + "per_class_f1": { + "FAVOR": 0.5143769968051118, + "AGAINST": 0.552956188855806, + "NONE": 0.7380638353996307 + }, + "accuracy_pct": 63.142144638403984, + "macro_f1_pct": 60.17990070201829, + "weighted_f1_pct": 64.86787591960807, + "macro_precision_pct": 60.617331163329666, + "macro_recall_pct": 63.29876154621205, + "per_class_f1_pct": { + "FAVOR": 51.43769968051119, + "AGAINST": 55.2956188855806, + "NONE": 73.80638353996306 + } + }, + { + "after_task": 7, + "best_epoch": 14, + "best_dev_macro_f1": 0.7326662497394204, + "losses": { + "current_ce": 0.38933320352364154, + "replay_ce": 0.18236051325996716, + "dark_distillation": 0.3250864769731249, + "supervised_contrastive": 4.069064423583803, + "total": 0.998428778563227, + "replay_multiplier": 0.5838337234251252, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09645133756624359, + "ema_update_magnitude": 0.05368551393621005, + "final_multiplier": 0.5566072518106697 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 154, + "1": 147, + "2": 147 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6327827191867853, + "macro_f1": 0.557708009911873, + "weighted_f1": 0.6552853074673407, + "macro_precision": 0.557199323445225, + "macro_recall": 0.5826685518951314, + "per_class_f1": { + "FAVOR": 0.40131578947368424, + "AGAINST": 0.4889975550122249, + "NONE": 0.7828106852497096 + }, + "accuracy_pct": 63.27827191867853, + "macro_f1_pct": 55.7708009911873, + "weighted_f1_pct": 65.52853074673408, + "macro_precision_pct": 55.719932344522505, + "macro_recall_pct": 58.266855189513144, + "per_class_f1_pct": { + "FAVOR": 40.131578947368425, + "AGAINST": 48.899755501222494, + "NONE": 78.28106852497096 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5800273597811217, + "macro_f1": 0.5440573535811631, + "weighted_f1": 0.6005443568077924, + "macro_precision": 0.5513765325638397, + "macro_recall": 0.5809299613702024, + "per_class_f1": { + "FAVOR": 0.4175084175084175, + "AGAINST": 0.5183673469387755, + "NONE": 0.6962962962962963 + }, + "accuracy_pct": 58.00273597811218, + "macro_f1_pct": 54.40573535811632, + "weighted_f1_pct": 60.054435680779235, + "macro_precision_pct": 55.13765325638397, + "macro_recall_pct": 58.092996137020236, + "per_class_f1_pct": { + "FAVOR": 41.75084175084175, + "AGAINST": 51.83673469387755, + "NONE": 69.62962962962963 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.7122736418511066, + "macro_f1": 0.6760579671838519, + "weighted_f1": 0.7142228313676188, + "macro_precision": 0.6742482494181813, + "macro_recall": 0.6790518087496716, + "per_class_f1": { + "FAVOR": 0.6048387096774194, + "AGAINST": 0.6024096385542169, + "NONE": 0.8209255533199196 + }, + "accuracy_pct": 71.22736418511066, + "macro_f1_pct": 67.60579671838519, + "weighted_f1_pct": 71.42228313676188, + "macro_precision_pct": 67.42482494181813, + "macro_recall_pct": 67.90518087496716, + "per_class_f1_pct": { + "FAVOR": 60.483870967741936, + "AGAINST": 60.24096385542169, + "NONE": 82.09255533199196 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6764295676429568, + "macro_f1": 0.6363153676548962, + "weighted_f1": 0.6925431477277065, + "macro_precision": 0.6310926738149693, + "macro_recall": 0.6769117392581016, + "per_class_f1": { + "FAVOR": 0.5117845117845118, + "AGAINST": 0.625615763546798, + "NONE": 0.771545827633379 + }, + "accuracy_pct": 67.64295676429568, + "macro_f1_pct": 63.631536765489614, + "weighted_f1_pct": 69.25431477277066, + "macro_precision_pct": 63.10926738149693, + "macro_recall_pct": 67.69117392581016, + "per_class_f1_pct": { + "FAVOR": 51.178451178451176, + "AGAINST": 62.5615763546798, + "NONE": 77.1545827633379 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.7027914614121511, + "macro_f1": 0.6932640677585229, + "weighted_f1": 0.7110026916481256, + "macro_precision": 0.6974469231994315, + "macro_recall": 0.713712505430203, + "per_class_f1": { + "FAVOR": 0.6136363636363636, + "AGAINST": 0.7062314540059347, + "NONE": 0.7599243856332704 + }, + "accuracy_pct": 70.27914614121511, + "macro_f1_pct": 69.32640677585229, + "weighted_f1_pct": 71.10026916481256, + "macro_precision_pct": 69.74469231994314, + "macro_recall_pct": 71.37125054302031, + "per_class_f1_pct": { + "FAVOR": 61.36363636363637, + "AGAINST": 70.62314540059347, + "NONE": 75.99243856332704 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6322869955156951, + "macro_f1": 0.5882016794591084, + "weighted_f1": 0.6502956969727246, + "macro_precision": 0.5994170805254248, + "macro_recall": 0.6161635922906187, + "per_class_f1": { + "FAVOR": 0.5576923076923077, + "AGAINST": 0.4460431654676259, + "NONE": 0.7608695652173914 + }, + "accuracy_pct": 63.22869955156951, + "macro_f1_pct": 58.82016794591084, + "weighted_f1_pct": 65.02956969727246, + "macro_precision_pct": 59.941708052542474, + "macro_recall_pct": 61.61635922906187, + "per_class_f1_pct": { + "FAVOR": 55.769230769230774, + "AGAINST": 44.60431654676259, + "NONE": 76.08695652173914 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7364185110663984, + "macro_f1": 0.7097149898502301, + "weighted_f1": 0.7460579842626739, + "macro_precision": 0.7143041427097154, + "macro_recall": 0.7266887577232405, + "per_class_f1": { + "FAVOR": 0.627177700348432, + "AGAINST": 0.676056338028169, + "NONE": 0.8259109311740891 + }, + "accuracy_pct": 73.64185110663985, + "macro_f1_pct": 70.971498985023, + "weighted_f1_pct": 74.60579842626738, + "macro_precision_pct": 71.43041427097154, + "macro_recall_pct": 72.66887577232404, + "per_class_f1_pct": { + "FAVOR": 62.7177700348432, + "AGAINST": 67.6056338028169, + "NONE": 82.5910931174089 + } + } + ], + "accuracy": 0.66074994453073, + "macro_f1": 0.6259542343009993, + "weighted_f1": 0.6752791658068897, + "macro_precision": 0.623995064976213, + "macro_recall": 0.6517280046874842, + "per_class_f1": { + "FAVOR": 0.53430258973194, + "AGAINST": 0.5726280436607892, + "NONE": 0.7709320695102686 + }, + "accuracy_pct": 66.074994453073, + "macro_f1_pct": 62.59542343009993, + "weighted_f1_pct": 67.52791658068898, + "macro_precision_pct": 62.3995064976213, + "macro_recall_pct": 65.17280046874842, + "per_class_f1_pct": { + "FAVOR": 53.430258973194, + "AGAINST": 57.262804366078925, + "NONE": 77.09320695102686 + } + }, + { + "after_task": 8, + "best_epoch": 8, + "best_dev_macro_f1": 0.7734156910627498, + "losses": { + "current_ce": 0.3188671692183205, + "replay_ce": 0.16653647843529196, + "dark_distillation": 0.32476043297963986, + "supervised_contrastive": 4.014220697739545, + "total": 0.9070791248012992, + "replay_multiplier": 0.5651531295176306, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10054323849101035, + "ema_update_magnitude": 0.053728406064887546, + "final_multiplier": 0.5343810968421456 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 176, + "1": 168, + "2": 168 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "abortion", + "accuracy": 0.6149936467598475, + "macro_f1": 0.5440200839137009, + "weighted_f1": 0.6406154897767269, + "macro_precision": 0.5495214315439035, + "macro_recall": 0.5699669168078754, + "per_class_f1": { + "FAVOR": 0.40540540540540543, + "AGAINST": 0.4583333333333333, + "NONE": 0.7683215130023641 + }, + "accuracy_pct": 61.499364675984744, + "macro_f1_pct": 54.402008391370096, + "weighted_f1_pct": 64.06154897767269, + "macro_precision_pct": 54.952143154390356, + "macro_recall_pct": 56.996691680787535, + "per_class_f1_pct": { + "FAVOR": 40.54054054054054, + "AGAINST": 45.83333333333333, + "NONE": 76.8321513002364 + } + }, + { + "task_index": 1, + "target": "death penalty", + "accuracy": 0.5813953488372093, + "macro_f1": 0.549122801488911, + "weighted_f1": 0.5981715214635795, + "macro_precision": 0.5631806108897742, + "macro_recall": 0.582280510720417, + "per_class_f1": { + "FAVOR": 0.42585551330798477, + "AGAINST": 0.5507246376811594, + "NONE": 0.6707882534775889 + }, + "accuracy_pct": 58.139534883720934, + "macro_f1_pct": 54.9122801488911, + "weighted_f1_pct": 59.817152146357955, + "macro_precision_pct": 56.31806108897742, + "macro_recall_pct": 58.228051072041694, + "per_class_f1_pct": { + "FAVOR": 42.585551330798474, + "AGAINST": 55.072463768115945, + "NONE": 67.07882534775888 + } + }, + { + "task_index": 2, + "target": "marijuana legalization", + "accuracy": 0.716297786720322, + "macro_f1": 0.6700716758231308, + "weighted_f1": 0.7140202129619635, + "macro_precision": 0.678380221653879, + "macro_recall": 0.6706699234333723, + "per_class_f1": { + "FAVOR": 0.5523809523809524, + "AGAINST": 0.6231884057971014, + "NONE": 0.8346456692913385 + }, + "accuracy_pct": 71.6297786720322, + "macro_f1_pct": 67.00716758231307, + "weighted_f1_pct": 71.40202129619635, + "macro_precision_pct": 67.8380221653879, + "macro_recall_pct": 67.06699234333723, + "per_class_f1_pct": { + "FAVOR": 55.23809523809524, + "AGAINST": 62.31884057971014, + "NONE": 83.46456692913385 + } + }, + { + "task_index": 3, + "target": "nuclear energy", + "accuracy": 0.6847977684797768, + "macro_f1": 0.6466862025461412, + "weighted_f1": 0.6977892409902622, + "macro_precision": 0.6461160691666847, + "macro_recall": 0.6881824948493515, + "per_class_f1": { + "FAVOR": 0.5338345864661654, + "AGAINST": 0.636568848758465, + "NONE": 0.7696551724137931 + }, + "accuracy_pct": 68.47977684797767, + "macro_f1_pct": 64.66862025461411, + "weighted_f1_pct": 69.77892409902621, + "macro_precision_pct": 64.61160691666848, + "macro_recall_pct": 68.81824948493515, + "per_class_f1_pct": { + "FAVOR": 53.383458646616546, + "AGAINST": 63.6568848758465, + "NONE": 76.96551724137932 + } + }, + { + "task_index": 4, + "target": "cloning", + "accuracy": 0.6896551724137931, + "macro_f1": 0.6768535934053176, + "weighted_f1": 0.6949662413158252, + "macro_precision": 0.6793693757392099, + "macro_recall": 0.6968793436950204, + "per_class_f1": { + "FAVOR": 0.6066666666666667, + "AGAINST": 0.6767676767676768, + "NONE": 0.7471264367816092 + }, + "accuracy_pct": 68.96551724137932, + "macro_f1_pct": 67.68535934053176, + "weighted_f1_pct": 69.49662413158252, + "macro_precision_pct": 67.93693757392099, + "macro_recall_pct": 69.68793436950203, + "per_class_f1_pct": { + "FAVOR": 60.66666666666667, + "AGAINST": 67.67676767676768, + "NONE": 74.71264367816092 + } + }, + { + "task_index": 5, + "target": "gun control", + "accuracy": 0.6158445440956651, + "macro_f1": 0.5819464960208162, + "weighted_f1": 0.6368768823908122, + "macro_precision": 0.5979635388632777, + "macro_recall": 0.6178344360667256, + "per_class_f1": { + "FAVOR": 0.5672371638141809, + "AGAINST": 0.44516129032258067, + "NONE": 0.7334410339256866 + }, + "accuracy_pct": 61.58445440956651, + "macro_f1_pct": 58.19464960208162, + "weighted_f1_pct": 63.68768823908122, + "macro_precision_pct": 59.79635388632777, + "macro_recall_pct": 61.78344360667256, + "per_class_f1_pct": { + "FAVOR": 56.723716381418086, + "AGAINST": 44.516129032258064, + "NONE": 73.34410339256866 + } + }, + { + "task_index": 6, + "target": "minimum wage", + "accuracy": 0.7283702213279678, + "macro_f1": 0.7062754692559148, + "weighted_f1": 0.7387901109222667, + "macro_precision": 0.706749032069545, + "macro_recall": 0.7273146709928319, + "per_class_f1": { + "FAVOR": 0.6175438596491228, + "AGAINST": 0.6905829596412556, + "NONE": 0.8106995884773662 + }, + "accuracy_pct": 72.83702213279678, + "macro_f1_pct": 70.62754692559147, + "weighted_f1_pct": 73.87901109222666, + "macro_precision_pct": 70.6749032069545, + "macro_recall_pct": 72.7314670992832, + "per_class_f1_pct": { + "FAVOR": 61.75438596491228, + "AGAINST": 69.05829596412556, + "NONE": 81.06995884773663 + } + }, + { + "task_index": 7, + "target": "school uniforms", + "accuracy": 0.7508305647840532, + "macro_f1": 0.7152910321439679, + "weighted_f1": 0.7578587159325979, + "macro_precision": 0.7023505547263792, + "macro_recall": 0.7382135148372551, + "per_class_f1": { + "FAVOR": 0.6611570247933884, + "AGAINST": 0.6518987341772152, + "NONE": 0.8328173374613003 + }, + "accuracy_pct": 75.08305647840533, + "macro_f1_pct": 71.5291032143968, + "weighted_f1_pct": 75.78587159325978, + "macro_precision_pct": 70.23505547263792, + "macro_recall_pct": 73.82135148372551, + "per_class_f1_pct": { + "FAVOR": 66.11570247933885, + "AGAINST": 65.18987341772153, + "NONE": 83.28173374613003 + } + } + ], + "accuracy": 0.6658837345860247, + "macro_f1": 0.6328978375046277, + "weighted_f1": 0.6796255736276681, + "macro_precision": 0.6307212593460513, + "macro_recall": 0.6595186280526214, + "per_class_f1": { + "FAVOR": 0.5460149713782475, + "AGAINST": 0.5841248303934871, + "NONE": 0.7685537107421484 + }, + "accuracy_pct": 66.58837345860246, + "macro_f1_pct": 63.28978375046277, + "weighted_f1_pct": 67.9625573627668, + "macro_precision_pct": 63.072125934605126, + "macro_recall_pct": 65.95186280526214, + "per_class_f1_pct": { + "FAVOR": 54.60149713782475, + "AGAINST": 58.41248303934871, + "NONE": 76.85537107421484 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__doc__s3__seed13/resolved_config.yaml b/outputs/trace_s8_full__doc__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..790968f50769217f97405083f54bddd8b0a0308c --- /dev/null +++ b/outputs/trace_s8_full__doc__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: doc + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__doc__s3__seed13/run.log b/outputs/trace_s8_full__doc__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..5283881a2cfd82cc57ccb53c07a0e2ddbbe6a604 --- /dev/null +++ b/outputs/trace_s8_full__doc__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 09:52:45,869 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__doc__s3__seed13 device=cuda +2026-09-02 10:23:24,173 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6362834193247375, 'average_forgetting': 0.014767535044749718, 'backward_transfer': -0.009418554045379455, 'final_mean_target_macro_f1_pct': 63.62834193247375, 'average_forgetting_pct': 1.4767535044749718, 'backward_transfer_pct': -0.9418554045379455, 'final_pooled_accuracy': 0.6658837345860247, 'final_pooled_accuracy_pct': 66.58837345860246, 'final_pooled_macro_f1': 0.6328978375046277, 'final_pooled_macro_f1_pct': 63.28978375046277, 'final_pooled_weighted_f1': 0.6796255736276681, 'final_pooled_weighted_f1_pct': 67.9625573627668, 'final_pooled_macro_precision': 0.6307212593460513, 'final_pooled_macro_precision_pct': 63.072125934605126, 'final_pooled_macro_recall': 0.6595186280526214, 'final_pooled_macro_recall_pct': 65.95186280526214, 'final_pooled_per_class_f1': {'FAVOR': 0.5460149713782475, 'AGAINST': 0.5841248303934871, 'NONE': 0.7685537107421484}, 'final_pooled_per_class_f1_pct': {'FAVOR': 54.60149713782475, 'AGAINST': 58.41248303934871, 'NONE': 76.85537107421484}} +2026-09-02 10:23:24,174 [INFO] __main__: efficiency={'training_wall_time_seconds': 1836.928849024931, 'training_wall_time_minutes': 30.615480817082183, 'peak_gpu_memory_allocated_gb': 10.430231094360352, 'peak_gpu_memory_reserved_gb': 10.6015625} diff --git a/outputs/trace_s8_full__x__s1__seed13/environment.json b/outputs/trace_s8_full__x__s1__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..790512e9944fd470c78efe5a81caf036da2e1965 --- /dev/null +++ b/outputs/trace_s8_full__x__s1__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T10:23:32.119222+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s1__seed13/metrics.json b/outputs/trace_s8_full__x__s1__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..999e412dc35572dfd49c6746f7b5a308f0cb4cff --- /dev/null +++ b/outputs/trace_s8_full__x__s1__seed13/metrics.json @@ -0,0 +1,2354 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s1", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7318205551557231, + 0.5648148148148148, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7173447356166123, + 0.6595375092107117, + 0.5578158279250632, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7174825174825173, + 0.717388660890713, + 0.5689905591866377, + 0.7085427062060318, + null, + null, + null, + null, + null, + null + ], + [ + 0.7194556738076251, + 0.6523679988840064, + 0.572389678924913, + 0.6996376811594204, + 0.7034933602264775, + null, + null, + null, + null, + null + ], + [ + 0.7074314574314574, + 0.6622805423285231, + 0.5641345427059713, + 0.6922274822859618, + 0.695649895178197, + 0.6611111111111111, + null, + null, + null, + null + ], + [ + 0.6949230057355615, + 0.682268668432931, + 0.59973604826546, + 0.6696580621451088, + 0.7177684013127051, + 0.6681646944112041, + 0.6506361730891784, + null, + null, + null + ], + [ + 0.6757990867579907, + 0.7210338040222507, + 0.6069831441348469, + 0.6681915943753309, + 0.7123694207027541, + 0.5987664340204512, + 0.650698440701381, + 0.7214211616301788, + null, + null + ], + [ + 0.7036659589525831, + 0.6841070494786874, + 0.5909484650060123, + 0.7091763088555391, + 0.7365091744112995, + 0.548970554405337, + 0.6873661113537572, + 0.6804738543823481, + 0.8225310317859337, + null + ], + [ + 0.661016059068805, + 0.6788232167181634, + 0.61440535574215, + 0.6747763347763348, + 0.7198731381138289, + 0.5931882019211573, + 0.6613453413493912, + 0.6709048527230346, + 0.7880539247662535, + 0.5996803513632235 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 0.6862624481284426, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10795797371276879, + "ema_update_magnitude": 0.06257687394518485, + "final_multiplier": 0.5796410565436867 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6971104775982825, + "losses": { + "current_ce": 0.7222974632467543, + "replay_ce": 0.26207199749492466, + "dark_distillation": 0.38265881666115353, + "supervised_contrastive": 4.120718887874058, + "total": 1.4649224775178091, + "replay_multiplier": 0.7074943545546575, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10556911174169166, + "ema_update_magnitude": 0.0666741709643145, + "final_multiplier": 0.6315689301947904 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7318205551557231, + "weighted_f1": 0.7633380559323518, + "macro_precision": 0.709848484848485, + "macro_recall": 0.8052723638180909, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7974683544303798, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 73.1820555155723, + "weighted_f1_pct": 76.33380559323518, + "macro_precision_pct": 70.9848484848485, + "macro_recall_pct": 80.52723638180909, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 79.74683544303798, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.5648148148148148, + "weighted_f1": 0.809095378564405, + "macro_precision": 0.570010055304173, + "macro_recall": 0.5600048537798811, + "per_class_f1": { + "FAVOR": 0.8444444444444444, + "AGAINST": 0.0, + "NONE": 0.85 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 56.481481481481474, + "weighted_f1_pct": 80.90953785644051, + "macro_precision_pct": 57.0010055304173, + "macro_recall_pct": 56.000485377988106, + "per_class_f1_pct": { + "FAVOR": 84.44444444444444, + "AGAINST": 0.0, + "NONE": 85.0 + } + } + ], + "accuracy": 0.7760617760617761, + "macro_f1": 0.7781180254954133, + "weighted_f1": 0.7740346056711571, + "macro_precision": 0.7867049538990467, + "macro_recall": 0.7859565430833921, + "per_class_f1": { + "FAVOR": 0.77, + "AGAINST": 0.7455621301775148, + "NONE": 0.8187919463087249 + }, + "accuracy_pct": 77.60617760617761, + "macro_f1_pct": 77.81180254954133, + "weighted_f1_pct": 77.40346056711572, + "macro_precision_pct": 78.67049538990467, + "macro_recall_pct": 78.59565430833922, + "per_class_f1_pct": { + "FAVOR": 77.0, + "AGAINST": 74.55621301775149, + "NONE": 81.87919463087249 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6345510835913313, + "losses": { + "current_ce": 0.6519526062589703, + "replay_ce": 0.2176734558109081, + "dark_distillation": 0.3667075966343735, + "supervised_contrastive": 4.30255932374434, + "total": 1.328178515217521, + "replay_multiplier": 0.5985787495634024, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10284690818145112, + "ema_update_magnitude": 0.05955740350883068, + "final_multiplier": 0.5790879333363579 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7173447356166123, + "weighted_f1": 0.7500929344853924, + "macro_precision": 0.6964646464646466, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7848101265822784, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.73447356166123, + "weighted_f1_pct": 75.00929344853924, + "macro_precision_pct": 69.64646464646465, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.48101265822784, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.6595375092107117, + "weighted_f1": 0.8054954024067332, + "macro_precision": 0.6628510378510378, + "macro_recall": 0.7900254823443756, + "per_class_f1": { + "FAVOR": 0.8067226890756303, + "AGAINST": 0.3076923076923077, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 65.95375092107118, + "weighted_f1_pct": 80.54954024067332, + "macro_precision_pct": 66.28510378510379, + "macro_recall_pct": 79.00254823443757, + "per_class_f1_pct": { + "FAVOR": 80.67226890756302, + "AGAINST": 30.76923076923077, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5578158279250632, + "weighted_f1": 0.5860789465765162, + "macro_precision": 0.5510912698412699, + "macro_recall": 0.5689905591866377, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.6565656565656566, + "NONE": 0.5263157894736842 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 55.78158279250632, + "weighted_f1_pct": 58.60789465765161, + "macro_precision_pct": 55.10912698412699, + "macro_recall_pct": 56.89905591866376, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 65.65656565656566, + "NONE": 52.63157894736842 + } + } + ], + "accuracy": 0.6815144766146993, + "macro_f1": 0.6843023005015123, + "weighted_f1": 0.6806176321910506, + "macro_precision": 0.6779371584699453, + "macro_recall": 0.6945797657059445, + "per_class_f1": { + "FAVOR": 0.6413793103448275, + "AGAINST": 0.6858638743455497, + "NONE": 0.7256637168141593 + }, + "accuracy_pct": 68.15144766146993, + "macro_f1_pct": 68.43023005015124, + "weighted_f1_pct": 68.06176321910506, + "macro_precision_pct": 67.79371584699453, + "macro_recall_pct": 69.45797657059445, + "per_class_f1_pct": { + "FAVOR": 64.13793103448275, + "AGAINST": 68.58638743455498, + "NONE": 72.56637168141593 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.6685871632680144, + "losses": { + "current_ce": 0.6198929184313976, + "replay_ce": 0.1971948093085578, + "dark_distillation": 0.3955822670098507, + "supervised_contrastive": 4.200918235200824, + "total": 1.3060550274270954, + "replay_multiplier": 0.6596039887383621, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09310331615171555, + "ema_update_magnitude": 0.055873039363045274, + "final_multiplier": 0.6001186818308162 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7174825174825173, + "weighted_f1": 0.7540952198486445, + "macro_precision": 0.7036974474474474, + "macro_recall": 0.7919940029985008, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7948717948717948, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.74825174825173, + "weighted_f1_pct": 75.40952198486444, + "macro_precision_pct": 70.36974474474475, + "macro_recall_pct": 79.19940029985007, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 79.48717948717949, + "NONE": 75.75757575757575 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.717388660890713, + "weighted_f1": 0.8753342291845735, + "macro_precision": 0.7028535980148884, + "macro_recall": 0.7812401407596165, + "per_class_f1": { + "FAVOR": 0.8992248062015504, + "AGAINST": 0.35294117647058826, + "NONE": 0.9 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 71.7388660890713, + "weighted_f1_pct": 87.53342291845735, + "macro_precision_pct": 70.28535980148885, + "macro_recall_pct": 78.12401407596164, + "per_class_f1_pct": { + "FAVOR": 89.92248062015504, + "AGAINST": 35.294117647058826, + "NONE": 90.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5689905591866377, + "weighted_f1": 0.5947368421052631, + "macro_precision": 0.5689905591866377, + "macro_recall": 0.5689905591866377, + "per_class_f1": { + "FAVOR": 0.48148148148148145, + "AGAINST": 0.6666666666666666, + "NONE": 0.5588235294117647 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 56.89905591866376, + "weighted_f1_pct": 59.473684210526315, + "macro_precision_pct": 56.89905591866376, + "macro_recall_pct": 56.89905591866376, + "per_class_f1_pct": { + "FAVOR": 48.148148148148145, + "AGAINST": 66.66666666666666, + "NONE": 55.88235294117647 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7085427062060318, + "weighted_f1": 0.7447299235260708, + "macro_precision": 0.6988402647655582, + "macro_recall": 0.7212121212121213, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.8038277511961722, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 70.85427062060317, + "weighted_f1_pct": 74.47299235260708, + "macro_precision_pct": 69.88402647655582, + "macro_recall_pct": 72.12121212121212, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 80.38277511961722, + "NONE": 67.3913043478261 + } + } + ], + "accuracy": 0.7181102362204724, + "macro_f1": 0.7144510692660329, + "weighted_f1": 0.7192033907492809, + "macro_precision": 0.7093334092021798, + "macro_recall": 0.7222968775717269, + "per_class_f1": { + "FAVOR": 0.6772486772486772, + "AGAINST": 0.7406143344709898, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 71.81102362204724, + "macro_f1_pct": 71.44510692660329, + "weighted_f1_pct": 71.92033907492808, + "macro_precision_pct": 70.93334092021799, + "macro_recall_pct": 72.22968775717268, + "per_class_f1_pct": { + "FAVOR": 67.72486772486772, + "AGAINST": 74.06143344709898, + "NONE": 72.54901960784314 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.7052934323576526, + "losses": { + "current_ce": 0.5566724649884484, + "replay_ce": 0.2035828510017106, + "dark_distillation": 0.38346654014153914, + "supervised_contrastive": 4.292866823889993, + "total": 1.239621711138523, + "replay_multiplier": 0.6319208953582468, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09431737218637819, + "ema_update_magnitude": 0.05158091185728228, + "final_multiplier": 0.5468866515423536 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7194556738076251, + "weighted_f1": 0.751432549042211, + "macro_precision": 0.702208687092408, + "macro_recall": 0.7998650674662668, + "per_class_f1": { + "FAVOR": 0.6176470588235294, + "AGAINST": 0.7870967741935484, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.94556738076251, + "weighted_f1_pct": 75.1432549042211, + "macro_precision_pct": 70.2208687092408, + "macro_recall_pct": 79.98650674662669, + "per_class_f1_pct": { + "FAVOR": 61.76470588235294, + "AGAINST": 78.70967741935485, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6523679988840064, + "weighted_f1": 0.798995235423414, + "macro_precision": 0.6594622697563873, + "macro_recall": 0.7850503579662663, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.2962962962962963, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 65.23679988840064, + "weighted_f1_pct": 79.8995235423414, + "macro_precision_pct": 65.94622697563874, + "macro_recall_pct": 78.50503579662663, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 29.629629629629626, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.572389678924913, + "weighted_f1": 0.6059856266383189, + "macro_precision": 0.5862924112924114, + "macro_recall": 0.5675381263616558, + "per_class_f1": { + "FAVOR": 0.44680851063829785, + "AGAINST": 0.7031963470319634, + "NONE": 0.5671641791044776 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 57.2389678924913, + "weighted_f1_pct": 60.59856266383189, + "macro_precision_pct": 58.62924112924114, + "macro_recall_pct": 56.75381263616558, + "per_class_f1_pct": { + "FAVOR": 44.680851063829785, + "AGAINST": 70.31963470319634, + "NONE": 56.71641791044776 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.6996376811594204, + "weighted_f1": 0.7384467040673213, + "macro_precision": 0.687018193031144, + "macro_recall": 0.7290684624017958, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.8, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 69.96376811594203, + "weighted_f1_pct": 73.84467040673212, + "macro_precision_pct": 68.70181930311439, + "macro_recall_pct": 72.90684624017958, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 80.0, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7034933602264775, + "weighted_f1": 0.726458146063373, + "macro_precision": 0.6968222768222768, + "macro_recall": 0.7117978325682647, + "per_class_f1": { + "FAVOR": 0.6857142857142857, + "AGAINST": 0.7713004484304933, + "NONE": 0.6534653465346535 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 70.34933602264775, + "weighted_f1_pct": 72.6458146063373, + "macro_precision_pct": 69.68222768222768, + "macro_recall_pct": 71.17978325682647, + "per_class_f1_pct": { + "FAVOR": 68.57142857142857, + "AGAINST": 77.13004484304933, + "NONE": 65.34653465346535 + } + } + ], + "accuracy": 0.7091346153846154, + "macro_f1": 0.6972907683468345, + "weighted_f1": 0.7097870264506873, + "macro_precision": 0.6937060154246093, + "macro_recall": 0.7014255765199161, + "per_class_f1": { + "FAVOR": 0.641860465116279, + "AGAINST": 0.7475728155339806, + "NONE": 0.7024390243902439 + }, + "accuracy_pct": 70.91346153846155, + "macro_f1_pct": 69.72907683468344, + "weighted_f1_pct": 70.97870264506872, + "macro_precision_pct": 69.37060154246093, + "macro_recall_pct": 70.14255765199161, + "per_class_f1_pct": { + "FAVOR": 64.18604651162791, + "AGAINST": 74.75728155339806, + "NONE": 70.24390243902438 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.6939713816561107, + "losses": { + "current_ce": 0.6734056080381076, + "replay_ce": 0.21065963606039684, + "dark_distillation": 0.42934948652982713, + "supervised_contrastive": 4.346980994939804, + "total": 1.3927509864171346, + "replay_multiplier": 0.6496362633416287, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09979670806904502, + "ema_update_magnitude": 0.0536505922089572, + "final_multiplier": 0.5375988171056573 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7074314574314574, + "weighted_f1": 0.7355823400343949, + "macro_precision": 0.6891082526729244, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.75 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.74314574314575, + "weighted_f1_pct": 73.5582340034395, + "macro_precision_pct": 68.91082526729244, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6622805423285231, + "weighted_f1": 0.8263719451881356, + "macro_precision": 0.6605973653319421, + "macro_recall": 0.7464142701128503, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.2727272727272727, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 66.2280542328523, + "weighted_f1_pct": 82.63719451881356, + "macro_precision_pct": 66.05973653319421, + "macro_recall_pct": 74.64142701128503, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 27.27272727272727, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5641345427059713, + "weighted_f1": 0.6010824680749492, + "macro_precision": 0.5662177328843995, + "macro_recall": 0.5668119099491649, + "per_class_f1": { + "FAVOR": 0.46938775510204084, + "AGAINST": 0.6952380952380952, + "NONE": 0.5277777777777778 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 56.41345427059713, + "weighted_f1_pct": 60.10824680749492, + "macro_precision_pct": 56.62177328843995, + "macro_recall_pct": 56.681190994916484, + "per_class_f1_pct": { + "FAVOR": 46.93877551020408, + "AGAINST": 69.52380952380952, + "NONE": 52.77777777777778 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.6922274822859618, + "weighted_f1": 0.7287274338774056, + "macro_precision": 0.6786111111111112, + "macro_recall": 0.7272166105499439, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7857142857142857, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 69.22274822859617, + "weighted_f1_pct": 72.87274338774056, + "macro_precision_pct": 67.86111111111111, + "macro_recall_pct": 72.72166105499439, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 78.57142857142857, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.695649895178197, + "weighted_f1": 0.7131572646298246, + "macro_precision": 0.6797048499176159, + "macro_recall": 0.7279878430320909, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.7452830188679245, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.5649895178197, + "weighted_f1_pct": 71.31572646298247, + "macro_precision_pct": 67.9704849917616, + "macro_recall_pct": 72.79878430320909, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 74.52830188679245, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.647887323943662, + "macro_f1": 0.6611111111111111, + "weighted_f1": 0.6438631790744467, + "macro_precision": 0.6598290598290598, + "macro_recall": 0.6666666666666666, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.5535714285714286, + "NONE": 0.6964285714285714 + }, + "accuracy_pct": 64.7887323943662, + "macro_f1_pct": 66.11111111111111, + "weighted_f1_pct": 64.38631790744466, + "macro_precision_pct": 65.98290598290598, + "macro_recall_pct": 66.66666666666666, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 55.35714285714286, + "NONE": 69.64285714285714 + } + } + ], + "accuracy": 0.6950718685831622, + "macro_f1": 0.6908680303646048, + "weighted_f1": 0.6958267322241201, + "macro_precision": 0.6835360646369644, + "macro_recall": 0.7023705671432944, + "per_class_f1": { + "FAVOR": 0.6627218934911243, + "AGAINST": 0.7108167770419426, + "NONE": 0.6990654205607477 + }, + "accuracy_pct": 69.50718685831622, + "macro_f1_pct": 69.08680303646048, + "weighted_f1_pct": 69.58267322241201, + "macro_precision_pct": 68.35360646369644, + "macro_recall_pct": 70.23705671432944, + "per_class_f1_pct": { + "FAVOR": 66.27218934911244, + "AGAINST": 71.08167770419426, + "NONE": 69.90654205607477 + } + }, + { + "after_task": 7, + "best_epoch": 5, + "best_dev_macro_f1": 0.6603893320231972, + "losses": { + "current_ce": 0.6171669213544755, + "replay_ce": 0.19434612852240365, + "dark_distillation": 0.3801094494168721, + "supervised_contrastive": 4.302487322640798, + "total": 1.2824477903426639, + "replay_multiplier": 0.6029697367374827, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09155379554533578, + "ema_update_magnitude": 0.05274556423727405, + "final_multiplier": 0.5761155386633359 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6949230057355615, + "weighted_f1": 0.7276430068794605, + "macro_precision": 0.6873538011695907, + "macro_recall": 0.7775012493753124, + "per_class_f1": { + "FAVOR": 0.5753424657534246, + "AGAINST": 0.7631578947368421, + "NONE": 0.746268656716418 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.49230057355615, + "weighted_f1_pct": 72.76430068794605, + "macro_precision_pct": 68.73538011695906, + "macro_recall_pct": 77.75012493753124, + "per_class_f1_pct": { + "FAVOR": 57.534246575342465, + "AGAINST": 76.31578947368422, + "NONE": 74.6268656716418 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.682268668432931, + "weighted_f1": 0.8326512287868091, + "macro_precision": 0.6708676156893474, + "macro_recall": 0.7500546050236622, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.3333333333333333, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 68.2268668432931, + "weighted_f1_pct": 83.26512287868091, + "macro_precision_pct": 67.08676156893473, + "macro_recall_pct": 75.00546050236622, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 33.33333333333333, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.59973604826546, + "weighted_f1": 0.6347165991902833, + "macro_precision": 0.6021235664076952, + "macro_recall": 0.5980392156862745, + "per_class_f1": { + "FAVOR": 0.5192307692307693, + "AGAINST": 0.7211538461538461, + "NONE": 0.5588235294117647 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 59.973604826546, + "weighted_f1_pct": 63.47165991902833, + "macro_precision_pct": 60.212356640769514, + "macro_recall_pct": 59.80392156862745, + "per_class_f1_pct": { + "FAVOR": 51.92307692307693, + "AGAINST": 72.11538461538461, + "NONE": 55.88235294117647 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6696580621451088, + "weighted_f1": 0.7039838191225455, + "macro_precision": 0.6662531251474126, + "macro_recall": 0.7128507295173963, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.7564766839378239, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 66.96580621451088, + "weighted_f1_pct": 70.39838191225455, + "macro_precision_pct": 66.62531251474127, + "macro_recall_pct": 71.28507295173962, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 75.64766839378238, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7177684013127051, + "weighted_f1": 0.737468497219188, + "macro_precision": 0.7039098775854792, + "macro_recall": 0.7463021679575558, + "per_class_f1": { + "FAVOR": 0.7088607594936709, + "AGAINST": 0.7777777777777778, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.77684013127052, + "weighted_f1_pct": 73.74684972191879, + "macro_precision_pct": 70.39098775854792, + "macro_recall_pct": 74.63021679575557, + "per_class_f1_pct": { + "FAVOR": 70.88607594936708, + "AGAINST": 77.77777777777779, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6549295774647887, + "macro_f1": 0.6681646944112041, + "weighted_f1": 0.6527581925538071, + "macro_precision": 