Prepare Solstice-1 175M private release package

Merged FP32 checkpoint, frozen tokenizer, native MLX runner, Apache-2.0 license, technical documentation, training provenance and benchmark evidence.

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+docs/figures/cooldown-detail.png filter=lfs diff=lfs merge=lfs -text
+docs/figures/pretraining-overview.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,6 @@
+__pycache__/
+*.pyc
+.venv/
+.DS_Store
+model.safetensors
+*.zip

BACKUP_AND_CLEANUP.md

@@ -0,0 +1,5 @@
+# Preservation and cleanup
+
+The selected checkpoint is stage-3 update 128. The inference export hash and ancestor identities are recorded in provenance/LINEAGE.json. Existing pretrained masters, parent 2000, parent 1800, prior attempt-5 checkpoints and checkpoint 167 remain protected locally; preparing this release does not delete or modify them.
+
+The GitHub source repository excludes the large merged weight tensor. Hugging Face holds the inference weight artifact. Verify both package manifests and load the independently downloaded artifact before treating a public upload as complete. Retain local originals until that verification has succeeded. Optimizer state and raw corpora are not backed up by the inference release.

CITATION.cff

@@ -0,0 +1,10 @@
+cff-version: 1.2.0
+message: "If you use Solstice-1, cite this model and its recorded release revision."
+title: "Solstice-1: a 175M English chat research model"
+authors:
+  - family-names: Zebrowski
+    given-names: William
+type: software
+version: "1.0.0"
+license: Apache-2.0
+url: "https://huggingface.co/OpenSML/Solstice-1"

FILE_MANIFEST.json

@@ -0,0 +1,1875 @@
+{
+  "format": "solstice-release-files-v1",
+  "files": {
+    ".gitignore": {
+      "bytes": 60,
+      "sha256": "70a6a34e99e8ec921af10056c133b8035f937a7a5e43131c06f8817ea04222e0"
+    },
+    "BACKUP_AND_CLEANUP.md": {
+      "bytes": 730,
+      "sha256": "f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859"
+    },
+    "CITATION.cff": {
+      "bytes": 323,
+      "sha256": "af646a190d135ec25a6ee2c76d1594c3f9ba8005cd3ca6df908606a15023273e"
+    },
+    "LICENSE": {
+      "bytes": 11358,
+      "sha256": "cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30"
+    },
+    "PACKAGING_VERIFICATION.json": {
+      "bytes": 1269,
+      "sha256": "0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375"
+    },
+    "README.md": {
+      "bytes": 7647,
+      "sha256": "2bcb95ff85d71d6f2c9fa6711bcbad62c52d68ab3367a8317db72b6d88bb04fd"
+    },
+    "REPOSITORIES.json": {
+      "bytes": 716,
+      "sha256": "152a87d0098752350d54fdd3cde3861a77eb497c53dc973ee1a8370d12a13ff6"
+    },
+    "RUN_GUIDE.md": {
+      "bytes": 3339,
+      "sha256": "fd05e1595bcdc52942e9425b6efb2c7e89553e6b0ad7d5f93d9075969cd75dba"
+    },
+    "SOURCE_NOTICE.md": {
+      "bytes": 1054,
+      "sha256": "87fa33719fe1182d467a05ab417be2a362f50eea59f0ac6496d40e677544e27c"
+    },
+    "THIRD_PARTY_NOTICES.md": {
+      "bytes": 4208,
+      "sha256": "85525fd4cfd74dbabeb8fe767074cf02a8834dc40a3fc2e648180ef2a3ff35d9"
+    },
+    "TRAINING_AND_REPRODUCIBILITY.md": {
+      "bytes": 4232,
+      "sha256": "a4bbaa7901cce3f88cdcac1ffd30ecdf60d4999c198054dce6bb7ef1d72607b2"
+    },
+    "config.json": {
+      "bytes": 913,
+      "sha256": "c98126fd2051388690f7d6ad709f961afbb654f0268e4fd4db07e0ead915ee46"
+    },
+    "configs/pretrain_20261022_121457_b3d5f398/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261022_121457_b3d5f398/recipe.json": {
+      "bytes": 1311,
+      "sha256": "ddf14c4cef8a4fb66392a27e87f8c5a807a88da35fe61599f252c9eba91bec29"
+    },
+    "configs/pretrain_20261024_111140_5f596b7a/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261024_111140_5f596b7a/recipe.json": {
+      "bytes": 2142,
+      "sha256": "900c630d3a315d63ed332a9e300a2d1a86f9a2dd8e9627a15ad6430c2077162e"
+    },
+    "configs/pretrain_20261027_222615_4d065da7/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261027_222615_4d065da7/recipe.json": {
+      "bytes": 3616,
+      "sha256": "a81a72c7bfe47b4b9e47ce0b63bb7c6b5a139ddba905c5d2a017bd87f249a0d6"
+    },
+    "configs/pretrain_20261028_093017_dd42e713/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261028_093017_dd42e713/recipe.json": {
+      "bytes": 4440,
+      "sha256": "0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1"
+    },
+    "configs/pretrain_20261102_234538_0ee851dc/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261102_234538_0ee851dc/recipe.json": {
+      "bytes": 4440,
+      "sha256": "0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1"
+    },
+    "configs/pretrain_20261103_000432_8ac2605e/corpus.json": {
+      "bytes": 1356,
+      "sha256": "c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632"
+    },
+    "configs/pretrain_20261103_000432_8ac2605e/recipe.json": {
+      "bytes": 4440,
+      "sha256": "0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1"
+    },
+    "configs/pretrain_20261104_170916_448e6f64/corpus.json": {
+      "bytes": 11740,
+      "sha256": "f27196edf4eab3270c3ec9a8e232509fcc4435baa68c0d976acdf6b5881a14cb"
+    },
+    "configs/pretrain_20261104_170916_448e6f64/recipe.json": {
+      "bytes": 5222,
+      "sha256": "218451ccbe74415b0e4b2e810200feaacf1e17bb582c0a9f1f84e381d6bf9ad9"
+    },
+    "configs/pretrain_20261107_142423_40700422/corpus.json": {
+      "bytes": 11740,
+      "sha256": "f27196edf4eab3270c3ec9a8e232509fcc4435baa68c0d976acdf6b5881a14cb"
+    },
+    "configs/pretrain_20261107_142423_40700422/recipe.json": {
+      "bytes": 6041,
+      "sha256": "c70c4491d8865d0db8af21d555c309dcfcbaa346760f0668e82a8c0ce54c8d6f"
+    },
+    "configs/sft1/plan.json": {
+      "bytes": 5727,
+      "sha256": "fb204be1d9e72eb1cae140bfdb03dbcadc1843c47b1894a731396ceeedb9a22f"
+    },
+    "configs/sft2/plan.json": {
+      "bytes": 6296,