0.6666666666666666, + "macro_recall": 0.6713675213675213, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.5739130434782609, + "NONE": 0.6972477064220184 + }, + "accuracy_pct": 65.49295774647888, + "macro_f1_pct": 66.8164694411204, + "weighted_f1_pct": 65.27581925538071, + "macro_precision_pct": 66.66666666666666, + "macro_recall_pct": 67.13675213675214, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 57.391304347826086, + "NONE": 69.72477064220183 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6506361730891784, + "weighted_f1": 0.6274828017770117, + "macro_precision": 0.6414846380928914, + "macro_recall": 0.6731830851072044, + "per_class_f1": { + "FAVOR": 0.6309523809523809, + "AGAINST": 0.5467625899280576, + "NONE": 0.7741935483870968 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 65.06361730891784, + "weighted_f1_pct": 62.74828017770117, + "macro_precision_pct": 64.14846380928914, + "macro_recall_pct": 67.31830851072044, + "per_class_f1_pct": { + "FAVOR": 63.095238095238095, + "AGAINST": 54.67625899280576, + "NONE": 77.41935483870968 + } + } + ], + "accuracy": 0.6908006814310051, + "macro_f1": 0.6903479330344465, + "weighted_f1": 0.6916487168982494, + "macro_precision": 0.6851647751604681, + "macro_recall": 0.7013558016090901, + "per_class_f1": { + "FAVOR": 0.6619318181818182, + "AGAINST": 0.6993275696445725, + "NONE": 0.7097844112769486 + }, + "accuracy_pct": 69.08006814310052, + "macro_f1_pct": 69.03479330344466, + "weighted_f1_pct": 69.16487168982493, + "macro_precision_pct": 68.51647751604682, + "macro_recall_pct": 70.135580160909, + "per_class_f1_pct": { + "FAVOR": 66.19318181818183, + "AGAINST": 69.93275696445725, + "NONE": 70.97844112769486 + } + }, + { + "after_task": 8, + "best_epoch": 7, + "best_dev_macro_f1": 0.7320440442322124, + "losses": { + "current_ce": 0.5774710106244986, + "replay_ce": 0.23240115370439446, + "dark_distillation": 0.35538035242453864, + "supervised_contrastive": 4.3191340605417885, + "total": 1.2524419537488964, + "replay_multiplier": 0.5883996931333431, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09464730558582933, + "ema_update_magnitude": 0.05410676883019801, + "final_multiplier": 0.5716672914807087 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6757990867579907, + "weighted_f1": 0.6943141302308126, + "macro_precision": 0.6713664421997755, + "macro_recall": 0.7787506246876562, + "per_class_f1": { + "FAVOR": 0.5753424657534246, + "AGAINST": 0.7123287671232876, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 67.57990867579908, + "weighted_f1_pct": 69.43141302308126, + "macro_precision_pct": 67.13664421997755, + "macro_recall_pct": 77.87506246876562, + "per_class_f1_pct": { + "FAVOR": 57.534246575342465, + "AGAINST": 71.23287671232876, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7210338040222507, + "weighted_f1": 0.8494979040521659, + "macro_precision": 0.6982658496535704, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.42105263157894735, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.10338040222507, + "weighted_f1_pct": 84.94979040521659, + "macro_precision_pct": 69.82658496535704, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 42.10526315789473, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6069831441348469, + "weighted_f1": 0.6346260387811634, + "macro_precision": 0.6022611943328056, + "macro_recall": 0.6183732752360204, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7156862745098039, + "NONE": 0.6052631578947368 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 60.69831441348469, + "weighted_f1_pct": 63.46260387811634, + "macro_precision_pct": 60.22611943328056, + "macro_recall_pct": 61.837327523602035, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 71.56862745098039, + "NONE": 60.526315789473685 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6681915943753309, + "weighted_f1": 0.7025724588093856, + "macro_precision": 0.6562790697674418, + "macro_recall": 0.7047699214365881, + "per_class_f1": { + "FAVOR": 0.5783132530120482, + "AGAINST": 0.7525773195876289, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 66.81915943753309, + "weighted_f1_pct": 70.25724588093856, + "macro_precision_pct": 65.62790697674417, + "macro_recall_pct": 70.4769921436588, + "per_class_f1_pct": { + "FAVOR": 57.831325301204814, + "AGAINST": 75.25773195876289, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7123694207027541, + "weighted_f1": 0.7290717601631307, + "macro_precision": 0.6935672514619883, + "macro_recall": 0.7475956861693458, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.7596153846153846, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 71.2369420702754, + "weighted_f1_pct": 72.90717601631307, + "macro_precision_pct": 69.35672514619883, + "macro_recall_pct": 74.75956861693459, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 75.96153846153845, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.5987664340204512, + "weighted_f1": 0.5807498782665151, + "macro_precision": 0.5986774568206511, + "macro_recall": 0.62008547008547, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.43564356435643564, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 59.87664340204512, + "weighted_f1_pct": 58.074987826651515, + "macro_precision_pct": 59.86774568206511, + "macro_recall_pct": 62.008547008547, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 43.56435643564357, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.645, + "macro_f1": 0.650698440701381, + "weighted_f1": 0.631162515226614, + "macro_precision": 0.6554705215419502, + "macro_recall": 0.6806221481018229, + "per_class_f1": { + "FAVOR": 0.6703910614525139, + "AGAINST": 0.5238095238095238, + "NONE": 0.7578947368421053 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 65.0698440701381, + "weighted_f1_pct": 63.1162515226614, + "macro_precision_pct": 65.54705215419501, + "macro_recall_pct": 68.06221481018228, + "per_class_f1_pct": { + "FAVOR": 67.0391061452514, + "AGAINST": 52.38095238095239, + "NONE": 75.78947368421053 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7214211616301788, + "weighted_f1": 0.7242115089974672, + "macro_precision": 0.7219030969030968, + "macro_recall": 0.7308773555761507, + "per_class_f1": { + "FAVOR": 0.7288135593220338, + "AGAINST": 0.6776859504132231, + "NONE": 0.7577639751552795 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 72.14211616301787, + "weighted_f1_pct": 72.42115089974672, + "macro_precision_pct": 72.19030969030969, + "macro_recall_pct": 73.08773555761508, + "per_class_f1_pct": { + "FAVOR": 72.88135593220339, + "AGAINST": 67.76859504132231, + "NONE": 75.77639751552795 + } + } + ], + "accuracy": 0.6885007278020379, + "macro_f1": 0.6903878421028259, + "weighted_f1": 0.6874230482545486, + "macro_precision": 0.68553188400363, + "macro_recall": 0.7067925237829612, + "per_class_f1": { + "FAVOR": 0.6707466340269278, + "AGAINST": 0.6738159070598749, + "NONE": 0.7266009852216748 + }, + "accuracy_pct": 68.85007278020379, + "macro_f1_pct": 69.03878421028259, + "weighted_f1_pct": 68.74230482545485, + "macro_precision_pct": 68.553188400363, + "macro_recall_pct": 70.67925237829613, + "per_class_f1_pct": { + "FAVOR": 67.07466340269278, + "AGAINST": 67.38159070598749, + "NONE": 72.66009852216749 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.659242191500256, + "losses": { + "current_ce": 0.42716561698665223, + "replay_ce": 0.22762162868554395, + "dark_distillation": 0.38987442714472614, + "supervised_contrastive": 4.077124718825022, + "total": 1.0807434156537057, + "replay_multiplier": 0.5772462882993676, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10126776346655016, + "ema_update_magnitude": 0.04927772559227879, + "final_multiplier": 0.5 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7036659589525831, + "weighted_f1": 0.7326835936363901, + "macro_precision": 0.6820194298455169, + "macro_recall": 0.7792303848075962, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7643312101910829, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.36659589525831, + "weighted_f1_pct": 73.26835936363901, + "macro_precision_pct": 68.20194298455169, + "macro_recall_pct": 77.92303848075962, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.43312101910828, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6841070494786874, + "weighted_f1": 0.8239914270905213, + "macro_precision": 0.6753989762119844, + "macro_recall": 0.8081058124014077, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.34782608695652173, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.41070494786874, + "weighted_f1_pct": 82.39914270905213, + "macro_precision_pct": 67.53989762119843, + "macro_recall_pct": 80.81058124014076, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 34.78260869565217, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.5909484650060123, + "weighted_f1": 0.6234737867904104, + "macro_precision": 0.6006330131177132, + "macro_recall": 0.5867828612926652, + "per_class_f1": { + "FAVOR": 0.4536082474226804, + "AGAINST": 0.7222222222222222, + "NONE": 0.5970149253731343 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 59.09484650060123, + "weighted_f1_pct": 62.34737867904104, + "macro_precision_pct": 60.06330131177132, + "macro_recall_pct": 58.678286129266525, + "per_class_f1_pct": { + "FAVOR": 45.36082474226804, + "AGAINST": 72.22222222222221, + "NONE": 59.70149253731343 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7473118279569892, + "macro_f1": 0.7091763088555391, + "weighted_f1": 0.7482902745390781, + "macro_precision": 0.7073214736766139, + "macro_recall": 0.7129629629629629, + "per_class_f1": { + "FAVOR": 0.6285714285714286, + "AGAINST": 0.8093023255813954, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 74.73118279569893, + "macro_f1_pct": 70.9176308855539, + "weighted_f1_pct": 74.82902745390781, + "macro_precision_pct": 70.7321473676614, + "macro_recall_pct": 71.29629629629629, + "per_class_f1_pct": { + "FAVOR": 62.857142857142854, + "AGAINST": 80.93023255813954, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.751269035532995, + "macro_f1": 0.7365091744112995, + "weighted_f1": 0.7515402117747891, + "macro_precision": 0.7296296296296295, + "macro_recall": 0.7444355054974524, + "per_class_f1": { + "FAVOR": 0.7536231884057971, + "AGAINST": 0.7892376681614349, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 75.1269035532995, + "macro_f1_pct": 73.65091744112995, + "weighted_f1_pct": 75.15402117747891, + "macro_precision_pct": 72.96296296296295, + "macro_recall_pct": 74.44355054974524, + "per_class_f1_pct": { + "FAVOR": 75.36231884057972, + "AGAINST": 78.9237668161435, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.548970554405337, + "weighted_f1": 0.5576137987344355, + "macro_precision": 0.5963383838383839, + "macro_recall": 0.5371794871794872, + "per_class_f1": { + "FAVOR": 0.4782608695652174, + "AGAINST": 0.5079365079365079, + "NONE": 0.6607142857142857 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 54.897055440533705, + "weighted_f1_pct": 55.76137987344355, + "macro_precision_pct": 59.63383838383839, + "macro_recall_pct": 53.717948717948715, + "per_class_f1_pct": { + "FAVOR": 47.82608695652174, + "AGAINST": 50.79365079365079, + "NONE": 66.07142857142857 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6873661113537572, + "weighted_f1": 0.665527001310289, + "macro_precision": 0.7195447798119742, + "macro_recall": 0.7106370493904369, + "per_class_f1": { + "FAVOR": 0.5645161290322581, + "AGAINST": 0.7021276595744681, + "NONE": 0.7954545454545454 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.73661113537572, + "weighted_f1_pct": 66.5527001310289, + "macro_precision_pct": 71.95447798119741, + "macro_recall_pct": 71.06370493904369, + "per_class_f1_pct": { + "FAVOR": 56.451612903225815, + "AGAINST": 70.2127659574468, + "NONE": 79.54545454545455 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6804738543823481, + "weighted_f1": 0.6897157142026625, + "macro_precision": 0.6810708898944192, + "macro_recall": 0.679935125115848, + "per_class_f1": { + "FAVOR": 0.6407766990291263, + "AGAINST": 0.6461538461538462, + "NONE": 0.7544910179640718 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 68.04738543823481, + "weighted_f1_pct": 68.97157142026626, + "macro_precision_pct": 68.10708898944192, + "macro_recall_pct": 67.9935125115848, + "per_class_f1_pct": { + "FAVOR": 64.07766990291263, + "AGAINST": 64.61538461538461, + "NONE": 75.44910179640718 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8225310317859337, + "weighted_f1": 0.8369075113122173, + "macro_precision": 0.8172728252517983, + "macro_recall": 0.8314750032391212, + "per_class_f1": { + "FAVOR": 0.8148148148148148, + "AGAINST": 0.784, + "NONE": 0.8687782805429864 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 82.25310317859336, + "weighted_f1_pct": 83.69075113122173, + "macro_precision_pct": 81.72728252517983, + "macro_recall_pct": 83.14750032391211, + "per_class_f1_pct": { + "FAVOR": 81.48148148148148, + "AGAINST": 78.4, + "NONE": 86.87782805429865 + } + } + ], + "accuracy": 0.7141041931385006, + "macro_f1": 0.7070091169773877, + "weighted_f1": 0.7123846900309271, + "macro_precision": 0.7142662297298114, + "macro_recall": 0.7031488067199048, + "per_class_f1": { + "FAVOR": 0.6424731182795699, + "AGAINST": 0.7213114754098361, + "NONE": 0.7572427572427572 + }, + "accuracy_pct": 71.41041931385006, + "macro_f1_pct": 70.70091169773876, + "weighted_f1_pct": 71.2384690030927, + "macro_precision_pct": 71.42662297298114, + "macro_recall_pct": 70.31488067199048, + "per_class_f1_pct": { + "FAVOR": 64.24731182795699, + "AGAINST": 72.1311475409836, + "NONE": 75.72427572427573 + } + }, + { + "after_task": 10, + "best_epoch": 6, + "best_dev_macro_f1": 0.6450079996179287, + "losses": { + "current_ce": 0.6229372417315459, + "replay_ce": 0.22543158317223574, + "dark_distillation": 0.39491736659636867, + "supervised_contrastive": 4.231151305712187, + "total": 1.3111426961727632, + "replay_multiplier": 0.6252223476117874, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09525205826978211, + "ema_update_magnitude": 0.05314993873928781, + "final_multiplier": 0.5579925484523537 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.661016059068805, + "weighted_f1": 0.6904563663430738, + "macro_precision": 0.6540098983555956, + "macro_recall": 0.7515142428785607, + "per_class_f1": { + "FAVOR": 0.5915492957746479, + "AGAINST": 0.7248322147651006, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.10160590688051, + "weighted_f1_pct": 69.04563663430739, + "macro_precision_pct": 65.40098983555957, + "macro_recall_pct": 75.15142428785607, + "per_class_f1_pct": { + "FAVOR": 59.154929577464785, + "AGAINST": 72.48322147651007, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6788232167181634, + "weighted_f1": 0.8389866537770491, + "macro_precision": 0.6736842105263158, + "macro_recall": 0.7532095619463658, + "per_class_f1": { + "FAVOR": 0.8503937007874016, + "AGAINST": 0.3, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 67.88232167181634, + "weighted_f1_pct": 83.8986653777049, + "macro_precision_pct": 67.36842105263158, + "macro_recall_pct": 75.32095619463658, + "per_class_f1_pct": { + "FAVOR": 85.03937007874016, + "AGAINST": 30.0, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.61440535574215, + "weighted_f1": 0.639646121379787, + "macro_precision": 0.6144746646795828, + "macro_recall": 0.6161946259985477, + "per_class_f1": { + "FAVOR": 0.5391304347826087, + "AGAINST": 0.7070707070707071, + "NONE": 0.5970149253731343 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 61.440535574215005, + "weighted_f1_pct": 63.964612137978705, + "macro_precision_pct": 61.44746646795828, + "macro_recall_pct": 61.61946259985477, + "per_class_f1_pct": { + "FAVOR": 53.91304347826087, + "AGAINST": 70.70707070707071, + "NONE": 59.70149253731343 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6747763347763348, + "weighted_f1": 0.7121610110319787, + "macro_precision": 0.6671904016072219, + "macro_recall": 0.7049943883277217, + "per_class_f1": { + "FAVOR": 0.5952380952380952, + "AGAINST": 0.77, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 67.47763347763347, + "weighted_f1_pct": 71.21610110319787, + "macro_precision_pct": 66.71904016072219, + "macro_recall_pct": 70.49943883277217, + "per_class_f1_pct": { + "FAVOR": 59.523809523809526, + "AGAINST": 77.0, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7198731381138289, + "weighted_f1": 0.7374200913529374, + "macro_precision": 0.7068069504224453, + "macro_recall": 0.7391720431803721, + "per_class_f1": { + "FAVOR": 0.7027027027027027, + "AGAINST": 0.7706422018348624, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.98731381138289, + "weighted_f1_pct": 73.74200913529374, + "macro_precision_pct": 70.68069504224454, + "macro_recall_pct": 73.91720431803721, + "per_class_f1_pct": { + "FAVOR": 70.27027027027027, + "AGAINST": 77.06422018348624, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5931882019211573, + "weighted_f1": 0.5788938962689602, + "macro_precision": 0.6066666666666666, + "macro_recall": 0.5837606837606838, + "per_class_f1": { + "FAVOR": 0.6545454545454545, + "AGAINST": 0.5081967213114754, + "NONE": 0.616822429906542 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 59.31882019211573, + "weighted_f1_pct": 57.88938962689601, + "macro_precision_pct": 60.66666666666666, + "macro_recall_pct": 58.37606837606838, + "per_class_f1_pct": { + "FAVOR": 65.45454545454545, + "AGAINST": 50.81967213114754, + "NONE": 61.6822429906542 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6613453413493912, + "weighted_f1": 0.6495685213844574, + "macro_precision": 0.6625646513286964, + "macro_recall": 0.662705857827809, + "per_class_f1": { + "FAVOR": 0.6352941176470588, + "AGAINST": 0.6258503401360545, + "NONE": 0.7228915662650602 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 66.13453413493912, + "weighted_f1_pct": 64.95685213844574, + "macro_precision_pct": 66.25646513286964, + "macro_recall_pct": 66.2705857827809, + "per_class_f1_pct": { + "FAVOR": 63.52941176470588, + "AGAINST": 62.585034013605444, + "NONE": 72.28915662650603 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.675, + "macro_f1": 0.6709048527230346, + "weighted_f1": 0.67541428268701, + "macro_precision": 0.6743865016583519, + "macro_recall": 0.6825919060858819, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.628099173553719, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 67.09048527230345, + "weighted_f1_pct": 67.541428268701, + "macro_precision_pct": 67.43865016583518, + "macro_recall_pct": 68.2591906085882, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.8099173553719, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7880539247662535, + "weighted_f1": 0.8084787345075015, + "macro_precision": 0.7780362115153675, + "macro_recall": 0.8026893960427194, + "per_class_f1": { + "FAVOR": 0.7857142857142857, + "AGAINST": 0.72, + "NONE": 0.8584474885844748 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 78.80539247662534, + "weighted_f1_pct": 80.84787345075016, + "macro_precision_pct": 77.80362115153675, + "macro_recall_pct": 80.26893960427194, + "per_class_f1_pct": { + "FAVOR": 78.57142857142857, + "AGAINST": 72.0, + "NONE": 85.84474885844749 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.5996803513632235, + "weighted_f1": 0.62904337396484, + "macro_precision": 0.5995670995670995, + "macro_recall": 0.6082741422547248, + "per_class_f1": { + "FAVOR": 0.6701570680628273, + "AGAINST": 0.40816326530612246, + "NONE": 0.7207207207207207 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 59.96803513632235, + "weighted_f1_pct": 62.904337396484, + "macro_precision_pct": 59.95670995670995, + "macro_recall_pct": 60.827414225472474, + "per_class_f1_pct": { + "FAVOR": 67.01570680628272, + "AGAINST": 40.816326530612244, + "NONE": 72.07207207207207 + } + } + ], + "accuracy": 0.6877113866967306, + "macro_f1": 0.689357550558008, + "weighted_f1": 0.6878745717010384, + "macro_precision": 0.687991306018859, + "macro_recall": 0.6925093749628486, + "per_class_f1": { + "FAVOR": 0.6679174484052532, + "AGAINST": 0.6695278969957081, + "NONE": 0.7306273062730627 + }, + "accuracy_pct": 68.77113866967306, + "macro_f1_pct": 68.9357550558008, + "weighted_f1_pct": 68.78745717010383, + "macro_precision_pct": 68.7991306018859, + "macro_recall_pct": 69.25093749628486, + "per_class_f1_pct": { + "FAVOR": 66.79174484052533, + "AGAINST": 66.95278969957081, + "NONE": 73.06273062730627 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6662066776542341, + "average_forgetting": 0.03806884006462382, + "backward_transfer": -0.004807116965571111, + "final_mean_target_macro_f1_pct": 66.62066776542342, + "average_forgetting_pct": 3.806884006462382, + "backward_transfer_pct": -0.4807116965571111, + "final_pooled_accuracy": 0.6877113866967306, + "final_pooled_accuracy_pct": 68.77113866967306, + "final_pooled_macro_f1": 0.689357550558008, + "final_pooled_macro_f1_pct": 68.9357550558008, + "final_pooled_weighted_f1": 0.6878745717010384, + "final_pooled_weighted_f1_pct": 68.78745717010383, + "final_pooled_macro_precision": 0.687991306018859, + "final_pooled_macro_precision_pct": 68.7991306018859, + "final_pooled_macro_recall": 0.6925093749628486, + "final_pooled_macro_recall_pct": 69.25093749628486, + "final_pooled_per_class_f1": { + "FAVOR": 0.6679174484052532, + "AGAINST": 0.6695278969957081, + "NONE": 0.7306273062730627 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 66.79174484052533, + "AGAINST": 66.95278969957081, + "NONE": 73.06273062730627 + } + }, + "efficiency": { + "training_wall_time_seconds": 977.7541670140345, + "training_wall_time_minutes": 16.29590278356724, + "peak_gpu_memory_allocated_gb": 10.431715965270996, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s1__seed13/metrics.partial.json b/outputs/trace_s8_full__x__s1__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..b3c0ed215763c7a6c2ddf3feb46d9a60e8825a4b --- /dev/null +++ b/outputs/trace_s8_full__x__s1__seed13/metrics.partial.json @@ -0,0 +1,2238 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7318205551557231, + 0.5648148148148148, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7173447356166123, + 0.6595375092107117, + 0.5578158279250632, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7174825174825173, + 0.717388660890713, + 0.5689905591866377, + 0.7085427062060318, + null, + null, + null, + null, + null, + null + ], + [ + 0.7194556738076251, + 0.6523679988840064, + 0.572389678924913, + 0.6996376811594204, + 0.7034933602264775, + null, + null, + null, + null, + null + ], + [ + 0.7074314574314574, + 0.6622805423285231, + 0.5641345427059713, + 0.6922274822859618, + 0.695649895178197, + 0.6611111111111111, + null, + null, + null, + null + ], + [ + 0.6949230057355615, + 0.682268668432931, + 0.59973604826546, + 0.6696580621451088, + 0.7177684013127051, + 0.6681646944112041, + 0.6506361730891784, + null, + null, + null + ], + [ + 0.6757990867579907, + 0.7210338040222507, + 0.6069831441348469, + 0.6681915943753309, + 0.7123694207027541, + 0.5987664340204512, + 0.650698440701381, + 0.7214211616301788, + null, + null + ], + [ + 0.7036659589525831, + 0.6841070494786874, + 0.5909484650060123, + 0.7091763088555391, + 0.7365091744112995, + 0.548970554405337, + 0.6873661113537572, + 0.6804738543823481, + 0.8225310317859337, + null + ], + [ + 0.661016059068805, + 0.6788232167181634, + 0.61440535574215, + 0.6747763347763348, + 0.7198731381138289, + 0.5931882019211573, + 0.6613453413493912, + 0.6709048527230346, + 0.7880539247662535, + 0.5996803513632235 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 0.6862624481284426, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10795797371276879, + "ema_update_magnitude": 0.06257687394518485, + "final_multiplier": 0.5796410565436867 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 12, + "best_dev_macro_f1": 0.6971104775982825, + "losses": { + "current_ce": 0.7222974632467543, + "replay_ce": 0.26207199749492466, + "dark_distillation": 0.38265881666115353, + "supervised_contrastive": 4.120718887874058, + "total": 1.4649224775178091, + "replay_multiplier": 0.7074943545546575, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10556911174169166, + "ema_update_magnitude": 0.0666741709643145, + "final_multiplier": 0.6315689301947904 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7534246575342466, + "macro_f1": 0.7318205551557231, + "weighted_f1": 0.7633380559323518, + "macro_precision": 0.709848484848485, + "macro_recall": 0.8052723638180909, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7974683544303798, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 75.34246575342466, + "macro_f1_pct": 73.1820555155723, + "weighted_f1_pct": 76.33380559323518, + "macro_precision_pct": 70.9848484848485, + "macro_recall_pct": 80.52723638180909, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 79.74683544303798, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.5648148148148148, + "weighted_f1": 0.809095378564405, + "macro_precision": 0.570010055304173, + "macro_recall": 0.5600048537798811, + "per_class_f1": { + "FAVOR": 0.8444444444444444, + "AGAINST": 0.0, + "NONE": 0.85 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 56.481481481481474, + "weighted_f1_pct": 80.90953785644051, + "macro_precision_pct": 57.0010055304173, + "macro_recall_pct": 56.000485377988106, + "per_class_f1_pct": { + "FAVOR": 84.44444444444444, + "AGAINST": 0.0, + "NONE": 85.0 + } + } + ], + "accuracy": 0.7760617760617761, + "macro_f1": 0.7781180254954133, + "weighted_f1": 0.7740346056711571, + "macro_precision": 0.7867049538990467, + "macro_recall": 0.7859565430833921, + "per_class_f1": { + "FAVOR": 0.77, + "AGAINST": 0.7455621301775148, + "NONE": 0.8187919463087249 + }, + "accuracy_pct": 77.60617760617761, + "macro_f1_pct": 77.81180254954133, + "weighted_f1_pct": 77.40346056711572, + "macro_precision_pct": 78.67049538990467, + "macro_recall_pct": 78.59565430833922, + "per_class_f1_pct": { + "FAVOR": 77.0, + "AGAINST": 74.55621301775149, + "NONE": 81.87919463087249 + } + }, + { + "after_task": 3, + "best_epoch": 15, + "best_dev_macro_f1": 0.6345510835913313, + "losses": { + "current_ce": 0.6519526062589703, + "replay_ce": 0.2176734558109081, + "dark_distillation": 0.3667075966343735, + "supervised_contrastive": 4.30255932374434, + "total": 1.328178515217521, + "replay_multiplier": 0.5985787495634024, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10284690818145112, + "ema_update_magnitude": 0.05955740350883068, + "final_multiplier": 0.5790879333363579 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7173447356166123, + "weighted_f1": 0.7500929344853924, + "macro_precision": 0.6964646464646466, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7848101265822784, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.73447356166123, + "weighted_f1_pct": 75.00929344853924, + "macro_precision_pct": 69.64646464646465, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.48101265822784, + "NONE": 78.26086956521739 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7699115044247787, + "macro_f1": 0.6595375092107117, + "weighted_f1": 0.8054954024067332, + "macro_precision": 0.6628510378510378, + "macro_recall": 0.7900254823443756, + "per_class_f1": { + "FAVOR": 0.8067226890756303, + "AGAINST": 0.3076923076923077, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.99115044247787, + "macro_f1_pct": 65.95375092107118, + "weighted_f1_pct": 80.54954024067332, + "macro_precision_pct": 66.28510378510379, + "macro_recall_pct": 79.00254823443757, + "per_class_f1_pct": { + "FAVOR": 80.67226890756302, + "AGAINST": 30.76923076923077, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5578158279250632, + "weighted_f1": 0.5860789465765162, + "macro_precision": 0.5510912698412699, + "macro_recall": 0.5689905591866377, + "per_class_f1": { + "FAVOR": 0.49056603773584906, + "AGAINST": 0.6565656565656566, + "NONE": 0.5263157894736842 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 55.78158279250632, + "weighted_f1_pct": 58.60789465765161, + "macro_precision_pct": 55.10912698412699, + "macro_recall_pct": 56.89905591866376, + "per_class_f1_pct": { + "FAVOR": 49.056603773584904, + "AGAINST": 65.65656565656566, + "NONE": 52.63157894736842 + } + } + ], + "accuracy": 0.6815144766146993, + "macro_f1": 0.6843023005015123, + "weighted_f1": 0.6806176321910506, + "macro_precision": 0.6779371584699453, + "macro_recall": 0.6945797657059445, + "per_class_f1": { + "FAVOR": 0.6413793103448275, + "AGAINST": 0.6858638743455497, + "NONE": 0.7256637168141593 + }, + "accuracy_pct": 68.15144766146993, + "macro_f1_pct": 68.43023005015124, + "weighted_f1_pct": 68.06176321910506, + "macro_precision_pct": 67.79371584699453, + "macro_recall_pct": 69.45797657059445, + "per_class_f1_pct": { + "FAVOR": 64.13793103448275, + "AGAINST": 68.58638743455498, + "NONE": 72.56637168141593 + } + }, + { + "after_task": 4, + "best_epoch": 10, + "best_dev_macro_f1": 0.6685871632680144, + "losses": { + "current_ce": 0.6198929184313976, + "replay_ce": 0.1971948093085578, + "dark_distillation": 0.3955822670098507, + "supervised_contrastive": 4.200918235200824, + "total": 1.3060550274270954, + "replay_multiplier": 0.6596039887383621, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09310331615171555, + "ema_update_magnitude": 0.055873039363045274, + "final_multiplier": 0.6001186818308162 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7174825174825173, + "weighted_f1": 0.7540952198486445, + "macro_precision": 0.7036974474474474, + "macro_recall": 0.7919940029985008, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7948717948717948, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.74825174825173, + "weighted_f1_pct": 75.40952198486444, + "macro_precision_pct": 70.36974474474475, + "macro_recall_pct": 79.19940029985007, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 79.48717948717949, + "NONE": 75.75757575757575 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.717388660890713, + "weighted_f1": 0.8753342291845735, + "macro_precision": 0.7028535980148884, + "macro_recall": 0.7812401407596165, + "per_class_f1": { + "FAVOR": 0.8992248062015504, + "AGAINST": 0.35294117647058826, + "NONE": 0.9 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 71.7388660890713, + "weighted_f1_pct": 87.53342291845735, + "macro_precision_pct": 70.28535980148885, + "macro_recall_pct": 78.12401407596164, + "per_class_f1_pct": { + "FAVOR": 89.92248062015504, + "AGAINST": 35.294117647058826, + "NONE": 90.0 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5689905591866377, + "weighted_f1": 0.5947368421052631, + "macro_precision": 0.5689905591866377, + "macro_recall": 0.5689905591866377, + "per_class_f1": { + "FAVOR": 0.48148148148148145, + "AGAINST": 0.6666666666666666, + "NONE": 0.5588235294117647 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 56.89905591866376, + "weighted_f1_pct": 59.473684210526315, + "macro_precision_pct": 56.89905591866376, + "macro_recall_pct": 56.89905591866376, + "per_class_f1_pct": { + "FAVOR": 48.148148148148145, + "AGAINST": 66.66666666666666, + "NONE": 55.88235294117647 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7419354838709677, + "macro_f1": 0.7085427062060318, + "weighted_f1": 0.7447299235260708, + "macro_precision": 0.6988402647655582, + "macro_recall": 0.7212121212121213, + "per_class_f1": { + "FAVOR": 0.647887323943662, + "AGAINST": 0.8038277511961722, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 74.19354838709677, + "macro_f1_pct": 70.85427062060317, + "weighted_f1_pct": 74.47299235260708, + "macro_precision_pct": 69.88402647655582, + "macro_recall_pct": 72.12121212121212, + "per_class_f1_pct": { + "FAVOR": 64.7887323943662, + "AGAINST": 80.38277511961722, + "NONE": 67.3913043478261 + } + } + ], + "accuracy": 0.7181102362204724, + "macro_f1": 0.7144510692660329, + "weighted_f1": 0.7192033907492809, + "macro_precision": 0.7093334092021798, + "macro_recall": 0.7222968775717269, + "per_class_f1": { + "FAVOR": 0.6772486772486772, + "AGAINST": 0.7406143344709898, + "NONE": 0.7254901960784313 + }, + "accuracy_pct": 71.81102362204724, + "macro_f1_pct": 71.44510692660329, + "weighted_f1_pct": 71.92033907492808, + "macro_precision_pct": 70.93334092021799, + "macro_recall_pct": 72.22968775717268, + "per_class_f1_pct": { + "FAVOR": 67.72486772486772, + "AGAINST": 74.06143344709898, + "NONE": 72.54901960784314 + } + }, + { + "after_task": 5, + "best_epoch": 14, + "best_dev_macro_f1": 0.7052934323576526, + "losses": { + "current_ce": 0.5566724649884484, + "replay_ce": 0.2035828510017106, + "dark_distillation": 0.38346654014153914, + "supervised_contrastive": 4.292866823889993, + "total": 1.239621711138523, + "replay_multiplier": 0.6319208953582468, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09431737218637819, + "ema_update_magnitude": 0.05158091185728228, + "final_multiplier": 0.5468866515423536 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7194556738076251, + "weighted_f1": 0.751432549042211, + "macro_precision": 0.702208687092408, + "macro_recall": 0.7998650674662668, + "per_class_f1": { + "FAVOR": 0.6176470588235294, + "AGAINST": 0.7870967741935484, + "NONE": 0.7536231884057971 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.94556738076251, + "weighted_f1_pct": 75.1432549042211, + "macro_precision_pct": 70.2208687092408, + "macro_recall_pct": 79.98650674662669, + "per_class_f1_pct": { + "FAVOR": 61.76470588235294, + "AGAINST": 78.70967741935485, + "NONE": 75.36231884057972 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7610619469026548, + "macro_f1": 0.6523679988840064, + "weighted_f1": 0.798995235423414, + "macro_precision": 0.6594622697563873, + "macro_recall": 0.7850503579662663, + "per_class_f1": { + "FAVOR": 0.7966101694915254, + "AGAINST": 0.2962962962962963, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 76.10619469026548, + "macro_f1_pct": 65.23679988840064, + "weighted_f1_pct": 79.8995235423414, + "macro_precision_pct": 65.94622697563874, + "macro_recall_pct": 78.50503579662663, + "per_class_f1_pct": { + "FAVOR": 79.66101694915254, + "AGAINST": 29.629629629629626, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.572389678924913, + "weighted_f1": 0.6059856266383189, + "macro_precision": 0.5862924112924114, + "macro_recall": 0.5675381263616558, + "per_class_f1": { + "FAVOR": 0.44680851063829785, + "AGAINST": 0.7031963470319634, + "NONE": 0.5671641791044776 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 57.2389678924913, + "weighted_f1_pct": 60.59856266383189, + "macro_precision_pct": 58.62924112924114, + "macro_recall_pct": 56.75381263616558, + "per_class_f1_pct": { + "FAVOR": 44.680851063829785, + "AGAINST": 70.31963470319634, + "NONE": 56.71641791044776 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.6996376811594204, + "weighted_f1": 0.7384467040673213, + "macro_precision": 0.687018193031144, + "macro_recall": 0.7290684624017958, + "per_class_f1": { + "FAVOR": 0.625, + "AGAINST": 0.8, + "NONE": 0.6739130434782609 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 69.96376811594203, + "weighted_f1_pct": 73.84467040673212, + "macro_precision_pct": 68.70181930311439, + "macro_recall_pct": 72.90684624017958, + "per_class_f1_pct": { + "FAVOR": 62.5, + "AGAINST": 80.0, + "NONE": 67.3913043478261 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7034933602264775, + "weighted_f1": 0.726458146063373, + "macro_precision": 0.6968222768222768, + "macro_recall": 0.7117978325682647, + "per_class_f1": { + "FAVOR": 0.6857142857142857, + "AGAINST": 0.7713004484304933, + "NONE": 0.6534653465346535 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 70.34933602264775, + "weighted_f1_pct": 72.6458146063373, + "macro_precision_pct": 69.68222768222768, + "macro_recall_pct": 71.17978325682647, + "per_class_f1_pct": { + "FAVOR": 68.57142857142857, + "AGAINST": 77.13004484304933, + "NONE": 65.34653465346535 + } + } + ], + "accuracy": 0.7091346153846154, + "macro_f1": 0.6972907683468345, + "weighted_f1": 0.7097870264506873, + "macro_precision": 0.6937060154246093, + "macro_recall": 0.7014255765199161, + "per_class_f1": { + "FAVOR": 0.641860465116279, + "AGAINST": 0.7475728155339806, + "NONE": 0.7024390243902439 + }, + "accuracy_pct": 70.91346153846155, + "macro_f1_pct": 69.72907683468344, + "weighted_f1_pct": 70.97870264506872, + "macro_precision_pct": 69.37060154246093, + "macro_recall_pct": 70.14255765199161, + "per_class_f1_pct": { + "FAVOR": 64.18604651162791, + "AGAINST": 74.75728155339806, + "NONE": 70.24390243902438 + } + }, + { + "after_task": 6, + "best_epoch": 14, + "best_dev_macro_f1": 0.6939713816561107, + "losses": { + "current_ce": 0.6734056080381076, + "replay_ce": 0.21065963606039684, + "dark_distillation": 0.42934948652982713, + "supervised_contrastive": 4.346980994939804, + "total": 1.3927509864171346, + "replay_multiplier": 0.6496362633416287, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09979670806904502, + "ema_update_magnitude": 0.0536505922089572, + "final_multiplier": 0.5375988171056573 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7074314574314574, + "weighted_f1": 