+      "sha256": "51d66108416988bad1015d7b52cf63db0571f87f202b2b18c7e98e67a29ec62b"
+    },
+    "configs/sft3/plan.json": {
+      "bytes": 6229,
+      "sha256": "ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284"
+    },
+    "docs/DATA.md": {
+      "bytes": 6922,
+      "sha256": "e0b0c730321360c40f20e53a22890c52cd5294bad2a5777a0a891096557a2f0a"
+    },
+    "docs/EVALUATION.md": {
+      "bytes": 6613,
+      "sha256": "5522817f8d698f05b43adfc1c5227bab4e917492c898fa4b792d521d42a512d8"
+    },
+    "docs/EXAMPLES.md": {
+      "bytes": 1117,
+      "sha256": "7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3"
+    },
+    "docs/TECHNICAL_REPORT.md": {
+      "bytes": 13515,
+      "sha256": "f98d45b39b9595510fa80d495a3758b211d8c770adc739d381763d3cfb2a2a0b"
+    },
+    "docs/figures/cooldown-detail.png": {
+      "bytes": 185666,
+      "sha256": "0fa840653d1a5305297ef6f51ba08dc70a4bbfaf8029c411627e39dfc3cdc153"
+    },
+    "docs/figures/cooldown-detail.svg": {
+      "bytes": 174630,
+      "sha256": "1071d21211f7d32d9f9f5874e62f46a2a17306f88f0f355b24aee5b4af4c4e0f"
+    },
+    "docs/figures/pretraining-overview.png": {
+      "bytes": 157953,
+      "sha256": "c46bba5d4fae1c55701def21829de08fd3756dc9659fee2a54a79f42eb1afcb0"
+    },
+    "docs/figures/pretraining-overview.svg": {
+      "bytes": 360608,
+      "sha256": "862a5400ecde1ececb915441766f378849c65024288522c84a4883d0c4b31a4e"
+    },
+    "evaluation/README.md": {
+      "bytes": 1816,
+      "sha256": "9e15b1b93bbc1160f3988a7956eee88b2467e5d826f097df35401562c116b90d"
+    },
+    "evaluation/benchmark_best.py": {
+      "bytes": 22604,
+      "sha256": "f3835c8cb0bdf580e4142a57ac3724c474dcc1d44aa01dbbd805c989a587c7c7"
+    },
+    "evaluation/benchmark_ifeval.py": {
+      "bytes": 12169,
+      "sha256": "10ae9d2c49236c377d6e9952674b14289f5f92c310c29cbe592a40c98a14c752"
+    },
+    "evaluation/chat_adapter_search_bench.py": {
+      "bytes": 7431,
+      "sha256": "48dcb288740c2451e6d4218684809a201c078d517bfbac94a3adec61c8d10c40"
+    },
+    "evaluation/chat_benchmark/__init__.py": {
+      "bytes": 84,
+      "sha256": "66883e6b6589b81927fdc331b25f38915e59422cdf4b8339c715cd85d74e2b1d"
+    },
+    "evaluation/chat_benchmark/panel.py": {
+      "bytes": 12138,
+      "sha256": "70d78840147614727a34a775f669f9316c8d589721d41fdf2dd8c98deb2475bd"
+    },
+    "evaluation/chat_benchmark/scoring.py": {
+      "bytes": 3080,
+      "sha256": "e63a8715c302e2b59b0fadc3598e912d3f695fb9352b541e33dba3b046017c54"
+    },
+    "evaluation/vendor/ifeval/LICENSE": {
+      "bytes": 11358,
+      "sha256": "cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30"
+    },
+    "evaluation/vendor/ifeval/NOTICE": {
+      "bytes": 248,
+      "sha256": "bc7bead98aea2499b1226aca21021d267a9e926d36711eb9e2230ebb0841abfd"
+    },
+    "evaluation/vendor/ifeval/instruction_following_eval/__init__.py": {
+      "bytes": 75,
+      "sha256": "8f77aceef943735bec21b7f42630794cad7ce20d32b72a83889e20274fb6df31"
+    },
+    "evaluation/vendor/ifeval/instruction_following_eval/evaluation_lib.py": {
+      "bytes": 6984,
+      "sha256": "35decc06000718487f44d7deafa6d3f48a8ec0886281edf40162c0265b7d248c"
+    },
+    "evaluation/vendor/ifeval/instruction_following_eval/instructions.py": {
+      "bytes": 55162,
+      "sha256": "60e086f5342a03ce8e18b64bbcccf86308f523c08aa826707a562150a52f3edf"
+    },
+    "evaluation/vendor/ifeval/instruction_following_eval/instructions_registry.py": {
+      "bytes": 7240,
+      "sha256": "ec92d72c264f6d906978613085db262356174300370a3fffe6fefd5969ce9cfc"
+    },
+    "evaluation/vendor/ifeval/instruction_following_eval/instructions_util.py": {
+      "bytes": 19538,
+      "sha256": "a73797261eee5bf447e279d82a2b700b1bdd3cb1193412dbab1270a85832bc6b"
+    },
+    "evaluation/vendor/ifeval/prepared_runtime.json": {
+      "bytes": 15907,
+      "sha256": "5e3542d6369138730a11db9b7793fe1a722021b66ea8ad32b2e93d49ade91555"
+    },
+    "evaluation/vendor/ifeval/source.json": {
+      "bytes": 649,
+      "sha256": "d1ccd2965b9b8752d51b1e814faa5c4879e0c630047fb01c4fa50dca75527d70"
+    },
+    "inference.py": {
+      "bytes": 7065,
+      "sha256": "e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410"
+    },
+    "inference_native/__init__.py": {
+      "bytes": 50,
+      "sha256": "0628a9b64e3e12824a303a720a8456b2ea4341bf4b45a12aea25cbc71b68deaf"
+    },
+    "inference_native/chat_format.py": {
+      "bytes": 4129,
+      "sha256": "5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8"
+    },
+    "inference_native/common.py": {
+      "bytes": 250,
+      "sha256": "d9bdeeb4b8ea60dacb9af1cafa1dfbd295bfb384c2070d282848ec7e5eb42168"
+    },
+    "inference_native/model.py": {
+      "bytes": 10552,
+      "sha256": "1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe"
+    },
+    "inference_native/tokenization.py": {
+      "bytes": 2797,
+      "sha256": "254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5"
+    },
+    "model.safetensors": {
+      "bytes": 696844242,
+      "sha256": "7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515"
+    },
+    "provenance/INFERENCE_EXPORT.json": {
+      "bytes": 353,
+      "sha256": "1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f"
+    },
+    "provenance/LINEAGE.json": {
+      "bytes": 5763,
+      "sha256": "490082ad7c96b776738e4d6bb7f11507af0c697facd3a7b95d944abd64e25b3a"
+    },
+    "provenance/PRETRAINING_PHASES.json": {
+      "bytes": 41733,
+      "sha256": "0db456c4bf1a5e46f831195346626b73d953d412b23c47bc6b27eaa230babf42"
+    },
+    "provenance/SELECTED_PREFIX.json": {
+      "bytes": 6968,
+      "sha256": "d8aaf1ef62597ac384b6f7c9510b24c2bbb63470150c3835266203ee6a567fd8"
+    },
+    "provenance/TENSOR_INVENTORY.json": {
+      "bytes": 38395,
+      "sha256": "834edc532edfe9a897b60a68046006db6700bdd21de72f055199eb827aef401a"
+    },
+    "provenance/checkpoint_metadata.json": {