0.7355823400343949, + "macro_precision": 0.6891082526729244, + "macro_recall": 0.7907796101949026, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7662337662337663, + "NONE": 0.75 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.74314574314575, + "weighted_f1_pct": 73.5582340034395, + "macro_precision_pct": 68.91082526729244, + "macro_recall_pct": 79.07796101949026, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 76.62337662337663, + "NONE": 75.0 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6622805423285231, + "weighted_f1": 0.8263719451881356, + "macro_precision": 0.6605973653319421, + "macro_recall": 0.7464142701128503, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.2727272727272727, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 66.2280542328523, + "weighted_f1_pct": 82.63719451881356, + "macro_precision_pct": 66.05973653319421, + "macro_recall_pct": 74.64142701128503, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 27.27272727272727, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6052631578947368, + "macro_f1": 0.5641345427059713, + "weighted_f1": 0.6010824680749492, + "macro_precision": 0.5662177328843995, + "macro_recall": 0.5668119099491649, + "per_class_f1": { + "FAVOR": 0.46938775510204084, + "AGAINST": 0.6952380952380952, + "NONE": 0.5277777777777778 + }, + "accuracy_pct": 60.526315789473685, + "macro_f1_pct": 56.41345427059713, + "weighted_f1_pct": 60.10824680749492, + "macro_precision_pct": 56.62177328843995, + "macro_recall_pct": 56.681190994916484, + "per_class_f1_pct": { + "FAVOR": 46.93877551020408, + "AGAINST": 69.52380952380952, + "NONE": 52.77777777777778 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.6922274822859618, + "weighted_f1": 0.7287274338774056, + "macro_precision": 0.6786111111111112, + "macro_recall": 0.7272166105499439, + "per_class_f1": { + "FAVOR": 0.6172839506172839, + "AGAINST": 0.7857142857142857, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 69.22274822859617, + "weighted_f1_pct": 72.87274338774056, + "macro_precision_pct": 67.86111111111111, + "macro_recall_pct": 72.72166105499439, + "per_class_f1_pct": { + "FAVOR": 61.72839506172839, + "AGAINST": 78.57142857142857, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.695649895178197, + "weighted_f1": 0.7131572646298246, + "macro_precision": 0.6797048499176159, + "macro_recall": 0.7279878430320909, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.7452830188679245, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.5649895178197, + "weighted_f1_pct": 71.31572646298247, + "macro_precision_pct": 67.9704849917616, + "macro_recall_pct": 72.79878430320909, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 74.52830188679245, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.647887323943662, + "macro_f1": 0.6611111111111111, + "weighted_f1": 0.6438631790744467, + "macro_precision": 0.6598290598290598, + "macro_recall": 0.6666666666666666, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.5535714285714286, + "NONE": 0.6964285714285714 + }, + "accuracy_pct": 64.7887323943662, + "macro_f1_pct": 66.11111111111111, + "weighted_f1_pct": 64.38631790744466, + "macro_precision_pct": 65.98290598290598, + "macro_recall_pct": 66.66666666666666, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 55.35714285714286, + "NONE": 69.64285714285714 + } + } + ], + "accuracy": 0.6950718685831622, + "macro_f1": 0.6908680303646048, + "weighted_f1": 0.6958267322241201, + "macro_precision": 0.6835360646369644, + "macro_recall": 0.7023705671432944, + "per_class_f1": { + "FAVOR": 0.6627218934911243, + "AGAINST": 0.7108167770419426, + "NONE": 0.6990654205607477 + }, + "accuracy_pct": 69.50718685831622, + "macro_f1_pct": 69.08680303646048, + "weighted_f1_pct": 69.58267322241201, + "macro_precision_pct": 68.35360646369644, + "macro_recall_pct": 70.23705671432944, + "per_class_f1_pct": { + "FAVOR": 66.27218934911244, + "AGAINST": 71.08167770419426, + "NONE": 69.90654205607477 + } + }, + { + "after_task": 7, + "best_epoch": 5, + "best_dev_macro_f1": 0.6603893320231972, + "losses": { + "current_ce": 0.6171669213544755, + "replay_ce": 0.19434612852240365, + "dark_distillation": 0.3801094494168721, + "supervised_contrastive": 4.302487322640798, + "total": 1.2824477903426639, + "replay_multiplier": 0.6029697367374827, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09155379554533578, + "ema_update_magnitude": 0.05274556423727405, + "final_multiplier": 0.5761155386633359 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7123287671232876, + "macro_f1": 0.6949230057355615, + "weighted_f1": 0.7276430068794605, + "macro_precision": 0.6873538011695907, + "macro_recall": 0.7775012493753124, + "per_class_f1": { + "FAVOR": 0.5753424657534246, + "AGAINST": 0.7631578947368421, + "NONE": 0.746268656716418 + }, + "accuracy_pct": 71.23287671232876, + "macro_f1_pct": 69.49230057355615, + "weighted_f1_pct": 72.76430068794605, + "macro_precision_pct": 68.73538011695906, + "macro_recall_pct": 77.75012493753124, + "per_class_f1_pct": { + "FAVOR": 57.534246575342465, + "AGAINST": 76.31578947368422, + "NONE": 74.6268656716418 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.682268668432931, + "weighted_f1": 0.8326512287868091, + "macro_precision": 0.6708676156893474, + "macro_recall": 0.7500546050236622, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.3333333333333333, + "NONE": 0.8607594936708861 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 68.2268668432931, + "weighted_f1_pct": 83.26512287868091, + "macro_precision_pct": 67.08676156893473, + "macro_recall_pct": 75.00546050236622, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 33.33333333333333, + "NONE": 86.07594936708861 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.59973604826546, + "weighted_f1": 0.6347165991902833, + "macro_precision": 0.6021235664076952, + "macro_recall": 0.5980392156862745, + "per_class_f1": { + "FAVOR": 0.5192307692307693, + "AGAINST": 0.7211538461538461, + "NONE": 0.5588235294117647 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 59.973604826546, + "weighted_f1_pct": 63.47165991902833, + "macro_precision_pct": 60.212356640769514, + "macro_recall_pct": 59.80392156862745, + "per_class_f1_pct": { + "FAVOR": 51.92307692307693, + "AGAINST": 72.11538461538461, + "NONE": 55.88235294117647 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6696580621451088, + "weighted_f1": 0.7039838191225455, + "macro_precision": 0.6662531251474126, + "macro_recall": 0.7128507295173963, + "per_class_f1": { + "FAVOR": 0.5934065934065934, + "AGAINST": 0.7564766839378239, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 66.96580621451088, + "weighted_f1_pct": 70.39838191225455, + "macro_precision_pct": 66.62531251474127, + "macro_recall_pct": 71.28507295173962, + "per_class_f1_pct": { + "FAVOR": 59.34065934065934, + "AGAINST": 75.64766839378238, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7177684013127051, + "weighted_f1": 0.737468497219188, + "macro_precision": 0.7039098775854792, + "macro_recall": 0.7463021679575558, + "per_class_f1": { + "FAVOR": 0.7088607594936709, + "AGAINST": 0.7777777777777778, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.77684013127052, + "weighted_f1_pct": 73.74684972191879, + "macro_precision_pct": 70.39098775854792, + "macro_recall_pct": 74.63021679575557, + "per_class_f1_pct": { + "FAVOR": 70.88607594936708, + "AGAINST": 77.77777777777779, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.6549295774647887, + "macro_f1": 0.6681646944112041, + "weighted_f1": 0.6527581925538071, + "macro_precision": 0.6666666666666666, + "macro_recall": 0.6713675213675213, + "per_class_f1": { + "FAVOR": 0.7333333333333333, + "AGAINST": 0.5739130434782609, + "NONE": 0.6972477064220184 + }, + "accuracy_pct": 65.49295774647888, + "macro_f1_pct": 66.8164694411204, + "weighted_f1_pct": 65.27581925538071, + "macro_precision_pct": 66.66666666666666, + "macro_recall_pct": 67.13675213675214, + "per_class_f1_pct": { + "FAVOR": 73.33333333333333, + "AGAINST": 57.391304347826086, + "NONE": 69.72477064220183 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.635, + "macro_f1": 0.6506361730891784, + "weighted_f1": 0.6274828017770117, + "macro_precision": 0.6414846380928914, + "macro_recall": 0.6731830851072044, + "per_class_f1": { + "FAVOR": 0.6309523809523809, + "AGAINST": 0.5467625899280576, + "NONE": 0.7741935483870968 + }, + "accuracy_pct": 63.5, + "macro_f1_pct": 65.06361730891784, + "weighted_f1_pct": 62.74828017770117, + "macro_precision_pct": 64.14846380928914, + "macro_recall_pct": 67.31830851072044, + "per_class_f1_pct": { + "FAVOR": 63.095238095238095, + "AGAINST": 54.67625899280576, + "NONE": 77.41935483870968 + } + } + ], + "accuracy": 0.6908006814310051, + "macro_f1": 0.6903479330344465, + "weighted_f1": 0.6916487168982494, + "macro_precision": 0.6851647751604681, + "macro_recall": 0.7013558016090901, + "per_class_f1": { + "FAVOR": 0.6619318181818182, + "AGAINST": 0.6993275696445725, + "NONE": 0.7097844112769486 + }, + "accuracy_pct": 69.08006814310052, + "macro_f1_pct": 69.03479330344466, + "weighted_f1_pct": 69.16487168982493, + "macro_precision_pct": 68.51647751604682, + "macro_recall_pct": 70.135580160909, + "per_class_f1_pct": { + "FAVOR": 66.19318181818183, + "AGAINST": 69.93275696445725, + "NONE": 70.97844112769486 + } + }, + { + "after_task": 8, + "best_epoch": 7, + "best_dev_macro_f1": 0.7320440442322124, + "losses": { + "current_ce": 0.5774710106244986, + "replay_ce": 0.23240115370439446, + "dark_distillation": 0.35538035242453864, + "supervised_contrastive": 4.3191340605417885, + "total": 1.2524419537488964, + "replay_multiplier": 0.5883996931333431, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09464730558582933, + "ema_update_magnitude": 0.05410676883019801, + "final_multiplier": 0.5716672914807087 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.684931506849315, + "macro_f1": 0.6757990867579907, + "weighted_f1": 0.6943141302308126, + "macro_precision": 0.6713664421997755, + "macro_recall": 0.7787506246876562, + "per_class_f1": { + "FAVOR": 0.5753424657534246, + "AGAINST": 0.7123287671232876, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 68.4931506849315, + "macro_f1_pct": 67.57990867579908, + "weighted_f1_pct": 69.43141302308126, + "macro_precision_pct": 67.13664421997755, + "macro_recall_pct": 77.87506246876562, + "per_class_f1_pct": { + "FAVOR": 57.534246575342465, + "AGAINST": 71.23287671232876, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7210338040222507, + "weighted_f1": 0.8494979040521659, + "macro_precision": 0.6982658496535704, + "macro_recall": 0.8280063099138454, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.42105263157894735, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 72.10338040222507, + "weighted_f1_pct": 84.94979040521659, + "macro_precision_pct": 69.82658496535704, + "macro_recall_pct": 82.80063099138454, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 42.10526315789473, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.6069831441348469, + "weighted_f1": 0.6346260387811634, + "macro_precision": 0.6022611943328056, + "macro_recall": 0.6183732752360204, + "per_class_f1": { + "FAVOR": 0.5, + "AGAINST": 0.7156862745098039, + "NONE": 0.6052631578947368 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 60.69831441348469, + "weighted_f1_pct": 63.46260387811634, + "macro_precision_pct": 60.22611943328056, + "macro_recall_pct": 61.837327523602035, + "per_class_f1_pct": { + "FAVOR": 50.0, + "AGAINST": 71.56862745098039, + "NONE": 60.526315789473685 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.6681915943753309, + "weighted_f1": 0.7025724588093856, + "macro_precision": 0.6562790697674418, + "macro_recall": 0.7047699214365881, + "per_class_f1": { + "FAVOR": 0.5783132530120482, + "AGAINST": 0.7525773195876289, + "NONE": 0.6736842105263158 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 66.81915943753309, + "weighted_f1_pct": 70.25724588093856, + "macro_precision_pct": 65.62790697674417, + "macro_recall_pct": 70.4769921436588, + "per_class_f1_pct": { + "FAVOR": 57.831325301204814, + "AGAINST": 75.25773195876289, + "NONE": 67.36842105263158 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7258883248730964, + "macro_f1": 0.7123694207027541, + "weighted_f1": 0.7290717601631307, + "macro_precision": 0.6935672514619883, + "macro_recall": 0.7475956861693458, + "per_class_f1": { + "FAVOR": 0.6923076923076923, + "AGAINST": 0.7596153846153846, + "NONE": 0.6851851851851852 + }, + "accuracy_pct": 72.58883248730965, + "macro_f1_pct": 71.2369420702754, + "weighted_f1_pct": 72.90717601631307, + "macro_precision_pct": 69.35672514619883, + "macro_recall_pct": 74.75956861693459, + "per_class_f1_pct": { + "FAVOR": 69.23076923076923, + "AGAINST": 75.96153846153845, + "NONE": 68.51851851851852 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.5987664340204512, + "weighted_f1": 0.5807498782665151, + "macro_precision": 0.5986774568206511, + "macro_recall": 0.62008547008547, + "per_class_f1": { + "FAVOR": 0.6557377049180327, + "AGAINST": 0.43564356435643564, + "NONE": 0.7049180327868853 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 59.87664340204512, + "weighted_f1_pct": 58.074987826651515, + "macro_precision_pct": 59.86774568206511, + "macro_recall_pct": 62.008547008547, + "per_class_f1_pct": { + "FAVOR": 65.57377049180327, + "AGAINST": 43.56435643564357, + "NONE": 70.49180327868852 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.645, + "macro_f1": 0.650698440701381, + "weighted_f1": 0.631162515226614, + "macro_precision": 0.6554705215419502, + "macro_recall": 0.6806221481018229, + "per_class_f1": { + "FAVOR": 0.6703910614525139, + "AGAINST": 0.5238095238095238, + "NONE": 0.7578947368421053 + }, + "accuracy_pct": 64.5, + "macro_f1_pct": 65.0698440701381, + "weighted_f1_pct": 63.1162515226614, + "macro_precision_pct": 65.54705215419501, + "macro_recall_pct": 68.06221481018228, + "per_class_f1_pct": { + "FAVOR": 67.0391061452514, + "AGAINST": 52.38095238095239, + "NONE": 75.78947368421053 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.725, + "macro_f1": 0.7214211616301788, + "weighted_f1": 0.7242115089974672, + "macro_precision": 0.7219030969030968, + "macro_recall": 0.7308773555761507, + "per_class_f1": { + "FAVOR": 0.7288135593220338, + "AGAINST": 0.6776859504132231, + "NONE": 0.7577639751552795 + }, + "accuracy_pct": 72.5, + "macro_f1_pct": 72.14211616301787, + "weighted_f1_pct": 72.42115089974672, + "macro_precision_pct": 72.19030969030969, + "macro_recall_pct": 73.08773555761508, + "per_class_f1_pct": { + "FAVOR": 72.88135593220339, + "AGAINST": 67.76859504132231, + "NONE": 75.77639751552795 + } + } + ], + "accuracy": 0.6885007278020379, + "macro_f1": 0.6903878421028259, + "weighted_f1": 0.6874230482545486, + "macro_precision": 0.68553188400363, + "macro_recall": 0.7067925237829612, + "per_class_f1": { + "FAVOR": 0.6707466340269278, + "AGAINST": 0.6738159070598749, + "NONE": 0.7266009852216748 + }, + "accuracy_pct": 68.85007278020379, + "macro_f1_pct": 69.03878421028259, + "weighted_f1_pct": 68.74230482545485, + "macro_precision_pct": 68.553188400363, + "macro_recall_pct": 70.67925237829613, + "per_class_f1_pct": { + "FAVOR": 67.07466340269278, + "AGAINST": 67.38159070598749, + "NONE": 72.66009852216749 + } + }, + { + "after_task": 9, + "best_epoch": 7, + "best_dev_macro_f1": 0.659242191500256, + "losses": { + "current_ce": 0.42716561698665223, + "replay_ce": 0.22762162868554395, + "dark_distillation": 0.38987442714472614, + "supervised_contrastive": 4.077124718825022, + "total": 1.0807434156537057, + "replay_multiplier": 0.5772462882993676, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10126776346655016, + "ema_update_magnitude": 0.04927772559227879, + "final_multiplier": 0.5 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.7036659589525831, + "weighted_f1": 0.7326835936363901, + "macro_precision": 0.6820194298455169, + "macro_recall": 0.7792303848075962, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.7643312101910829, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 70.36659589525831, + "weighted_f1_pct": 73.26835936363901, + "macro_precision_pct": 68.20194298455169, + "macro_recall_pct": 77.92303848075962, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 76.43312101910828, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7964601769911505, + "macro_f1": 0.6841070494786874, + "weighted_f1": 0.8239914270905213, + "macro_precision": 0.6753989762119844, + "macro_recall": 0.8081058124014077, + "per_class_f1": { + "FAVOR": 0.8264462809917356, + "AGAINST": 0.34782608695652173, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 79.64601769911505, + "macro_f1_pct": 68.41070494786874, + "weighted_f1_pct": 82.39914270905213, + "macro_precision_pct": 67.53989762119843, + "macro_recall_pct": 80.81058124014076, + "per_class_f1_pct": { + "FAVOR": 82.64462809917356, + "AGAINST": 34.78260869565217, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.5909484650060123, + "weighted_f1": 0.6234737867904104, + "macro_precision": 0.6006330131177132, + "macro_recall": 0.5867828612926652, + "per_class_f1": { + "FAVOR": 0.4536082474226804, + "AGAINST": 0.7222222222222222, + "NONE": 0.5970149253731343 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 59.09484650060123, + "weighted_f1_pct": 62.34737867904104, + "macro_precision_pct": 60.06330131177132, + "macro_recall_pct": 58.678286129266525, + "per_class_f1_pct": { + "FAVOR": 45.36082474226804, + "AGAINST": 72.22222222222221, + "NONE": 59.70149253731343 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7473118279569892, + "macro_f1": 0.7091763088555391, + "weighted_f1": 0.7482902745390781, + "macro_precision": 0.7073214736766139, + "macro_recall": 0.7129629629629629, + "per_class_f1": { + "FAVOR": 0.6285714285714286, + "AGAINST": 0.8093023255813954, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 74.73118279569893, + "macro_f1_pct": 70.9176308855539, + "weighted_f1_pct": 74.82902745390781, + "macro_precision_pct": 70.7321473676614, + "macro_recall_pct": 71.29629629629629, + "per_class_f1_pct": { + "FAVOR": 62.857142857142854, + "AGAINST": 80.93023255813954, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.751269035532995, + "macro_f1": 0.7365091744112995, + "weighted_f1": 0.7515402117747891, + "macro_precision": 0.7296296296296295, + "macro_recall": 0.7444355054974524, + "per_class_f1": { + "FAVOR": 0.7536231884057971, + "AGAINST": 0.7892376681614349, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 75.1269035532995, + "macro_f1_pct": 73.65091744112995, + "weighted_f1_pct": 75.15402117747891, + "macro_precision_pct": 72.96296296296295, + "macro_recall_pct": 74.44355054974524, + "per_class_f1_pct": { + "FAVOR": 75.36231884057972, + "AGAINST": 78.9237668161435, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5633802816901409, + "macro_f1": 0.548970554405337, + "weighted_f1": 0.5576137987344355, + "macro_precision": 0.5963383838383839, + "macro_recall": 0.5371794871794872, + "per_class_f1": { + "FAVOR": 0.4782608695652174, + "AGAINST": 0.5079365079365079, + "NONE": 0.6607142857142857 + }, + "accuracy_pct": 56.33802816901409, + "macro_f1_pct": 54.897055440533705, + "weighted_f1_pct": 55.76137987344355, + "macro_precision_pct": 59.63383838383839, + "macro_recall_pct": 53.717948717948715, + "per_class_f1_pct": { + "FAVOR": 47.82608695652174, + "AGAINST": 50.79365079365079, + "NONE": 66.07142857142857 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.68, + "macro_f1": 0.6873661113537572, + "weighted_f1": 0.665527001310289, + "macro_precision": 0.7195447798119742, + "macro_recall": 0.7106370493904369, + "per_class_f1": { + "FAVOR": 0.5645161290322581, + "AGAINST": 0.7021276595744681, + "NONE": 0.7954545454545454 + }, + "accuracy_pct": 68.0, + "macro_f1_pct": 68.73661113537572, + "weighted_f1_pct": 66.5527001310289, + "macro_precision_pct": 71.95447798119741, + "macro_recall_pct": 71.06370493904369, + "per_class_f1_pct": { + "FAVOR": 56.451612903225815, + "AGAINST": 70.2127659574468, + "NONE": 79.54545454545455 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6804738543823481, + "weighted_f1": 0.6897157142026625, + "macro_precision": 0.6810708898944192, + "macro_recall": 0.679935125115848, + "per_class_f1": { + "FAVOR": 0.6407766990291263, + "AGAINST": 0.6461538461538462, + "NONE": 0.7544910179640718 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 68.04738543823481, + "weighted_f1_pct": 68.97157142026626, + "macro_precision_pct": 68.10708898944192, + "macro_recall_pct": 67.9935125115848, + "per_class_f1_pct": { + "FAVOR": 64.07766990291263, + "AGAINST": 64.61538461538461, + "NONE": 75.44910179640718 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.835, + "macro_f1": 0.8225310317859337, + "weighted_f1": 0.8369075113122173, + "macro_precision": 0.8172728252517983, + "macro_recall": 0.8314750032391212, + "per_class_f1": { + "FAVOR": 0.8148148148148148, + "AGAINST": 0.784, + "NONE": 0.8687782805429864 + }, + "accuracy_pct": 83.5, + "macro_f1_pct": 82.25310317859336, + "weighted_f1_pct": 83.69075113122173, + "macro_precision_pct": 81.72728252517983, + "macro_recall_pct": 83.14750032391211, + "per_class_f1_pct": { + "FAVOR": 81.48148148148148, + "AGAINST": 78.4, + "NONE": 86.87782805429865 + } + } + ], + "accuracy": 0.7141041931385006, + "macro_f1": 0.7070091169773877, + "weighted_f1": 0.7123846900309271, + "macro_precision": 0.7142662297298114, + "macro_recall": 0.7031488067199048, + "per_class_f1": { + "FAVOR": 0.6424731182795699, + "AGAINST": 0.7213114754098361, + "NONE": 0.7572427572427572 + }, + "accuracy_pct": 71.41041931385006, + "macro_f1_pct": 70.70091169773876, + "weighted_f1_pct": 71.2384690030927, + "macro_precision_pct": 71.42662297298114, + "macro_recall_pct": 70.31488067199048, + "per_class_f1_pct": { + "FAVOR": 64.24731182795699, + "AGAINST": 72.1311475409836, + "NONE": 75.72427572427573 + } + }, + { + "after_task": 10, + "best_epoch": 6, + "best_dev_macro_f1": 0.6450079996179287, + "losses": { + "current_ce": 0.6229372417315459, + "replay_ce": 0.22543158317223574, + "dark_distillation": 0.39491736659636867, + "supervised_contrastive": 4.231151305712187, + "total": 1.3111426961727632, + "replay_multiplier": 0.6252223476117874, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09525205826978211, + "ema_update_magnitude": 0.05314993873928781, + "final_multiplier": 0.5579925484523537 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 23, + "1": 18, + "2": 23 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.661016059068805, + "weighted_f1": 0.6904563663430738, + "macro_precision": 0.6540098983555956, + "macro_recall": 0.7515142428785607, + "per_class_f1": { + "FAVOR": 0.5915492957746479, + "AGAINST": 0.7248322147651006, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.10160590688051, + "weighted_f1_pct": 69.04563663430739, + "macro_precision_pct": 65.40098983555957, + "macro_recall_pct": 75.15142428785607, + "per_class_f1_pct": { + "FAVOR": 59.154929577464785, + "AGAINST": 72.48322147651007, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 1, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8141592920353983, + "macro_f1": 0.6788232167181634, + "weighted_f1": 0.8389866537770491, + "macro_precision": 0.6736842105263158, + "macro_recall": 0.7532095619463658, + "per_class_f1": { + "FAVOR": 0.8503937007874016, + "AGAINST": 0.3, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 81.41592920353983, + "macro_f1_pct": 67.88232167181634, + "weighted_f1_pct": 83.8986653777049, + "macro_precision_pct": 67.36842105263158, + "macro_recall_pct": 75.32095619463658, + "per_class_f1_pct": { + "FAVOR": 85.03937007874016, + "AGAINST": 30.0, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 2, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.61440535574215, + "weighted_f1": 0.639646121379787, + "macro_precision": 0.6144746646795828, + "macro_recall": 0.6161946259985477, + "per_class_f1": { + "FAVOR": 0.5391304347826087, + "AGAINST": 0.7070707070707071, + "NONE": 0.5970149253731343 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 61.440535574215005, + "weighted_f1_pct": 63.964612137978705, + "macro_precision_pct": 61.44746646795828, + "macro_recall_pct": 61.61946259985477, + "per_class_f1_pct": { + "FAVOR": 53.91304347826087, + "AGAINST": 70.70707070707071, + "NONE": 59.70149253731343 + } + }, + { + "task_index": 3, + "target": "Legalization of Abortion", + "accuracy": 0.7043010752688172, + "macro_f1": 0.6747763347763348, + "weighted_f1": 0.7121610110319787, + "macro_precision": 0.6671904016072219, + "macro_recall": 0.7049943883277217, + "per_class_f1": { + "FAVOR": 0.5952380952380952, + "AGAINST": 0.77, + "NONE": 0.6590909090909091 + }, + "accuracy_pct": 70.43010752688173, + "macro_f1_pct": 67.47763347763347, + "weighted_f1_pct": 71.21610110319787, + "macro_precision_pct": 66.71904016072219, + "macro_recall_pct": 70.49943883277217, + "per_class_f1_pct": { + "FAVOR": 59.523809523809526, + "AGAINST": 77.0, + "NONE": 65.9090909090909 + } + }, + { + "task_index": 4, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7198731381138289, + "weighted_f1": 0.7374200913529374, + "macro_precision": 0.7068069504224453, + "macro_recall": 0.7391720431803721, + "per_class_f1": { + "FAVOR": 0.7027027027027027, + "AGAINST": 0.7706422018348624, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.98731381138289, + "weighted_f1_pct": 73.74200913529374, + "macro_precision_pct": 70.68069504224454, + "macro_recall_pct": 73.91720431803721, + "per_class_f1_pct": { + "FAVOR": 70.27027027027027, + "AGAINST": 77.06422018348624, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 5, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5931882019211573, + "weighted_f1": 0.5788938962689602, + "macro_precision": 0.6066666666666666, + "macro_recall": 0.5837606837606838, + "per_class_f1": { + "FAVOR": 0.6545454545454545, + "AGAINST": 0.5081967213114754, + "NONE": 0.616822429906542 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 59.31882019211573, + "weighted_f1_pct": 57.88938962689601, + "macro_precision_pct": 60.66666666666666, + "macro_recall_pct": 58.37606837606838, + "per_class_f1_pct": { + "FAVOR": 65.45454545454545, + "AGAINST": 50.81967213114754, + "NONE": 61.6822429906542 + } + }, + { + "task_index": 6, + "target": "face masks", + "accuracy": 0.65, + "macro_f1": 0.6613453413493912, + "weighted_f1": 0.6495685213844574, + "macro_precision": 0.6625646513286964, + "macro_recall": 0.662705857827809, + "per_class_f1": { + "FAVOR": 0.6352941176470588, + "AGAINST": 0.6258503401360545, + "NONE": 0.7228915662650602 + }, + "accuracy_pct": 65.0, + "macro_f1_pct": 66.13453413493912, + "weighted_f1_pct": 64.95685213844574, + "macro_precision_pct": 66.25646513286964, + "macro_recall_pct": 66.2705857827809, + "per_class_f1_pct": { + "FAVOR": 63.52941176470588, + "AGAINST": 62.585034013605444, + "NONE": 72.28915662650603 + } + }, + { + "task_index": 7, + "target": "fauci", + "accuracy": 0.675, + "macro_f1": 0.6709048527230346, + "weighted_f1": 0.67541428268701, + "macro_precision": 0.6743865016583519, + "macro_recall": 0.6825919060858819, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.628099173553719, + "NONE": 0.717948717948718 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 67.09048527230345, + "weighted_f1_pct": 67.541428268701, + "macro_precision_pct": 67.43865016583518, + "macro_recall_pct": 68.2591906085882, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 62.8099173553719, + "NONE": 71.7948717948718 + } + }, + { + "task_index": 8, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7880539247662535, + "weighted_f1": 0.8084787345075015, + "macro_precision": 0.7780362115153675, + "macro_recall": 0.8026893960427194, + "per_class_f1": { + "FAVOR": 0.7857142857142857, + "AGAINST": 0.72, + "NONE": 0.8584474885844748 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 78.80539247662534, + "weighted_f1_pct": 80.84787345075016, + "macro_precision_pct": 77.80362115153675, + "macro_recall_pct": 80.26893960427194, + "per_class_f1_pct": { + "FAVOR": 78.57142857142857, + "AGAINST": 72.0, + "NONE": 85.84474885844749 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.5996803513632235, + "weighted_f1": 0.62904337396484, + "macro_precision": 0.5995670995670995, + "macro_recall": 0.6082741422547248, + "per_class_f1": { + "FAVOR": 0.6701570680628273, + "AGAINST": 0.40816326530612246, + "NONE": 0.7207207207207207 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 59.96803513632235, + "weighted_f1_pct": 62.904337396484, + "macro_precision_pct": 59.95670995670995, + "macro_recall_pct": 60.827414225472474, + "per_class_f1_pct": { + "FAVOR": 67.01570680628272, + "AGAINST": 40.816326530612244, + "NONE": 72.07207207207207 + } + } + ], + "accuracy": 0.6877113866967306, + "macro_f1": 0.689357550558008, + "weighted_f1": 0.6878745717010384, + "macro_precision": 0.687991306018859, + "macro_recall": 0.6925093749628486, + "per_class_f1": { + "FAVOR": 0.6679174484052532, + "AGAINST": 0.6695278969957081, + "NONE": 0.7306273062730627 + }, + "accuracy_pct": 68.77113866967306, + "macro_f1_pct": 68.9357550558008, + "weighted_f1_pct": 68.78745717010383, + "macro_precision_pct": 68.7991306018859, + "macro_recall_pct": 69.25093749628486, + "per_class_f1_pct": { + "FAVOR": 66.79174484052533, + "AGAINST": 66.95278969957081, + "NONE": 73.06273062730627 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s1__seed13/resolved_config.yaml b/outputs/trace_s8_full__x__s1__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c818cd1d9f46afed1e96b2c9e283a602d0322703 --- /dev/null +++ b/outputs/trace_s8_full__x__s1__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s1 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__x__s1__seed13/run.log b/outputs/trace_s8_full__x__s1__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..5ab1d233e71473e7efa08efba336ca4b2e4084f1 --- /dev/null +++ b/outputs/trace_s8_full__x__s1__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 10:23:32,173 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__x__s1__seed13 device=cuda +2026-09-02 10:39:51,285 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6662066776542341, 'average_forgetting': 0.03806884006462382, 'backward_transfer': -0.004807116965571111, 'final_mean_target_macro_f1_pct': 66.62066776542342, 'average_forgetting_pct': 3.806884006462382, 'backward_transfer_pct': -0.4807116965571111, 'final_pooled_accuracy': 0.6877113866967306, 'final_pooled_accuracy_pct': 68.77113866967306, 'final_pooled_macro_f1': 0.689357550558008, 'final_pooled_macro_f1_pct': 68.9357550558008, 'final_pooled_weighted_f1': 0.6878745717010384, 'final_pooled_weighted_f1_pct': 68.78745717010383, 'final_pooled_macro_precision': 0.687991306018859, 'final_pooled_macro_precision_pct': 68.7991306018859, 'final_pooled_macro_recall': 0.6925093749628486, 'final_pooled_macro_recall_pct': 69.25093749628486, 'final_pooled_per_class_f1': {'FAVOR': 0.6679174484052532, 'AGAINST': 0.6695278969957081, 'NONE': 0.7306273062730627}, 'final_pooled_per_class_f1_pct': {'FAVOR': 66.79174484052533, 'AGAINST': 66.95278969957081, 'NONE': 73.06273062730627}} +2026-09-02 10:39:51,285 [INFO] __main__: efficiency={'training_wall_time_seconds': 977.7541670140345, 'training_wall_time_minutes': 16.29590278356724, 'peak_gpu_memory_allocated_gb': 10.431715965270996, 'peak_gpu_memory_reserved_gb': 10.58984375} diff --git a/outputs/trace_s8_full__x__s2__seed13/environment.json b/outputs/trace_s8_full__x__s2__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..5cff666a538e8705ce6fe24427f7a101fea897fd --- /dev/null +++ b/outputs/trace_s8_full__x__s2__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T10:39:59.300755+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s2__seed13/metrics.json b/outputs/trace_s8_full__x__s2__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..3f4f6a17e6b32fb0cb7bee95fa5cc49adbb9d5e9 --- /dev/null +++ b/outputs/trace_s8_full__x__s2__seed13/metrics.json @@ -0,0 +1,2354 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s2", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7420705909078001, + 0.6927570711075864, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7416944750774578, + 0.6439955106621773, + 0.5689407443418139, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7382595826884623, + 0.6771909302986444, + 0.5831924424017, + 0.7207945042617001, + null, + null, + null, + null, + null, + null + ], + [ + 0.7448627450980392, + 0.680855614973262, + 0.6074954013072029, + 0.7377131090040958, + 0.648070576288398, + null, + null, + null, + null, + null + ], + [ + 0.6673721340388007, + 0.6891100055279159, + 0.5767310789049919, + 0.6999440990887891, + 0.6478919812620556, + 0.712037037037037, + null, + null, + null, + null + ], + [ + 0.7190925780225544, + 0.7008117308195864, + 0.5807626770080497, + 0.7097267060477602, + 0.6048672282958633, + 0.7181333859510656, + 0.5903609109086035, + null, + null, + null + ], + [ + 0.7151868938894642, + 0.6764994389994391, + 0.5842889750570635, + 0.6803735359290916, + 0.6658755909891536, + 0.7007246799168464, + 0.5794731679977582, + 0.7182850767928212, + null, + null + ], + [ + 0.7070527435780515, + 0.6647196374915189, + 0.6038681710970867, + 0.6594512816036775, + 0.6245580324645243, + 0.730108162550359, + 0.5748842592592592, + 0.6550837484483565, + 0.6749414006966611, + null + ], + [ + 0.6721263265567062, + 0.6487946111463073, + 0.5626202531645569, + 0.6227004940144377, + 0.628191368546783, + 0.7153287258550417, + 0.5752970838369739, + 0.642294390389595, + 0.6460240385243323, + 0.7399907370975146 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 0.7207898354223875, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10138365834772536, + "ema_update_magnitude": 0.06471023128442037, + "final_multiplier": 0.6382708252889969 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.6251337247900822, + "losses": { + "current_ce": 0.6757452957557909, + "replay_ce": 0.1390355658802119, + "dark_distillation": 0.38217191443298804, + "supervised_contrastive": 4.239523621761438, + "total": 1.3066114400372362, + "replay_multiplier": 0.598462906802525, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10362827898120411, + "ema_update_magnitude": 0.05418670752128535, + "final_multiplier": 0.5228949863300695 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.7420705909078001, + "weighted_f1": 0.8640357721444364, + "macro_precision": 0.7182733183295825, + "macro_recall": 0.7843950976823201, + "per_class_f1": { + "FAVOR": 0.8837209302325582, + "AGAINST": 0.46153846153846156, + "NONE": 0.8809523809523809 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 74.20705909078002, + "weighted_f1_pct": 86.40357721444364, + "macro_precision_pct": 71.82733183295825, + "macro_recall_pct": 78.43950976823201, + "per_class_f1_pct": { + "FAVOR": 88.37209302325581, + "AGAINST": 46.15384615384615, + "NONE": 88.09523809523809 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6927570711075864, + "weighted_f1": 0.7215023918549025, + "macro_precision": 0.6767206953253465, + "macro_recall": 0.7300785634118968, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7628865979381443, + "NONE": 0.7 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 69.27570711075865, + "weighted_f1_pct": 72.15023918549025, + "macro_precision_pct": 67.67206953253465, + "macro_recall_pct": 73.00785634118968, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.28865979381443, + "NONE": 70.0 + } + } + ], + "accuracy": 0.7692307692307693, + "macro_f1": 0.7697262479871175, + "weighted_f1": 0.7680028436172104, + "macro_precision": 0.7702840531718752, + "macro_recall": 0.7762084105097071, + "per_class_f1": { + "FAVOR": 0.782608695652174, + "AGAINST": 0.7439613526570048, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 76.92307692307693, + "macro_f1_pct": 76.97262479871175, + "weighted_f1_pct": 76.80028436172104, + "macro_precision_pct": 77.02840531718752, + "macro_recall_pct": 77.62084105097071, + "per_class_f1_pct": { + "FAVOR": 78.26086956521739, + "AGAINST": 74.39613526570048, + "NONE": 78.26086956521739 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.6063492063492063, + "losses": { + "current_ce": 0.7316161394119263, + "replay_ce": 0.23026843952635925, + "dark_distillation": 0.38139258945981663, + "supervised_contrastive": 4.419250331322352, + "total": 1.4633842170238496, + "replay_multiplier": 0.6681105596417665, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10096908425856493, + "ema_update_magnitude": 0.05408146850315572, + "final_multiplier": 0.5356240368057823 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.7416944750774578, + "weighted_f1": 0.8640348370132873, + "macro_precision": 0.7197662601626016, + "macro_recall": 0.7812401407596165, + "per_class_f1": { + "FAVOR": 0.8854961832061069, + "AGAINST": 0.46153846153846156, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 74.16944750774577, + "weighted_f1_pct": 86.40348370132872, + "macro_precision_pct": 71.97662601626016, + "macro_recall_pct": 78.12401407596164, + "per_class_f1_pct": { + "FAVOR": 88.54961832061069, + "AGAINST": 46.15384615384615, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6451612903225806, + "macro_f1": 0.6439955106621773, + "weighted_f1": 0.6472140762463343, + "macro_precision": 0.6574643679906838, + "macro_recall": 0.7379910213243547, + "per_class_f1": { + "FAVOR": 0.5858585858585859, + "AGAINST": 0.6424242424242425, + "NONE": 0.7037037037037037 + }, + "accuracy_pct": 64.51612903225806, + "macro_f1_pct": 64.39955106621773, + "weighted_f1_pct": 64.72140762463343, + "macro_precision_pct": 65.74643679906838, + "macro_recall_pct": 73.79910213243546, + "per_class_f1_pct": { + "FAVOR": 58.58585858585859, + "AGAINST": 64.24242424242425, + "NONE": 70.37037037037037 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5689407443418139, + "weighted_f1": 0.5403055180866433, + "macro_precision": 0.5749516908212561, + "macro_recall": 0.6423076923076924, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.32941176470588235, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.894074434181384, + "weighted_f1_pct": 54.03055180866433, + "macro_precision_pct": 57.49516908212561, + "macro_recall_pct": 64.23076923076924, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 32.94117647058823, + "NONE": 71.07438016528926 + } + } + ], + "accuracy": 0.6802721088435374, + "macro_f1": 0.6728324003213052, + "weighted_f1": 0.660575198571689, + "macro_precision": 0.6963032721622057, + "macro_recall": 0.7072270445479795, + "per_class_f1": { + "FAVOR": 0.7337662337662337, + "AGAINST": 0.532319391634981, + "NONE": 0.752411575562701 + }, + "accuracy_pct": 68.02721088435374, + "macro_f1_pct": 67.28324003213052, + "weighted_f1_pct": 66.0575198571689, + "macro_precision_pct": 69.63032721622056, + "macro_recall_pct": 70.72270445479795, + "per_class_f1_pct": { + "FAVOR": 73.37662337662337, + "AGAINST": 53.2319391634981, + "NONE": 75.2411575562701 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7402164209466441, + "losses": { + "current_ce": 0.6443170191585154, + "replay_ce": 0.2376939520455789, + "dark_distillation": 0.38270324814146844, + "supervised_contrastive": 4.376125502931899, + "total": 1.3424699099167534, + "replay_multiplier": 0.5999306783767159, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0931216502846001, + "ema_update_magnitude": 0.053369868727746024, + "final_multiplier": 0.5731198766842732 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7382595826884623, + "weighted_f1": 0.8529596215881656, + "macro_precision": 0.709349593495935, + "macro_recall": 0.8329814342919549, + "per_class_f1": { + "FAVOR": 0.8661417322834646, + "AGAINST": 0.47058823529411764, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.82595826884622, + "weighted_f1_pct": 85.29596215881656, + "macro_precision_pct": 70.9349593495935, + "macro_recall_pct": 83.29814342919549, + "per_class_f1_pct": { + "FAVOR": 86.61417322834646, + "AGAINST": 47.05882352941176, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6771909302986444, + "weighted_f1": 0.6939867959317195, + "macro_precision": 0.6639633675383427, + "macro_recall": 0.7356902356902357, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.717391304347826, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.71909302986444, + "weighted_f1_pct": 69.39867959317195, + "macro_precision_pct": 66.39633675383428, + "macro_recall_pct": 73.56902356902357, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 71.73913043478261, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.5831924424017, + "weighted_f1": 0.5559465321780778, + "macro_precision": 0.5939209726443769, + "macro_recall": 0.6542735042735043, + "per_class_f1": { + "FAVOR": 0.6753246753246753, + "AGAINST": 0.35294117647058826, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 58.31924424017, + "weighted_f1_pct": 55.594653217807775, + "macro_precision_pct": 59.392097264437695, + "macro_recall_pct": 65.42735042735043, + "per_class_f1_pct": { + "FAVOR": 67.53246753246754, + "AGAINST": 35.294117647058826, + "NONE": 72.1311475409836 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.73, + "macro_f1": 0.7207945042617001, + "weighted_f1": 0.7250960189865779, + "macro_precision": 0.7419189466299715, + "macro_recall": 0.7330089589125733, + "per_class_f1": { + "FAVOR": 0.7317073170731707, + "AGAINST": 0.6542056074766355, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 72.07945042617001, + "weighted_f1_pct": 72.50960189865779, + "macro_precision_pct": 74.19189466299714, + "macro_recall_pct": 73.30089589125734, + "per_class_f1_pct": { + "FAVOR": 73.17073170731707, + "AGAINST": 65.42056074766354, + "NONE": 77.64705882352942 + } + } + ], + "accuracy": 0.7082683307332294, + "macro_f1": 0.7040371071445786, + "weighted_f1": 0.69931731302485, + "macro_precision": 0.7154450440977955, + "macro_recall": 0.7212885154061626, + "per_class_f1": { + "FAVOR": 0.7409200968523002, + "AGAINST": 0.6106870229007634, + "NONE": 0.7605042016806722 + }, + "accuracy_pct": 70.82683307332293, + "macro_f1_pct": 70.40371071445786, + "weighted_f1_pct": 69.931731302485, + "macro_precision_pct": 71.54450440977955, + "macro_recall_pct": 72.12885154061625, + "per_class_f1_pct": { + "FAVOR": 74.09200968523002, + "AGAINST": 61.06870229007634, + "NONE": 76.05042016806722 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6444444444444444, + "losses": { + "current_ce": 0.656370164339359, + "replay_ce": 0.2301974947635944, + "dark_distillation": 0.44339590607545315, + "supervised_contrastive": 4.260124096503625, + "total": 1.3673993850365662, + "replay_multiplier": 0.6230831125425607, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09820985698124235, + "ema_update_magnitude": 0.056970408125201336, + "final_multiplier": 0.580088494946922 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7448627450980392, + "weighted_f1": 0.8502842269651223, + "macro_precision": 0.7118599791013583, + "macro_recall": 0.8361363912146583, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.5, + "NONE": 0.8705882352941177 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 74.48627450980392, + "weighted_f1_pct": 85.02842269651222, + "macro_precision_pct": 71.18599791013584, + "macro_recall_pct": 83.61363912146584, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 50.0, + "NONE": 87.05882352941177 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.680855614973262, + "weighted_f1": 0.6995169915473521, + "macro_precision": 0.6660573603611578, + "macro_recall": 0.7344556677890012, + "per_class_f1": { + "FAVOR": 0.6352941176470588, + "AGAINST": 0.7272727272727273, + "NONE": 0.68 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 68.0855614973262, + "weighted_f1_pct": 69.95169915473521, + "macro_precision_pct": 66.60573603611579, + "macro_recall_pct": 73.44556677890012, + "per_class_f1_pct": { + "FAVOR": 63.52941176470588, + "AGAINST": 72.72727272727273, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6074954013072029, + "weighted_f1": 0.5854366072076317, + "macro_precision": 0.6149814659066131, + "macro_recall": 0.6615384615384615, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.4044943820224719, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.74954013072029, + "weighted_f1_pct": 58.54366072076317, + "macro_precision_pct": 61.49814659066131, + "macro_recall_pct": 66.15384615384615, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 40.44943820224719, + "NONE": 74.19354838709677 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.75, + "macro_f1": 0.7377131090040958, + "weighted_f1": 0.7449259713516326, + "macro_precision": 0.7601323796975971, + "macro_recall": 0.7453969725054063, + "per_class_f1": { + "FAVOR": 0.7288135593220338, + "AGAINST": 0.6728971962616822, + "NONE": 0.8114285714285714 + }, + "accuracy_pct": 75.0, + "macro_f1_pct": 73.77131090040959, + "weighted_f1_pct": 74.49259713516327, + "macro_precision_pct": 76.0132379697597, + "macro_recall_pct": 74.53969725054063, + "per_class_f1_pct": { + "FAVOR": 72.88135593220339, + "AGAINST": 67.28971962616822, + "NONE": 81.14285714285714 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.648070576288398, + "weighted_f1": 0.6804998095963443, + "macro_precision": 0.6465941465941466, + "macro_recall": 0.6550666162316648, + "per_class_f1": { + "FAVOR": 0.7326732673267327, + "AGAINST": 0.46153846153846156, + "NONE": 0.75 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 64.8070576288398, + "weighted_f1_pct": 68.04998095963442, + "macro_precision_pct": 64.65941465941467, + "macro_recall_pct": 65.50666162316648, + "per_class_f1_pct": { + "FAVOR": 73.26732673267327, + "AGAINST": 46.15384615384615, + "NONE": 75.0 + } + } + ], + "accuracy": 0.7122473246135553, + "macro_f1": 0.7046884941718682, + "weighted_f1": 0.7045732294686651, + "macro_precision": 0.7146206554110776, + "macro_recall": 0.7124786604191637, + "per_class_f1": { + "FAVOR": 0.7387687188019967, + "AGAINST": 0.6037735849056604, + "NONE": 0.7715231788079471 + }, + "accuracy_pct": 71.22473246135553, + "macro_f1_pct": 70.46884941718682, + "weighted_f1_pct": 70.45732294686651, + "macro_precision_pct": 71.46206554110776, + "macro_recall_pct": 71.24786604191638, + "per_class_f1_pct": { + "FAVOR": 73.87687188019967, + "AGAINST": 60.37735849056604, + "NONE": 77.15231788079471 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.8609723609723611, + "losses": { + "current_ce": 0.48282817933639444, + "replay_ce": 0.25347888149597025, + "dark_distillation": 0.4362580687566488, + "supervised_contrastive": 3.7781283725191046, + "total": 1.1614947441206487, + "replay_multiplier": 0.6257369823982759, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10432290151951146, + "ema_update_magnitude": 0.05566255211860548, + "final_multiplier": 0.533560237568689 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6673721340388007, + "weighted_f1": 0.833113264971672, + "macro_precision": 0.6639433551198257, + "macro_recall": 0.6901832301905109, + "per_class_f1": { + "FAVOR": 0.8592592592592593, + "AGAINST": 0.2857142857142857, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 66.73721340388006, + "weighted_f1_pct": 83.3113264971672, + "macro_precision_pct": 66.39433551198258, + "macro_recall_pct": 69.01832301905108, + "per_class_f1_pct": { + "FAVOR": 85.92592592592592, + "AGAINST": 28.57142857142857, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6891100055279159, + "weighted_f1": 0.7132883967260472, + "macro_precision": 0.6747162485065711, + "macro_recall": 0.7221099887766554, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7562189054726368, + "NONE": 0.6444444444444445 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.91100055279159, + "weighted_f1_pct": 71.32883967260472, + "macro_precision_pct": 67.47162485065711, + "macro_recall_pct": 72.21099887766555, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 75.62189054726367, + "NONE": 64.44444444444444 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5767310789049919, + "weighted_f1": 0.5566544192692386, + "macro_precision": 0.5759220354808591, + "macro_recall": 0.6247863247863248, + "per_class_f1": { + "FAVOR": 0.6388888888888888, + "AGAINST": 0.391304347826087, + "NONE": 0.7 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 57.67310789049919, + "weighted_f1_pct": 55.66544192692386, + "macro_precision_pct": 57.592203548085905, + "macro_recall_pct": 62.47863247863248, + "per_class_f1_pct": { + "FAVOR": 63.888888888888886, + "AGAINST": 39.130434782608695, + "NONE": 70.0 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6999440990887891, + "weighted_f1": 0.7037065698220367, + "macro_precision": 0.7047619047619048, + "macro_recall": 0.7114766759345073, + "per_class_f1": { + "FAVOR": 0.7049180327868853, + "AGAINST": 0.6495726495726496, + "NONE": 0.7453416149068323 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.99440990887891, + "weighted_f1_pct": 70.37065698220367, + "macro_precision_pct": 70.47619047619048, + "macro_recall_pct": 71.14766759345072, + "per_class_f1_pct": { + "FAVOR": 70.49180327868852, + "AGAINST": 64.95726495726495, + "NONE": 74.53416149068323 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6478919812620556, + "weighted_f1": 0.6700457426288233, + "macro_precision": 0.6420194106075142, + "macro_recall": 0.6618501239860463, + "per_class_f1": { + "FAVOR": 0.7015706806282722, + "AGAINST": 0.5052631578947369, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.78919812620556, + "weighted_f1_pct": 67.00457426288233, + "macro_precision_pct": 64.20194106075142, + "macro_recall_pct": 66.18501239860463, + "per_class_f1_pct": { + "FAVOR": 70.15706806282722, + "AGAINST": 50.526315789473685, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.712037037037037, + "weighted_f1": 0.7738394216133943, + "macro_precision": 0.7048560135516656, + "macro_recall": 0.7245577211394303, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.8444444444444444, + "NONE": 0.75 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 71.2037037037037, + "weighted_f1_pct": 77.38394216133942, + "macro_precision_pct": 70.48560135516657, + "macro_recall_pct": 72.45577211394303, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 84.44444444444444, + "NONE": 75.0 + } + } + ], + "accuracy": 0.7041540020263425, + "macro_f1": 0.7054055283085097, + "weighted_f1": 0.7030027857496425, + "macro_precision": 0.7045077209685201, + "macro_recall": 0.7094306363476114, + "per_class_f1": { + "FAVOR": 0.711864406779661, + "AGAINST": 0.6695278969957081, + "NONE": 0.7348242811501597 + }, + "accuracy_pct": 70.41540020263425, + "macro_f1_pct": 70.54055283085097, + "weighted_f1_pct": 70.30027857496425, + "macro_precision_pct": 70.450772096852, + "macro_recall_pct": 70.94306363476115, + "per_class_f1_pct": { + "FAVOR": 71.1864406779661, + "AGAINST": 66.95278969957081, + "NONE": 73.48242811501598 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.5563271604938271, + "losses": { + "current_ce": 0.5817336595419681, + "replay_ce": 0.22941977074652006, + "dark_distillation": 0.4572236171274474, + "supervised_contrastive": 4.228520169402614, + "total": 1.2951659498792707, + "replay_multiplier": 0.6223886659368644, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10125105852992596, + "ema_update_magnitude": 0.053141566485488054, + "final_multiplier": 0.5248494905342785 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7190925780225544, + "weighted_f1": 0.8519840137025407, + "macro_precision": 0.6996434174868065, + "macro_recall": 0.831161266836549, + "per_class_f1": { + "FAVOR": 0.8548387096774194, + "AGAINST": 0.4, + "NONE": 0.9024390243902439 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 71.90925780225544, + "weighted_f1_pct": 85.19840137025408, + "macro_precision_pct": 69.96434174868065, + "macro_recall_pct": 83.1161266836549, + "per_class_f1_pct": { + "FAVOR": 85.48387096774194, + "AGAINST": 40.0, + "NONE": 90.2439024390244 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7008117308195864, + "weighted_f1": 0.730785799356359, + "macro_precision": 0.6856164041010252, + "macro_recall": 0.7276094276094276, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7761194029850746, + "NONE": 0.6947368421052632 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.08117308195864, + "weighted_f1_pct": 73.0785799356359, + "macro_precision_pct": 68.56164041010253, + "macro_recall_pct": 72.76094276094275, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 77.61194029850746, + "NONE": 69.47368421052632 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5807626770080497, + "weighted_f1": 0.5565428530995599, + "macro_precision": 0.5783993783993785, + "macro_recall": 0.6247863247863248, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.3870967741935484, + "NONE": 0.6885245901639344 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.076267700804976, + "weighted_f1_pct": 55.65428530995599, + "macro_precision_pct": 57.839937839937846, + "macro_recall_pct": 62.47863247863248, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 38.70967741935484, + "NONE": 68.85245901639344 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7097267060477602, + "weighted_f1": 0.7158708516883039, + "macro_precision": 0.7323499545721767, + "macro_recall": 0.7262588816805685, + "per_class_f1": { + "FAVOR": 0.7131782945736435, + "AGAINST": 0.6355140186915887, + "NONE": 0.7804878048780488 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 70.97267060477603, + "weighted_f1_pct": 71.58708516883038, + "macro_precision_pct": 73.23499545721766, + "macro_recall_pct": 72.62588816805685, + "per_class_f1_pct": { + "FAVOR": 71.31782945736434, + "AGAINST": 63.55140186915887, + "NONE": 78.04878048780488 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.6048672282958633, + "weighted_f1": 0.6271031260429414, + "macro_precision": 0.6042937853107344, + "macro_recall": 0.6289721066420096, + "per_class_f1": { + "FAVOR": 0.651685393258427, + "AGAINST": 0.43564356435643564, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 60.48672282958633, + "weighted_f1_pct": 62.71031260429414, + "macro_precision_pct": 60.42937853107344, + "macro_recall_pct": 62.89721066420096, + "per_class_f1_pct": { + "FAVOR": 65.1685393258427, + "AGAINST": 43.56435643564357, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7808219178082192, + "macro_f1": 0.7181333859510656, + "weighted_f1": 0.7802270401084069, + "macro_precision": 0.7120894278571224, + "macro_recall": 0.7281809095452273, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.850828729281768, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 78.08219178082192, + "macro_f1_pct": 71.81333859510656, + "weighted_f1_pct": 78.0227040108407, + "macro_precision_pct": 71.20894278571224, + "macro_recall_pct": 72.81809095452273, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 85.0828729281768, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5903609109086035, + "weighted_f1": 0.6175335608241325, + "macro_precision": 0.5848731481938209, + "macro_recall": 0.5973129992737837, + "per_class_f1": { + "FAVOR": 0.5321100917431193, + "AGAINST": 0.6834170854271356, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.03609109086035, + "weighted_f1_pct": 61.75335608241325, + "macro_precision_pct": 58.487314819382085, + "macro_recall_pct": 59.731299927378366, + "per_class_f1_pct": { + "FAVOR": 53.21100917431193, + "AGAINST": 68.34170854271356, + "NONE": 55.55555555555556 + } + } + ], + "accuracy": 0.6890399320305862, + "macro_f1": 0.6909910738875026, + "weighted_f1": 0.6878658072394095, + "macro_precision": 0.6882248456031044, + "macro_recall": 0.696535444939164, + "per_class_f1": { + "FAVOR": 0.6712141882673943, + "AGAINST": 0.6674057649667405, + "NONE": 0.7343532684283728 + }, + "accuracy_pct": 68.90399320305862, + "macro_f1_pct": 69.09910738875026, + "weighted_f1_pct": 68.78658072394094, + "macro_precision_pct": 68.82248456031044, + "macro_recall_pct": 69.6535444939164, + "per_class_f1_pct": { + "FAVOR": 67.12141882673943, + "AGAINST": 66.74057649667405, + "NONE": 73.43532684283728 + } + }, + { + "after_task": 8, + "best_epoch": 14, + "best_dev_macro_f1": 0.7061420877554253, + "losses": { + "current_ce": 0.5435022296327533, + "replay_ce": 0.22035438978310787, + "dark_distillation": 0.433873850107193, + "supervised_contrastive": 4.268619733868223, + "total": 1.2484463727835453, + "replay_multiplier": 0.6279489699287756, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09643344680537284, + "ema_update_magnitude": 0.054508257284998744, + "final_multiplier": 0.5652422379447893 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7151868938894642, + "weighted_f1": 0.8405366004178842, + "macro_precision": 0.6972433262755843, + "macro_recall": 0.8167212716903288, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.42105263157894735, + "NONE": 0.8717948717948718 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.51868938894643, + "weighted_f1_pct": 84.05366004178842, + "macro_precision_pct": 69.72433262755843, + "macro_recall_pct": 81.67212716903288, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 42.10526315789473, + "NONE": 87.17948717948718 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6764994389994391, + "weighted_f1": 0.6938052085229506, + "macro_precision": 0.6650714622317057, + "macro_recall": 0.7394500561167229, + "per_class_f1": { + "FAVOR": 0.6593406593406593, + "AGAINST": 0.7243243243243244, + "NONE": 0.6458333333333334 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.6499438999439, + "weighted_f1_pct": 69.38052085229506, + "macro_precision_pct": 66.50714622317057, + "macro_recall_pct": 73.94500561167229, + "per_class_f1_pct": { + "FAVOR": 65.93406593406593, + "AGAINST": 72.43243243243244, + "NONE": 64.58333333333334 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5842889750570635, + "weighted_f1": 0.5533534512910313, + "macro_precision": 0.5820540935672515, + "macro_recall": 0.6303418803418804, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.3695652173913043, + "NONE": 0.6774193548387096 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.428897505706345, + "weighted_f1_pct": 55.33534512910313, + "macro_precision_pct": 58.20540935672515, + "macro_recall_pct": 63.03418803418804, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 36.95652173913043, + "NONE": 67.74193548387096 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6803735359290916, + "weighted_f1": 0.687141975308642, + "macro_precision": 0.7012321077672858, + "macro_recall": 0.6962619709607661, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.6111111111111112, + "NONE": 0.7530864197530864 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 68.03735359290916, + "weighted_f1_pct": 68.7141975308642, + "macro_precision_pct": 70.12321077672858, + "macro_recall_pct": 69.62619709607661, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 61.111111111111114, + "NONE": 75.30864197530865 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.7, + "macro_f1": 0.6658755909891536, + "weighted_f1": 0.6998875189271037, + "macro_precision": 0.6616780045351475, + "macro_recall": 0.6713515179534597, + "per_class_f1": { + "FAVOR": 0.7562189054726368, + "AGAINST": 0.47619047619047616, + "NONE": 0.7652173913043478 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 66.58755909891536, + "weighted_f1_pct": 69.98875189271037, + "macro_precision_pct": 66.16780045351474, + "macro_recall_pct": 67.13515179534598, + "per_class_f1_pct": { + "FAVOR": 75.62189054726367, + "AGAINST": 47.61904761904761, + "NONE": 76.52173913043478 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7465753424657534, + "macro_f1": 0.7007246799168464, + "weighted_f1": 0.7543620120706354, + "macro_precision": 0.6832107843137255, + "macro_recall": 0.7294852573713143, + "per_class_f1": { + "FAVOR": 0.5263157894736842, + "AGAINST": 0.813953488372093, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 74.65753424657534, + "macro_f1_pct": 70.07246799168463, + "weighted_f1_pct": 75.43620120706353, + "macro_precision_pct": 68.32107843137256, + "macro_recall_pct": 72.94852573713143, + "per_class_f1_pct": { + "FAVOR": 52.63157894736842, + "AGAINST": 81.3953488372093, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5794731679977582, + "weighted_f1": 0.5946979049136082, + "macro_precision": 0.5720211161387632, + "macro_recall": 0.6056644880174292, + "per_class_f1": { + "FAVOR": 0.5409836065573771, + "AGAINST": 0.6333333333333333, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 57.94731679977582, + "weighted_f1_pct": 59.46979049136082, + "macro_precision_pct": 57.20211161387632, + "macro_recall_pct": 60.56644880174292, + "per_class_f1_pct": { + "FAVOR": 54.09836065573771, + "AGAINST": 63.33333333333333, + "NONE": 56.41025641025641 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7182850767928212, + "weighted_f1": 0.737953049811892, + "macro_precision": 0.705325206219432, + "macro_recall": 0.7356280136082324, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.7685185185185185, + "NONE": 0.7102803738317757 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.82850767928211, + "weighted_f1_pct": 73.7953049811892, + "macro_precision_pct": 70.53252062194319, + "macro_recall_pct": 73.56280136082324, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 76.85185185185185, + "NONE": 71.02803738317756 + } + } + ], + "accuracy": 0.6899563318777293, + "macro_f1": 0.6922808591785552, + "weighted_f1": 0.6883392172674279, + "macro_precision": 0.6899051628826104, + "macro_recall": 0.7050388213146378, + "per_class_f1": { + "FAVOR": 0.6927502876869965, + "AGAINST": 0.6647727272727273, + "NONE": 0.7193195625759417 + }, + "accuracy_pct": 68.99563318777294, + "macro_f1_pct": 69.22808591785552, + "weighted_f1_pct": 68.83392172674279, + "macro_precision_pct": 68.99051628826103, + "macro_recall_pct": 70.50388213146378, + "per_class_f1_pct": { + "FAVOR": 69.27502876869966, + "AGAINST": 66.47727272727273, + "NONE": 71.93195625759417 + } + }, + { + "after_task": 9, + "best_epoch": 13, + "best_dev_macro_f1": 0.6423910840932118, + "losses": { + "current_ce": 0.520153154929479, + "replay_ce": 0.20112030771043565, + "dark_distillation": 0.40791797637939453, + "supervised_contrastive": 4.208267084000602, + "total": 1.1806766307543195, + "replay_multiplier": 0.586306728213941, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09318459210348767, + "ema_update_magnitude": 0.05341735456724901, + "final_multiplier": 0.573242350064971 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7070527435780515, + "weighted_f1": 0.8569491019293084, + "macro_precision": 0.6921009098428454, + "macro_recall": 0.7712898920033976, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.35294117647058826, + "NONE": 0.9 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 70.70527435780515, + "weighted_f1_pct": 85.69491019293083, + "macro_precision_pct": 69.21009098428455, + "macro_recall_pct": 77.12898920033976, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 35.294117647058826, + "NONE": 90.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6647196374915189, + "weighted_f1": 0.6788558541023741, + "macro_precision": 0.6592714858811398, + "macro_recall": 0.7318181818181818, + "per_class_f1": { + "FAVOR": 0.6105263157894737, + "AGAINST": 0.6961325966850829, + "NONE": 0.6875 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.47196374915188, + "weighted_f1_pct": 67.88558541023741, + "macro_precision_pct": 65.92714858811398, + "macro_recall_pct": 73.18181818181819, + "per_class_f1_pct": { + "FAVOR": 61.05263157894737, + "AGAINST": 69.61325966850829, + "NONE": 68.75 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6038681710970867, + "weighted_f1": 0.5722183165929984, + "macro_precision": 0.6245985115550332, + "macro_recall": 0.6782051282051281, + "per_class_f1": { + "FAVOR": 0.72, + "AGAINST": 0.3614457831325301, + "NONE": 0.7301587301587301 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.386817109708666, + "weighted_f1_pct": 57.22183165929984, + "macro_precision_pct": 62.45985115550332, + "macro_recall_pct": 67.82051282051282, + "per_class_f1_pct": { + "FAVOR": 72.0, + "AGAINST": 36.144578313253014, + "NONE": 73.01587301587301 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.685, + "macro_f1": 0.6594512816036775, + "weighted_f1": 0.6718573797678274, + "macro_precision": 0.7142153239714215, + "macro_recall": 0.683348779734322, + "per_class_f1": { + "FAVOR": 0.6567164179104478, + "AGAINST": 0.5263157894736842, + "NONE": 0.7953216374269005 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 65.94512816036774, + "weighted_f1_pct": 67.18573797678275, + "macro_precision_pct": 71.42153239714216, + "macro_recall_pct": 68.3348779734322, + "per_class_f1_pct": { + "FAVOR": 65.67164179104478, + "AGAINST": 52.63157894736842, + "NONE": 79.53216374269006 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.69, + "macro_f1": 0.6245580324645243, + "weighted_f1": 0.6745694893502369, + "macro_precision": 0.6381158835786785, + "macro_recall": 0.624452570083638, + "per_class_f1": { + "FAVOR": 0.7579908675799086, + "AGAINST": 0.34782608695652173, + "NONE": 0.7678571428571429 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 62.45580324645243, + "weighted_f1_pct": 67.45694893502369, + "macro_precision_pct": 63.811588357867855, + "macro_recall_pct": 62.4452570083638, + "per_class_f1_pct": { + "FAVOR": 75.79908675799086, + "AGAINST": 34.78260869565217, + "NONE": 76.78571428571429 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.730108162550359, + "weighted_f1": 0.7792938240989353, + "macro_precision": 0.7105820105820105, + "macro_recall": 0.759720139930035, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8323699421965318, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.0108162550359, + "weighted_f1_pct": 77.92938240989353, + "macro_precision_pct": 71.05820105820105, + "macro_recall_pct": 75.9720139930035, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 83.23699421965318, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5748842592592592, + "weighted_f1": 0.5888779239766082, + "macro_precision": 0.572412493465125, + "macro_recall": 0.5951343500363109, + "per_class_f1": { + "FAVOR": 0.546875, + "AGAINST": 0.6222222222222222, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 57.488425925925924, + "weighted_f1_pct": 58.88779239766082, + "macro_precision_pct": 57.241249346512504, + "macro_recall_pct": 59.51343500363109, + "per_class_f1_pct": { + "FAVOR": 54.6875, + "AGAINST": 62.22222222222222, + "NONE": 55.55555555555556 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6548223350253807, + "macro_f1": 0.6550837484483565, + "weighted_f1": 0.6542131341403148, + "macro_precision": 0.6429426157775356, + "macro_recall": 0.7313583518685028, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.6451612903225806, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 65.48223350253807, + "macro_f1_pct": 65.50837484483564, + "weighted_f1_pct": 65.42131341403147, + "macro_precision_pct": 64.29426157775356, + "macro_recall_pct": 73.13583518685029, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 64.51612903225806, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6749414006966611, + "weighted_f1": 0.6574842437449702, + "macro_precision": 0.6905348205348205, + "macro_recall": 0.6701449208224277, + "per_class_f1": { + "FAVOR": 0.6740331491712708, + "AGAINST": 0.5815602836879432, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 67.4941400696661, + "weighted_f1_pct": 65.74842437449702, + "macro_precision_pct": 69.05348205348206, + "macro_recall_pct": 67.01449208224277, + "per_class_f1_pct": { + "FAVOR": 67.40331491712708, + "AGAINST": 58.156028368794324, + "NONE": 76.92307692307693 + } + } + ], + "accuracy": 0.6785260482846251, + "macro_f1": 0.6828484653203835, + "weighted_f1": 0.673080365477855, + "macro_precision": 0.6875155659740639, + "macro_recall": 0.7027316736151069, + "per_class_f1": { + "FAVOR": 0.6841155234657039, + "AGAINST": 0.6168888888888889, + "NONE": 0.7475409836065574 + }, + "accuracy_pct": 67.85260482846252, + "macro_f1_pct": 68.28484653203834, + "weighted_f1_pct": 67.3080365477855, + "macro_precision_pct": 68.75155659740639, + "macro_recall_pct": 70.27316736151069, + "per_class_f1_pct": { + "FAVOR": 68.41155234657039, + "AGAINST": 61.68888888888889, + "NONE": 74.75409836065575 + } + }, + { + "after_task": 10, + "best_epoch": 10, + "best_dev_macro_f1": 0.7144959799827056, + "losses": { + "current_ce": 0.3933915218959252, + "replay_ce": 0.1967419909313321, + "dark_distillation": 0.4252892947445313, + "supervised_contrastive": 4.03174215555191, + "total": 1.0384533358116945, + "replay_multiplier": 0.5866537625997507, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09619238924636181, + "ema_update_magnitude": 0.053022193494251656, + "final_multiplier": 0.5512098608805173 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6721263265567062, + "weighted_f1": 0.8453635718674366, + "macro_precision": 0.6659919028340081, + "macro_recall": 0.7014682684140273, + "per_class_f1": { + "FAVOR": 0.8636363636363636, + "AGAINST": 0.26666666666666666, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 67.21263265567062, + "weighted_f1_pct": 84.53635718674366, + "macro_precision_pct": 66.59919028340082, + "macro_recall_pct": 70.14682684140273, + "per_class_f1_pct": { + "FAVOR": 86.36363636363636, + "AGAINST": 26.666666666666668, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6487946111463073, + "weighted_f1": 0.6645052953722506, + "macro_precision": 0.6479045843689589, + "macro_recall": 0.7155443322109988, + "per_class_f1": { + "FAVOR": 0.5773195876288659, + "AGAINST": 0.6815642458100558, + "NONE": 0.6875 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 64.87946111463073, + "weighted_f1_pct": 66.45052953722505, + "macro_precision_pct": 64.79045843689589, + "macro_recall_pct": 71.55443322109988, + "per_class_f1_pct": { + "FAVOR": 57.73195876288659, + "AGAINST": 68.15642458100558, + "NONE": 68.75 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5626202531645569, + "weighted_f1": 0.5237593153859869, + "macro_precision": 0.5747227658186562, + "macro_recall": 0.6542735042735043, + "per_class_f1": { + "FAVOR": 0.7088607594936709, + "AGAINST": 0.275, + "NONE": 0.704 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.26202531645569, + "weighted_f1_pct": 52.375931538598685, + "macro_precision_pct": 57.47227658186562, + "macro_recall_pct": 65.42735042735043, + "per_class_f1_pct": { + "FAVOR": 70.88607594936708, + "AGAINST": 27.500000000000004, + "NONE": 70.39999999999999 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.64, + "macro_f1": 0.6227004940144377, + "weighted_f1": 0.6328140244424865, + "macro_precision": 0.6755163817663817, + "macro_recall": 0.6472042014210689, + "per_class_f1": { + "FAVOR": 0.6197183098591549, + "AGAINST": 0.5154639175257731, + "NONE": 0.7329192546583851 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 62.27004940144377, + "weighted_f1_pct": 63.28140244424865, + "macro_precision_pct": 67.55163817663818, + "macro_recall_pct": 64.72042014210689, + "per_class_f1_pct": { + "FAVOR": 61.97183098591549, + "AGAINST": 51.546391752577314, + "NONE": 73.29192546583852 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.67, + "macro_f1": 0.628191368546783, + "weighted_f1": 0.6675777020187178, + "macro_precision": 0.6338312553908884, + "macro_recall": 0.6237114557502906, + "per_class_f1": { + "FAVOR": 0.7358490566037735, + "AGAINST": 0.41975308641975306, + "NONE": 0.7289719626168224 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 62.819136854678305, + "weighted_f1_pct": 66.75777020187178, + "macro_precision_pct": 63.38312553908884, + "macro_recall_pct": 62.37114557502906, + "per_class_f1_pct": { + "FAVOR": 73.58490566037736, + "AGAINST": 41.9753086419753, + "NONE": 72.89719626168224 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7153287258550417, + "weighted_f1": 0.7655902937085346, + "macro_precision": 0.6941574485878284, + "macro_recall": 0.7506346826586707, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.8187134502923976, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 71.53287258550417, + "weighted_f1_pct": 76.55902937085347, + "macro_precision_pct": 69.41574485878283, + "macro_recall_pct": 75.06346826586707, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 81.87134502923976, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5752970838369739, + "weighted_f1": 0.5838560925098216, + "macro_precision": 0.5767562840347651, + "macro_recall": 0.6013071895424836, + "per_class_f1": { + "FAVOR": 0.5413533834586466, + "AGAINST": 0.6091954022988506, + "NONE": 0.5753424657534246 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.52970838369739, + "weighted_f1_pct": 58.385609250982164, + "macro_precision_pct": 57.67562840347651, + "macro_recall_pct": 60.130718954248366, + "per_class_f1_pct": { + "FAVOR": 54.13533834586466, + "AGAINST": 60.91954022988506, + "NONE": 57.534246575342465 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6446700507614214, + "macro_f1": 0.642294390389595, + "weighted_f1": 0.6468664974010052, + "macro_precision": 0.6298504167356627, + "macro_recall": 0.7075281708285353, + "per_class_f1": { + "FAVOR": 0.6170212765957447, + "AGAINST": 0.6494845360824743, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 64.46700507614213, + "macro_f1_pct": 64.2294390389595, + "weighted_f1_pct": 64.68664974010052, + "macro_precision_pct": 62.985041673566265, + "macro_recall_pct": 70.75281708285353, + "per_class_f1_pct": { + "FAVOR": 61.702127659574465, + "AGAINST": 64.94845360824742, + "NONE": 66.0377358490566 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6460240385243323, + "weighted_f1": 0.6229879465779312, + "macro_precision": 0.6613284713284714, + "macro_recall": 0.6408635026521206, + "per_class_f1": { + "FAVOR": 0.6298342541436464, + "AGAINST": 0.5390070921985816, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 64.60240385243323, + "weighted_f1_pct": 62.298794657793124, + "macro_precision_pct": 66.13284713284713, + "macro_recall_pct": 64.08635026521206, + "per_class_f1_pct": { + "FAVOR": 62.98342541436463, + "AGAINST": 53.90070921985816, + "NONE": 76.92307692307693 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7399907370975146, + "weighted_f1": 0.7845166301888647, + "macro_precision": 0.7245527553372638, + "macro_recall": 