+      "bytes": 35799,
+      "sha256": "f1adf7d2c3a91ee6a4643beb26ce8d9cdfbdb69156e96439d01be657abf4ad20"
+    },
+    "provenance/pretrained_base_manifest.json": {
+      "bytes": 731,
+      "sha256": "f1f8519ca75a35343cfec7e88c8197bb84aa7d97d809353861b97c1cf3d77513"
+    },
+    "provenance/pretrained_base_metadata.json": {
+      "bytes": 7501,
+      "sha256": "8bc81bacb9ef74be0b3c0118e4d60f0d00ace0d17139040be8454b873f2cc448"
+    },
+    "provenance/sft1/checkpoint_manifest.json": {
+      "bytes": 988,
+      "sha256": "f9273450799963b2fd8e61a5e9cb6892be7ba4a2308849623e288302a687a1f7"
+    },
+    "provenance/sft1/contract.json": {
+      "bytes": 31985,
+      "sha256": "df241d47b313c08bd5f5d11436bcd05f8f3b53c0c6c7bdc065c071ee0c1dbf67"
+    },
+    "provenance/sft1/corpus_manifest.json": {
+      "bytes": 7120,
+      "sha256": "d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a"
+    },
+    "provenance/sft1/state.json": {
+      "bytes": 35335,
+      "sha256": "c6e0e82a2b4f80aa16b803cac7149e7984ad66a076d0105331cca08134a4fdd8"
+    },
+    "provenance/sft2/checkpoint_manifest.json": {
+      "bytes": 987,
+      "sha256": "dcab4ad8e4484ad1cc4b32ffbc6352b6a6ff9ea0a4ddccdb8d367d23db098a3f"
+    },
+    "provenance/sft2/contract.json": {
+      "bytes": 32961,
+      "sha256": "c7d30352bfe3bae2d52937c257d9ab48fccddd90054e4f8b83ee2f78f7d55d87"
+    },
+    "provenance/sft2/corpus_manifest.json": {
+      "bytes": 7120,
+      "sha256": "d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a"
+    },
+    "provenance/sft2/fresh_manifest.json": {
+      "bytes": 1683,
+      "sha256": "c11c2bcdbd109e60178da47b9ac830734c5a15557f96c33ae7405009a3159265"
+    },
+    "provenance/sft2/selected_indices.json": {
+      "bytes": 103574,
+      "sha256": "904bcd42c0497f4b0e0b99fdb8dbd5f2d6155d2dd9e9cab31d9c5021d0b7a4ec"
+    },
+    "provenance/sft2/state.json": {
+      "bytes": 36625,
+      "sha256": "de16e15cfd6687ddc0480dd113f1f01b6900a17737c99b7a4c46863a763df194"
+    },
+    "provenance/sft3/checkpoint_manifest.json": {
+      "bytes": 987,
+      "sha256": "2118c7446c412a4ee7a4b7a356a211ad91574983e213da45980707b5fa1ca124"
+    },
+    "provenance/sft3/contract.json": {
+      "bytes": 33562,
+      "sha256": "5776cf85c123d0560a40d4ebd255c10aae8ac1817006d8cf09f405731604ceae"
+    },
+    "provenance/sft3/corpus_manifest.json": {
+      "bytes": 7120,
+      "sha256": "d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a"
+    },
+    "provenance/sft3/fresh_manifest.json": {
+      "bytes": 1585,
+      "sha256": "741da3f1897cb5b092cdcab91051630b07262c285933efab3cc00a844a6c3d01"
+    },
+    "provenance/sft3/selected_indices.json": {
+      "bytes": 53066,
+      "sha256": "eee29c2b37c883cb6df6387ee1c734bb1a7011d2106fe97d7351a42d174a2531"
+    },
+    "provenance/sft3/state.json": {
+      "bytes": 35488,
+      "sha256": "78371afe2868d7e6278f4bf3a1de4c2033ef1673e6c590c3c1c6b671b8b56cc9"
+    },
+    "requirements.txt": {
+      "bytes": 44,
+      "sha256": "cfeac63c4d14d73a7b247bf9e39fd0b4642754aa4b69e7ea6cf94e79a594e3d8"
+    },
+    "results/FINAL_COMPARISON.json": {
+      "bytes": 9415,
+      "sha256": "28fdc175a0019e6a34590256c3c1da8fc3a39813699afae8a477d63b00249a7b"
+    },
+    "results/baselines/167/choices/results.json": {
+      "bytes": 3274,
+      "sha256": "f76237736d1aab8f04517ad9e3336925ca9773399a9ad888f07c15d3ceaa4f87"
+    },
+    "results/baselines/167/ifeval/results.json": {
+      "bytes": 2768,
+      "sha256": "187b395945254a4d9359694d5f9ead19a8b95caffcb42477cb9a349cf6ad625e"
+    },
+    "results/baselines/attempt5.json": {
+      "bytes": 4785,
+      "sha256": "a5d49e3dc60a6c4dda4799090881cc2aab033b922ac0b5dafd52c0c6c7e2e6d8"
+    },
+    "results/baselines/parent2000.json": {
+      "bytes": 4788,
+      "sha256": "2b59f728ebc9bef6deaf378e6908f7b55723b368f60da83f9af5b2b1731d0346"
+    },
+    "results/package_inference_verification.json": {
+      "bytes": 26015,
+      "sha256": "fed55313b61008e8f379a7b4bc49535cc501a31a942c7560687de4a4bd1cd47f"
+    },
+    "results/pretraining_steps.csv": {
+      "bytes": 2563275,
+      "sha256": "ba2cfb7c8b446dede8e7f33bb78885855e74bd75f861b0c201d6e01e88bc987f"
+    },
+    "results/pretraining_validation.csv": {
+      "bytes": 27917,
+      "sha256": "c4e3210780814051e3ad1bd99a9c143b9c5dbe910224239a134fddc546e9c9a8"
+    },
+    "results/public/chat/protocol.json": {
+      "bytes": 1496,
+      "sha256": "06f43fbe7a4f38dbafb7d82e778d25fe45c009b5939ee9c81919122758d02000"
+    },
+    "results/public/chat/responses.jsonl": {
+      "bytes": 139340,
+      "sha256": "fea56aceda29f7e78a9f8ce730c050c89b5e856753bb5609e97d3a2cafce159e"
+    },
+    "results/public/chat/results.json": {
+      "bytes": 3216,
+      "sha256": "784d4efc1d63d5b35a201f44525144755e586f27ad812085ee563a4ec2396eaf"
+    },
+    "results/public/choices/arc_challenge.jsonl": {
+      "bytes": 265180,
+      "sha256": "d467b86959d28256f44af33d3f3dfe877ea5849ac7d103eaa9c3c348cb1af180"
+    },
+    "results/public/choices/arc_easy.jsonl": {
+      "bytes": 536730,
+      "sha256": "92bead9bc0361159c46d047ef6ed04cf93f59a36e38bab0d6eb638f89d995388"
+    },
+    "results/public/choices/hellaswag.jsonl": {
+      "bytes": 2272459,
+      "sha256": "b7ed4f70e9be865b1288e9df0f2f10e5d79ac393c0b6887e672b134d72819067"
+    },
+    "results/public/choices/piqa.jsonl": {
+      "bytes": 328258,
+      "sha256": "a095ec405560fbf0649664058b3705a850475981eafe8c82efbba377b284b92e"
+    },
+    "results/public/choices/results.json": {
+      "bytes": 3075,
+      "sha256": "3321b5b1cc771c69c06ec01d7ea62d1aeb78764f455e2f514b23c73d44ab0938"
+    },
+    "results/public/ifeval/inputs.sha256.json": {
+      "bytes": 102,
+      "sha256": "501160470cb0f9f3f06ed853d1107cb819ab8f8295220cb2ccae0f065b267bdb"
+    },
+    "results/public/ifeval/loose.jsonl": {
+      "bytes": 722945,