0.7673552112832472, + "per_class_f1": { + "FAVOR": 0.65625, + "AGAINST": 0.7130434782608696, + "NONE": 0.8506787330316742 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 73.99907370975146, + "weighted_f1_pct": 78.45166301888648, + "macro_precision_pct": 72.45527553372638, + "macro_recall_pct": 76.73552112832472, + "per_class_f1_pct": { + "FAVOR": 65.625, + "AGAINST": 71.30434782608695, + "NONE": 85.06787330316742 + } + } + ], + "accuracy": 0.6708004509582863, + "macro_f1": 0.6734813960008085, + "weighted_f1": 0.6675799464291333, + "macro_precision": 0.6789825894845579, + "macro_recall": 0.6889816139518289, + "per_class_f1": { + "FAVOR": 0.6593776282590412, + "AGAINST": 0.6110665597433841, + "NONE": 0.75 + }, + "accuracy_pct": 67.08004509582864, + "macro_f1_pct": 67.34813960008084, + "weighted_f1_pct": 66.75799464291333, + "macro_precision_pct": 67.89825894845579, + "macro_recall_pct": 68.89816139518288, + "per_class_f1_pct": { + "FAVOR": 65.93776282590412, + "AGAINST": 61.10665597433841, + "NONE": 75.0 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6453368029132248, + "average_forgetting": 0.05343025872625525, + "backward_transfer": -0.04238536298915944, + "final_mean_target_macro_f1_pct": 64.53368029132248, + "average_forgetting_pct": 5.3430258726255255, + "backward_transfer_pct": -4.238536298915943, + "final_pooled_accuracy": 0.6708004509582863, + "final_pooled_accuracy_pct": 67.08004509582864, + "final_pooled_macro_f1": 0.6734813960008085, + "final_pooled_macro_f1_pct": 67.34813960008084, + "final_pooled_weighted_f1": 0.6675799464291333, + "final_pooled_weighted_f1_pct": 66.75799464291333, + "final_pooled_macro_precision": 0.6789825894845579, + "final_pooled_macro_precision_pct": 67.89825894845579, + "final_pooled_macro_recall": 0.6889816139518289, + "final_pooled_macro_recall_pct": 68.89816139518288, + "final_pooled_per_class_f1": { + "FAVOR": 0.6593776282590412, + "AGAINST": 0.6110665597433841, + "NONE": 0.75 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 65.93776282590412, + "AGAINST": 61.10665597433841, + "NONE": 75.0 + } + }, + "efficiency": { + "training_wall_time_seconds": 985.6295762639493, + "training_wall_time_minutes": 16.427159604399154, + "peak_gpu_memory_allocated_gb": 10.428888320922852, + "peak_gpu_memory_reserved_gb": 10.6328125 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s2__seed13/metrics.partial.json b/outputs/trace_s8_full__x__s2__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..99aef422d92a67828ec8044c875094ba434acafc --- /dev/null +++ b/outputs/trace_s8_full__x__s2__seed13/metrics.partial.json @@ -0,0 +1,2238 @@ +{ + "matrix": [ + [ + 0.7686582375025478, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7420705909078001, + 0.6927570711075864, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7416944750774578, + 0.6439955106621773, + 0.5689407443418139, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7382595826884623, + 0.6771909302986444, + 0.5831924424017, + 0.7207945042617001, + null, + null, + null, + null, + null, + null + ], + [ + 0.7448627450980392, + 0.680855614973262, + 0.6074954013072029, + 0.7377131090040958, + 0.648070576288398, + null, + null, + null, + null, + null + ], + [ + 0.6673721340388007, + 0.6891100055279159, + 0.5767310789049919, + 0.6999440990887891, + 0.6478919812620556, + 0.712037037037037, + null, + null, + null, + null + ], + [ + 0.7190925780225544, + 0.7008117308195864, + 0.5807626770080497, + 0.7097267060477602, + 0.6048672282958633, + 0.7181333859510656, + 0.5903609109086035, + null, + null, + null + ], + [ + 0.7151868938894642, + 0.6764994389994391, + 0.5842889750570635, + 0.6803735359290916, + 0.6658755909891536, + 0.7007246799168464, + 0.5794731679977582, + 0.7182850767928212, + null, + null + ], + [ + 0.7070527435780515, + 0.6647196374915189, + 0.6038681710970867, + 0.6594512816036775, + 0.6245580324645243, + 0.730108162550359, + 0.5748842592592592, + 0.6550837484483565, + 0.6749414006966611, + null + ], + [ + 0.6721263265567062, + 0.6487946111463073, + 0.5626202531645569, + 0.6227004940144377, + 0.628191368546783, + 0.7153287258550417, + 0.5752970838369739, + 0.642294390389595, + 0.6460240385243323, + 0.7399907370975146 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 15, + "best_dev_macro_f1": 0.7597477727640864, + "losses": { + "current_ce": 0.6096288577431724, + "total": 0.6096288577431724, + "replay_multiplier": 0.7207898354223875, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10138365834772536, + "ema_update_magnitude": 0.06471023128442037, + "final_multiplier": 0.6382708252889969 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 23, + "1": 18, + "2": 23 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + } + ], + "accuracy": 0.911504424778761, + "macro_f1": 0.7686582375025478, + "weighted_f1": 0.9093690590338819, + "macro_precision": 0.7806712962962963, + "macro_recall": 0.7620191724305303, + "per_class_f1": { + "FAVOR": 0.9312977099236641, + "AGAINST": 0.4444444444444444, + "NONE": 0.9302325581395349 + }, + "accuracy_pct": 91.1504424778761, + "macro_f1_pct": 76.86582375025478, + "weighted_f1_pct": 90.9369059033882, + "macro_precision_pct": 78.06712962962963, + "macro_recall_pct": 76.20191724305303, + "per_class_f1_pct": { + "FAVOR": 93.12977099236642, + "AGAINST": 44.44444444444444, + "NONE": 93.02325581395348 + } + }, + { + "after_task": 2, + "best_epoch": 11, + "best_dev_macro_f1": 0.6251337247900822, + "losses": { + "current_ce": 0.6757452957557909, + "replay_ce": 0.1390355658802119, + "dark_distillation": 0.38217191443298804, + "supervised_contrastive": 4.239523621761438, + "total": 1.3066114400372362, + "replay_multiplier": 0.598462906802525, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10362827898120411, + "ema_update_magnitude": 0.05418670752128535, + "final_multiplier": 0.5228949863300695 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 45, + "1": 39, + "2": 44 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.7420705909078001, + "weighted_f1": 0.8640357721444364, + "macro_precision": 0.7182733183295825, + "macro_recall": 0.7843950976823201, + "per_class_f1": { + "FAVOR": 0.8837209302325582, + "AGAINST": 0.46153846153846156, + "NONE": 0.8809523809523809 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 74.20705909078002, + "weighted_f1_pct": 86.40357721444364, + "macro_precision_pct": 71.82733183295825, + "macro_recall_pct": 78.43950976823201, + "per_class_f1_pct": { + "FAVOR": 88.37209302325581, + "AGAINST": 46.15384615384615, + "NONE": 88.09523809523809 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7150537634408602, + "macro_f1": 0.6927570711075864, + "weighted_f1": 0.7215023918549025, + "macro_precision": 0.6767206953253465, + "macro_recall": 0.7300785634118968, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7628865979381443, + "NONE": 0.7 + }, + "accuracy_pct": 71.50537634408603, + "macro_f1_pct": 69.27570711075865, + "weighted_f1_pct": 72.15023918549025, + "macro_precision_pct": 67.67206953253465, + "macro_recall_pct": 73.00785634118968, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.28865979381443, + "NONE": 70.0 + } + } + ], + "accuracy": 0.7692307692307693, + "macro_f1": 0.7697262479871175, + "weighted_f1": 0.7680028436172104, + "macro_precision": 0.7702840531718752, + "macro_recall": 0.7762084105097071, + "per_class_f1": { + "FAVOR": 0.782608695652174, + "AGAINST": 0.7439613526570048, + "NONE": 0.782608695652174 + }, + "accuracy_pct": 76.92307692307693, + "macro_f1_pct": 76.97262479871175, + "weighted_f1_pct": 76.80028436172104, + "macro_precision_pct": 77.02840531718752, + "macro_recall_pct": 77.62084105097071, + "per_class_f1_pct": { + "FAVOR": 78.26086956521739, + "AGAINST": 74.39613526570048, + "NONE": 78.26086956521739 + } + }, + { + "after_task": 3, + "best_epoch": 7, + "best_dev_macro_f1": 0.6063492063492063, + "losses": { + "current_ce": 0.7316161394119263, + "replay_ce": 0.23026843952635925, + "dark_distillation": 0.38139258945981663, + "supervised_contrastive": 4.419250331322352, + "total": 1.4633842170238496, + "replay_multiplier": 0.6681105596417665, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10096908425856493, + "ema_update_magnitude": 0.05408146850315572, + "final_multiplier": 0.5356240368057823 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 67, + "1": 60, + "2": 65 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8584070796460177, + "macro_f1": 0.7416944750774578, + "weighted_f1": 0.8640348370132873, + "macro_precision": 0.7197662601626016, + "macro_recall": 0.7812401407596165, + "per_class_f1": { + "FAVOR": 0.8854961832061069, + "AGAINST": 0.46153846153846156, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 85.84070796460178, + "macro_f1_pct": 74.16944750774577, + "weighted_f1_pct": 86.40348370132872, + "macro_precision_pct": 71.97662601626016, + "macro_recall_pct": 78.12401407596164, + "per_class_f1_pct": { + "FAVOR": 88.54961832061069, + "AGAINST": 46.15384615384615, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6451612903225806, + "macro_f1": 0.6439955106621773, + "weighted_f1": 0.6472140762463343, + "macro_precision": 0.6574643679906838, + "macro_recall": 0.7379910213243547, + "per_class_f1": { + "FAVOR": 0.5858585858585859, + "AGAINST": 0.6424242424242425, + "NONE": 0.7037037037037037 + }, + "accuracy_pct": 64.51612903225806, + "macro_f1_pct": 64.39955106621773, + "weighted_f1_pct": 64.72140762463343, + "macro_precision_pct": 65.74643679906838, + "macro_recall_pct": 73.79910213243546, + "per_class_f1_pct": { + "FAVOR": 58.58585858585859, + "AGAINST": 64.24242424242425, + "NONE": 70.37037037037037 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5689407443418139, + "weighted_f1": 0.5403055180866433, + "macro_precision": 0.5749516908212561, + "macro_recall": 0.6423076923076924, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.32941176470588235, + "NONE": 0.7107438016528925 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.894074434181384, + "weighted_f1_pct": 54.03055180866433, + "macro_precision_pct": 57.49516908212561, + "macro_recall_pct": 64.23076923076924, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 32.94117647058823, + "NONE": 71.07438016528926 + } + } + ], + "accuracy": 0.6802721088435374, + "macro_f1": 0.6728324003213052, + "weighted_f1": 0.660575198571689, + "macro_precision": 0.6963032721622057, + "macro_recall": 0.7072270445479795, + "per_class_f1": { + "FAVOR": 0.7337662337662337, + "AGAINST": 0.532319391634981, + "NONE": 0.752411575562701 + }, + "accuracy_pct": 68.02721088435374, + "macro_f1_pct": 67.28324003213052, + "weighted_f1_pct": 66.0575198571689, + "macro_precision_pct": 69.63032721622056, + "macro_recall_pct": 70.72270445479795, + "per_class_f1_pct": { + "FAVOR": 73.37662337662337, + "AGAINST": 53.2319391634981, + "NONE": 75.2411575562701 + } + }, + { + "after_task": 4, + "best_epoch": 14, + "best_dev_macro_f1": 0.7402164209466441, + "losses": { + "current_ce": 0.6443170191585154, + "replay_ce": 0.2376939520455789, + "dark_distillation": 0.38270324814146844, + "supervised_contrastive": 4.376125502931899, + "total": 1.3424699099167534, + "replay_multiplier": 0.5999306783767159, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.0931216502846001, + "ema_update_magnitude": 0.053369868727746024, + "final_multiplier": 0.5731198766842732 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 89, + "1": 81, + "2": 86 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7382595826884623, + "weighted_f1": 0.8529596215881656, + "macro_precision": 0.709349593495935, + "macro_recall": 0.8329814342919549, + "per_class_f1": { + "FAVOR": 0.8661417322834646, + "AGAINST": 0.47058823529411764, + "NONE": 0.8780487804878049 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 73.82595826884622, + "weighted_f1_pct": 85.29596215881656, + "macro_precision_pct": 70.9349593495935, + "macro_recall_pct": 83.29814342919549, + "per_class_f1_pct": { + "FAVOR": 86.61417322834646, + "AGAINST": 47.05882352941176, + "NONE": 87.8048780487805 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6771909302986444, + "weighted_f1": 0.6939867959317195, + "macro_precision": 0.6639633675383427, + "macro_recall": 0.7356902356902357, + "per_class_f1": { + "FAVOR": 0.627906976744186, + "AGAINST": 0.717391304347826, + "NONE": 0.6862745098039216 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.71909302986444, + "weighted_f1_pct": 69.39867959317195, + "macro_precision_pct": 66.39633675383428, + "macro_recall_pct": 73.56902356902357, + "per_class_f1_pct": { + "FAVOR": 62.7906976744186, + "AGAINST": 71.73913043478261, + "NONE": 68.62745098039215 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5985915492957746, + "macro_f1": 0.5831924424017, + "weighted_f1": 0.5559465321780778, + "macro_precision": 0.5939209726443769, + "macro_recall": 0.6542735042735043, + "per_class_f1": { + "FAVOR": 0.6753246753246753, + "AGAINST": 0.35294117647058826, + "NONE": 0.7213114754098361 + }, + "accuracy_pct": 59.859154929577464, + "macro_f1_pct": 58.31924424017, + "weighted_f1_pct": 55.594653217807775, + "macro_precision_pct": 59.392097264437695, + "macro_recall_pct": 65.42735042735043, + "per_class_f1_pct": { + "FAVOR": 67.53246753246754, + "AGAINST": 35.294117647058826, + "NONE": 72.1311475409836 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.73, + "macro_f1": 0.7207945042617001, + "weighted_f1": 0.7250960189865779, + "macro_precision": 0.7419189466299715, + "macro_recall": 0.7330089589125733, + "per_class_f1": { + "FAVOR": 0.7317073170731707, + "AGAINST": 0.6542056074766355, + "NONE": 0.7764705882352941 + }, + "accuracy_pct": 73.0, + "macro_f1_pct": 72.07945042617001, + "weighted_f1_pct": 72.50960189865779, + "macro_precision_pct": 74.19189466299714, + "macro_recall_pct": 73.30089589125734, + "per_class_f1_pct": { + "FAVOR": 73.17073170731707, + "AGAINST": 65.42056074766354, + "NONE": 77.64705882352942 + } + } + ], + "accuracy": 0.7082683307332294, + "macro_f1": 0.7040371071445786, + "weighted_f1": 0.69931731302485, + "macro_precision": 0.7154450440977955, + "macro_recall": 0.7212885154061626, + "per_class_f1": { + "FAVOR": 0.7409200968523002, + "AGAINST": 0.6106870229007634, + "NONE": 0.7605042016806722 + }, + "accuracy_pct": 70.82683307332293, + "macro_f1_pct": 70.40371071445786, + "weighted_f1_pct": 69.931731302485, + "macro_precision_pct": 71.54450440977955, + "macro_recall_pct": 72.12885154061625, + "per_class_f1_pct": { + "FAVOR": 74.09200968523002, + "AGAINST": 61.06870229007634, + "NONE": 76.05042016806722 + } + }, + { + "after_task": 5, + "best_epoch": 10, + "best_dev_macro_f1": 0.6444444444444444, + "losses": { + "current_ce": 0.656370164339359, + "replay_ce": 0.2301974947635944, + "dark_distillation": 0.44339590607545315, + "supervised_contrastive": 4.260124096503625, + "total": 1.3673993850365662, + "replay_multiplier": 0.6230831125425607, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09820985698124235, + "ema_update_magnitude": 0.056970408125201336, + "final_multiplier": 0.580088494946922 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 111, + "1": 102, + "2": 107 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7448627450980392, + "weighted_f1": 0.8502842269651223, + "macro_precision": 0.7118599791013583, + "macro_recall": 0.8361363912146583, + "per_class_f1": { + "FAVOR": 0.864, + "AGAINST": 0.5, + "NONE": 0.8705882352941177 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 74.48627450980392, + "weighted_f1_pct": 85.02842269651222, + "macro_precision_pct": 71.18599791013584, + "macro_recall_pct": 83.61363912146584, + "per_class_f1_pct": { + "FAVOR": 86.4, + "AGAINST": 50.0, + "NONE": 87.05882352941177 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6935483870967742, + "macro_f1": 0.680855614973262, + "weighted_f1": 0.6995169915473521, + "macro_precision": 0.6660573603611578, + "macro_recall": 0.7344556677890012, + "per_class_f1": { + "FAVOR": 0.6352941176470588, + "AGAINST": 0.7272727272727273, + "NONE": 0.68 + }, + "accuracy_pct": 69.35483870967742, + "macro_f1_pct": 68.0855614973262, + "weighted_f1_pct": 69.95169915473521, + "macro_precision_pct": 66.60573603611579, + "macro_recall_pct": 73.44556677890012, + "per_class_f1_pct": { + "FAVOR": 63.52941176470588, + "AGAINST": 72.72727272727273, + "NONE": 68.0 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6074954013072029, + "weighted_f1": 0.5854366072076317, + "macro_precision": 0.6149814659066131, + "macro_recall": 0.6615384615384615, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.4044943820224719, + "NONE": 0.7419354838709677 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.74954013072029, + "weighted_f1_pct": 58.54366072076317, + "macro_precision_pct": 61.49814659066131, + "macro_recall_pct": 66.15384615384615, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 40.44943820224719, + "NONE": 74.19354838709677 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.75, + "macro_f1": 0.7377131090040958, + "weighted_f1": 0.7449259713516326, + "macro_precision": 0.7601323796975971, + "macro_recall": 0.7453969725054063, + "per_class_f1": { + "FAVOR": 0.7288135593220338, + "AGAINST": 0.6728971962616822, + "NONE": 0.8114285714285714 + }, + "accuracy_pct": 75.0, + "macro_f1_pct": 73.77131090040959, + "weighted_f1_pct": 74.49259713516327, + "macro_precision_pct": 76.0132379697597, + "macro_recall_pct": 74.53969725054063, + "per_class_f1_pct": { + "FAVOR": 72.88135593220339, + "AGAINST": 67.28971962616822, + "NONE": 81.14285714285714 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.685, + "macro_f1": 0.648070576288398, + "weighted_f1": 0.6804998095963443, + "macro_precision": 0.6465941465941466, + "macro_recall": 0.6550666162316648, + "per_class_f1": { + "FAVOR": 0.7326732673267327, + "AGAINST": 0.46153846153846156, + "NONE": 0.75 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 64.8070576288398, + "weighted_f1_pct": 68.04998095963442, + "macro_precision_pct": 64.65941465941467, + "macro_recall_pct": 65.50666162316648, + "per_class_f1_pct": { + "FAVOR": 73.26732673267327, + "AGAINST": 46.15384615384615, + "NONE": 75.0 + } + } + ], + "accuracy": 0.7122473246135553, + "macro_f1": 0.7046884941718682, + "weighted_f1": 0.7045732294686651, + "macro_precision": 0.7146206554110776, + "macro_recall": 0.7124786604191637, + "per_class_f1": { + "FAVOR": 0.7387687188019967, + "AGAINST": 0.6037735849056604, + "NONE": 0.7715231788079471 + }, + "accuracy_pct": 71.22473246135553, + "macro_f1_pct": 70.46884941718682, + "weighted_f1_pct": 70.45732294686651, + "macro_precision_pct": 71.46206554110776, + "macro_recall_pct": 71.24786604191638, + "per_class_f1_pct": { + "FAVOR": 73.87687188019967, + "AGAINST": 60.37735849056604, + "NONE": 77.15231788079471 + } + }, + { + "after_task": 6, + "best_epoch": 12, + "best_dev_macro_f1": 0.8609723609723611, + "losses": { + "current_ce": 0.48282817933639444, + "replay_ce": 0.25347888149597025, + "dark_distillation": 0.4362580687566488, + "supervised_contrastive": 3.7781283725191046, + "total": 1.1614947441206487, + "replay_multiplier": 0.6257369823982759, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10432290151951146, + "ema_update_magnitude": 0.05566255211860548, + "final_multiplier": 0.533560237568689 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.6673721340388007, + "weighted_f1": 0.833113264971672, + "macro_precision": 0.6639433551198257, + "macro_recall": 0.6901832301905109, + "per_class_f1": { + "FAVOR": 0.8592592592592593, + "AGAINST": 0.2857142857142857, + "NONE": 0.8571428571428571 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 66.73721340388006, + "weighted_f1_pct": 83.3113264971672, + "macro_precision_pct": 66.39433551198258, + "macro_recall_pct": 69.01832301905108, + "per_class_f1_pct": { + "FAVOR": 85.92592592592592, + "AGAINST": 28.57142857142857, + "NONE": 85.71428571428571 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7096774193548387, + "macro_f1": 0.6891100055279159, + "weighted_f1": 0.7132883967260472, + "macro_precision": 0.6747162485065711, + "macro_recall": 0.7221099887766554, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7562189054726368, + "NONE": 0.6444444444444445 + }, + "accuracy_pct": 70.96774193548387, + "macro_f1_pct": 68.91100055279159, + "weighted_f1_pct": 71.32883967260472, + "macro_precision_pct": 67.47162485065711, + "macro_recall_pct": 72.21099887766555, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 75.62189054726367, + "NONE": 64.44444444444444 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5767310789049919, + "weighted_f1": 0.5566544192692386, + "macro_precision": 0.5759220354808591, + "macro_recall": 0.6247863247863248, + "per_class_f1": { + "FAVOR": 0.6388888888888888, + "AGAINST": 0.391304347826087, + "NONE": 0.7 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 57.67310789049919, + "weighted_f1_pct": 55.66544192692386, + "macro_precision_pct": 57.592203548085905, + "macro_recall_pct": 62.47863247863248, + "per_class_f1_pct": { + "FAVOR": 63.888888888888886, + "AGAINST": 39.130434782608695, + "NONE": 70.0 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.705, + "macro_f1": 0.6999440990887891, + "weighted_f1": 0.7037065698220367, + "macro_precision": 0.7047619047619048, + "macro_recall": 0.7114766759345073, + "per_class_f1": { + "FAVOR": 0.7049180327868853, + "AGAINST": 0.6495726495726496, + "NONE": 0.7453416149068323 + }, + "accuracy_pct": 70.5, + "macro_f1_pct": 69.99440990887891, + "weighted_f1_pct": 70.37065698220367, + "macro_precision_pct": 70.47619047619048, + "macro_recall_pct": 71.14766759345072, + "per_class_f1_pct": { + "FAVOR": 70.49180327868852, + "AGAINST": 64.95726495726495, + "NONE": 74.53416149068323 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6478919812620556, + "weighted_f1": 0.6700457426288233, + "macro_precision": 0.6420194106075142, + "macro_recall": 0.6618501239860463, + "per_class_f1": { + "FAVOR": 0.7015706806282722, + "AGAINST": 0.5052631578947369, + "NONE": 0.7368421052631579 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.78919812620556, + "weighted_f1_pct": 67.00457426288233, + "macro_precision_pct": 64.20194106075142, + "macro_recall_pct": 66.18501239860463, + "per_class_f1_pct": { + "FAVOR": 70.15706806282722, + "AGAINST": 50.526315789473685, + "NONE": 73.68421052631578 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.712037037037037, + "weighted_f1": 0.7738394216133943, + "macro_precision": 0.7048560135516656, + "macro_recall": 0.7245577211394303, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.8444444444444444, + "NONE": 0.75 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 71.2037037037037, + "weighted_f1_pct": 77.38394216133942, + "macro_precision_pct": 70.48560135516657, + "macro_recall_pct": 72.45577211394303, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 84.44444444444444, + "NONE": 75.0 + } + } + ], + "accuracy": 0.7041540020263425, + "macro_f1": 0.7054055283085097, + "weighted_f1": 0.7030027857496425, + "macro_precision": 0.7045077209685201, + "macro_recall": 0.7094306363476114, + "per_class_f1": { + "FAVOR": 0.711864406779661, + "AGAINST": 0.6695278969957081, + "NONE": 0.7348242811501597 + }, + "accuracy_pct": 70.41540020263425, + "macro_f1_pct": 70.54055283085097, + "weighted_f1_pct": 70.30027857496425, + "macro_precision_pct": 70.450772096852, + "macro_recall_pct": 70.94306363476115, + "per_class_f1_pct": { + "FAVOR": 71.1864406779661, + "AGAINST": 66.95278969957081, + "NONE": 73.48242811501598 + } + }, + { + "after_task": 7, + "best_epoch": 11, + "best_dev_macro_f1": 0.5563271604938271, + "losses": { + "current_ce": 0.5817336595419681, + "replay_ce": 0.22941977074652006, + "dark_distillation": 0.4572236171274474, + "supervised_contrastive": 4.228520169402614, + "total": 1.2951659498792707, + "replay_multiplier": 0.6223886659368644, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10125105852992596, + "ema_update_magnitude": 0.053141566485488054, + "final_multiplier": 0.5248494905342785 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.7190925780225544, + "weighted_f1": 0.8519840137025407, + "macro_precision": 0.6996434174868065, + "macro_recall": 0.831161266836549, + "per_class_f1": { + "FAVOR": 0.8548387096774194, + "AGAINST": 0.4, + "NONE": 0.9024390243902439 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 71.90925780225544, + "weighted_f1_pct": 85.19840137025408, + "macro_precision_pct": 69.96434174868065, + "macro_recall_pct": 83.1161266836549, + "per_class_f1_pct": { + "FAVOR": 85.48387096774194, + "AGAINST": 40.0, + "NONE": 90.2439024390244 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.7258064516129032, + "macro_f1": 0.7008117308195864, + "weighted_f1": 0.730785799356359, + "macro_precision": 0.6856164041010252, + "macro_recall": 0.7276094276094276, + "per_class_f1": { + "FAVOR": 0.631578947368421, + "AGAINST": 0.7761194029850746, + "NONE": 0.6947368421052632 + }, + "accuracy_pct": 72.58064516129032, + "macro_f1_pct": 70.08117308195864, + "weighted_f1_pct": 73.0785799356359, + "macro_precision_pct": 68.56164041010253, + "macro_recall_pct": 72.76094276094275, + "per_class_f1_pct": { + "FAVOR": 63.1578947368421, + "AGAINST": 77.61194029850746, + "NONE": 69.47368421052632 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5807626770080497, + "weighted_f1": 0.5565428530995599, + "macro_precision": 0.5783993783993785, + "macro_recall": 0.6247863247863248, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.3870967741935484, + "NONE": 0.6885245901639344 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.076267700804976, + "weighted_f1_pct": 55.65428530995599, + "macro_precision_pct": 57.839937839937846, + "macro_recall_pct": 62.47863247863248, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 38.70967741935484, + "NONE": 68.85245901639344 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.72, + "macro_f1": 0.7097267060477602, + "weighted_f1": 0.7158708516883039, + "macro_precision": 0.7323499545721767, + "macro_recall": 0.7262588816805685, + "per_class_f1": { + "FAVOR": 0.7131782945736435, + "AGAINST": 0.6355140186915887, + "NONE": 0.7804878048780488 + }, + "accuracy_pct": 72.0, + "macro_f1_pct": 70.97267060477603, + "weighted_f1_pct": 71.58708516883038, + "macro_precision_pct": 73.23499545721766, + "macro_recall_pct": 72.62588816805685, + "per_class_f1_pct": { + "FAVOR": 71.31782945736434, + "AGAINST": 63.55140186915887, + "NONE": 78.04878048780488 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.62, + "macro_f1": 0.6048672282958633, + "weighted_f1": 0.6271031260429414, + "macro_precision": 0.6042937853107344, + "macro_recall": 0.6289721066420096, + "per_class_f1": { + "FAVOR": 0.651685393258427, + "AGAINST": 0.43564356435643564, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 62.0, + "macro_f1_pct": 60.48672282958633, + "weighted_f1_pct": 62.71031260429414, + "macro_precision_pct": 60.42937853107344, + "macro_recall_pct": 62.89721066420096, + "per_class_f1_pct": { + "FAVOR": 65.1685393258427, + "AGAINST": 43.56435643564357, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7808219178082192, + "macro_f1": 0.7181333859510656, + "weighted_f1": 0.7802270401084069, + "macro_precision": 0.7120894278571224, + "macro_recall": 0.7281809095452273, + "per_class_f1": { + "FAVOR": 0.5416666666666666, + "AGAINST": 0.850828729281768, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 78.08219178082192, + "macro_f1_pct": 71.81333859510656, + "weighted_f1_pct": 78.0227040108407, + "macro_precision_pct": 71.20894278571224, + "macro_recall_pct": 72.81809095452273, + "per_class_f1_pct": { + "FAVOR": 54.166666666666664, + "AGAINST": 85.0828729281768, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.6157894736842106, + "macro_f1": 0.5903609109086035, + "weighted_f1": 0.6175335608241325, + "macro_precision": 0.5848731481938209, + "macro_recall": 0.5973129992737837, + "per_class_f1": { + "FAVOR": 0.5321100917431193, + "AGAINST": 0.6834170854271356, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 61.578947368421055, + "macro_f1_pct": 59.03609109086035, + "weighted_f1_pct": 61.75335608241325, + "macro_precision_pct": 58.487314819382085, + "macro_recall_pct": 59.731299927378366, + "per_class_f1_pct": { + "FAVOR": 53.21100917431193, + "AGAINST": 68.34170854271356, + "NONE": 55.55555555555556 + } + } + ], + "accuracy": 0.6890399320305862, + "macro_f1": 0.6909910738875026, + "weighted_f1": 0.6878658072394095, + "macro_precision": 0.6882248456031044, + "macro_recall": 0.696535444939164, + "per_class_f1": { + "FAVOR": 0.6712141882673943, + "AGAINST": 0.6674057649667405, + "NONE": 0.7343532684283728 + }, + "accuracy_pct": 68.90399320305862, + "macro_f1_pct": 69.09910738875026, + "weighted_f1_pct": 68.78658072394094, + "macro_precision_pct": 68.82248456031044, + "macro_recall_pct": 69.6535444939164, + "per_class_f1_pct": { + "FAVOR": 67.12141882673943, + "AGAINST": 66.74057649667405, + "NONE": 73.43532684283728 + } + }, + { + "after_task": 8, + "best_epoch": 14, + "best_dev_macro_f1": 0.7061420877554253, + "losses": { + "current_ce": 0.5435022296327533, + "replay_ce": 0.22035438978310787, + "dark_distillation": 0.433873850107193, + "supervised_contrastive": 4.268619733868223, + "total": 1.2484463727835453, + "replay_multiplier": 0.6279489699287756, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09643344680537284, + "ema_update_magnitude": 0.054508257284998744, + "final_multiplier": 0.5652422379447893 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8230088495575221, + "macro_f1": 0.7151868938894642, + "weighted_f1": 0.8405366004178842, + "macro_precision": 0.6972433262755843, + "macro_recall": 0.8167212716903288, + "per_class_f1": { + "FAVOR": 0.8527131782945736, + "AGAINST": 0.42105263157894735, + "NONE": 0.8717948717948718 + }, + "accuracy_pct": 82.30088495575221, + "macro_f1_pct": 71.51868938894643, + "weighted_f1_pct": 84.05366004178842, + "macro_precision_pct": 69.72433262755843, + "macro_recall_pct": 81.67212716903288, + "per_class_f1_pct": { + "FAVOR": 85.27131782945736, + "AGAINST": 42.10526315789473, + "NONE": 87.17948717948718 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6881720430107527, + "macro_f1": 0.6764994389994391, + "weighted_f1": 0.6938052085229506, + "macro_precision": 0.6650714622317057, + "macro_recall": 0.7394500561167229, + "per_class_f1": { + "FAVOR": 0.6593406593406593, + "AGAINST": 0.7243243243243244, + "NONE": 0.6458333333333334 + }, + "accuracy_pct": 68.81720430107528, + "macro_f1_pct": 67.6499438999439, + "weighted_f1_pct": 69.38052085229506, + "macro_precision_pct": 66.50714622317057, + "macro_recall_pct": 73.94500561167229, + "per_class_f1_pct": { + "FAVOR": 65.93406593406593, + "AGAINST": 72.43243243243244, + "NONE": 64.58333333333334 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5842889750570635, + "weighted_f1": 0.5533534512910313, + "macro_precision": 0.5820540935672515, + "macro_recall": 0.6303418803418804, + "per_class_f1": { + "FAVOR": 0.7058823529411765, + "AGAINST": 0.3695652173913043, + "NONE": 0.6774193548387096 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 58.428897505706345, + "weighted_f1_pct": 55.33534512910313, + "macro_precision_pct": 58.20540935672515, + "macro_recall_pct": 63.03418803418804, + "per_class_f1_pct": { + "FAVOR": 70.58823529411765, + "AGAINST": 36.95652173913043, + "NONE": 67.74193548387096 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.69, + "macro_f1": 0.6803735359290916, + "weighted_f1": 0.687141975308642, + "macro_precision": 0.7012321077672858, + "macro_recall": 0.6962619709607661, + "per_class_f1": { + "FAVOR": 0.676923076923077, + "AGAINST": 0.6111111111111112, + "NONE": 0.7530864197530864 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 68.03735359290916, + "weighted_f1_pct": 68.7141975308642, + "macro_precision_pct": 70.12321077672858, + "macro_recall_pct": 69.62619709607661, + "per_class_f1_pct": { + "FAVOR": 67.6923076923077, + "AGAINST": 61.111111111111114, + "NONE": 75.30864197530865 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.7, + "macro_f1": 0.6658755909891536, + "weighted_f1": 0.6998875189271037, + "macro_precision": 0.6616780045351475, + "macro_recall": 0.6713515179534597, + "per_class_f1": { + "FAVOR": 0.7562189054726368, + "AGAINST": 0.47619047619047616, + "NONE": 0.7652173913043478 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 66.58755909891536, + "weighted_f1_pct": 69.98875189271037, + "macro_precision_pct": 66.16780045351474, + "macro_recall_pct": 67.13515179534598, + "per_class_f1_pct": { + "FAVOR": 75.62189054726367, + "AGAINST": 47.61904761904761, + "NONE": 76.52173913043478 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7465753424657534, + "macro_f1": 0.7007246799168464, + "weighted_f1": 0.7543620120706354, + "macro_precision": 0.6832107843137255, + "macro_recall": 0.7294852573713143, + "per_class_f1": { + "FAVOR": 0.5263157894736842, + "AGAINST": 0.813953488372093, + "NONE": 0.7619047619047619 + }, + "accuracy_pct": 74.65753424657534, + "macro_f1_pct": 70.07246799168463, + "weighted_f1_pct": 75.43620120706353, + "macro_precision_pct": 68.32107843137256, + "macro_recall_pct": 72.94852573713143, + "per_class_f1_pct": { + "FAVOR": 52.63157894736842, + "AGAINST": 81.3953488372093, + "NONE": 76.19047619047619 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5794731679977582, + "weighted_f1": 0.5946979049136082, + "macro_precision": 0.5720211161387632, + "macro_recall": 0.6056644880174292, + "per_class_f1": { + "FAVOR": 0.5409836065573771, + "AGAINST": 0.6333333333333333, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 57.94731679977582, + "weighted_f1_pct": 59.46979049136082, + "macro_precision_pct": 57.20211161387632, + "macro_recall_pct": 60.56644880174292, + "per_class_f1_pct": { + "FAVOR": 54.09836065573771, + "AGAINST": 63.33333333333333, + "NONE": 56.41025641025641 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.7360406091370558, + "macro_f1": 0.7182850767928212, + "weighted_f1": 0.737953049811892, + "macro_precision": 0.705325206219432, + "macro_recall": 0.7356280136082324, + "per_class_f1": { + "FAVOR": 0.676056338028169, + "AGAINST": 0.7685185185185185, + "NONE": 0.7102803738317757 + }, + "accuracy_pct": 73.60406091370558, + "macro_f1_pct": 71.82850767928211, + "weighted_f1_pct": 73.7953049811892, + "macro_precision_pct": 70.53252062194319, + "macro_recall_pct": 73.56280136082324, + "per_class_f1_pct": { + "FAVOR": 67.6056338028169, + "AGAINST": 76.85185185185185, + "NONE": 71.02803738317756 + } + } + ], + "accuracy": 0.6899563318777293, + "macro_f1": 0.6922808591785552, + "weighted_f1": 0.6883392172674279, + "macro_precision": 0.6899051628826104, + "macro_recall": 0.7050388213146378, + "per_class_f1": { + "FAVOR": 0.6927502876869965, + "AGAINST": 0.6647727272727273, + "NONE": 0.7193195625759417 + }, + "accuracy_pct": 68.99563318777294, + "macro_f1_pct": 69.22808591785552, + "weighted_f1_pct": 68.83392172674279, + "macro_precision_pct": 68.99051628826103, + "macro_recall_pct": 70.50388213146378, + "per_class_f1_pct": { + "FAVOR": 69.27502876869966, + "AGAINST": 66.47727272727273, + "NONE": 71.93195625759417 + } + }, + { + "after_task": 9, + "best_epoch": 13, + "best_dev_macro_f1": 0.6423910840932118, + "losses": { + "current_ce": 0.520153154929479, + "replay_ce": 0.20112030771043565, + "dark_distillation": 0.40791797637939453, + "supervised_contrastive": 4.208267084000602, + "total": 1.1806766307543195, + "replay_multiplier": 0.586306728213941, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09318459210348767, + "ema_update_magnitude": 0.05341735456724901, + "final_multiplier": 0.573242350064971 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8407079646017699, + "macro_f1": 0.7070527435780515, + "weighted_f1": 0.8569491019293084, + "macro_precision": 0.6921009098428454, + "macro_recall": 0.7712898920033976, + "per_class_f1": { + "FAVOR": 0.8682170542635659, + "AGAINST": 0.35294117647058826, + "NONE": 0.9 + }, + "accuracy_pct": 84.070796460177, + "macro_f1_pct": 70.70527435780515, + "weighted_f1_pct": 85.69491019293083, + "macro_precision_pct": 69.21009098428455, + "macro_recall_pct": 77.12898920033976, + "per_class_f1_pct": { + "FAVOR": 86.82170542635659, + "AGAINST": 35.294117647058826, + "NONE": 90.0 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6720430107526881, + "macro_f1": 0.6647196374915189, + "weighted_f1": 0.6788558541023741, + "macro_precision": 0.6592714858811398, + "macro_recall": 0.7318181818181818, + "per_class_f1": { + "FAVOR": 0.6105263157894737, + "AGAINST": 0.6961325966850829, + "NONE": 0.6875 + }, + "accuracy_pct": 67.20430107526882, + "macro_f1_pct": 66.47196374915188, + "weighted_f1_pct": 67.88558541023741, + "macro_precision_pct": 65.92714858811398, + "macro_recall_pct": 73.18181818181819, + "per_class_f1_pct": { + "FAVOR": 61.05263157894737, + "AGAINST": 69.61325966850829, + "NONE": 68.75 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.6197183098591549, + "macro_f1": 0.6038681710970867, + "weighted_f1": 0.5722183165929984, + "macro_precision": 0.6245985115550332, + "macro_recall": 0.6782051282051281, + "per_class_f1": { + "FAVOR": 0.72, + "AGAINST": 0.3614457831325301, + "NONE": 0.7301587301587301 + }, + "accuracy_pct": 61.97183098591549, + "macro_f1_pct": 