+      "sha256": "40e0e8f2ded6b9f95893aff33697252752f574f1aac090e01136a3784026ee83"
+    },
+    "results/public/ifeval/responses.jsonl": {
+      "bytes": 700622,
+      "sha256": "ba051d49356da173c78f5079eb6450acbfe510ca0215625ed32050441c7a2aff"
+    },
+    "results/public/ifeval/results.json": {
+      "bytes": 2710,
+      "sha256": "859d5c8359f15179d021fc8c1676483efd7ea5320dd336105f24608077e24668"
+    },
+    "results/public/ifeval/strict.jsonl": {
+      "bytes": 722972,
+      "sha256": "fdc6f6e6962f76f5603670b6dff4f3ea7c28e6cd5a085f83ceddc3f09196065a"
+    },
+    "results/public/summary.json": {
+      "bytes": 4787,
+      "sha256": "b9a155a97a22e09027507737c5a4e9afe2a312fdce39c95df11a3022756e1b73"
+    },
+    "results/search/trial_01.json": {
+      "bytes": 9996,
+      "sha256": "8d07aeaecc61f0acbcf4d18d413472e8440561634b15df95c3a3444dc433f376"
+    },
+    "results/search/trial_02.json": {
+      "bytes": 10075,
+      "sha256": "4d4271ce322e29ed1fcdfb64e4631dfd358b0a43f907200a1f2fa41652dc0770"
+    },
+    "results/search/trial_03.json": {
+      "bytes": 10136,
+      "sha256": "6dc15f24abbf6b24528178be803d97c2885939835558f353ff1bf34427f524cd"
+    },
+    "results/search/trial_04.json": {
+      "bytes": 10077,
+      "sha256": "be38e4323b1b120fa92717e2439c3d231675e70d860a4673d296e202e62a2342"
+    },
+    "results/search/trial_05.json": {
+      "bytes": 10239,
+      "sha256": "da83222ee38d33eea6c7062c28fd310e61da89a9a0315dc62559c4a9181759d9"
+    },
+    "results/selection/batch_order_audit.json": {
+      "bytes": 1718,
+      "sha256": "8ba998638ee42e06e5fb6d5d26e78ec51be5ce5b8565889d5880457cca614c6e"
+    },
+    "results/selection/comparison.json": {
+      "bytes": 9415,
+      "sha256": "28fdc175a0019e6a34590256c3c1da8fc3a39813699afae8a477d63b00249a7b"
+    },
+    "results/selection/data_audit.json": {
+      "bytes": 2697,
+      "sha256": "6f4978a0a95ad88e089de1eda6026a4a1f61d3d03f398ac1234a045f4b150ec6"
+    },
+    "results/selection/design.json": {
+      "bytes": 2527,
+      "sha256": "a478bc790345129fa960060490d4c2148351ccfc17ac433ec06364969787a8fe"
+    },
+    "results/selection/selection_frozen.json": {
+      "bytes": 11194,
+      "sha256": "0567b916bdf27de7f6eafe8c61da7fb3686a69e08361ac6cb5645ac567cf8f69"
+    },
+    "scripts/materialize_training_snapshot.py": {
+      "bytes": 1823,
+      "sha256": "cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a"
+    },
+    "scripts/plot_training.py": {
+      "bytes": 2877,
+      "sha256": "c47f496b10f813dcb61697a0d49ed14043137ee1e2a96911a68fccb31fa9df26"
+    },
+    "scripts/verify_inference.py": {
+      "bytes": 4559,
+      "sha256": "002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3"
+    },
+    "scripts/verify_package.py": {
+      "bytes": 1763,
+      "sha256": "ea7d6388ad1f17fcf244119b2bf15018d981c4e2c127bedf2af73dc74eaa0528"
+    },
+    "tokenizer/corpus.json": {
+      "bytes": 1549,
+      "sha256": "e4c47b7b6aba08f948de7c2c7cb94c58dbd09dde2ded6ac185c60856ee686d27"
+    },
+    "tokenizer/manifest.json": {
+      "bytes": 988,
+      "sha256": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "tokenizer/report.json": {
+      "bytes": 889,
+      "sha256": "e871f690eabebf82da173f0e406f38b79a10cd7d2647bb3592c3fc35ae9a2f16"
+    },
+    "tokenizer/tokenizer.json": {
+      "bytes": 2288703,
+      "sha256": "e095f0413a11fa91fbb58f37d00e4c6bcda59beb89b28f6f680db1436a7e8891"
+    },
+    "tokenizer/tokenizer_config.json": {
+      "bytes": 200,
+      "sha256": "619f500553367eb76f5be00a98625130ec857b9c367561334984096c029ba2a9"
+    },
+    "training/SNAPSHOTS.json": {
+      "bytes": 1363,
+      "sha256": "1d16aa9630669fc9b5d03eea415c7c2a9b7423d048b3ba0ebe75d62193cb3066"
+    },
+    "training/launchers/SOURCE.json": {
+      "bytes": 2977,
+      "sha256": "71f9d16d8a05d4e944ef0863ec52c673e1342cb5378ddfec57a3f7cc7962772c"
+    },
+    "training/launchers/five_mac_cluster.py": {
+      "bytes": 8856,
+      "sha256": "27544f082d142bb4ee436f3a527667bd3c68559d1f70c67114a2ac16dccca2bc"
+    },
+    "training/launchers/launch_adapter_trial_09_replica_20261108.py": {
+      "bytes": 13325,
+      "sha256": "f289117941679701cb4e29dc79c3dc859f5739df416fcbc2447441d607d00695"
+    },
+    "training/launchers/launch_code_pretraining.py": {
+      "bytes": 9935,
+      "sha256": "5f74006eae3c6b34cd1901b634dda1f6b8712ab740ef118bd203214767124e76"
+    },
+    "training/launchers/launch_deep44_cooldown.py": {
+      "bytes": 10893,
+      "sha256": "7887737805af333126e2755fc04b6e16b1ced7e1e7c8091e358867d3d6176a21"
+    },
+    "training/launchers/launch_research.py": {
+      "bytes": 13948,
+      "sha256": "b70ea382d073a0a2ef37c1acb4b5925ea1fe4ef509e0cfdb31282ab054bf5247"
+    },
+    "training/launchers/launch_stable_pretraining.py": {
+      "bytes": 10482,
+      "sha256": "770dc6f9bb9c60114838bfda93582fb917cd93754a3af33757580f75395031bf"
+    },
+    "training/launchers/preflight_chat_repair_search.py": {
+      "bytes": 7243,
+      "sha256": "58f99d20ee7d37bde6e0ee260f6a31c39ac76296771fa3afcf466a311dfd2327"
+    },
+    "training/launchers/reproduce_parent2000_attempt5.py": {
+      "bytes": 9659,
+      "sha256": "6ba5201968f384b1823e878ff0b3b173e0977916508891ee33e6065dfd2bc98b"
+    },
+    "training/launchers/run_chat_repair_search.py": {
+      "bytes": 16970,
+      "sha256": "4e4c405eb89b9de3e3d1eecec682ab225bf4288768537ba766bad1d5e4ddfbbf"
+    },
+    "training/snapshots/pretrain_20261022_121457_b3d5f398.json": {
+      "bytes": 13450,
+      "sha256": "94ffce3491a40e8f4c70adfad0a564c7cfa0cf77a4eba6348378a2adf7b9512a"
+    },
+    "training/snapshots/pretrain_20261024_111140_5f596b7a.json": {
+      "bytes": 13450,
+      "sha256": "d5dd508491a9dd6151c972353ecfe898aa5fafbe9056f152884daa41995b04ee"
+    },
+    "training/snapshots/pretrain_20261027_222615_4d065da7.json": {
+      "bytes": 13450,
+      "sha256": "d2c70ebe2b052eafab538244460084263f13f3b4d455eb8cd76a2e92836e47f7"
+    },
+    "training/snapshots/pretrain_20261028_093017_dd42e713.json": {
+      "bytes": 13450,
+      "sha256": "cd7a2ad3988fd6ba53c0f291f39239625ab390eff0a890a6915ced6bfb145493"
+    },
+    "training/snapshots/pretrain_20261102_234538_0ee851dc.json": {
+      "bytes": 13450,
+      "sha256": "70f212b330a9e25d245629ebc9d1909537fae8c8639bfb65dda17a0f17e4e39c"
+    },
+    "training/snapshots/pretrain_20261103_000432_8ac2605e.json": {
+      "bytes": 13450,
+      "sha256": "d47efd77e209ad2ca131064c5e943519d63564bee55a2b866f8c6a8ea74676cb"
+    },
+    "training/snapshots/pretrain_20261104_170916_448e6f64.json": {
+      "bytes": 84640,
+      "sha256": "45adf703c6004b8269a652775d04328af7811783a5211e9d151bdc2885b0d6c5"
+    },
+    "training/snapshots/pretrain_20261107_142423_40700422.json": {
+      "bytes": 84640,
+      "sha256": "227ca4a97a3581ce39781551ba30558178c1d7b0358a54b923d23dd5d6f9aee3"
+    },
+    "training/snapshots/sft1.json": {
+      "bytes": 104751,
+      "sha256": "95b573e73247cb8628179dc010c34f73527500e702c7b5e042fe883a7be702cd"
+    },
+    "training/snapshots/sft2.json": {
+      "bytes": 107618,
+      "sha256": "ffda31c2dc6d0afe2112ffaf8dd3d28f6dab40119eb4360203103c00658e85f1"
+    },
+    "training/snapshots/sft3.json": {
+      "bytes": 109772,
+      "sha256": "dfbd41d4b60ae84bfc5fe32a0051a244e3489b7c2f24cea333d362aba98f79cc"
+    },
+    "training/source_blobs/0132642f0e4338c1f38a72a74407430afbd0d269c02e458f83df2366baeb85d6.py": {
+      "bytes": 8600,
+      "sha256": "0132642f0e4338c1f38a72a74407430afbd0d269c02e458f83df2366baeb85d6"
+    },
+    "training/source_blobs/01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b.py": {
+      "bytes": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    "training/source_blobs/01e7b0b2f162a9cdc242aecf6dbb7c22806365308973b4faf2c90d86a796b9e9.py": {
+      "bytes": 41354,
+      "sha256": "01e7b0b2f162a9cdc242aecf6dbb7c22806365308973b4faf2c90d86a796b9e9"
+    },
+    "training/source_blobs/023fda52d696e3b8349afb663e085a467f10f3d3ba869e4a85ac2c4b9efe8c16.py": {
+      "bytes": 2725,
+      "sha256": "023fda52d696e3b8349afb663e085a467f10f3d3ba869e4a85ac2c4b9efe8c16"
+    },
+    "training/source_blobs/0348ad7d8aeb3e6749be7ce4df4a5b3029d9db869de9a334f2c5c6058bedb711.py": {
+      "bytes": 10762,
+      "sha256": "0348ad7d8aeb3e6749be7ce4df4a5b3029d9db869de9a334f2c5c6058bedb711"
+    },
+    "training/source_blobs/03bceb2e4c8a3ca7c63ea11d31b5e89dece07cbb326b23e53ad8d6972e1214fd.py": {
+      "bytes": 12719,
+      "sha256": "03bceb2e4c8a3ca7c63ea11d31b5e89dece07cbb326b23e53ad8d6972e1214fd"
+    },
+    "training/source_blobs/04945e1ff569c5cd580028d77db3d8f35d6fce9817f6aa52e13768c26e978d62.py": {
+      "bytes": 6657,
+      "sha256": "04945e1ff569c5cd580028d77db3d8f35d6fce9817f6aa52e13768c26e978d62"
+    },
+    "training/source_blobs/04c8c482003ed8ffa1f2238337f351ec2748e2ef5ae60b608e6d6acdeda5f516.py": {
+      "bytes": 6188,
+      "sha256": "04c8c482003ed8ffa1f2238337f351ec2748e2ef5ae60b608e6d6acdeda5f516"
+    },
+    "training/source_blobs/0548da651db528075ca56db9570e19812f7824da024d7f6350e585d1058d166a.py": {
+      "bytes": 4397,
+      "sha256": "0548da651db528075ca56db9570e19812f7824da024d7f6350e585d1058d166a"
+    },
+    "training/source_blobs/064b0370ae2b6c0f6a6fa7c684ded54bd97b4e39958aebb0c2aff7dac1f9b9e4.py": {
+      "bytes": 6837,
+      "sha256": "064b0370ae2b6c0f6a6fa7c684ded54bd97b4e39958aebb0c2aff7dac1f9b9e4"
+    },
+    "training/source_blobs/06c96307cb6749a4c9f169de9a4e049fe07ec8f753aff30344a0aa00921e5530.py": {
+      "bytes": 2052,
+      "sha256": "06c96307cb6749a4c9f169de9a4e049fe07ec8f753aff30344a0aa00921e5530"
+    },
+    "training/source_blobs/071b6416f3f0785dde7c5a06e90afe405fc0885a39853c2f2027d670f060556e.py": {
+      "bytes": 1077,
+      "sha256": "071b6416f3f0785dde7c5a06e90afe405fc0885a39853c2f2027d670f060556e"
+    },
+    "training/source_blobs/072851f20850e4548c0b3a48e0a2e3ceab961d7b956e76a8944ed409cd460898.py": {
+      "bytes": 489,
+      "sha256": "072851f20850e4548c0b3a48e0a2e3ceab961d7b956e76a8944ed409cd460898"
+    },
+    "training/source_blobs/072fed811b8a3d13878ab1283991246877a12e7755fbca5a0c88bf2ed95bd382.py": {
+      "bytes": 30723,
+      "sha256": "072fed811b8a3d13878ab1283991246877a12e7755fbca5a0c88bf2ed95bd382"
+    },
+    "training/source_blobs/0769ac3b1bec8cacf5a24c63b3049271d0907e20bba94a2475bbcfdeac329ae9.py": {
+      "bytes": 3795,
+      "sha256": "0769ac3b1bec8cacf5a24c63b3049271d0907e20bba94a2475bbcfdeac329ae9"
+    },
+    "training/source_blobs/07751cc81cf60a44c571c0a5ac8194690a6805211eaab10236590849b7c3b90f.py": {
+      "bytes": 4510,
+      "sha256": "07751cc81cf60a44c571c0a5ac8194690a6805211eaab10236590849b7c3b90f"
+    },
+    "training/source_blobs/077ab6db2082770a4683b37912c0b43e34330101fec97aab1a1cac4bb4d7b2ed.py": {
+      "bytes": 432,
+      "sha256": "077ab6db2082770a4683b37912c0b43e34330101fec97aab1a1cac4bb4d7b2ed"
+    },
+    "training/source_blobs/0aaa7019ab53fd8b5e6b0e397f970026b4e8d38e12bf5d5f0fcb7f686de9b82c.py": {
+      "bytes": 2004,
+      "sha256": "0aaa7019ab53fd8b5e6b0e397f970026b4e8d38e12bf5d5f0fcb7f686de9b82c"
+    },
+    "training/source_blobs/0bb4999f8a50190afebd71df014e9ea0e1db8bea44144d93f47c34b63f646c60.py": {
+      "bytes": 48423,
+      "sha256": "0bb4999f8a50190afebd71df014e9ea0e1db8bea44144d93f47c34b63f646c60"
+    },
+    "training/source_blobs/0bd916e6d5eafb4be3d5478a4a551cf1b41e9f6dbe11d86c3928dac1fe83acc5.py": {
+      "bytes": 46520,
+      "sha256": "0bd916e6d5eafb4be3d5478a4a551cf1b41e9f6dbe11d86c3928dac1fe83acc5"
+    },
+    "training/source_blobs/0ca74fa61e1748e3d689826ef2c134d82c5d4cf7f47ca761f15963d586e8c942.py": {
+      "bytes": 3823,
+      "sha256": "0ca74fa61e1748e3d689826ef2c134d82c5d4cf7f47ca761f15963d586e8c942"
+    },
+    "training/source_blobs/0d0a8d68c9278f292c5e8c59ce8f3e25ea585ae18d07dfe7e6c7e2fac6449f46.py": {