60.386817109708666, + "weighted_f1_pct": 57.22183165929984, + "macro_precision_pct": 62.45985115550332, + "macro_recall_pct": 67.82051282051282, + "per_class_f1_pct": { + "FAVOR": 72.0, + "AGAINST": 36.144578313253014, + "NONE": 73.01587301587301 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.685, + "macro_f1": 0.6594512816036775, + "weighted_f1": 0.6718573797678274, + "macro_precision": 0.7142153239714215, + "macro_recall": 0.683348779734322, + "per_class_f1": { + "FAVOR": 0.6567164179104478, + "AGAINST": 0.5263157894736842, + "NONE": 0.7953216374269005 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 65.94512816036774, + "weighted_f1_pct": 67.18573797678275, + "macro_precision_pct": 71.42153239714216, + "macro_recall_pct": 68.3348779734322, + "per_class_f1_pct": { + "FAVOR": 65.67164179104478, + "AGAINST": 52.63157894736842, + "NONE": 79.53216374269006 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.69, + "macro_f1": 0.6245580324645243, + "weighted_f1": 0.6745694893502369, + "macro_precision": 0.6381158835786785, + "macro_recall": 0.624452570083638, + "per_class_f1": { + "FAVOR": 0.7579908675799086, + "AGAINST": 0.34782608695652173, + "NONE": 0.7678571428571429 + }, + "accuracy_pct": 69.0, + "macro_f1_pct": 62.45580324645243, + "weighted_f1_pct": 67.45694893502369, + "macro_precision_pct": 63.811588357867855, + "macro_recall_pct": 62.4452570083638, + "per_class_f1_pct": { + "FAVOR": 75.79908675799086, + "AGAINST": 34.78260869565217, + "NONE": 76.78571428571429 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.773972602739726, + "macro_f1": 0.730108162550359, + "weighted_f1": 0.7792938240989353, + "macro_precision": 0.7105820105820105, + "macro_recall": 0.759720139930035, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.8323699421965318, + "NONE": 0.8125 + }, + "accuracy_pct": 77.3972602739726, + "macro_f1_pct": 73.0108162550359, + "weighted_f1_pct": 77.92938240989353, + "macro_precision_pct": 71.05820105820105, + "macro_recall_pct": 75.9720139930035, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 83.23699421965318, + "NONE": 81.25 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5842105263157895, + "macro_f1": 0.5748842592592592, + "weighted_f1": 0.5888779239766082, + "macro_precision": 0.572412493465125, + "macro_recall": 0.5951343500363109, + "per_class_f1": { + "FAVOR": 0.546875, + "AGAINST": 0.6222222222222222, + "NONE": 0.5555555555555556 + }, + "accuracy_pct": 58.42105263157895, + "macro_f1_pct": 57.488425925925924, + "weighted_f1_pct": 58.88779239766082, + "macro_precision_pct": 57.241249346512504, + "macro_recall_pct": 59.51343500363109, + "per_class_f1_pct": { + "FAVOR": 54.6875, + "AGAINST": 62.22222222222222, + "NONE": 55.55555555555556 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6548223350253807, + "macro_f1": 0.6550837484483565, + "weighted_f1": 0.6542131341403148, + "macro_precision": 0.6429426157775356, + "macro_recall": 0.7313583518685028, + "per_class_f1": { + "FAVOR": 0.6304347826086957, + "AGAINST": 0.6451612903225806, + "NONE": 0.6896551724137931 + }, + "accuracy_pct": 65.48223350253807, + "macro_f1_pct": 65.50837484483564, + "weighted_f1_pct": 65.42131341403147, + "macro_precision_pct": 64.29426157775356, + "macro_recall_pct": 73.13583518685029, + "per_class_f1_pct": { + "FAVOR": 63.04347826086957, + "AGAINST": 64.51612903225806, + "NONE": 68.96551724137932 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6749414006966611, + "weighted_f1": 0.6574842437449702, + "macro_precision": 0.6905348205348205, + "macro_recall": 0.6701449208224277, + "per_class_f1": { + "FAVOR": 0.6740331491712708, + "AGAINST": 0.5815602836879432, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 67.4941400696661, + "weighted_f1_pct": 65.74842437449702, + "macro_precision_pct": 69.05348205348206, + "macro_recall_pct": 67.01449208224277, + "per_class_f1_pct": { + "FAVOR": 67.40331491712708, + "AGAINST": 58.156028368794324, + "NONE": 76.92307692307693 + } + } + ], + "accuracy": 0.6785260482846251, + "macro_f1": 0.6828484653203835, + "weighted_f1": 0.673080365477855, + "macro_precision": 0.6875155659740639, + "macro_recall": 0.7027316736151069, + "per_class_f1": { + "FAVOR": 0.6841155234657039, + "AGAINST": 0.6168888888888889, + "NONE": 0.7475409836065574 + }, + "accuracy_pct": 67.85260482846252, + "macro_f1_pct": 68.28484653203834, + "weighted_f1_pct": 67.3080365477855, + "macro_precision_pct": 68.75155659740639, + "macro_recall_pct": 70.27316736151069, + "per_class_f1_pct": { + "FAVOR": 68.41155234657039, + "AGAINST": 61.68888888888889, + "NONE": 74.75409836065575 + } + }, + { + "after_task": 10, + "best_epoch": 10, + "best_dev_macro_f1": 0.7144959799827056, + "losses": { + "current_ce": 0.3933915218959252, + "replay_ce": 0.1967419909313321, + "dark_distillation": 0.4252892947445313, + "supervised_contrastive": 4.03174215555191, + "total": 1.0384533358116945, + "replay_multiplier": 0.5866537625997507, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09619238924636181, + "ema_update_magnitude": 0.053022193494251656, + "final_multiplier": 0.5512098608805173 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 23, + "1": 18, + "2": 23 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 22, + "1": 21, + "2": 21 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Climate Change is a Real Concern", + "accuracy": 0.831858407079646, + "macro_f1": 0.6721263265567062, + "weighted_f1": 0.8453635718674366, + "macro_precision": 0.6659919028340081, + "macro_recall": 0.7014682684140273, + "per_class_f1": { + "FAVOR": 0.8636363636363636, + "AGAINST": 0.26666666666666666, + "NONE": 0.8860759493670886 + }, + "accuracy_pct": 83.1858407079646, + "macro_f1_pct": 67.21263265567062, + "weighted_f1_pct": 84.53635718674366, + "macro_precision_pct": 66.59919028340082, + "macro_recall_pct": 70.14682684140273, + "per_class_f1_pct": { + "FAVOR": 86.36363636363636, + "AGAINST": 26.666666666666668, + "NONE": 88.60759493670885 + } + }, + { + "task_index": 1, + "target": "Legalization of Abortion", + "accuracy": 0.6559139784946236, + "macro_f1": 0.6487946111463073, + "weighted_f1": 0.6645052953722506, + "macro_precision": 0.6479045843689589, + "macro_recall": 0.7155443322109988, + "per_class_f1": { + "FAVOR": 0.5773195876288659, + "AGAINST": 0.6815642458100558, + "NONE": 0.6875 + }, + "accuracy_pct": 65.59139784946237, + "macro_f1_pct": 64.87946111463073, + "weighted_f1_pct": 66.45052953722505, + "macro_precision_pct": 64.79045843689589, + "macro_recall_pct": 71.55443322109988, + "per_class_f1_pct": { + "FAVOR": 57.73195876288659, + "AGAINST": 68.15642458100558, + "NONE": 68.75 + } + }, + { + "task_index": 2, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5626202531645569, + "weighted_f1": 0.5237593153859869, + "macro_precision": 0.5747227658186562, + "macro_recall": 0.6542735042735043, + "per_class_f1": { + "FAVOR": 0.7088607594936709, + "AGAINST": 0.275, + "NONE": 0.704 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.26202531645569, + "weighted_f1_pct": 52.375931538598685, + "macro_precision_pct": 57.47227658186562, + "macro_recall_pct": 65.42735042735043, + "per_class_f1_pct": { + "FAVOR": 70.88607594936708, + "AGAINST": 27.500000000000004, + "NONE": 70.39999999999999 + } + }, + { + "task_index": 3, + "target": "fauci", + "accuracy": 0.64, + "macro_f1": 0.6227004940144377, + "weighted_f1": 0.6328140244424865, + "macro_precision": 0.6755163817663817, + "macro_recall": 0.6472042014210689, + "per_class_f1": { + "FAVOR": 0.6197183098591549, + "AGAINST": 0.5154639175257731, + "NONE": 0.7329192546583851 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 62.27004940144377, + "weighted_f1_pct": 63.28140244424865, + "macro_precision_pct": 67.55163817663818, + "macro_recall_pct": 64.72042014210689, + "per_class_f1_pct": { + "FAVOR": 61.97183098591549, + "AGAINST": 51.546391752577314, + "NONE": 73.29192546583852 + } + }, + { + "task_index": 4, + "target": "school closures", + "accuracy": 0.67, + "macro_f1": 0.628191368546783, + "weighted_f1": 0.6675777020187178, + "macro_precision": 0.6338312553908884, + "macro_recall": 0.6237114557502906, + "per_class_f1": { + "FAVOR": 0.7358490566037735, + "AGAINST": 0.41975308641975306, + "NONE": 0.7289719626168224 + }, + "accuracy_pct": 67.0, + "macro_f1_pct": 62.819136854678305, + "weighted_f1_pct": 66.75777020187178, + "macro_precision_pct": 63.38312553908884, + "macro_recall_pct": 62.37114557502906, + "per_class_f1_pct": { + "FAVOR": 73.58490566037736, + "AGAINST": 41.9753086419753, + "NONE": 72.89719626168224 + } + }, + { + "task_index": 5, + "target": "Atheism", + "accuracy": 0.7602739726027398, + "macro_f1": 0.7153287258550417, + "weighted_f1": 0.7655902937085346, + "macro_precision": 0.6941574485878284, + "macro_recall": 0.7506346826586707, + "per_class_f1": { + "FAVOR": 0.509090909090909, + "AGAINST": 0.8187134502923976, + "NONE": 0.8181818181818182 + }, + "accuracy_pct": 76.02739726027397, + "macro_f1_pct": 71.53287258550417, + "weighted_f1_pct": 76.55902937085347, + "macro_precision_pct": 69.41574485878283, + "macro_recall_pct": 75.06346826586707, + "per_class_f1_pct": { + "FAVOR": 50.90909090909091, + "AGAINST": 81.87134502923976, + "NONE": 81.81818181818183 + } + }, + { + "task_index": 6, + "target": "Feminist Movement", + "accuracy": 0.5789473684210527, + "macro_f1": 0.5752970838369739, + "weighted_f1": 0.5838560925098216, + "macro_precision": 0.5767562840347651, + "macro_recall": 0.6013071895424836, + "per_class_f1": { + "FAVOR": 0.5413533834586466, + "AGAINST": 0.6091954022988506, + "NONE": 0.5753424657534246 + }, + "accuracy_pct": 57.89473684210527, + "macro_f1_pct": 57.52970838369739, + "weighted_f1_pct": 58.385609250982164, + "macro_precision_pct": 57.67562840347651, + "macro_recall_pct": 60.130718954248366, + "per_class_f1_pct": { + "FAVOR": 54.13533834586466, + "AGAINST": 60.91954022988506, + "NONE": 57.534246575342465 + } + }, + { + "task_index": 7, + "target": "Hillary Clinton", + "accuracy": 0.6446700507614214, + "macro_f1": 0.642294390389595, + "weighted_f1": 0.6468664974010052, + "macro_precision": 0.6298504167356627, + "macro_recall": 0.7075281708285353, + "per_class_f1": { + "FAVOR": 0.6170212765957447, + "AGAINST": 0.6494845360824743, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 64.46700507614213, + "macro_f1_pct": 64.2294390389595, + "weighted_f1_pct": 64.68664974010052, + "macro_precision_pct": 62.985041673566265, + "macro_recall_pct": 70.75281708285353, + "per_class_f1_pct": { + "FAVOR": 61.702127659574465, + "AGAINST": 64.94845360824742, + "NONE": 66.0377358490566 + } + }, + { + "task_index": 8, + "target": "face masks", + "accuracy": 0.625, + "macro_f1": 0.6460240385243323, + "weighted_f1": 0.6229879465779312, + "macro_precision": 0.6613284713284714, + "macro_recall": 0.6408635026521206, + "per_class_f1": { + "FAVOR": 0.6298342541436464, + "AGAINST": 0.5390070921985816, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 62.5, + "macro_f1_pct": 64.60240385243323, + "weighted_f1_pct": 62.298794657793124, + "macro_precision_pct": 66.13284713284713, + "macro_recall_pct": 64.08635026521206, + "per_class_f1_pct": { + "FAVOR": 62.98342541436463, + "AGAINST": 53.90070921985816, + "NONE": 76.92307692307693 + } + }, + { + "task_index": 9, + "target": "stay at home orders", + "accuracy": 0.78, + "macro_f1": 0.7399907370975146, + "weighted_f1": 0.7845166301888647, + "macro_precision": 0.7245527553372638, + "macro_recall": 0.7673552112832472, + "per_class_f1": { + "FAVOR": 0.65625, + "AGAINST": 0.7130434782608696, + "NONE": 0.8506787330316742 + }, + "accuracy_pct": 78.0, + "macro_f1_pct": 73.99907370975146, + "weighted_f1_pct": 78.45166301888648, + "macro_precision_pct": 72.45527553372638, + "macro_recall_pct": 76.73552112832472, + "per_class_f1_pct": { + "FAVOR": 65.625, + "AGAINST": 71.30434782608695, + "NONE": 85.06787330316742 + } + } + ], + "accuracy": 0.6708004509582863, + "macro_f1": 0.6734813960008085, + "weighted_f1": 0.6675799464291333, + "macro_precision": 0.6789825894845579, + "macro_recall": 0.6889816139518289, + "per_class_f1": { + "FAVOR": 0.6593776282590412, + "AGAINST": 0.6110665597433841, + "NONE": 0.75 + }, + "accuracy_pct": 67.08004509582864, + "macro_f1_pct": 67.34813960008084, + "weighted_f1_pct": 66.75799464291333, + "macro_precision_pct": 67.89825894845579, + "macro_recall_pct": 68.89816139518288, + "per_class_f1_pct": { + "FAVOR": 65.93776282590412, + "AGAINST": 61.10665597433841, + "NONE": 75.0 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s2__seed13/resolved_config.yaml b/outputs/trace_s8_full__x__s2__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..50ef1888c1b896800d58f7d17e4335cd7e7a2025 --- /dev/null +++ b/outputs/trace_s8_full__x__s2__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s2 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__x__s2__seed13/run.log b/outputs/trace_s8_full__x__s2__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..0dce77c9a6e213f7073f695fc6a64ba2a412a378 --- /dev/null +++ b/outputs/trace_s8_full__x__s2__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 10:39:59,357 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__x__s2__seed13 device=cuda +2026-09-02 10:56:26,406 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6453368029132248, 'average_forgetting': 0.05343025872625525, 'backward_transfer': -0.04238536298915944, 'final_mean_target_macro_f1_pct': 64.53368029132248, 'average_forgetting_pct': 5.3430258726255255, 'backward_transfer_pct': -4.238536298915943, 'final_pooled_accuracy': 0.6708004509582863, 'final_pooled_accuracy_pct': 67.08004509582864, 'final_pooled_macro_f1': 0.6734813960008085, 'final_pooled_macro_f1_pct': 67.34813960008084, 'final_pooled_weighted_f1': 0.6675799464291333, 'final_pooled_weighted_f1_pct': 66.75799464291333, 'final_pooled_macro_precision': 0.6789825894845579, 'final_pooled_macro_precision_pct': 67.89825894845579, 'final_pooled_macro_recall': 0.6889816139518289, 'final_pooled_macro_recall_pct': 68.89816139518288, 'final_pooled_per_class_f1': {'FAVOR': 0.6593776282590412, 'AGAINST': 0.6110665597433841, 'NONE': 0.75}, 'final_pooled_per_class_f1_pct': {'FAVOR': 65.93776282590412, 'AGAINST': 61.10665597433841, 'NONE': 75.0}} +2026-09-02 10:56:26,407 [INFO] __main__: efficiency={'training_wall_time_seconds': 985.6295762639493, 'training_wall_time_minutes': 16.427159604399154, 'peak_gpu_memory_allocated_gb': 10.428888320922852, 'peak_gpu_memory_reserved_gb': 10.6328125} diff --git a/outputs/trace_s8_full__x__s3__seed13/environment.json b/outputs/trace_s8_full__x__s3__seed13/environment.json new file mode 100644 index 0000000000000000000000000000000000000000..5258930cb2ef9b500620ce39b688e2bf6d302a9d --- /dev/null +++ b/outputs/trace_s8_full__x__s3__seed13/environment.json @@ -0,0 +1,11 @@ +{ + "started_at_utc": "2026-09-02T10:56:34.423305+00:00", + "python": "3.12.14", + "platform": "Linux-5.15.0-171-generic-x86_64-with-glibc2.39", + "torch": "2.6.0+cu124", + "transformers": "4.56.0", + "cuda_available": true, + "cuda_version": "12.4", + "device": "cuda", + "gpu": "NVIDIA GeForce RTX 3090" +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s3__seed13/metrics.json b/outputs/trace_s8_full__x__s3__seed13/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c7df6398ce7fb80eb58a0ca88c3352ccdf2080e9 --- /dev/null +++ b/outputs/trace_s8_full__x__s3__seed13/metrics.json @@ -0,0 +1,2354 @@ +{ + "experiment": "trace_s8_full", + "protocol": "paper_target_aware", + "sequence": "s3", + "corpus": "x", + "seed": 13, + "architecture": { + "stage": 8, + "target_aware_fusion": true, + "class_balanced_objective": true, + "lora": true, + "lora_module_count": 24, + "replay": true, + "dark_distillation": true, + "supervised_contrastive": true, + "update_aware_controller": true + }, + "parameters": { + "total_parameters": 114599555, + "trainable_parameters": 5117315, + "trainable_fraction": 0.04465388194570215 + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.596579966072294, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6990980034458295, + 0.6226705479316422, + 0.7030940878766966, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6983831958843316, + 0.5887419960412661, + 0.6953357100415923, + 0.6910028993447938, + null, + null, + null, + null, + null, + null + ], + [ + 0.6898943919398476, + 0.5924510399490873, + 0.6839596533391595, + 0.6901105162171447, + 0.7829186016958943, + null, + null, + null, + null, + null + ], + [ + 0.6679569892473118, + 0.6158155458878143, + 0.693494671177624, + 0.6758620689655173, + 0.7675482787036007, + 0.694829278080315, + null, + null, + null, + null + ], + [ + 0.665124716553288, + 0.6320258090829, + 0.6824479213828535, + 0.6841540113514343, + 0.7585303838077792, + 0.6946198630786444, + 0.6972286972286973, + null, + null, + null + ], + [ + 0.715940355940356, + 0.6206646003256173, + 0.6910035867217273, + 0.6525543883250572, + 0.8251524238284986, + 0.6380121816168328, + 0.7118306758753024, + 0.609984111148062, + null, + null + ], + [ + 0.6985476859346945, + 0.620980870980871, + 0.6612307023265926, + 0.7049329501915708, + 0.7979534767805374, + 0.6124622565095191, + 0.7306397306397306, + 0.5686327841291522, + 0.6771068992042775, + null + ], + [ + 0.6892608201217474, + 0.5987190197716513, + 0.6796858396858397, + 0.6920975323149235, + 0.8079288302412034, + 0.6887736414206246, + 0.7340101522842639, + 0.5623136123136123, + 0.6845790852219841, + 0.6447329538384969 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 0.6862624481284426, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10795797371276879, + "ema_update_magnitude": 0.06257687394518485, + "final_multiplier": 0.5796410565436867 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6269608881868326, + "losses": { + "current_ce": 0.7093145791328315, + "replay_ce": 0.22077340695894127, + "dark_distillation": 0.38159919065056425, + "supervised_contrastive": 4.350307623545329, + "total": 1.4104225379047972, + "replay_multiplier": 0.6142165815115603, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.11817399906990574, + "ema_update_magnitude": 0.05939854852669155, + "final_multiplier": 0.5026363582022335 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.596579966072294, + "weighted_f1": 0.6147010297132592, + "macro_precision": 0.5931935931935932, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.59375, + "AGAINST": 0.6480446927374302, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 59.6579966072294, + "weighted_f1_pct": 61.47010297132592, + "macro_precision_pct": 59.31935931935932, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 59.375, + "AGAINST": 64.80446927374301, + "NONE": 54.794520547945204 + } + } + ], + "accuracy": 0.6666666666666666, + "macro_f1": 0.654621435594887, + "weighted_f1": 0.6743244310998735, + "macro_precision": 0.6442073608079101, + "macro_recall": 0.6928679485772008, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.7138643067846607, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.4621435594887, + "weighted_f1_pct": 67.43244310998735, + "macro_precision_pct": 64.420736080791, + "macro_recall_pct": 69.28679485772008, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 71.38643067846607, + "NONE": 66.66666666666666 + } + }, + { + "after_task": 3, + "best_epoch": 9, + "best_dev_macro_f1": 0.7630211318735909, + "losses": { + "current_ce": 0.6028909276832234, + "replay_ce": 0.1996306974779476, + "dark_distillation": 0.3657441600705638, + "supervised_contrastive": 4.320868700200861, + "total": 1.2792593280474345, + "replay_multiplier": 0.6254400971576703, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09795596788232877, + "ema_update_magnitude": 0.05818585290700524, + "final_multiplier": 0.5940000815152167 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6990980034458295, + "weighted_f1": 0.7393428489318901, + "macro_precision": 0.6821707709830722, + "macro_recall": 0.7556171914042977, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.782608695652174, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.90980034458295, + "weighted_f1_pct": 73.934284893189, + "macro_precision_pct": 68.21707709830721, + "macro_recall_pct": 75.56171914042977, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 78.26086956521739, + "NONE": 76.92307692307693 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6226705479316422, + "weighted_f1": 0.6366292276762126, + "macro_precision": 0.6158044075919964, + "macro_recall": 0.6470588235294118, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6740331491712708, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 62.26705479316422, + "weighted_f1_pct": 63.662922767621254, + "macro_precision_pct": 61.58044075919964, + "macro_recall_pct": 64.70588235294117, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 67.40331491712708, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.7030940878766966, + "weighted_f1": 0.7182840958056143, + "macro_precision": 0.7044970334412083, + "macro_recall": 0.7095262883914627, + "per_class_f1": { + "FAVOR": 0.6984126984126984, + "AGAINST": 0.75, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 70.30940878766965, + "weighted_f1_pct": 71.82840958056143, + "macro_precision_pct": 70.44970334412083, + "macro_recall_pct": 70.95262883914627, + "per_class_f1_pct": { + "FAVOR": 69.84126984126983, + "AGAINST": 75.0, + "NONE": 66.08695652173913 + } + } + ], + "accuracy": 0.6885553470919324, + "macro_f1": 0.6700224476150556, + "weighted_f1": 0.6937263262644933, + "macro_precision": 0.6552230793139499, + "macro_recall": 0.7002366121601508, + "per_class_f1": { + "FAVOR": 0.6007905138339921, + "AGAINST": 0.7347670250896058, + "NONE": 0.6745098039215687 + }, + "accuracy_pct": 68.85553470919325, + "macro_f1_pct": 67.00224476150555, + "weighted_f1_pct": 69.37263262644933, + "macro_precision_pct": 65.52230793139499, + "macro_recall_pct": 70.02366121601507, + "per_class_f1_pct": { + "FAVOR": 60.079051383399204, + "AGAINST": 73.47670250896059, + "NONE": 67.45098039215686 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.6808621014902849, + "losses": { + "current_ce": 0.6186725724311102, + "replay_ce": 0.2190132564968533, + "dark_distillation": 0.3627972994058851, + "supervised_contrastive": 4.313315141011798, + "total": 1.2969302307991755, + "replay_multiplier": 0.6125457563861113, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09427406151524267, + "ema_update_magnitude": 0.06213868220273308, + "final_multiplier": 0.6591280910570105 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6983831958843316, + "weighted_f1": 0.7347945922161876, + "macro_precision": 0.6876130789174267, + "macro_recall": 0.7714142928535731, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.7741935483870968, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.83831958843317, + "weighted_f1_pct": 73.47945922161875, + "macro_precision_pct": 68.76130789174267, + "macro_recall_pct": 77.1414292853573, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 77.41935483870968, + "NONE": 75.75757575757575 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5887419960412661, + "weighted_f1": 0.5987730492071637, + "macro_precision": 0.597150506789061, + "macro_recall": 0.6292665214233842, + "per_class_f1": { + "FAVOR": 0.583941605839416, + "AGAINST": 0.6181818181818182, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.87419960412661, + "weighted_f1_pct": 59.87730492071637, + "macro_precision_pct": 59.7150506789061, + "macro_recall_pct": 62.926652142338426, + "per_class_f1_pct": { + "FAVOR": 58.3941605839416, + "AGAINST": 61.81818181818181, + "NONE": 56.41025641025641 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6953357100415923, + "weighted_f1": 0.7151392998362243, + "macro_precision": 0.6774747070455689, + "macro_recall": 0.73518106625863, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.7450980392156863, + "NONE": 0.6909090909090909 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.53357100415923, + "weighted_f1_pct": 71.51392998362243, + "macro_precision_pct": 67.74747070455689, + "macro_recall_pct": 73.518106625863, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 74.50980392156863, + "NONE": 69.0909090909091 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6910028993447938, + "weighted_f1": 0.6833319635641394, + "macro_precision": 0.6818863397810766, + "macro_recall": 0.7082049737876296, + "per_class_f1": { + "FAVOR": 0.6751592356687898, + "AGAINST": 0.6666666666666666, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.10028993447938, + "weighted_f1_pct": 68.33319635641394, + "macro_precision_pct": 68.18863397810766, + "macro_recall_pct": 70.82049737876297, + "per_class_f1_pct": { + "FAVOR": 67.51592356687898, + "AGAINST": 66.66666666666666, + "NONE": 73.11827956989248 + } + } + ], + "accuracy": 0.6753069577080492, + "macro_f1": 0.671287386205036, + "weighted_f1": 0.6789068243762634, + "macro_precision": 0.6638160574687392, + "macro_recall": 0.701177868318099, + "per_class_f1": { + "FAVOR": 0.6247191011235955, + "AGAINST": 0.7032640949554896, + "NONE": 0.6858789625360231 + }, + "accuracy_pct": 67.53069577080491, + "macro_f1_pct": 67.1287386205036, + "weighted_f1_pct": 67.89068243762634, + "macro_precision_pct": 66.38160574687392, + "macro_recall_pct": 70.11778683180991, + "per_class_f1_pct": { + "FAVOR": 62.47191011235955, + "AGAINST": 70.32640949554896, + "NONE": 68.58789625360231 + } + }, + { + "after_task": 5, + "best_epoch": 6, + "best_dev_macro_f1": 0.6654093524831955, + "losses": { + "current_ce": 0.4358942994537453, + "replay_ce": 0.21232219041946035, + "dark_distillation": 0.3923066610470414, + "supervised_contrastive": 4.081723755598068, + "total": 1.0873236504693826, + "replay_multiplier": 0.5900478693791011, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10032757414191823, + "ema_update_magnitude": 0.05437592970889881, + "final_multiplier": 0.541983897985826 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6898943919398476, + "weighted_f1": 0.7150434776951248, + "macro_precision": 0.6759683950263375, + "macro_recall": 0.7799100449775113, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.7417218543046358, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.98943919398477, + "weighted_f1_pct": 71.50434776951248, + "macro_precision_pct": 67.59683950263376, + "macro_recall_pct": 77.99100449775112, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 74.17218543046357, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5924510399490873, + "weighted_f1": 0.5918574567112913, + "macro_precision": 0.5981640310908604, + "macro_recall": 0.6296296296296297, + "per_class_f1": { + "FAVOR": 0.5652173913043478, + "AGAINST": 0.5988023952095808, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 59.24510399490873, + "weighted_f1_pct": 59.18574567112913, + "macro_precision_pct": 59.81640310908604, + "macro_recall_pct": 62.96296296296296, + "per_class_f1_pct": { + "FAVOR": 56.52173913043478, + "AGAINST": 59.88023952095808, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6954314720812182, + "macro_f1": 0.6839596533391595, + "weighted_f1": 0.6995439013651351, + "macro_precision": 0.6665919305624278, + "macro_recall": 0.7299176039415497, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.7236180904522613, + "NONE": 0.6782608695652174 + }, + "accuracy_pct": 69.54314720812182, + "macro_f1_pct": 68.39596533391595, + "weighted_f1_pct": 69.95439013651351, + "macro_precision_pct": 66.65919305624278, + "macro_recall_pct": 72.99176039415497, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 72.36180904522614, + "NONE": 67.82608695652173 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6901105162171447, + "weighted_f1": 0.6843955961088908, + "macro_precision": 0.6838810641627543, + "macro_recall": 0.7003335417969564, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6845637583892618, + "NONE": 0.7191011235955056 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.01105162171447, + "weighted_f1_pct": 68.43955961088908, + "macro_precision_pct": 68.38810641627543, + "macro_recall_pct": 70.03335417969564, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 68.45637583892618, + "NONE": 71.91011235955057 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7829186016958943, + "weighted_f1": 0.809463159353989, + "macro_precision": 0.7716959064327487, + "macro_recall": 0.8073907490699095, + "per_class_f1": { + "FAVOR": 0.7619047619047619, + "AGAINST": 0.7222222222222222, + "NONE": 0.8646288209606987 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 78.29186016958943, + "weighted_f1_pct": 80.9463159353989, + "macro_precision_pct": 77.16959064327486, + "macro_recall_pct": 80.73907490699095, + "per_class_f1_pct": { + "FAVOR": 76.19047619047619, + "AGAINST": 72.22222222222221, + "NONE": 86.46288209606988 + } + } + ], + "accuracy": 0.6977491961414791, + "macro_f1": 0.6959287363606136, + "weighted_f1": 0.6984004879695633, + "macro_precision": 0.6924002097176563, + "macro_recall": 0.7202300152262531, + "per_class_f1": { + "FAVOR": 0.6379647749510763, + "AGAINST": 0.6925064599483204, + "NONE": 0.7573149741824441 + }, + "accuracy_pct": 69.77491961414792, + "macro_f1_pct": 69.59287363606136, + "weighted_f1_pct": 69.84004879695634, + "macro_precision_pct": 69.24002097176563, + "macro_recall_pct": 72.02300152262531, + "per_class_f1_pct": { + "FAVOR": 63.79647749510763, + "AGAINST": 69.25064599483204, + "NONE": 75.7314974182444 + } + }, + { + "after_task": 6, + "best_epoch": 5, + "best_dev_macro_f1": 0.6718427273858404, + "losses": { + "current_ce": 0.3773283552201021, + "replay_ce": 0.23725067909274783, + "dark_distillation": 0.4199893119789305, + "supervised_contrastive": 3.9003202427001225, + "total": 1.0871713672365462, + "replay_multiplier": 0.70590457921482, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09754419038195396, + "ema_update_magnitude": 0.059936500760013264, + "final_multiplier": 0.6144548488774144 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6917808219178082, + "macro_f1": 0.6679569892473118, + "weighted_f1": 0.7003711886875827, + "macro_precision": 0.6503926286534982, + "macro_recall": 0.7435332333833083, + "per_class_f1": { + "FAVOR": 0.5483870967741935, + "AGAINST": 0.7354838709677419, + "NONE": 0.72 + }, + "accuracy_pct": 69.17808219178082, + "macro_f1_pct": 66.79569892473118, + "weighted_f1_pct": 70.03711886875827, + "macro_precision_pct": 65.03926286534983, + "macro_recall_pct": 74.35332333833084, + "per_class_f1_pct": { + "FAVOR": 54.83870967741935, + "AGAINST": 73.54838709677419, + "NONE": 72.0 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6158155458878143, + "weighted_f1": 0.6328455644558489, + "macro_precision": 0.6131015103617844, + "macro_recall": 0.6372549019607843, + "per_class_f1": { + "FAVOR": 0.5669291338582677, + "AGAINST": 0.6777777777777778, + "NONE": 0.6027397260273972 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 61.58155458878143, + "weighted_f1_pct": 63.2845564455849, + "macro_precision_pct": 61.31015103617844, + "macro_recall_pct": 63.725490196078425, + "per_class_f1_pct": { + "FAVOR": 56.69291338582677, + "AGAINST": 67.77777777777779, + "NONE": 60.273972602739725 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.693494671177624, + "weighted_f1": 0.7186250584640558, + "macro_precision": 0.6820776151404955, + "macro_recall": 0.709505255575011, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.7627906976744186, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.3494671177624, + "weighted_f1_pct": 71.86250584640558, + "macro_precision_pct": 68.20776151404955, + "macro_recall_pct": 70.9505255575011, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 76.27906976744187, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6758620689655173, + "weighted_f1": 0.6548505747126437, + "macro_precision": 0.6725342190016104, + "macro_recall": 0.6885785097167211, + "per_class_f1": { + "FAVOR": 0.5793103448275863, + "AGAINST": 0.6666666666666666, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 67.58620689655173, + "weighted_f1_pct": 65.48505747126437, + "macro_precision_pct": 67.25342190016104, + "macro_recall_pct": 68.85785097167211, + "per_class_f1_pct": { + "FAVOR": 57.931034482758626, + "AGAINST": 66.66666666666666, + "NONE": 78.16091954022988 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7675482787036007, + "weighted_f1": 0.8063573108421765, + "macro_precision": 0.7557389937106919, + "macro_recall": 0.7884465174819998, + "per_class_f1": { + "FAVOR": 0.7096774193548387, + "AGAINST": 0.7207207207207207, + "NONE": 0.8722466960352423 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 76.75482787036006, + "weighted_f1_pct": 80.63573108421765, + "macro_precision_pct": 75.57389937106919, + "macro_recall_pct": 78.84465174819998, + "per_class_f1_pct": { + "FAVOR": 70.96774193548387, + "AGAINST": 72.07207207207207, + "NONE": 87.22466960352423 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.694829278080315, + "weighted_f1": 0.8273223044922638, + "macro_precision": 0.6781385281385282, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.38095238095238093, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.4829278080315, + "weighted_f1_pct": 82.73223044922638, + "macro_precision_pct": 67.81385281385282, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 38.095238095238095, + "NONE": 86.74698795180723 + } + } + ], + "accuracy": 0.7122370936902486, + "macro_f1": 0.7107668755575625, + "weighted_f1": 0.7115291974315995, + "macro_precision": 0.7071149776913791, + "macro_recall": 0.7172894469528067, + "per_class_f1": { + "FAVOR": 0.6496598639455783, + "AGAINST": 0.7058823529411765, + "NONE": 0.7767584097859327 + }, + "accuracy_pct": 71.22370936902486, + "macro_f1_pct": 71.07668755575625, + "weighted_f1_pct": 71.15291974315994, + "macro_precision_pct": 70.7114977691379, + "macro_recall_pct": 71.72894469528067, + "per_class_f1_pct": { + "FAVOR": 64.96598639455783, + "AGAINST": 70.58823529411765, + "NONE": 77.67584097859327 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6171121065448096, + "losses": { + "current_ce": 0.5663822464870684, + "replay_ce": 0.24519189681970713, + "dark_distillation": 0.4146255061481938, + "supervised_contrastive": 4.1774978970036365, + "total": 1.2812954089858315, + "replay_multiplier": 0.6523877583343842, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09583490140424801, + "ema_update_magnitude": 0.05929910813088242, + "final_multiplier": 0.6187631777357253 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.665124716553288, + "weighted_f1": 0.6921535737582706, + "macro_precision": 0.6644641537324464, + "macro_recall": 0.7593853073463267, + "per_class_f1": { + "FAVOR": 0.56, + "AGAINST": 0.7210884353741497, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.5124716553288, + "weighted_f1_pct": 69.21535737582705, + "macro_precision_pct": 66.44641537324463, + "macro_recall_pct": 75.93853073463266, + "per_class_f1_pct": { + "FAVOR": 56.00000000000001, + "AGAINST": 72.10884353741497, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6320258090829, + "weighted_f1": 0.6351152630077549, + "macro_precision": 0.6254312531709793, + "macro_recall": 0.6695715323166304, + "per_class_f1": { + "FAVOR": 0.5826771653543307, + "AGAINST": 0.6551724137931034, + "NONE": 0.6582278481012658 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 63.202580908289995, + "weighted_f1_pct": 63.511526300775486, + "macro_precision_pct": 62.543125317097925, + "macro_recall_pct": 66.95715323166304, + "per_class_f1_pct": { + "FAVOR": 58.26771653543307, + "AGAINST": 65.51724137931035, + "NONE": 65.82278481012658 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6824479213828535, + "weighted_f1": 0.7031195798501539, + "macro_precision": 0.667969235947985, + "macro_recall": 0.7042207604414789, + "per_class_f1": { + "FAVOR": 0.6486486486486487, + "AGAINST": 0.7383177570093458, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.24479213828535, + "weighted_f1_pct": 70.3119579850154, + "macro_precision_pct": 66.7969235947985, + "macro_recall_pct": 70.42207604414789, + "per_class_f1_pct": { + "FAVOR": 64.86486486486487, + "AGAINST": 73.83177570093457, + "NONE": 66.0377358490566 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.675, + "macro_f1": 0.6841540113514343, + "weighted_f1": 0.6745283018867925, + "macro_precision": 0.6839151012592747, + "macro_recall": 0.6847064909395533, + "per_class_f1": { + "FAVOR": 0.6289308176100629, + "AGAINST": 0.6918238993710691, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 68.41540113514343, + "weighted_f1_pct": 67.45283018867924, + "macro_precision_pct": 68.39151012592747, + "macro_recall_pct": 68.47064909395533, + "per_class_f1_pct": { + "FAVOR": 62.893081761006286, + "AGAINST": 69.18238993710692, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.785, + "macro_f1": 0.7585303838077792, + "weighted_f1": 0.7862199039734378, + "macro_precision": 0.741741502382344, + "macro_recall": 0.7994965480222852, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.6972477064220184, + "NONE": 0.8430493273542601 + }, + "accuracy_pct": 78.5, + "macro_f1_pct": 75.85303838077792, + "weighted_f1_pct": 78.62199039734378, + "macro_precision_pct": 74.1741502382344, + "macro_recall_pct": 79.94965480222852, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 69.72477064220183, + "NONE": 84.30493273542601 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6946198630786444, + "weighted_f1": 0.8277027350202723, + "macro_precision": 0.6790935672514621, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.38095238095238093, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.46198630786444, + "weighted_f1_pct": 82.77027350202722, + "macro_precision_pct": 67.90935672514621, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 38.095238095238095, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.6972286972286973, + "weighted_f1": 0.7268967591548237, + "macro_precision": 0.6811834823329077, + "macro_recall": 0.7331649831649831, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7692307692307693, + "NONE": 0.7070707070707071 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 69.72286972286973, + "weighted_f1_pct": 72.68967591548237, + "macro_precision_pct": 68.11834823329077, + "macro_recall_pct": 73.31649831649831, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.92307692307693, + "NONE": 70.70707070707071 + } + } + ], + "accuracy": 0.7094155844155844, + "macro_f1": 0.7087624822325092, + "weighted_f1": 0.7101218691325805, + "macro_precision": 0.7046863653505122, + "macro_recall": 0.7213074764907176, + "per_class_f1": { + "FAVOR": 0.6609929078014184, + "AGAINST": 0.7085377821393523, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 70.94155844155844, + "macro_f1_pct": 70.87624822325091, + "weighted_f1_pct": 71.01218691325805, + "macro_precision_pct": 70.46863653505122, + "macro_recall_pct": 72.13074764907176, + "per_class_f1_pct": { + "FAVOR": 66.09929078014184, + "AGAINST": 70.85377821393523, + "NONE": 75.67567567567568 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.6597064989517819, + "losses": { + "current_ce": 0.6451548337936401, + "replay_ce": 0.2370354709525903, + "dark_distillation": 0.4551077346007029, + "supervised_contrastive": 4.321132810910543, + "total": 1.4016534114877384, + "replay_multiplier": 0.685048702314224, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10017886626050582, + "ema_update_magnitude": 0.058735631177916615, + "final_multiplier": 0.5863076052905218 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.715940355940356, + "weighted_f1": 0.7398620253414775, + "macro_precision": 0.6955414482911345, + "macro_recall": 0.8022738630684657, + "per_class_f1": { + "FAVOR": 0.6349206349206349, + "AGAINST": 0.7662337662337663, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.5940355940356, + "weighted_f1_pct": 73.98620253414775, + "macro_precision_pct": 69.55414482911345, + "macro_recall_pct": 80.22738630684657, + "per_class_f1_pct": { + "FAVOR": 63.49206349206349, + "AGAINST": 76.62337662337663, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6206646003256173, + "weighted_f1": 0.6301769749405789, + "macro_precision": 0.6094611002148128, + "macro_recall": 0.6452432824981845, + "per_class_f1": { + "FAVOR": 0.5423728813559322, + "AGAINST": 0.6702702702702703, + "NONE": 0.6493506493506493 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 62.066460032561736, + "weighted_f1_pct": 63.01769749405789, + "macro_precision_pct": 60.94611002148128, + "macro_recall_pct": 64.52432824981845, + "per_class_f1_pct": { + "FAVOR": 54.23728813559322, + "AGAINST": 67.02702702702703, + "NONE": 64.93506493506493 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6954314720812182, + "macro_f1": 0.6910035867217273, + "weighted_f1": 0.6977446464503572, + "macro_precision": 0.6778919330289194, + "macro_recall": 0.7335247319630452, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.7106598984771574, + "NONE": 0.6774193548387096 + }, + "accuracy_pct": 69.54314720812182, + "macro_f1_pct": 69.10035867217273, + "weighted_f1_pct": 69.77446464503572, + "macro_precision_pct": 67.78919330289193, + "macro_recall_pct": 73.35247319630453, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 71.06598984771574, + "NONE": 67.74193548387096 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6525543883250572, + "weighted_f1": 0.635968108888757, + "macro_precision": 0.6485507246376812, + "macro_recall": 0.6637713385680865, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6506024096385542, + "NONE": 0.735632183908046 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 65.25543883250572, + "weighted_f1_pct": 63.5968108888757, + "macro_precision_pct": 64.85507246376811, + "macro_recall_pct": 66.37713385680864, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 65.06024096385542, + "NONE": 73.5632183908046 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.845, + "macro_f1": 0.8251524238284986, + "weighted_f1": 0.8437551627425782, + "macro_precision": 0.8209604519774011, + "macro_recall": 0.8362263312788051, + "per_class_f1": { + "FAVOR": 0.8135593220338984, + "AGAINST": 0.7777777777777778, + "NONE": 0.8841201716738197 + }, + "accuracy_pct": 84.5, + "macro_f1_pct": 82.51524238284986, + "weighted_f1_pct": 84.37551627425782, + "macro_precision_pct": 82.09604519774011, + "macro_recall_pct": 83.6226331278805, + "per_class_f1_pct": { + "FAVOR": 81.35593220338984, + "AGAINST": 77.77777777777779, + "NONE": 88.41201716738198 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6380121816168328, + "weighted_f1": 0.7801245112162997, + "macro_precision": 0.6505636070853462, + "macro_recall": 0.7814100230554545, + "per_class_f1": { + "FAVOR": 0.7678571428571429, + "AGAINST": 0.2857142857142857, + "NONE": 0.8604651162790697 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.80121816168328, + "weighted_f1_pct": 78.01245112162997, + "macro_precision_pct": 65.05636070853463, + "macro_recall_pct": 78.14100230554544, + "per_class_f1_pct": { + "FAVOR": 76.78571428571429, + "AGAINST": 28.57142857142857, + "NONE": 86.04651162790698 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.7118306758753024, + "weighted_f1": 0.7361186585507917, + "macro_precision": 0.6962532299741602, + "macro_recall": 0.7479797979797981, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.7731958762886598, + "NONE": 0.7047619047619048 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 71.18306758753023, + "weighted_f1_pct": 73.61186585507917, + "macro_precision_pct": 69.62532299741602, + "macro_recall_pct": 74.7979797979798, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 77.31958762886599, + "NONE": 70.47619047619048 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.609984111148062, + "weighted_f1": 0.5923923989564257, + "macro_precision": 0.614023114023114, + "macro_recall": 0.6376068376068376, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.4536082474226804, + "NONE": 0.7096774193548387 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 60.998411114806196, + "weighted_f1_pct": 59.23923989564257, + "macro_precision_pct": 61.402311402311405, + "macro_recall_pct": 63.760683760683754, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 45.36082474226804, + "NONE": 70.96774193548387 + } + } + ], + "accuracy": 0.7037845705967977, + "macro_f1": 0.7012906353094626, + "weighted_f1": 0.7012356158279139, + "macro_precision": 0.6978118619532511, + "macro_recall": 0.714270435446906, + "per_class_f1": { + "FAVOR": 0.652542372881356, + "AGAINST": 0.687333923826395, + "NONE": 0.7639956092206367 + }, + "accuracy_pct": 70.37845705967976, + "macro_f1_pct": 70.12906353094625, + "weighted_f1_pct": 70.1235615827914, + "macro_precision_pct": 69.7811861953251, + "macro_recall_pct": 71.4270435446906, + "per_class_f1_pct": { + "FAVOR": 65.2542372881356, + "AGAINST": 68.73339238263951, + "NONE": 76.39956092206367 + } + }, + { + "after_task": 9, + "best_epoch": 3, + "best_dev_macro_f1": 0.7190994704094557, + "losses": { + "current_ce": 0.5703916194214337, + "replay_ce": 0.22581045247506404, + "dark_distillation": 0.409260602619337, + "supervised_contrastive": 4.283192065142203, + "total": 1.2568184225455574, + "replay_multiplier": 0.5946455046275589, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09400880390005131, + "ema_update_magnitude": 0.05512870784228559, + "final_multiplier": 0.5864206920545183 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6985476859346945, + "weighted_f1": 0.7305364126544737, + "macro_precision": 0.682283666154634, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7662337662337663, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.85476859346946, + "weighted_f1_pct": 73.05364126544737, + "macro_precision_pct": 68.2283666154634, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 76.62337662337663, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.620980870980871, + "weighted_f1": 0.6429631771737035, + "macro_precision": 0.6180068226120858, + "macro_recall": 0.640159767610748, + "per_class_f1": { + "FAVOR": 0.5873015873015873, + "AGAINST": 0.6923076923076923, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 62.0980870980871, + "weighted_f1_pct": 64.29631771737036, + "macro_precision_pct": 61.800682261208586, + "macro_recall_pct": 64.0159767610748, + "per_class_f1_pct": { + "FAVOR": 58.730158730158735, + "AGAINST": 69.23076923076923, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6612307023265926, + "weighted_f1": 0.683272571945126, + "macro_precision": 0.6466208476517755, + "macro_recall": 0.6852912256347966, + "per_class_f1": { + "FAVOR": 0.6027397260273972, + "AGAINST": 0.7142857142857143, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.12307023265926, + "weighted_f1_pct": 68.3272571945126, + "macro_precision_pct": 64.66208476517755, + "macro_recall_pct": 68.52912256347966, + "per_class_f1_pct": { + "FAVOR": 60.273972602739725, + "AGAINST": 71.42857142857143, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.7, + "macro_f1": 0.7049329501915708, + "weighted_f1": 0.6986795977011494, + "macro_precision": 0.6997024861552963, + "macro_recall": 0.7136288884256364, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7125, + "NONE": 0.735632183908046 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 70.49329501915707, + "weighted_f1_pct": 69.86795977011494, + "macro_precision_pct": 69.97024861552963, + "macro_recall_pct": 71.36288884256365, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 71.25, + "NONE": 73.5632183908046 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7979534767805374, + "weighted_f1": 0.8130973821068488, + "macro_precision": 0.7767106842737096, + "macro_recall": 0.8330279304792048, + "per_class_f1": { + "FAVOR": 0.7868852459016393, + "AGAINST": 0.7619047619047619, + "NONE": 0.8450704225352113 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 79.79534767805374, + "weighted_f1_pct": 81.30973821068488, + "macro_precision_pct": 77.67106842737095, + "macro_recall_pct": 83.30279304792047, + "per_class_f1_pct": { + "FAVOR": 78.68852459016394, + "AGAINST": 76.19047619047619, + "NONE": 84.50704225352112 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7079646017699115, + "macro_f1": 0.6124622565095191, + "weighted_f1": 0.7514140850194144, + "macro_precision": 0.6329363460858688, + "macro_recall": 0.7551996116976095, + "per_class_f1": { + "FAVOR": 0.7256637168141593, + "AGAINST": 0.25806451612903225, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 70.79646017699115, + "macro_f1_pct": 61.246225650951914, + "weighted_f1_pct": 75.14140850194144, + "macro_precision_pct": 63.29363460858688, + "macro_recall_pct": 75.51996116976095, + "per_class_f1_pct": { + "FAVOR": 72.56637168141593, + "AGAINST": 25.806451612903224, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7526881720430108, + "macro_f1": 0.7306397306397306, + "weighted_f1": 0.7575757575757576, + "macro_precision": 0.7132275132275132, + "macro_recall": 0.7630190796857464, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.797979797979798, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 75.26881720430107, + "macro_f1_pct": 73.06397306397307, + "weighted_f1_pct": 75.75757575757575, + "macro_precision_pct": 71.32275132275132, + "macro_recall_pct": 76.30190796857464, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 79.7979797979798, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5686327841291522, + "weighted_f1": 0.5701792135649518, + "macro_precision": 0.5998650472334682, + "macro_recall": 0.5619658119658121, + "per_class_f1": { + "FAVOR": 0.5306122448979592, + "AGAINST": 0.4915254237288136, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 56.86327841291522, + "weighted_f1_pct": 57.01792135649518, + "macro_precision_pct": 59.986504723346826, + "macro_recall_pct": 56.19658119658121, + "per_class_f1_pct": { + "FAVOR": 53.06122448979592, + "AGAINST": 49.152542372881356, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.695, + "macro_f1": 0.6771068992042775, + "weighted_f1": 0.6902864790224342, + "macro_precision": 0.696494651333361, + "macro_recall": 0.6718103181958605, + "per_class_f1": { + "FAVOR": 0.6067415730337079, + "AGAINST": 0.6518518518518519, + "NONE": 0.7727272727272727 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 67.71068992042774, + "weighted_f1_pct": 69.02864790224342, + "macro_precision_pct": 69.6494651333361, + "macro_recall_pct": 67.18103181958604, + "per_class_f1_pct": { + "FAVOR": 60.67415730337079, + "AGAINST": 65.18518518518519, + "NONE": 77.27272727272727 + } + } + ], + "accuracy": 0.7007623888182973, + "macro_f1": 0.6970988648825823, + "weighted_f1": 0.7000367254497073, + "macro_precision": 0.6946830190975358, + "macro_recall": 0.701140201299623, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.6971080669710806, + "NONE": 0.7490272373540856 + }, + "accuracy_pct": 70.07623888182974, + "macro_f1_pct": 69.70988648825823, + "weighted_f1_pct": 70.00367254497073, + "macro_precision_pct": 69.46830190975358, + "macro_recall_pct": 70.1140201299623, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 69.71080669710807, + "NONE": 74.90272373540856 + } + }, + { + "after_task": 10, + "best_epoch": 3, + "best_dev_macro_f1": 0.6670726163568798, + "losses": { + "current_ce": 0.6257781086823879, + "replay_ce": 0.25476715205571593, + "dark_distillation": 0.40199438226528655, + "supervised_contrastive": 4.2510028924697485, + "total": 1.3297316633738003, + "replay_multiplier": 0.6097074989263762, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09897892460005277, + "ema_update_magnitude": 0.054363232901227826, + "final_multiplier": 0.5492404885271793 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6892608201217474, + "weighted_f1": 0.7141775736132989, + "macro_precision": 0.6746379807202642, + "macro_recall": 0.7799100449775113, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7417218543046358, + "NONE": 0.72 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.92608201217473, + "weighted_f1_pct": 71.41775736132989, + "macro_precision_pct": 67.46379807202642, + "macro_recall_pct": 77.99100449775112, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 74.17218543046357, + "NONE": 72.0 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5987190197716513, + "weighted_f1": 0.5976520247988392, + "macro_precision": 0.5978020634047051, + "macro_recall": 0.6336238198983297, + "per_class_f1": { + "FAVOR": 0.556390977443609, + "AGAINST": 0.6081871345029239, + "NONE": 0.631578947368421 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 59.87190197716513, + "weighted_f1_pct": 59.76520247988392, + "macro_precision_pct": 59.78020634047051, + "macro_recall_pct": 63.36238198983297, + "per_class_f1_pct": { + "FAVOR": 55.639097744360896, + "AGAINST": 60.81871345029239, + "NONE": 63.1578947368421 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6796858396858397, + "weighted_f1": 0.6905775123034005, + "macro_precision": 0.6661173368490442, + "macro_recall": 0.7270308498835307, + "per_class_f1": { + "FAVOR": 0.6216216216216216, + "AGAINST": 0.6974358974358974, + "NONE": 0.72 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 67.96858396858397, + "weighted_f1_pct": 69.05775123034005, + "macro_precision_pct": 66.61173368490442, + "macro_recall_pct": 72.70308498835307, + "per_class_f1_pct": { + "FAVOR": 62.16216216216216, + "AGAINST": 69.74358974358974, + "NONE": 72.0 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6920975323149235, + "weighted_f1": 0.683139982373678, + "macro_precision": 0.6834036568213783, + "macro_recall": 0.7080466958515738, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.6621621621621622, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.20975323149236, + "weighted_f1_pct": 68.3139982373678, + "macro_precision_pct": 68.34036568213783, + "macro_recall_pct": 70.80466958515737, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 66.21621621621621, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.83, + "macro_f1": 0.8079288302412034, + "weighted_f1": 0.83240465416936, + "macro_precision": 0.7884768307670065, + "macro_recall": 0.8389249819534678, + "per_class_f1": { + "FAVOR": 0.7619047619047619, + "AGAINST": 0.7931034482758621, + "NONE": 0.8687782805429864 + }, + "accuracy_pct": 83.0, + "macro_f1_pct": 80.79288302412034, + "weighted_f1_pct": 83.240465416936, + "macro_precision_pct": 78.84768307670065, + "macro_recall_pct": 83.89249819534677, + "per_class_f1_pct": { + "FAVOR": 76.19047619047619, + "AGAINST": 79.3103448275862, + "NONE": 86.87782805429865 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6887736414206246, + "weighted_f1": 0.8033386120167676, + "macro_precision": 0.6647643077415996, + "macro_recall": 0.796820774177891, + "per_class_f1": { + "FAVOR": 0.816, + "AGAINST": 0.42105263157894735, + "NONE": 0.8292682926829268 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 68.87736414206246, + "weighted_f1_pct": 80.33386120167675, + "macro_precision_pct": 66.47643077415995, + "macro_recall_pct": 79.6820774177891, + "per_class_f1_pct": { + "FAVOR": 81.6, + "AGAINST": 42.10526315789473, + "NONE": 82.92682926829268 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7580645161290323, + "macro_f1": 0.7340101522842639, + "weighted_f1": 0.7620084680343401, + "macro_precision": 0.7197540236866079, + "macro_recall": 0.7650392817059485, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.8020304568527918, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 75.80645161290323, + "macro_f1_pct": 73.40101522842639, + "weighted_f1_pct": 76.200846803434, + "macro_precision_pct": 71.9754023686608, + "macro_recall_pct": 76.50392817059485, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 80.20304568527918, + "NONE": 74.28571428571429 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5623136123136123, + "weighted_f1": 0.5440013820295511, + "macro_precision": 0.5652831402831403, + "macro_recall": 0.617948717948718, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.3409090909090909, + "NONE": 0.746031746031746 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.231361231361234, + "weighted_f1_pct": 54.400138202955105, + "macro_precision_pct": 56.52831402831403, + "macro_recall_pct": 61.794871794871796, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 34.090909090909086, + "NONE": 74.60317460317461 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6845790852219841, + "weighted_f1": 0.6980872134554192, + "macro_precision": 0.690795839380745, + "macro_recall": 0.6846308310163732, + "per_class_f1": { + "FAVOR": 0.6111111111111112, + "AGAINST": 0.6610169491525424, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 68.45790852219841, + "weighted_f1_pct": 69.80872134554193, + "macro_precision_pct": 69.07958393807449, + "macro_recall_pct": 68.46308310163732, + "per_class_f1_pct": { + "FAVOR": 61.111111111111114, + "AGAINST": 66.10169491525424, + "NONE": 78.16091954022988 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6447329538384969, + "weighted_f1": 0.6703621337252705, + "macro_precision": 0.6397890023389928, + "macro_recall": 0.6562224183583408, + "per_class_f1": { + "FAVOR": 0.6979166666666666, + "AGAINST": 0.44680851063829785, + "NONE": 0.7894736842105263 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.47329538384969, + "weighted_f1_pct": 67.03621337252706, + "macro_precision_pct": 63.97890023389928, + "macro_recall_pct": 65.62224183583409, + "per_class_f1_pct": { + "FAVOR": 69.79166666666666, + "AGAINST": 44.680851063829785, + "NONE": 78.94736842105263 + } + } + ], + "accuracy": 0.6995490417136415, + "macro_f1": 0.6994550539497663, + "weighted_f1": 0.6964575969159251, + "macro_precision": 0.698337255055825, + "macro_recall": 0.7110100791666873, + "per_class_f1": { + "FAVOR": 0.6654099905749293, + "AGAINST": 0.661526599845798, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 69.95490417136415, + "macro_f1_pct": 69.94550539497662, + "weighted_f1_pct": 69.64575969159252, + "macro_precision_pct": 69.8337255055825, + "macro_recall_pct": 71.10100791666873, + "per_class_f1_pct": { + "FAVOR": 66.54099905749294, + "AGAINST": 66.1526599845798, + "NONE": 77.14285714285715 + } + } + ], + "summary": { + "final_mean_target_macro_f1": 0.6782101487214348, + "average_forgetting": 0.01737079029072852, + "backward_transfer": -0.003406699817294393, + "final_mean_target_macro_f1_pct": 67.82101487214348, + "average_forgetting_pct": 1.7370790290728522, + "backward_transfer_pct": -0.3406699817294393, + "final_pooled_accuracy": 0.6995490417136415, + "final_pooled_accuracy_pct": 69.95490417136415, + "final_pooled_macro_f1": 0.6994550539497663, + "final_pooled_macro_f1_pct": 69.94550539497662, + "final_pooled_weighted_f1": 0.6964575969159251, + "final_pooled_weighted_f1_pct": 69.64575969159252, + "final_pooled_macro_precision": 0.698337255055825, + "final_pooled_macro_precision_pct": 69.8337255055825, + "final_pooled_macro_recall": 0.7110100791666873, + "final_pooled_macro_recall_pct": 71.10100791666873, + "final_pooled_per_class_f1": { + "FAVOR": 0.6654099905749293, + "AGAINST": 0.661526599845798, + "NONE": 0.7714285714285715 + }, + "final_pooled_per_class_f1_pct": { + "FAVOR": 66.54099905749294, + "AGAINST": 66.1526599845798, + "NONE": 77.14285714285715 + } + }, + "efficiency": { + "training_wall_time_seconds": 976.2795598909725, + "training_wall_time_minutes": 16.271325998182874, + "peak_gpu_memory_allocated_gb": 10.431715965270996, + "peak_gpu_memory_reserved_gb": 10.58984375 + } +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s3__seed13/metrics.partial.json b/outputs/trace_s8_full__x__s3__seed13/metrics.partial.json new file mode 100644 index 0000000000000000000000000000000000000000..76be236e59445fbd58ecdb133ad940d55968c252 --- /dev/null +++ b/outputs/trace_s8_full__x__s3__seed13/metrics.partial.json @@ -0,0 +1,2238 @@ +{ + "matrix": [ + [ + 0.7152842910804694, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.7147058823529412, + 0.596579966072294, + null, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6990980034458295, + 0.6226705479316422, + 0.7030940878766966, + null, + null, + null, + null, + null, + null, + null + ], + [ + 0.6983831958843316, + 0.5887419960412661, + 0.6953357100415923, + 0.6910028993447938, + null, + null, + null, + null, + null, + null + ], + [ + 0.6898943919398476, + 0.5924510399490873, + 0.6839596533391595, + 0.6901105162171447, + 0.7829186016958943, + null, + null, + null, + null, + null + ], + [ + 0.6679569892473118, + 0.6158155458878143, + 0.693494671177624, + 0.6758620689655173, + 0.7675482787036007, + 0.694829278080315, + null, + null, + null, + null + ], + [ + 0.665124716553288, + 0.6320258090829, + 0.6824479213828535, + 0.6841540113514343, + 0.7585303838077792, + 0.6946198630786444, + 0.6972286972286973, + null, + null, + null + ], + [ + 0.715940355940356, + 0.6206646003256173, + 0.6910035867217273, + 0.6525543883250572, + 0.8251524238284986, + 0.6380121816168328, + 0.7118306758753024, + 0.609984111148062, + null, + null + ], + [ + 0.6985476859346945, + 0.620980870980871, + 0.6612307023265926, + 0.7049329501915708, + 0.7979534767805374, + 0.6124622565095191, + 0.7306397306397306, + 0.5686327841291522, + 0.6771068992042775, + null + ], + [ + 0.6892608201217474, + 0.5987190197716513, + 0.6796858396858397, + 0.6920975323149235, + 0.8079288302412034, + 0.6887736414206246, + 0.7340101522842639, + 0.5623136123136123, + 0.6845790852219841, + 0.6447329538384969 + ] + ], + "history": [ + { + "after_task": 1, + "best_epoch": 5, + "best_dev_macro_f1": 0.8193153566647542, + "losses": { + "current_ce": 0.5442819329358085, + "total": 0.5442819329358085, + "replay_multiplier": 0.6862624481284426, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10795797371276879, + "ema_update_magnitude": 0.06257687394518485, + "final_multiplier": 0.5796410565436867 + } + }, + "memory": { + "size": 64, + "samples_per_task": 64, + "label_counts": { + "0": 22, + "1": 21, + "2": 21 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + } + ], + "accuracy": 0.7397260273972602, + "macro_f1": 0.7152842910804694, + "weighted_f1": 0.7510221477327315, + "macro_precision": 0.6957942464040024, + "macro_recall": 0.7883158420789605, + "per_class_f1": { + "FAVOR": 0.5846153846153846, + "AGAINST": 0.7898089171974523, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.52842910804694, + "weighted_f1_pct": 75.10221477327315, + "macro_precision_pct": 69.57942464040025, + "macro_recall_pct": 78.83158420789606, + "per_class_f1_pct": { + "FAVOR": 58.46153846153847, + "AGAINST": 78.98089171974523, + "NONE": 77.14285714285715 + } + }, + { + "after_task": 2, + "best_epoch": 13, + "best_dev_macro_f1": 0.6269608881868326, + "losses": { + "current_ce": 0.7093145791328315, + "replay_ce": 0.22077340695894127, + "dark_distillation": 0.38159919065056425, + "supervised_contrastive": 4.350307623545329, + "total": 1.4104225379047972, + "replay_multiplier": 0.6142165815115603, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.11817399906990574, + "ema_update_magnitude": 0.05939854852669155, + "final_multiplier": 0.5026363582022335 + } + }, + "memory": { + "size": 128, + "samples_per_task": 64, + "label_counts": { + "0": 44, + "1": 42, + "2": 42 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7397260273972602, + "macro_f1": 0.7147058823529412, + "weighted_f1": 0.7502870668815471, + "macro_precision": 0.6934389140271492, + "macro_recall": 0.7786056971514242, + "per_class_f1": { + "FAVOR": 0.5625, + "AGAINST": 0.7875, + "NONE": 0.7941176470588235 + }, + "accuracy_pct": 73.97260273972603, + "macro_f1_pct": 71.47058823529412, + "weighted_f1_pct": 75.02870668815471, + "macro_precision_pct": 69.34389140271492, + "macro_recall_pct": 77.86056971514242, + "per_class_f1_pct": { + "FAVOR": 56.25, + "AGAINST": 78.75, + "NONE": 79.41176470588235 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6105263157894737, + "macro_f1": 0.596579966072294, + "weighted_f1": 0.6147010297132592, + "macro_precision": 0.5931935931935932, + "macro_recall": 0.6201888162672476, + "per_class_f1": { + "FAVOR": 0.59375, + "AGAINST": 0.6480446927374302, + "NONE": 0.547945205479452 + }, + "accuracy_pct": 61.05263157894737, + "macro_f1_pct": 59.6579966072294, + "weighted_f1_pct": 61.47010297132592, + "macro_precision_pct": 59.31935931935932, + "macro_recall_pct": 62.01888162672476, + "per_class_f1_pct": { + "FAVOR": 59.375, + "AGAINST": 64.80446927374301, + "NONE": 54.794520547945204 + } + } + ], + "accuracy": 0.6666666666666666, + "macro_f1": 0.654621435594887, + "weighted_f1": 0.6743244310998735, + "macro_precision": 0.6442073608079101, + "macro_recall": 0.6928679485772008, + "per_class_f1": { + "FAVOR": 0.5833333333333334, + "AGAINST": 0.7138643067846607, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 66.66666666666666, + "macro_f1_pct": 65.4621435594887, + "weighted_f1_pct": 67.43244310998735, + "macro_precision_pct": 64.420736080791, + "macro_recall_pct": 69.28679485772008, + "per_class_f1_pct": { + "FAVOR": 58.333333333333336, + "AGAINST": 71.38643067846607, + "NONE": 66.66666666666666 + } + }, + { + "after_task": 3, + "best_epoch": 9, + "best_dev_macro_f1": 0.7630211318735909, + "losses": { + "current_ce": 0.6028909276832234, + "replay_ce": 0.1996306974779476, + "dark_distillation": 0.3657441600705638, + "supervised_contrastive": 4.320868700200861, + "total": 1.2792593280474345, + "replay_multiplier": 0.6254400971576703, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09795596788232877, + "ema_update_magnitude": 0.05818585290700524, + "final_multiplier": 0.5940000815152167 + } + }, + "memory": { + "size": 192, + "samples_per_task": 64, + "label_counts": { + "0": 66, + "1": 63, + "2": 63 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.726027397260274, + "macro_f1": 0.6990980034458295, + "weighted_f1": 0.7393428489318901, + "macro_precision": 0.6821707709830722, + "macro_recall": 0.7556171914042977, + "per_class_f1": { + "FAVOR": 0.5454545454545454, + "AGAINST": 0.782608695652174, + "NONE": 0.7692307692307693 + }, + "accuracy_pct": 72.6027397260274, + "macro_f1_pct": 69.90980034458295, + "weighted_f1_pct": 73.934284893189, + "macro_precision_pct": 68.21707709830721, + "macro_recall_pct": 75.56171914042977, + "per_class_f1_pct": { + "FAVOR": 54.54545454545454, + "AGAINST": 78.26086956521739, + "NONE": 76.92307692307693 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6226705479316422, + "weighted_f1": 0.6366292276762126, + "macro_precision": 0.6158044075919964, + "macro_recall": 0.6470588235294118, + "per_class_f1": { + "FAVOR": 0.5806451612903226, + "AGAINST": 0.6740331491712708, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 62.26705479316422, + "weighted_f1_pct": 63.662922767621254, + "macro_precision_pct": 61.58044075919964, + "macro_recall_pct": 64.70588235294117, + "per_class_f1_pct": { + "FAVOR": 58.06451612903226, + "AGAINST": 67.40331491712708, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.7030940878766966, + "weighted_f1": 0.7182840958056143, + "macro_precision": 0.7044970334412083, + "macro_recall": 0.7095262883914627, + "per_class_f1": { + "FAVOR": 0.6984126984126984, + "AGAINST": 0.75, + "NONE": 0.6608695652173913 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 70.30940878766965, + "weighted_f1_pct": 71.82840958056143, + "macro_precision_pct": 70.44970334412083, + "macro_recall_pct": 70.95262883914627, + "per_class_f1_pct": { + "FAVOR": 69.84126984126983, + "AGAINST": 75.0, + "NONE": 66.08695652173913 + } + } + ], + "accuracy": 0.6885553470919324, + "macro_f1": 0.6700224476150556, + "weighted_f1": 0.6937263262644933, + "macro_precision": 0.6552230793139499, + "macro_recall": 0.7002366121601508, + "per_class_f1": { + "FAVOR": 0.6007905138339921, + "AGAINST": 0.7347670250896058, + "NONE": 0.6745098039215687 + }, + "accuracy_pct": 68.85553470919325, + "macro_f1_pct": 67.00224476150555, + "weighted_f1_pct": 69.37263262644933, + "macro_precision_pct": 65.52230793139499, + "macro_recall_pct": 70.02366121601507, + "per_class_f1_pct": { + "FAVOR": 60.079051383399204, + "AGAINST": 73.47670250896059, + "NONE": 67.45098039215686 + } + }, + { + "after_task": 4, + "best_epoch": 11, + "best_dev_macro_f1": 0.6808621014902849, + "losses": { + "current_ce": 0.6186725724311102, + "replay_ce": 0.2190132564968533, + "dark_distillation": 0.3627972994058851, + "supervised_contrastive": 4.313315141011798, + "total": 1.2969302307991755, + "replay_multiplier": 0.6125457563861113, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09427406151524267, + "ema_update_magnitude": 0.06213868220273308, + "final_multiplier": 0.6591280910570105 + } + }, + "memory": { + "size": 256, + "samples_per_task": 64, + "label_counts": { + "0": 88, + "1": 84, + "2": 84 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6983831958843316, + "weighted_f1": 0.7347945922161876, + "macro_precision": 0.6876130789174267, + "macro_recall": 0.7714142928535731, + "per_class_f1": { + "FAVOR": 0.5633802816901409, + "AGAINST": 0.7741935483870968, + "NONE": 0.7575757575757576 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.83831958843317, + "weighted_f1_pct": 73.47945922161875, + "macro_precision_pct": 68.76130789174267, + "macro_recall_pct": 77.1414292853573, + "per_class_f1_pct": { + "FAVOR": 56.33802816901409, + "AGAINST": 77.41935483870968, + "NONE": 75.75757575757575 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5887419960412661, + "weighted_f1": 0.5987730492071637, + "macro_precision": 0.597150506789061, + "macro_recall": 0.6292665214233842, + "per_class_f1": { + "FAVOR": 0.583941605839416, + "AGAINST": 0.6181818181818182, + "NONE": 0.5641025641025641 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 58.87419960412661, + "weighted_f1_pct": 59.87730492071637, + "macro_precision_pct": 59.7150506789061, + "macro_recall_pct": 62.926652142338426, + "per_class_f1_pct": { + "FAVOR": 58.3941605839416, + "AGAINST": 61.81818181818181, + "NONE": 56.41025641025641 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7106598984771574, + "macro_f1": 0.6953357100415923, + "weighted_f1": 0.7151392998362243, + "macro_precision": 0.6774747070455689, + "macro_recall": 0.73518106625863, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.7450980392156863, + "NONE": 0.6909090909090909 + }, + "accuracy_pct": 71.06598984771574, + "macro_f1_pct": 69.53357100415923, + "weighted_f1_pct": 71.51392998362243, + "macro_precision_pct": 67.74747070455689, + "macro_recall_pct": 73.518106625863, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 74.50980392156863, + "NONE": 69.0909090909091 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6910028993447938, + "weighted_f1": 0.6833319635641394, + "macro_precision": 0.6818863397810766, + "macro_recall": 0.7082049737876296, + "per_class_f1": { + "FAVOR": 0.6751592356687898, + "AGAINST": 0.6666666666666666, + "NONE": 0.7311827956989247 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.10028993447938, + "weighted_f1_pct": 68.33319635641394, + "macro_precision_pct": 68.18863397810766, + "macro_recall_pct": 70.82049737876297, + "per_class_f1_pct": { + "FAVOR": 67.51592356687898, + "AGAINST": 66.66666666666666, + "NONE": 73.11827956989248 + } + } + ], + "accuracy": 0.6753069577080492, + "macro_f1": 0.671287386205036, + "weighted_f1": 0.6789068243762634, + "macro_precision": 0.6638160574687392, + "macro_recall": 0.701177868318099, + "per_class_f1": { + "FAVOR": 0.6247191011235955, + "AGAINST": 0.7032640949554896, + "NONE": 0.6858789625360231 + }, + "accuracy_pct": 67.53069577080491, + "macro_f1_pct": 67.1287386205036, + "weighted_f1_pct": 67.89068243762634, + "macro_precision_pct": 66.38160574687392, + "macro_recall_pct": 70.11778683180991, + "per_class_f1_pct": { + "FAVOR": 62.47191011235955, + "AGAINST": 70.32640949554896, + "NONE": 68.58789625360231 + } + }, + { + "after_task": 5, + "best_epoch": 6, + "best_dev_macro_f1": 0.6654093524831955, + "losses": { + "current_ce": 0.4358942994537453, + "replay_ce": 0.21232219041946035, + "dark_distillation": 0.3923066610470414, + "supervised_contrastive": 4.081723755598068, + "total": 1.0873236504693826, + "replay_multiplier": 0.5900478693791011, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10032757414191823, + "ema_update_magnitude": 0.05437592970889881, + "final_multiplier": 0.541983897985826 + } + }, + "memory": { + "size": 320, + "samples_per_task": 64, + "label_counts": { + "0": 110, + "1": 105, + "2": 105 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6898943919398476, + "weighted_f1": 0.7150434776951248, + "macro_precision": 0.6759683950263375, + "macro_recall": 0.7799100449775113, + "per_class_f1": { + "FAVOR": 0.5882352941176471, + "AGAINST": 0.7417218543046358, + "NONE": 0.7397260273972602 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.98943919398477, + "weighted_f1_pct": 71.50434776951248, + "macro_precision_pct": 67.59683950263376, + "macro_recall_pct": 77.99100449775112, + "per_class_f1_pct": { + "FAVOR": 58.82352941176471, + "AGAINST": 74.17218543046357, + "NONE": 73.97260273972603 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5894736842105263, + "macro_f1": 0.5924510399490873, + "weighted_f1": 0.5918574567112913, + "macro_precision": 0.5981640310908604, + "macro_recall": 0.6296296296296297, + "per_class_f1": { + "FAVOR": 0.5652173913043478, + "AGAINST": 0.5988023952095808, + "NONE": 0.6133333333333333 + }, + "accuracy_pct": 58.94736842105262, + "macro_f1_pct": 59.24510399490873, + "weighted_f1_pct": 59.18574567112913, + "macro_precision_pct": 59.81640310908604, + "macro_recall_pct": 62.96296296296296, + "per_class_f1_pct": { + "FAVOR": 56.52173913043478, + "AGAINST": 59.88023952095808, + "NONE": 61.33333333333333 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6954314720812182, + "macro_f1": 0.6839596533391595, + "weighted_f1": 0.6995439013651351, + "macro_precision": 0.6665919305624278, + "macro_recall": 0.7299176039415497, + "per_class_f1": { + "FAVOR": 0.65, + "AGAINST": 0.7236180904522613, + "NONE": 0.6782608695652174 + }, + "accuracy_pct": 69.54314720812182, + "macro_f1_pct": 68.39596533391595, + "weighted_f1_pct": 69.95439013651351, + "macro_precision_pct": 66.65919305624278, + "macro_recall_pct": 72.99176039415497, + "per_class_f1_pct": { + "FAVOR": 65.0, + "AGAINST": 72.36180904522614, + "NONE": 67.82608695652173 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6901105162171447, + "weighted_f1": 0.6843955961088908, + "macro_precision": 0.6838810641627543, + "macro_recall": 