+      "bytes": 57961,
+      "sha256": "0d0a8d68c9278f292c5e8c59ce8f3e25ea585ae18d07dfe7e6c7e2fac6449f46"
+    },
+    "training/source_blobs/0eafba0490c4db294eb44057a200a7cffaed51146b102845d03095973544bf5c.py": {
+      "bytes": 1280,
+      "sha256": "0eafba0490c4db294eb44057a200a7cffaed51146b102845d03095973544bf5c"
+    },
+    "training/source_blobs/0f1d2108827df13fc4420c9ab05f2fc22ce33c85d2fc54008e955c79fe7ddbc4.py": {
+      "bytes": 31483,
+      "sha256": "0f1d2108827df13fc4420c9ab05f2fc22ce33c85d2fc54008e955c79fe7ddbc4"
+    },
+    "training/source_blobs/0f452657e4e5a32f6df03d696874c3a959b9e6935ccc6e0ff86c329207fda7db.py": {
+      "bytes": 38731,
+      "sha256": "0f452657e4e5a32f6df03d696874c3a959b9e6935ccc6e0ff86c329207fda7db"
+    },
+    "training/source_blobs/1012b31fa62d5b56c959b69cb2d63dd9996e7518541842bdb119699ac00ed2c4.py": {
+      "bytes": 8826,
+      "sha256": "1012b31fa62d5b56c959b69cb2d63dd9996e7518541842bdb119699ac00ed2c4"
+    },
+    "training/source_blobs/10b5f6451f238387ba7b0342b380453808236f63824ff0ead4e22df3083662c5.py": {
+      "bytes": 12933,
+      "sha256": "10b5f6451f238387ba7b0342b380453808236f63824ff0ead4e22df3083662c5"
+    },
+    "training/source_blobs/10ca8b5430b7f97d53ea00e50ed9dab53e62208247d310105e3e7671b0198455.py": {
+      "bytes": 24477,
+      "sha256": "10ca8b5430b7f97d53ea00e50ed9dab53e62208247d310105e3e7671b0198455"
+    },
+    "training/source_blobs/112ccdbbd682d661902d606be4bfc9d6fa41ede284fe48855e704688ca6cc88d.py": {
+      "bytes": 12323,
+      "sha256": "112ccdbbd682d661902d606be4bfc9d6fa41ede284fe48855e704688ca6cc88d"
+    },
+    "training/source_blobs/116bf40b02262743bb5ddac629c8d223537c0c12ce4631d32347a459844bf4aa.py": {
+      "bytes": 5273,
+      "sha256": "116bf40b02262743bb5ddac629c8d223537c0c12ce4631d32347a459844bf4aa"
+    },
+    "training/source_blobs/1173c8c94dc11c6b8e446c26804a543b68e366986f01d249384082629dca369b.py": {
+      "bytes": 2205,
+      "sha256": "1173c8c94dc11c6b8e446c26804a543b68e366986f01d249384082629dca369b"
+    },
+    "training/source_blobs/11766ff21cd6deda9fc4f1ad4ba45007222ebf7534a1c5639d6ad26df4359c03.py": {
+      "bytes": 3225,
+      "sha256": "11766ff21cd6deda9fc4f1ad4ba45007222ebf7534a1c5639d6ad26df4359c03"
+    },
+    "training/source_blobs/14c0be0a65ff87803a56e1ec031308cdd99a96f6c2732fdf28055fab5fa0a4cc.py": {
+      "bytes": 19893,
+      "sha256": "14c0be0a65ff87803a56e1ec031308cdd99a96f6c2732fdf28055fab5fa0a4cc"
+    },
+    "training/source_blobs/1829527deaeb7c2ccd8c8a7b2a05f7a7bb26cf159576981d9a52be54a6995901.py": {
+      "bytes": 36222,
+      "sha256": "1829527deaeb7c2ccd8c8a7b2a05f7a7bb26cf159576981d9a52be54a6995901"
+    },
+    "training/source_blobs/197c0f0f54ac691a52a244425f96dc82d250a2373bb7612fccfc5188d2dd8cad.py": {
+      "bytes": 4500,
+      "sha256": "197c0f0f54ac691a52a244425f96dc82d250a2373bb7612fccfc5188d2dd8cad"
+    },
+    "training/source_blobs/1ac442f888499a13fab376be1be819b12688ae5f9eac90b67f8800a6454ed08c.py": {
+      "bytes": 31929,
+      "sha256": "1ac442f888499a13fab376be1be819b12688ae5f9eac90b67f8800a6454ed08c"
+    },
+    "training/source_blobs/1c032c459b232a7b426a05fca18bb6007fe7d8d56b73b291630f570b5575b23e.py": {
+      "bytes": 10280,
+      "sha256": "1c032c459b232a7b426a05fca18bb6007fe7d8d56b73b291630f570b5575b23e"
+    },
+    "training/source_blobs/1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe.py": {
+      "bytes": 10552,
+      "sha256": "1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe"
+    },
+    "training/source_blobs/1feb34473bfa8cf3e4b47eddbd767cccf2b3f1775e04f194d2b4556abe730e84.py": {
+      "bytes": 14653,
+      "sha256": "1feb34473bfa8cf3e4b47eddbd767cccf2b3f1775e04f194d2b4556abe730e84"
+    },
+    "training/source_blobs/210e8f84fce6a32e60cc535296ea4a05d98456a47af977898f02017561a73eec.py": {
+      "bytes": 2753,
+      "sha256": "210e8f84fce6a32e60cc535296ea4a05d98456a47af977898f02017561a73eec"
+    },
+    "training/source_blobs/24fb7b1b6ab2adb5575308c7d386687389627a0018b4eb730254dc4f9e708219.py": {
+      "bytes": 2148,
+      "sha256": "24fb7b1b6ab2adb5575308c7d386687389627a0018b4eb730254dc4f9e708219"
+    },
+    "training/source_blobs/254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5.py": {
+      "bytes": 2797,
+      "sha256": "254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5"
+    },
+    "training/source_blobs/25536570e721a16c61795744882e7868b2aad63f35f9b88c8ea4ba845d92cae1.py": {
+      "bytes": 11658,
+      "sha256": "25536570e721a16c61795744882e7868b2aad63f35f9b88c8ea4ba845d92cae1"
+    },
+    "training/source_blobs/26d3ff514c47d34363982d6a2b6e02c0c28d53e0df7bc589281149effae859ea.py": {
+      "bytes": 16960,
+      "sha256": "26d3ff514c47d34363982d6a2b6e02c0c28d53e0df7bc589281149effae859ea"
+    },
+    "training/source_blobs/27392bfdaa565c8150f217ac2d8835135a56d26da3056427545f6a2efd898f4d.py": {
+      "bytes": 4085,
+      "sha256": "27392bfdaa565c8150f217ac2d8835135a56d26da3056427545f6a2efd898f4d"
+    },
+    "training/source_blobs/274c2760db9514c13e91ea3c78df9ff3717054dd0997b35260238cb83dbefbe8.py": {
+      "bytes": 11269,
+      "sha256": "274c2760db9514c13e91ea3c78df9ff3717054dd0997b35260238cb83dbefbe8"
+    },
+    "training/source_blobs/28968ef3e5de86c1b523ec5c3cfffdbcfe3c0f523a8ca18bd463dfdba3e709ee.py": {
+      "bytes": 15442,
+      "sha256": "28968ef3e5de86c1b523ec5c3cfffdbcfe3c0f523a8ca18bd463dfdba3e709ee"
+    },
+    "training/source_blobs/292a9338f1169410e4450138740beea1588be4f1883bd164b8cd83a56d5fc5c0.py": {
+      "bytes": 7790,
+      "sha256": "292a9338f1169410e4450138740beea1588be4f1883bd164b8cd83a56d5fc5c0"
+    },
+    "training/source_blobs/29310a52ea67aac363214dd8f3d3f26663d1371fa0d10172861df51ada623bd1.py": {
+      "bytes": 57435,
+      "sha256": "29310a52ea67aac363214dd8f3d3f26663d1371fa0d10172861df51ada623bd1"
+    },
+    "training/source_blobs/2abf5926a22a33b4f29c699855b24f7e459b6e631aa6ef780876ddc205f17758.py": {
+      "bytes": 9205,
+      "sha256": "2abf5926a22a33b4f29c699855b24f7e459b6e631aa6ef780876ddc205f17758"
+    },
+    "training/source_blobs/2ac80ce082cb5fca838b98c027f352f21dc666a9270005bec79d9c055ca3b99b.py": {
+      "bytes": 3971,
+      "sha256": "2ac80ce082cb5fca838b98c027f352f21dc666a9270005bec79d9c055ca3b99b"
+    },
+    "training/source_blobs/2b0a53ffa8f1100a9c24ec5991ffbe2c5ed79e884de9486376fd4dd8eac6da93.py": {
+      "bytes": 5984,
+      "sha256": "2b0a53ffa8f1100a9c24ec5991ffbe2c5ed79e884de9486376fd4dd8eac6da93"
+    },
+    "training/source_blobs/2b6aa65bd43786a1b815bc21e73229b4304d0f4275aed56ccdfbce4c6ea08f68.py": {
+      "bytes": 15620,
+      "sha256": "2b6aa65bd43786a1b815bc21e73229b4304d0f4275aed56ccdfbce4c6ea08f68"
+    },
+    "training/source_blobs/2beddc756f979309b46b115101191fa17a3c2c8e12579b672154a45a513b4e6d.py": {
+      "bytes": 2154,
+      "sha256": "2beddc756f979309b46b115101191fa17a3c2c8e12579b672154a45a513b4e6d"
+    },
+    "training/source_blobs/2cca8d901c9fe4f49016a01b1e89ba2b519fe37b42b265bcf9333db64d8ba2d9.py": {
+      "bytes": 13500,
+      "sha256": "2cca8d901c9fe4f49016a01b1e89ba2b519fe37b42b265bcf9333db64d8ba2d9"
+    },
+    "training/source_blobs/2cd4fe99866d6c85fa1a2481a2058009e96824c8622e915caca56fbbf544e829.py": {
+      "bytes": 17595,
+      "sha256": "2cd4fe99866d6c85fa1a2481a2058009e96824c8622e915caca56fbbf544e829"
+    },
+    "training/source_blobs/2d901af050a73117f6f2c76f16060377912b765b0f2f4bafe24f3092d01f19c4.py": {
+      "bytes": 1310,
+      "sha256": "2d901af050a73117f6f2c76f16060377912b765b0f2f4bafe24f3092d01f19c4"
+    },
+    "training/source_blobs/2e41c214043ae68cc2fc533ecf1e7c3358e8390c6edc9ff8f20a962acbf0d5b2.py": {
+      "bytes": 1634,
+      "sha256": "2e41c214043ae68cc2fc533ecf1e7c3358e8390c6edc9ff8f20a962acbf0d5b2"
+    },
+    "training/source_blobs/2ef091c51f28ea44bce60988693903abae0dd973aee6892e26f69126b9d68db0.py": {
+      "bytes": 58,
+      "sha256": "2ef091c51f28ea44bce60988693903abae0dd973aee6892e26f69126b9d68db0"
+    },
+    "training/source_blobs/2f643abaa6ac3f86e47b37632df4d25fa34d0b25e77b59ffecc0290b2385e7dc.py": {
+      "bytes": 5269,
+      "sha256": "2f643abaa6ac3f86e47b37632df4d25fa34d0b25e77b59ffecc0290b2385e7dc"
+    },
+    "training/source_blobs/2fbd333bb25f0dd02907141f0a16d59ad49d4db08cc5453ea8655f61cadb65f1.py": {
+      "bytes": 9622,
+      "sha256": "2fbd333bb25f0dd02907141f0a16d59ad49d4db08cc5453ea8655f61cadb65f1"
+    },
+    "training/source_blobs/30b8ecb4b97d2520addb522f77c2e955d01ebc62fb9a2ef9a6d8a3f41dc3d0b3.py": {
+      "bytes": 2555,
+      "sha256": "30b8ecb4b97d2520addb522f77c2e955d01ebc62fb9a2ef9a6d8a3f41dc3d0b3"
+    },
+    "training/source_blobs/317024d64eca0650b791f08ef92001d05037c6dd81b0f31ad8eac15c509bd8bb.py": {
+      "bytes": 980,
+      "sha256": "317024d64eca0650b791f08ef92001d05037c6dd81b0f31ad8eac15c509bd8bb"
+    },
+    "training/source_blobs/317c05d4d17239e62d1316797892288f5e4eae467b3e66a4ca4f15d189960a9b.py": {
+      "bytes": 4938,
+      "sha256": "317c05d4d17239e62d1316797892288f5e4eae467b3e66a4ca4f15d189960a9b"
+    },
+    "training/source_blobs/321c85734de3ce593cd7401b0082722d155ab71344196a73216c02d04ff0ba6a.py": {
+      "bytes": 92003,
+      "sha256": "321c85734de3ce593cd7401b0082722d155ab71344196a73216c02d04ff0ba6a"
+    },
+    "training/source_blobs/32527d75848cdc0e90f1ca93cd4ea54aeb2368374a3d0edd8653296b359d09f7.py": {
+      "bytes": 30302,
+      "sha256": "32527d75848cdc0e90f1ca93cd4ea54aeb2368374a3d0edd8653296b359d09f7"
+    },
+    "training/source_blobs/345331fc34634f7ea7d9ebfb1cf8c91a17123033c52c7d0bbfbe2aa3ed6ab49d.py": {
+      "bytes": 18206,
+      "sha256": "345331fc34634f7ea7d9ebfb1cf8c91a17123033c52c7d0bbfbe2aa3ed6ab49d"
+    },
+    "training/source_blobs/35c8ad532a20f77fc92edd49aaf877db0b2a4750c7212687e21b5d475b73534e.py": {
+      "bytes": 44642,
+      "sha256": "35c8ad532a20f77fc92edd49aaf877db0b2a4750c7212687e21b5d475b73534e"
+    },
+    "training/source_blobs/392f3cbeb2c4ad10a8b7173ed4331f63368a5cedda6fa84d23eb0d8dbd434657.py": {
+      "bytes": 29393,
+      "sha256": "392f3cbeb2c4ad10a8b7173ed4331f63368a5cedda6fa84d23eb0d8dbd434657"
+    },
+    "training/source_blobs/3972d7d2ef83ddbbd883b08f30204c870c4f91fdca47326b4a9e969bb0283f20.py": {
+      "bytes": 28486,
+      "sha256": "3972d7d2ef83ddbbd883b08f30204c870c4f91fdca47326b4a9e969bb0283f20"
+    },
+    "training/source_blobs/3ad40e26678dbe103cbf895a9f885bf9a06734807afa341336f48725178dc0ad.py": {
+      "bytes": 3164,
+      "sha256": "3ad40e26678dbe103cbf895a9f885bf9a06734807afa341336f48725178dc0ad"
+    },
+    "training/source_blobs/3b8f68cfd6fce0828c205723e98a31902455a71c27413d8a7bd18b1aec42d11b.py": {
+      "bytes": 2801,
+      "sha256": "3b8f68cfd6fce0828c205723e98a31902455a71c27413d8a7bd18b1aec42d11b"
+    },
+    "training/source_blobs/3d78ed45dbab52a671d889fe937a210a832e55426bbdc4f905ddbd202ccf2336.py": {
+      "bytes": 6637,
+      "sha256": "3d78ed45dbab52a671d889fe937a210a832e55426bbdc4f905ddbd202ccf2336"
+    },
+    "training/source_blobs/3e1c35e75a1f27dff42c6e90246972d811df32c6a3666dfe5f2a58e8daae5ea4.py": {
+      "bytes": 12404,
+      "sha256": "3e1c35e75a1f27dff42c6e90246972d811df32c6a3666dfe5f2a58e8daae5ea4"
+    },
+    "training/source_blobs/3e82b28d76197292a35d4d479b77fe14e5482bff0f7fd04d2a34e8d7cec4c9c4.py": {
+      "bytes": 1865,
+      "sha256": "3e82b28d76197292a35d4d479b77fe14e5482bff0f7fd04d2a34e8d7cec4c9c4"
+    },
+    "training/source_blobs/3eed2836d8c4111cf409815d9e749806c0f54fb8164e704bbcc9297b19e829ae.py": {
+      "bytes": 507,
+      "sha256": "3eed2836d8c4111cf409815d9e749806c0f54fb8164e704bbcc9297b19e829ae"
+    },
+    "training/source_blobs/3f6164c3f80f5fc8794b8c1776a6ceb7bc80af9cd36e03f6dcf3c0442ca662b5.py": {
+      "bytes": 7980,
+      "sha256": "3f6164c3f80f5fc8794b8c1776a6ceb7bc80af9cd36e03f6dcf3c0442ca662b5"