0.7003335417969564, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.6845637583892618, + "NONE": 0.7191011235955056 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.01105162171447, + "weighted_f1_pct": 68.43955961088908, + "macro_precision_pct": 68.38810641627543, + "macro_recall_pct": 70.03335417969564, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 68.45637583892618, + "NONE": 71.91011235955057 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7829186016958943, + "weighted_f1": 0.809463159353989, + "macro_precision": 0.7716959064327487, + "macro_recall": 0.8073907490699095, + "per_class_f1": { + "FAVOR": 0.7619047619047619, + "AGAINST": 0.7222222222222222, + "NONE": 0.8646288209606987 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 78.29186016958943, + "weighted_f1_pct": 80.9463159353989, + "macro_precision_pct": 77.16959064327486, + "macro_recall_pct": 80.73907490699095, + "per_class_f1_pct": { + "FAVOR": 76.19047619047619, + "AGAINST": 72.22222222222221, + "NONE": 86.46288209606988 + } + } + ], + "accuracy": 0.6977491961414791, + "macro_f1": 0.6959287363606136, + "weighted_f1": 0.6984004879695633, + "macro_precision": 0.6924002097176563, + "macro_recall": 0.7202300152262531, + "per_class_f1": { + "FAVOR": 0.6379647749510763, + "AGAINST": 0.6925064599483204, + "NONE": 0.7573149741824441 + }, + "accuracy_pct": 69.77491961414792, + "macro_f1_pct": 69.59287363606136, + "weighted_f1_pct": 69.84004879695634, + "macro_precision_pct": 69.24002097176563, + "macro_recall_pct": 72.02300152262531, + "per_class_f1_pct": { + "FAVOR": 63.79647749510763, + "AGAINST": 69.25064599483204, + "NONE": 75.7314974182444 + } + }, + { + "after_task": 6, + "best_epoch": 5, + "best_dev_macro_f1": 0.6718427273858404, + "losses": { + "current_ce": 0.3773283552201021, + "replay_ce": 0.23725067909274783, + "dark_distillation": 0.4199893119789305, + "supervised_contrastive": 3.9003202427001225, + "total": 1.0871713672365462, + "replay_multiplier": 0.70590457921482, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09754419038195396, + "ema_update_magnitude": 0.059936500760013264, + "final_multiplier": 0.6144548488774144 + } + }, + "memory": { + "size": 384, + "samples_per_task": 64, + "label_counts": { + "0": 133, + "1": 123, + "2": 128 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.6917808219178082, + "macro_f1": 0.6679569892473118, + "weighted_f1": 0.7003711886875827, + "macro_precision": 0.6503926286534982, + "macro_recall": 0.7435332333833083, + "per_class_f1": { + "FAVOR": 0.5483870967741935, + "AGAINST": 0.7354838709677419, + "NONE": 0.72 + }, + "accuracy_pct": 69.17808219178082, + "macro_f1_pct": 66.79569892473118, + "weighted_f1_pct": 70.03711886875827, + "macro_precision_pct": 65.03926286534983, + "macro_recall_pct": 74.35332333833084, + "per_class_f1_pct": { + "FAVOR": 54.83870967741935, + "AGAINST": 73.54838709677419, + "NONE": 72.0 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6158155458878143, + "weighted_f1": 0.6328455644558489, + "macro_precision": 0.6131015103617844, + "macro_recall": 0.6372549019607843, + "per_class_f1": { + "FAVOR": 0.5669291338582677, + "AGAINST": 0.6777777777777778, + "NONE": 0.6027397260273972 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 61.58155458878143, + "weighted_f1_pct": 63.2845564455849, + "macro_precision_pct": 61.31015103617844, + "macro_recall_pct": 63.725490196078425, + "per_class_f1_pct": { + "FAVOR": 56.69291338582677, + "AGAINST": 67.77777777777779, + "NONE": 60.273972602739725 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.7157360406091371, + "macro_f1": 0.693494671177624, + "weighted_f1": 0.7186250584640558, + "macro_precision": 0.6820776151404955, + "macro_recall": 0.709505255575011, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.7627906976744186, + "NONE": 0.6605504587155964 + }, + "accuracy_pct": 71.57360406091371, + "macro_f1_pct": 69.3494671177624, + "weighted_f1_pct": 71.86250584640558, + "macro_precision_pct": 68.20776151404955, + "macro_recall_pct": 70.9505255575011, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 76.27906976744187, + "NONE": 66.05504587155964 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.66, + "macro_f1": 0.6758620689655173, + "weighted_f1": 0.6548505747126437, + "macro_precision": 0.6725342190016104, + "macro_recall": 0.6885785097167211, + "per_class_f1": { + "FAVOR": 0.5793103448275863, + "AGAINST": 0.6666666666666666, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 66.0, + "macro_f1_pct": 67.58620689655173, + "weighted_f1_pct": 65.48505747126437, + "macro_precision_pct": 67.25342190016104, + "macro_recall_pct": 68.85785097167211, + "per_class_f1_pct": { + "FAVOR": 57.931034482758626, + "AGAINST": 66.66666666666666, + "NONE": 78.16091954022988 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.805, + "macro_f1": 0.7675482787036007, + "weighted_f1": 0.8063573108421765, + "macro_precision": 0.7557389937106919, + "macro_recall": 0.7884465174819998, + "per_class_f1": { + "FAVOR": 0.7096774193548387, + "AGAINST": 0.7207207207207207, + "NONE": 0.8722466960352423 + }, + "accuracy_pct": 80.5, + "macro_f1_pct": 76.75482787036006, + "weighted_f1_pct": 80.63573108421765, + "macro_precision_pct": 75.57389937106919, + "macro_recall_pct": 78.84465174819998, + "per_class_f1_pct": { + "FAVOR": 70.96774193548387, + "AGAINST": 72.07207207207207, + "NONE": 87.22466960352423 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.694829278080315, + "weighted_f1": 0.8273223044922638, + "macro_precision": 0.6781385281385282, + "macro_recall": 0.813080936779517, + "per_class_f1": { + "FAVOR": 0.8360655737704918, + "AGAINST": 0.38095238095238093, + "NONE": 0.8674698795180723 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.4829278080315, + "weighted_f1_pct": 82.73223044922638, + "macro_precision_pct": 67.81385281385282, + "macro_recall_pct": 81.3080936779517, + "per_class_f1_pct": { + "FAVOR": 83.60655737704919, + "AGAINST": 38.095238095238095, + "NONE": 86.74698795180723 + } + } + ], + "accuracy": 0.7122370936902486, + "macro_f1": 0.7107668755575625, + "weighted_f1": 0.7115291974315995, + "macro_precision": 0.7071149776913791, + "macro_recall": 0.7172894469528067, + "per_class_f1": { + "FAVOR": 0.6496598639455783, + "AGAINST": 0.7058823529411765, + "NONE": 0.7767584097859327 + }, + "accuracy_pct": 71.22370936902486, + "macro_f1_pct": 71.07668755575625, + "weighted_f1_pct": 71.15291974315994, + "macro_precision_pct": 70.7114977691379, + "macro_recall_pct": 71.72894469528067, + "per_class_f1_pct": { + "FAVOR": 64.96598639455783, + "AGAINST": 70.58823529411765, + "NONE": 77.67584097859327 + } + }, + { + "after_task": 7, + "best_epoch": 12, + "best_dev_macro_f1": 0.6171121065448096, + "losses": { + "current_ce": 0.5663822464870684, + "replay_ce": 0.24519189681970713, + "dark_distillation": 0.4146255061481938, + "supervised_contrastive": 4.1774978970036365, + "total": 1.2812954089858315, + "replay_multiplier": 0.6523877583343842, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09583490140424801, + "ema_update_magnitude": 0.05929910813088242, + "final_multiplier": 0.6187631777357253 + } + }, + "memory": { + "size": 448, + "samples_per_task": 64, + "label_counts": { + "0": 155, + "1": 144, + "2": 149 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.678082191780822, + "macro_f1": 0.665124716553288, + "weighted_f1": 0.6921535737582706, + "macro_precision": 0.6644641537324464, + "macro_recall": 0.7593853073463267, + "per_class_f1": { + "FAVOR": 0.56, + "AGAINST": 0.7210884353741497, + "NONE": 0.7142857142857143 + }, + "accuracy_pct": 67.8082191780822, + "macro_f1_pct": 66.5124716553288, + "weighted_f1_pct": 69.21535737582705, + "macro_precision_pct": 66.44641537324463, + "macro_recall_pct": 75.93853073463266, + "per_class_f1_pct": { + "FAVOR": 56.00000000000001, + "AGAINST": 72.10884353741497, + "NONE": 71.42857142857143 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.631578947368421, + "macro_f1": 0.6320258090829, + "weighted_f1": 0.6351152630077549, + "macro_precision": 0.6254312531709793, + "macro_recall": 0.6695715323166304, + "per_class_f1": { + "FAVOR": 0.5826771653543307, + "AGAINST": 0.6551724137931034, + "NONE": 0.6582278481012658 + }, + "accuracy_pct": 63.1578947368421, + "macro_f1_pct": 63.202580908289995, + "weighted_f1_pct": 63.511526300775486, + "macro_precision_pct": 62.543125317097925, + "macro_recall_pct": 66.95715323166304, + "per_class_f1_pct": { + "FAVOR": 58.26771653543307, + "AGAINST": 65.51724137931035, + "NONE": 65.82278481012658 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.700507614213198, + "macro_f1": 0.6824479213828535, + "weighted_f1": 0.7031195798501539, + "macro_precision": 0.667969235947985, + "macro_recall": 0.7042207604414789, + "per_class_f1": { + "FAVOR": 0.6486486486486487, + "AGAINST": 0.7383177570093458, + "NONE": 0.660377358490566 + }, + "accuracy_pct": 70.05076142131979, + "macro_f1_pct": 68.24479213828535, + "weighted_f1_pct": 70.3119579850154, + "macro_precision_pct": 66.7969235947985, + "macro_recall_pct": 70.42207604414789, + "per_class_f1_pct": { + "FAVOR": 64.86486486486487, + "AGAINST": 73.83177570093457, + "NONE": 66.0377358490566 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.675, + "macro_f1": 0.6841540113514343, + "weighted_f1": 0.6745283018867925, + "macro_precision": 0.6839151012592747, + "macro_recall": 0.6847064909395533, + "per_class_f1": { + "FAVOR": 0.6289308176100629, + "AGAINST": 0.6918238993710691, + "NONE": 0.7317073170731707 + }, + "accuracy_pct": 67.5, + "macro_f1_pct": 68.41540113514343, + "weighted_f1_pct": 67.45283018867924, + "macro_precision_pct": 68.39151012592747, + "macro_recall_pct": 68.47064909395533, + "per_class_f1_pct": { + "FAVOR": 62.893081761006286, + "AGAINST": 69.18238993710692, + "NONE": 73.17073170731707 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.785, + "macro_f1": 0.7585303838077792, + "weighted_f1": 0.7862199039734378, + "macro_precision": 0.741741502382344, + "macro_recall": 0.7994965480222852, + "per_class_f1": { + "FAVOR": 0.7352941176470589, + "AGAINST": 0.6972477064220184, + "NONE": 0.8430493273542601 + }, + "accuracy_pct": 78.5, + "macro_f1_pct": 75.85303838077792, + "weighted_f1_pct": 78.62199039734378, + "macro_precision_pct": 74.1741502382344, + "macro_recall_pct": 79.94965480222852, + "per_class_f1_pct": { + "FAVOR": 73.52941176470588, + "AGAINST": 69.72477064220183, + "NONE": 84.30493273542601 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.8053097345132744, + "macro_f1": 0.6946198630786444, + "weighted_f1": 0.8277027350202723, + "macro_precision": 0.6790935672514621, + "macro_recall": 0.8099259798568136, + "per_class_f1": { + "FAVOR": 0.8387096774193549, + "AGAINST": 0.38095238095238093, + "NONE": 0.8641975308641975 + }, + "accuracy_pct": 80.53097345132744, + "macro_f1_pct": 69.46198630786444, + "weighted_f1_pct": 82.77027350202722, + "macro_precision_pct": 67.90935672514621, + "macro_recall_pct": 80.99259798568136, + "per_class_f1_pct": { + "FAVOR": 83.87096774193549, + "AGAINST": 38.095238095238095, + "NONE": 86.41975308641975 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7204301075268817, + "macro_f1": 0.6972286972286973, + "weighted_f1": 0.7268967591548237, + "macro_precision": 0.6811834823329077, + "macro_recall": 0.7331649831649831, + "per_class_f1": { + "FAVOR": 0.6153846153846154, + "AGAINST": 0.7692307692307693, + "NONE": 0.7070707070707071 + }, + "accuracy_pct": 72.04301075268818, + "macro_f1_pct": 69.72286972286973, + "weighted_f1_pct": 72.68967591548237, + "macro_precision_pct": 68.11834823329077, + "macro_recall_pct": 73.31649831649831, + "per_class_f1_pct": { + "FAVOR": 61.53846153846154, + "AGAINST": 76.92307692307693, + "NONE": 70.70707070707071 + } + } + ], + "accuracy": 0.7094155844155844, + "macro_f1": 0.7087624822325092, + "weighted_f1": 0.7101218691325805, + "macro_precision": 0.7046863653505122, + "macro_recall": 0.7213074764907176, + "per_class_f1": { + "FAVOR": 0.6609929078014184, + "AGAINST": 0.7085377821393523, + "NONE": 0.7567567567567568 + }, + "accuracy_pct": 70.94155844155844, + "macro_f1_pct": 70.87624822325091, + "weighted_f1_pct": 71.01218691325805, + "macro_precision_pct": 70.46863653505122, + "macro_recall_pct": 72.13074764907176, + "per_class_f1_pct": { + "FAVOR": 66.09929078014184, + "AGAINST": 70.85377821393523, + "NONE": 75.67567567567568 + } + }, + { + "after_task": 8, + "best_epoch": 9, + "best_dev_macro_f1": 0.6597064989517819, + "losses": { + "current_ce": 0.6451548337936401, + "replay_ce": 0.2370354709525903, + "dark_distillation": 0.4551077346007029, + "supervised_contrastive": 4.321132810910543, + "total": 1.4016534114877384, + "replay_multiplier": 0.685048702314224, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.10017886626050582, + "ema_update_magnitude": 0.058735631177916615, + "final_multiplier": 0.5863076052905218 + } + }, + "memory": { + "size": 512, + "samples_per_task": 64, + "label_counts": { + "0": 177, + "1": 165, + "2": 170 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7328767123287672, + "macro_f1": 0.715940355940356, + "weighted_f1": 0.7398620253414775, + "macro_precision": 0.6955414482911345, + "macro_recall": 0.8022738630684657, + "per_class_f1": { + "FAVOR": 0.6349206349206349, + "AGAINST": 0.7662337662337663, + "NONE": 0.7466666666666667 + }, + "accuracy_pct": 73.28767123287672, + "macro_f1_pct": 71.5940355940356, + "weighted_f1_pct": 73.98620253414775, + "macro_precision_pct": 69.55414482911345, + "macro_recall_pct": 80.22738630684657, + "per_class_f1_pct": { + "FAVOR": 63.49206349206349, + "AGAINST": 76.62337662337663, + "NONE": 74.66666666666667 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6263157894736842, + "macro_f1": 0.6206646003256173, + "weighted_f1": 0.6301769749405789, + "macro_precision": 0.6094611002148128, + "macro_recall": 0.6452432824981845, + "per_class_f1": { + "FAVOR": 0.5423728813559322, + "AGAINST": 0.6702702702702703, + "NONE": 0.6493506493506493 + }, + "accuracy_pct": 62.63157894736842, + "macro_f1_pct": 62.066460032561736, + "weighted_f1_pct": 63.01769749405789, + "macro_precision_pct": 60.94611002148128, + "macro_recall_pct": 64.52432824981845, + "per_class_f1_pct": { + "FAVOR": 54.23728813559322, + "AGAINST": 67.02702702702703, + "NONE": 64.93506493506493 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6954314720812182, + "macro_f1": 0.6910035867217273, + "weighted_f1": 0.6977446464503572, + "macro_precision": 0.6778919330289194, + "macro_recall": 0.7335247319630452, + "per_class_f1": { + "FAVOR": 0.684931506849315, + "AGAINST": 0.7106598984771574, + "NONE": 0.6774193548387096 + }, + "accuracy_pct": 69.54314720812182, + "macro_f1_pct": 69.10035867217273, + "weighted_f1_pct": 69.77446464503572, + "macro_precision_pct": 67.78919330289193, + "macro_recall_pct": 73.35247319630453, + "per_class_f1_pct": { + "FAVOR": 68.4931506849315, + "AGAINST": 71.06598984771574, + "NONE": 67.74193548387096 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.64, + "macro_f1": 0.6525543883250572, + "weighted_f1": 0.635968108888757, + "macro_precision": 0.6485507246376812, + "macro_recall": 0.6637713385680865, + "per_class_f1": { + "FAVOR": 0.5714285714285714, + "AGAINST": 0.6506024096385542, + "NONE": 0.735632183908046 + }, + "accuracy_pct": 64.0, + "macro_f1_pct": 65.25543883250572, + "weighted_f1_pct": 63.5968108888757, + "macro_precision_pct": 64.85507246376811, + "macro_recall_pct": 66.37713385680864, + "per_class_f1_pct": { + "FAVOR": 57.14285714285714, + "AGAINST": 65.06024096385542, + "NONE": 73.5632183908046 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.845, + "macro_f1": 0.8251524238284986, + "weighted_f1": 0.8437551627425782, + "macro_precision": 0.8209604519774011, + "macro_recall": 0.8362263312788051, + "per_class_f1": { + "FAVOR": 0.8135593220338984, + "AGAINST": 0.7777777777777778, + "NONE": 0.8841201716738197 + }, + "accuracy_pct": 84.5, + "macro_f1_pct": 82.51524238284986, + "weighted_f1_pct": 84.37551627425782, + "macro_precision_pct": 82.09604519774011, + "macro_recall_pct": 83.6226331278805, + "per_class_f1_pct": { + "FAVOR": 81.35593220338984, + "AGAINST": 77.77777777777779, + "NONE": 88.41201716738198 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7433628318584071, + "macro_f1": 0.6380121816168328, + "weighted_f1": 0.7801245112162997, + "macro_precision": 0.6505636070853462, + "macro_recall": 0.7814100230554545, + "per_class_f1": { + "FAVOR": 0.7678571428571429, + "AGAINST": 0.2857142857142857, + "NONE": 0.8604651162790697 + }, + "accuracy_pct": 74.33628318584071, + "macro_f1_pct": 63.80121816168328, + "weighted_f1_pct": 78.01245112162997, + "macro_precision_pct": 65.05636070853463, + "macro_recall_pct": 78.14100230554544, + "per_class_f1_pct": { + "FAVOR": 76.78571428571429, + "AGAINST": 28.57142857142857, + "NONE": 86.04651162790698 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7311827956989247, + "macro_f1": 0.7118306758753024, + "weighted_f1": 0.7361186585507917, + "macro_precision": 0.6962532299741602, + "macro_recall": 0.7479797979797981, + "per_class_f1": { + "FAVOR": 0.6575342465753424, + "AGAINST": 0.7731958762886598, + "NONE": 0.7047619047619048 + }, + "accuracy_pct": 73.11827956989248, + "macro_f1_pct": 71.18306758753023, + "weighted_f1_pct": 73.61186585507917, + "macro_precision_pct": 69.62532299741602, + "macro_recall_pct": 74.7979797979798, + "per_class_f1_pct": { + "FAVOR": 65.75342465753424, + "AGAINST": 77.31958762886599, + "NONE": 70.47619047619048 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.6126760563380281, + "macro_f1": 0.609984111148062, + "weighted_f1": 0.5923923989564257, + "macro_precision": 0.614023114023114, + "macro_recall": 0.6376068376068376, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.4536082474226804, + "NONE": 0.7096774193548387 + }, + "accuracy_pct": 61.267605633802816, + "macro_f1_pct": 60.998411114806196, + "weighted_f1_pct": 59.23923989564257, + "macro_precision_pct": 61.402311402311405, + "macro_recall_pct": 63.760683760683754, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 45.36082474226804, + "NONE": 70.96774193548387 + } + } + ], + "accuracy": 0.7037845705967977, + "macro_f1": 0.7012906353094626, + "weighted_f1": 0.7012356158279139, + "macro_precision": 0.6978118619532511, + "macro_recall": 0.714270435446906, + "per_class_f1": { + "FAVOR": 0.652542372881356, + "AGAINST": 0.687333923826395, + "NONE": 0.7639956092206367 + }, + "accuracy_pct": 70.37845705967976, + "macro_f1_pct": 70.12906353094625, + "weighted_f1_pct": 70.1235615827914, + "macro_precision_pct": 69.7811861953251, + "macro_recall_pct": 71.4270435446906, + "per_class_f1_pct": { + "FAVOR": 65.2542372881356, + "AGAINST": 68.73339238263951, + "NONE": 76.39956092206367 + } + }, + { + "after_task": 9, + "best_epoch": 3, + "best_dev_macro_f1": 0.7190994704094557, + "losses": { + "current_ce": 0.5703916194214337, + "replay_ce": 0.22581045247506404, + "dark_distillation": 0.409260602619337, + "supervised_contrastive": 4.283192065142203, + "total": 1.2568184225455574, + "replay_multiplier": 0.5946455046275589, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09400880390005131, + "ema_update_magnitude": 0.05512870784228559, + "final_multiplier": 0.5864206920545183 + } + }, + "memory": { + "size": 576, + "samples_per_task": 64, + "label_counts": { + "0": 199, + "1": 186, + "2": 191 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7191780821917808, + "macro_f1": 0.6985476859346945, + "weighted_f1": 0.7305364126544737, + "macro_precision": 0.682283666154634, + "macro_recall": 0.7792853573213394, + "per_class_f1": { + "FAVOR": 0.5970149253731343, + "AGAINST": 0.7662337662337663, + "NONE": 0.7323943661971831 + }, + "accuracy_pct": 71.91780821917808, + "macro_f1_pct": 69.85476859346946, + "weighted_f1_pct": 73.05364126544737, + "macro_precision_pct": 68.2283666154634, + "macro_recall_pct": 77.92853573213394, + "per_class_f1_pct": { + "FAVOR": 59.70149253731343, + "AGAINST": 76.62337662337663, + "NONE": 73.23943661971832 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.6368421052631579, + "macro_f1": 0.620980870980871, + "weighted_f1": 0.6429631771737035, + "macro_precision": 0.6180068226120858, + "macro_recall": 0.640159767610748, + "per_class_f1": { + "FAVOR": 0.5873015873015873, + "AGAINST": 0.6923076923076923, + "NONE": 0.5833333333333334 + }, + "accuracy_pct": 63.68421052631579, + "macro_f1_pct": 62.0980870980871, + "weighted_f1_pct": 64.29631771737036, + "macro_precision_pct": 61.800682261208586, + "macro_recall_pct": 64.0159767610748, + "per_class_f1_pct": { + "FAVOR": 58.730158730158735, + "AGAINST": 69.23076923076923, + "NONE": 58.333333333333336 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6802030456852792, + "macro_f1": 0.6612307023265926, + "weighted_f1": 0.683272571945126, + "macro_precision": 0.6466208476517755, + "macro_recall": 0.6852912256347966, + "per_class_f1": { + "FAVOR": 0.6027397260273972, + "AGAINST": 0.7142857142857143, + "NONE": 0.6666666666666666 + }, + "accuracy_pct": 68.02030456852792, + "macro_f1_pct": 66.12307023265926, + "weighted_f1_pct": 68.3272571945126, + "macro_precision_pct": 64.66208476517755, + "macro_recall_pct": 68.52912256347966, + "per_class_f1_pct": { + "FAVOR": 60.273972602739725, + "AGAINST": 71.42857142857143, + "NONE": 66.66666666666666 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.7, + "macro_f1": 0.7049329501915708, + "weighted_f1": 0.6986795977011494, + "macro_precision": 0.6997024861552963, + "macro_recall": 0.7136288884256364, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.7125, + "NONE": 0.735632183908046 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 70.49329501915707, + "weighted_f1_pct": 69.86795977011494, + "macro_precision_pct": 69.97024861552963, + "macro_recall_pct": 71.36288884256365, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 71.25, + "NONE": 73.5632183908046 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.81, + "macro_f1": 0.7979534767805374, + "weighted_f1": 0.8130973821068488, + "macro_precision": 0.7767106842737096, + "macro_recall": 0.8330279304792048, + "per_class_f1": { + "FAVOR": 0.7868852459016393, + "AGAINST": 0.7619047619047619, + "NONE": 0.8450704225352113 + }, + "accuracy_pct": 81.0, + "macro_f1_pct": 79.79534767805374, + "weighted_f1_pct": 81.30973821068488, + "macro_precision_pct": 77.67106842737095, + "macro_recall_pct": 83.30279304792047, + "per_class_f1_pct": { + "FAVOR": 78.68852459016394, + "AGAINST": 76.19047619047619, + "NONE": 84.50704225352112 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7079646017699115, + "macro_f1": 0.6124622565095191, + "weighted_f1": 0.7514140850194144, + "macro_precision": 0.6329363460858688, + "macro_recall": 0.7551996116976095, + "per_class_f1": { + "FAVOR": 0.7256637168141593, + "AGAINST": 0.25806451612903225, + "NONE": 0.8536585365853658 + }, + "accuracy_pct": 70.79646017699115, + "macro_f1_pct": 61.246225650951914, + "weighted_f1_pct": 75.14140850194144, + "macro_precision_pct": 63.29363460858688, + "macro_recall_pct": 75.51996116976095, + "per_class_f1_pct": { + "FAVOR": 72.56637168141593, + "AGAINST": 25.806451612903224, + "NONE": 85.36585365853658 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7526881720430108, + "macro_f1": 0.7306397306397306, + "weighted_f1": 0.7575757575757576, + "macro_precision": 0.7132275132275132, + "macro_recall": 0.7630190796857464, + "per_class_f1": { + "FAVOR": 0.6666666666666666, + "AGAINST": 0.797979797979798, + "NONE": 0.7272727272727273 + }, + "accuracy_pct": 75.26881720430107, + "macro_f1_pct": 73.06397306397307, + "weighted_f1_pct": 75.75757575757575, + "macro_precision_pct": 71.32275132275132, + "macro_recall_pct": 76.30190796857464, + "per_class_f1_pct": { + "FAVOR": 66.66666666666666, + "AGAINST": 79.7979797979798, + "NONE": 72.72727272727273 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5774647887323944, + "macro_f1": 0.5686327841291522, + "weighted_f1": 0.5701792135649518, + "macro_precision": 0.5998650472334682, + "macro_recall": 0.5619658119658121, + "per_class_f1": { + "FAVOR": 0.5306122448979592, + "AGAINST": 0.4915254237288136, + "NONE": 0.6837606837606838 + }, + "accuracy_pct": 57.74647887323944, + "macro_f1_pct": 56.86327841291522, + "weighted_f1_pct": 57.01792135649518, + "macro_precision_pct": 59.986504723346826, + "macro_recall_pct": 56.19658119658121, + "per_class_f1_pct": { + "FAVOR": 53.06122448979592, + "AGAINST": 49.152542372881356, + "NONE": 68.37606837606837 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.695, + "macro_f1": 0.6771068992042775, + "weighted_f1": 0.6902864790224342, + "macro_precision": 0.696494651333361, + "macro_recall": 0.6718103181958605, + "per_class_f1": { + "FAVOR": 0.6067415730337079, + "AGAINST": 0.6518518518518519, + "NONE": 0.7727272727272727 + }, + "accuracy_pct": 69.5, + "macro_f1_pct": 67.71068992042774, + "weighted_f1_pct": 69.02864790224342, + "macro_precision_pct": 69.6494651333361, + "macro_recall_pct": 67.18103181958604, + "per_class_f1_pct": { + "FAVOR": 60.67415730337079, + "AGAINST": 65.18518518518519, + "NONE": 77.27272727272727 + } + } + ], + "accuracy": 0.7007623888182973, + "macro_f1": 0.6970988648825823, + "weighted_f1": 0.7000367254497073, + "macro_precision": 0.6946830190975358, + "macro_recall": 0.701140201299623, + "per_class_f1": { + "FAVOR": 0.6451612903225806, + "AGAINST": 0.6971080669710806, + "NONE": 0.7490272373540856 + }, + "accuracy_pct": 70.07623888182974, + "macro_f1_pct": 69.70988648825823, + "weighted_f1_pct": 70.00367254497073, + "macro_precision_pct": 69.46830190975358, + "macro_recall_pct": 70.1140201299623, + "per_class_f1_pct": { + "FAVOR": 64.51612903225806, + "AGAINST": 69.71080669710807, + "NONE": 74.90272373540856 + } + }, + { + "after_task": 10, + "best_epoch": 3, + "best_dev_macro_f1": 0.6670726163568798, + "losses": { + "current_ce": 0.6257781086823879, + "replay_ce": 0.25476715205571593, + "dark_distillation": 0.40199438226528655, + "supervised_contrastive": 4.2510028924697485, + "total": 1.3297316633738003, + "replay_multiplier": 0.6097074989263762, + "update_controller": { + "warmup_steps": 16, + "baseline_update_magnitude": 0.09897892460005277, + "ema_update_magnitude": 0.054363232901227826, + "final_multiplier": 0.5492404885271793 + } + }, + "memory": { + "size": 640, + "samples_per_task": 64, + "label_counts": { + "0": 221, + "1": 207, + "2": 212 + }, + "task_label_counts": { + "0": { + "0": 22, + "1": 21, + "2": 21 + }, + "1": { + "0": 22, + "1": 21, + "2": 21 + }, + "2": { + "0": 22, + "1": 21, + "2": 21 + }, + "3": { + "0": 22, + "1": 21, + "2": 21 + }, + "4": { + "0": 22, + "1": 21, + "2": 21 + }, + "5": { + "0": 23, + "1": 18, + "2": 23 + }, + "6": { + "0": 22, + "1": 21, + "2": 21 + }, + "7": { + "0": 22, + "1": 21, + "2": 21 + }, + "8": { + "0": 22, + "1": 21, + "2": 21 + }, + "9": { + "0": 22, + "1": 21, + "2": 21 + } + } + }, + "per_target": [ + { + "task_index": 0, + "target": "Atheism", + "accuracy": 0.7054794520547946, + "macro_f1": 0.6892608201217474, + "weighted_f1": 0.7141775736132989, + "macro_precision": 0.6746379807202642, + "macro_recall": 0.7799100449775113, + "per_class_f1": { + "FAVOR": 0.6060606060606061, + "AGAINST": 0.7417218543046358, + "NONE": 0.72 + }, + "accuracy_pct": 70.54794520547945, + "macro_f1_pct": 68.92608201217473, + "weighted_f1_pct": 71.41775736132989, + "macro_precision_pct": 67.46379807202642, + "macro_recall_pct": 77.99100449775112, + "per_class_f1_pct": { + "FAVOR": 60.60606060606061, + "AGAINST": 74.17218543046357, + "NONE": 72.0 + } + }, + { + "task_index": 1, + "target": "Feminist Movement", + "accuracy": 0.5947368421052631, + "macro_f1": 0.5987190197716513, + "weighted_f1": 0.5976520247988392, + "macro_precision": 0.5978020634047051, + "macro_recall": 0.6336238198983297, + "per_class_f1": { + "FAVOR": 0.556390977443609, + "AGAINST": 0.6081871345029239, + "NONE": 0.631578947368421 + }, + "accuracy_pct": 59.473684210526315, + "macro_f1_pct": 59.87190197716513, + "weighted_f1_pct": 59.76520247988392, + "macro_precision_pct": 59.78020634047051, + "macro_recall_pct": 63.36238198983297, + "per_class_f1_pct": { + "FAVOR": 55.639097744360896, + "AGAINST": 60.81871345029239, + "NONE": 63.1578947368421 + } + }, + { + "task_index": 2, + "target": "Hillary Clinton", + "accuracy": 0.6903553299492385, + "macro_f1": 0.6796858396858397, + "weighted_f1": 0.6905775123034005, + "macro_precision": 0.6661173368490442, + "macro_recall": 0.7270308498835307, + "per_class_f1": { + "FAVOR": 0.6216216216216216, + "AGAINST": 0.6974358974358974, + "NONE": 0.72 + }, + "accuracy_pct": 69.03553299492386, + "macro_f1_pct": 67.96858396858397, + "weighted_f1_pct": 69.05775123034005, + "macro_precision_pct": 66.61173368490442, + "macro_recall_pct": 72.70308498835307, + "per_class_f1_pct": { + "FAVOR": 62.16216216216216, + "AGAINST": 69.74358974358974, + "NONE": 72.0 + } + }, + { + "task_index": 3, + "target": "face masks", + "accuracy": 0.685, + "macro_f1": 0.6920975323149235, + "weighted_f1": 0.683139982373678, + "macro_precision": 0.6834036568213783, + "macro_recall": 0.7080466958515738, + "per_class_f1": { + "FAVOR": 0.675, + "AGAINST": 0.6621621621621622, + "NONE": 0.7391304347826086 + }, + "accuracy_pct": 68.5, + "macro_f1_pct": 69.20975323149236, + "weighted_f1_pct": 68.3139982373678, + "macro_precision_pct": 68.34036568213783, + "macro_recall_pct": 70.80466958515737, + "per_class_f1_pct": { + "FAVOR": 67.5, + "AGAINST": 66.21621621621621, + "NONE": 73.91304347826086 + } + }, + { + "task_index": 4, + "target": "stay at home orders", + "accuracy": 0.83, + "macro_f1": 0.8079288302412034, + "weighted_f1": 0.83240465416936, + "macro_precision": 0.7884768307670065, + "macro_recall": 0.8389249819534678, + "per_class_f1": { + "FAVOR": 0.7619047619047619, + "AGAINST": 0.7931034482758621, + "NONE": 0.8687782805429864 + }, + "accuracy_pct": 83.0, + "macro_f1_pct": 80.79288302412034, + "weighted_f1_pct": 83.240465416936, + "macro_precision_pct": 78.84768307670065, + "macro_recall_pct": 83.89249819534677, + "per_class_f1_pct": { + "FAVOR": 76.19047619047619, + "AGAINST": 79.3103448275862, + "NONE": 86.87782805429865 + } + }, + { + "task_index": 5, + "target": "Climate Change is a Real Concern", + "accuracy": 0.7876106194690266, + "macro_f1": 0.6887736414206246, + "weighted_f1": 0.8033386120167676, + "macro_precision": 0.6647643077415996, + "macro_recall": 0.796820774177891, + "per_class_f1": { + "FAVOR": 0.816, + "AGAINST": 0.42105263157894735, + "NONE": 0.8292682926829268 + }, + "accuracy_pct": 78.76106194690266, + "macro_f1_pct": 68.87736414206246, + "weighted_f1_pct": 80.33386120167675, + "macro_precision_pct": 66.47643077415995, + "macro_recall_pct": 79.6820774177891, + "per_class_f1_pct": { + "FAVOR": 81.6, + "AGAINST": 42.10526315789473, + "NONE": 82.92682926829268 + } + }, + { + "task_index": 6, + "target": "Legalization of Abortion", + "accuracy": 0.7580645161290323, + "macro_f1": 0.7340101522842639, + "weighted_f1": 0.7620084680343401, + "macro_precision": 0.7197540236866079, + "macro_recall": 0.7650392817059485, + "per_class_f1": { + "FAVOR": 0.6571428571428571, + "AGAINST": 0.8020304568527918, + "NONE": 0.7428571428571429 + }, + "accuracy_pct": 75.80645161290323, + "macro_f1_pct": 73.40101522842639, + "weighted_f1_pct": 76.200846803434, + "macro_precision_pct": 71.9754023686608, + "macro_recall_pct": 76.50392817059485, + "per_class_f1_pct": { + "FAVOR": 65.71428571428571, + "AGAINST": 80.20304568527918, + "NONE": 74.28571428571429 + } + }, + { + "task_index": 7, + "target": "Donald Trump", + "accuracy": 0.5845070422535211, + "macro_f1": 0.5623136123136123, + "weighted_f1": 0.5440013820295511, + "macro_precision": 0.5652831402831403, + "macro_recall": 0.617948717948718, + "per_class_f1": { + "FAVOR": 0.6, + "AGAINST": 0.3409090909090909, + "NONE": 0.746031746031746 + }, + "accuracy_pct": 58.45070422535211, + "macro_f1_pct": 56.231361231361234, + "weighted_f1_pct": 54.400138202955105, + "macro_precision_pct": 56.52831402831403, + "macro_recall_pct": 61.794871794871796, + "per_class_f1_pct": { + "FAVOR": 60.0, + "AGAINST": 34.090909090909086, + "NONE": 74.60317460317461 + } + }, + { + "task_index": 8, + "target": "fauci", + "accuracy": 0.7, + "macro_f1": 0.6845790852219841, + "weighted_f1": 0.6980872134554192, + "macro_precision": 0.690795839380745, + "macro_recall": 0.6846308310163732, + "per_class_f1": { + "FAVOR": 0.6111111111111112, + "AGAINST": 0.6610169491525424, + "NONE": 0.7816091954022989 + }, + "accuracy_pct": 70.0, + "macro_f1_pct": 68.45790852219841, + "weighted_f1_pct": 69.80872134554193, + "macro_precision_pct": 69.07958393807449, + "macro_recall_pct": 68.46308310163732, + "per_class_f1_pct": { + "FAVOR": 61.111111111111114, + "AGAINST": 66.10169491525424, + "NONE": 78.16091954022988 + } + }, + { + "task_index": 9, + "target": "school closures", + "accuracy": 0.665, + "macro_f1": 0.6447329538384969, + "weighted_f1": 0.6703621337252705, + "macro_precision": 0.6397890023389928, + "macro_recall": 0.6562224183583408, + "per_class_f1": { + "FAVOR": 0.6979166666666666, + "AGAINST": 0.44680851063829785, + "NONE": 0.7894736842105263 + }, + "accuracy_pct": 66.5, + "macro_f1_pct": 64.47329538384969, + "weighted_f1_pct": 67.03621337252706, + "macro_precision_pct": 63.97890023389928, + "macro_recall_pct": 65.62224183583409, + "per_class_f1_pct": { + "FAVOR": 69.79166666666666, + "AGAINST": 44.680851063829785, + "NONE": 78.94736842105263 + } + } + ], + "accuracy": 0.6995490417136415, + "macro_f1": 0.6994550539497663, + "weighted_f1": 0.6964575969159251, + "macro_precision": 0.698337255055825, + "macro_recall": 0.7110100791666873, + "per_class_f1": { + "FAVOR": 0.6654099905749293, + "AGAINST": 0.661526599845798, + "NONE": 0.7714285714285715 + }, + "accuracy_pct": 69.95490417136415, + "macro_f1_pct": 69.94550539497662, + "weighted_f1_pct": 69.64575969159252, + "macro_precision_pct": 69.8337255055825, + "macro_recall_pct": 71.10100791666873, + "per_class_f1_pct": { + "FAVOR": 66.54099905749294, + "AGAINST": 66.1526599845798, + "NONE": 77.14285714285715 + } + } + ] +} \ No newline at end of file diff --git a/outputs/trace_s8_full__x__s3__seed13/resolved_config.yaml b/outputs/trace_s8_full__x__s3__seed13/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..46acfb799d940c698b2ae92df0c8723e3ad65d29 --- /dev/null +++ b/outputs/trace_s8_full__x__s3__seed13/resolved_config.yaml @@ -0,0 +1,57 @@ +run: + name: trace_s8_full + output_root: /root/BIRD/outputs + device: auto + deterministic: true + overwrite: false +seed: 13 +data: + root: /root/BIRD/data/constance + corpus: x + sequence: s3 + max_length: 256 + tokenization_mode: text_pair + timestamp_field: null +protocol: + mode: paper_target_aware +model: + family: trace + stage: 8 + backbone: bert-base-uncased + num_labels: 3 + dropout: 0.1 + label_smoothing: 0.0 + class_balance_beta: 0.999 + lora_rank: 8 + lora_alpha: 16.0 + lora_dropout: 0.05 + lora_last_n_layers: null + lora_targets: + - query + - value + contrastive_projection_size: 128 +training: + batch_size: 64 + max_epochs: 15 + learning_rate: 0.0001 + weight_decay: 0.0 + gradient_clip: 1.0 + num_workers: 2 + amp: true + amp_dtype: bf16 + max_train_batches: null +replay: + samples_per_task: 64 + batch_ratio: 1.0 + classification_weight: 1.0 + distillation_weight: 0.5 + contrastive_weight: 0.1 + contrastive_temperature: 0.1 +update_controller: + warmup_steps: 16 + smoothing: 0.05 + sensitivity: 1.0 + minimum: 0.5 + maximum: 3.0 +evaluation: + max_eval_batches: null diff --git a/outputs/trace_s8_full__x__s3__seed13/run.log b/outputs/trace_s8_full__x__s3__seed13/run.log new file mode 100644 index 0000000000000000000000000000000000000000..2eca03adeb3f7240b0cf7fb0235e930f6246018f --- /dev/null +++ b/outputs/trace_s8_full__x__s3__seed13/run.log @@ -0,0 +1,3 @@ +2026-09-02 10:56:34,477 [INFO] __main__: run_dir=/root/BIRD/outputs/trace_s8_full__x__s3__seed13 device=cuda +2026-09-02 11:12:52,095 [INFO] __main__: completed summary={'final_mean_target_macro_f1': 0.6782101487214348, 'average_forgetting': 0.01737079029072852, 'backward_transfer': -0.003406699817294393, 'final_mean_target_macro_f1_pct': 67.82101487214348, 'average_forgetting_pct': 1.7370790290728522, 'backward_transfer_pct': -0.3406699817294393, 'final_pooled_accuracy': 0.6995490417136415, 'final_pooled_accuracy_pct': 69.95490417136415, 'final_pooled_macro_f1': 0.6994550539497663, 'final_pooled_macro_f1_pct': 69.94550539497662, 'final_pooled_weighted_f1': 0.6964575969159251, 'final_pooled_weighted_f1_pct': 69.64575969159252, 'final_pooled_macro_precision': 0.698337255055825, 'final_pooled_macro_precision_pct': 69.8337255055825, 'final_pooled_macro_recall': 0.7110100791666873, 'final_pooled_macro_recall_pct': 71.10100791666873, 'final_pooled_per_class_f1': {'FAVOR': 0.6654099905749293, 'AGAINST': 0.661526599845798, 'NONE': 0.7714285714285715}, 'final_pooled_per_class_f1_pct': {'FAVOR': 66.54099905749294, 'AGAINST': 66.1526599845798, 'NONE': 77.14285714285715}} +2026-09-02 11:12:52,095 [INFO] __main__: efficiency={'training_wall_time_seconds': 976.2795598909725, 'training_wall_time_minutes': 16.271325998182874, 'peak_gpu_memory_allocated_gb': 10.431715965270996, 'peak_gpu_memory_reserved_gb': 10.58984375}