+    },
+    "training/source_blobs/40dbfada2856c91b0a7ffe5df7b9584175dbd479a1c6b581d1935d38dd6228aa.py": {
+      "bytes": 9289,
+      "sha256": "40dbfada2856c91b0a7ffe5df7b9584175dbd479a1c6b581d1935d38dd6228aa"
+    },
+    "training/source_blobs/41618bc3978700ba85bdae4a179ac27ac4faafee2f81c67f6887a72722e91f99.py": {
+      "bytes": 7592,
+      "sha256": "41618bc3978700ba85bdae4a179ac27ac4faafee2f81c67f6887a72722e91f99"
+    },
+    "training/source_blobs/41c77ec50fa0cfae3de7ccffa0ba792a5eaf754f0342159adad1316c4d8c8fe0.py": {
+      "bytes": 7344,
+      "sha256": "41c77ec50fa0cfae3de7ccffa0ba792a5eaf754f0342159adad1316c4d8c8fe0"
+    },
+    "training/source_blobs/41e6a98d9b41dc844063fc28ce81a08d83b919306b9daf0716b76a0bc21365fc.py": {
+      "bytes": 16152,
+      "sha256": "41e6a98d9b41dc844063fc28ce81a08d83b919306b9daf0716b76a0bc21365fc"
+    },
+    "training/source_blobs/4255b93bbc18d120318d83fb26b2fda1fe16d19ca1423250c77188940cc4f008.py": {
+      "bytes": 1494,
+      "sha256": "4255b93bbc18d120318d83fb26b2fda1fe16d19ca1423250c77188940cc4f008"
+    },
+    "training/source_blobs/42d96626fffbee32722c4f685a4aa121ec32a046283d51a9ac325fcea68e7996.py": {
+      "bytes": 5698,
+      "sha256": "42d96626fffbee32722c4f685a4aa121ec32a046283d51a9ac325fcea68e7996"
+    },
+    "training/source_blobs/441089f82a4a8e5037cd8f3f02f24f11852b6722ba8d32fff619618c6efba211.py": {
+      "bytes": 5677,
+      "sha256": "441089f82a4a8e5037cd8f3f02f24f11852b6722ba8d32fff619618c6efba211"
+    },
+    "training/source_blobs/44dfd6367ab2cd890d9630c47e3a49fe36b7f56107657196946343033fe2bebf.py": {
+      "bytes": 6557,
+      "sha256": "44dfd6367ab2cd890d9630c47e3a49fe36b7f56107657196946343033fe2bebf"
+    },
+    "training/source_blobs/464357ccafaa6674fad3980792773e1c2369b3fcaef2bd41fe2407eb888d6a26.py": {
+      "bytes": 17682,
+      "sha256": "464357ccafaa6674fad3980792773e1c2369b3fcaef2bd41fe2407eb888d6a26"
+    },
+    "training/source_blobs/467c462e44925f2a1810cb54778ad496b194a8a8ba4e788521e9bbc09164c000.py": {
+      "bytes": 47586,
+      "sha256": "467c462e44925f2a1810cb54778ad496b194a8a8ba4e788521e9bbc09164c000"
+    },
+    "training/source_blobs/4883847e90e5db21bf4305ad9cab06d031b486d11483492cfc2691c559b503bc.py": {
+      "bytes": 6978,
+      "sha256": "4883847e90e5db21bf4305ad9cab06d031b486d11483492cfc2691c559b503bc"
+    },
+    "training/source_blobs/489371cd03eec40cc26b9ad872177d022f6844e84b982b77bc35b4692decc511.py": {
+      "bytes": 8097,
+      "sha256": "489371cd03eec40cc26b9ad872177d022f6844e84b982b77bc35b4692decc511"
+    },
+    "training/source_blobs/4a177d4a54761cd46cf044f6a6cf9c30831665a9e0aa2b84bd52f9e563f4ebfc.py": {
+      "bytes": 782,
+      "sha256": "4a177d4a54761cd46cf044f6a6cf9c30831665a9e0aa2b84bd52f9e563f4ebfc"
+    },
+    "training/source_blobs/4a7da603e70e60435b6bece218bc2b7a8ec06a54483a37e3088e51f92d0d0058.py": {
+      "bytes": 5023,
+      "sha256": "4a7da603e70e60435b6bece218bc2b7a8ec06a54483a37e3088e51f92d0d0058"
+    },
+    "training/source_blobs/4b3b6862e0bc364f6d27adfdfbd187742173bbc45163109f6af0be8a81564c28.py": {
+      "bytes": 4638,
+      "sha256": "4b3b6862e0bc364f6d27adfdfbd187742173bbc45163109f6af0be8a81564c28"
+    },
+    "training/source_blobs/4c0c1486aeeaa2c2640f3fce552568a71c6c32c7bf194fc40247eda71ae2ae16.py": {
+      "bytes": 13301,
+      "sha256": "4c0c1486aeeaa2c2640f3fce552568a71c6c32c7bf194fc40247eda71ae2ae16"
+    },
+    "training/source_blobs/4ca1d4825452fa5ff01024f07d3bb718d1bd1919695f4450101a33fd1bb6d312.py": {
+      "bytes": 1999,
+      "sha256": "4ca1d4825452fa5ff01024f07d3bb718d1bd1919695f4450101a33fd1bb6d312"
+    },
+    "training/source_blobs/50087489251785aaf70b00a66fcad27689f154fb08089ccf54f959e0abf2700b.py": {
+      "bytes": 701,
+      "sha256": "50087489251785aaf70b00a66fcad27689f154fb08089ccf54f959e0abf2700b"
+    },
+    "training/source_blobs/5041b972c2b7e3c80a5483329c5b52512e566391959f74c2aa4c4ea0fa1f83d7.py": {
+      "bytes": 124,
+      "sha256": "5041b972c2b7e3c80a5483329c5b52512e566391959f74c2aa4c4ea0fa1f83d7"
+    },
+    "training/source_blobs/50b208a8b67a2fc7f5626ab42ef597a0844f70a11a42e048faa35ffdeadaddf0.py": {
+      "bytes": 1327,
+      "sha256": "50b208a8b67a2fc7f5626ab42ef597a0844f70a11a42e048faa35ffdeadaddf0"
+    },
+    "training/source_blobs/5105b1b526a8f8d734634231243e054649b0cc0abbfd28e8783d95891e29fcb2.py": {
+      "bytes": 18237,
+      "sha256": "5105b1b526a8f8d734634231243e054649b0cc0abbfd28e8783d95891e29fcb2"
+    },
+    "training/source_blobs/51060d603ab35f695729e585233b9e2d3fe4c11e27145251eb0a6f15a4862fbf.py": {
+      "bytes": 8489,
+      "sha256": "51060d603ab35f695729e585233b9e2d3fe4c11e27145251eb0a6f15a4862fbf"
+    },
+    "training/source_blobs/51411b86701c4b0d1c38691094232c43ffefea05c38dc8629ca1fc70316efaf4.py": {
+      "bytes": 18115,
+      "sha256": "51411b86701c4b0d1c38691094232c43ffefea05c38dc8629ca1fc70316efaf4"
+    },
+    "training/source_blobs/5234bc6b824e4b07f63dda48e5353bbdb5f5abac4c98c299b642d1815ce85d94.py": {
+      "bytes": 18131,
+      "sha256": "5234bc6b824e4b07f63dda48e5353bbdb5f5abac4c98c299b642d1815ce85d94"
+    },
+    "training/source_blobs/53d795e591f3e67f26ecd1f9eac289b0e3a66c4d83183cf47859168bb93c3817.py": {
+      "bytes": 6216,
+      "sha256": "53d795e591f3e67f26ecd1f9eac289b0e3a66c4d83183cf47859168bb93c3817"
+    },
+    "training/source_blobs/548f1142be10f46b302886abae76790f5802c0f159412994eee66a7a5a3366a9.py": {
+      "bytes": 607,
+      "sha256": "548f1142be10f46b302886abae76790f5802c0f159412994eee66a7a5a3366a9"
+    },
+    "training/source_blobs/554977e377d8a247404479de2909a6d544ffe98c00f3f0cacfcea741dce0a3d2.py": {
+      "bytes": 16993,
+      "sha256": "554977e377d8a247404479de2909a6d544ffe98c00f3f0cacfcea741dce0a3d2"
+    },
+    "training/source_blobs/55b04ac7875ad148fecf1c7cdfb5ab48e171abe3754ed9b19809cb9b93cd9ae0.py": {
+      "bytes": 2203,
+      "sha256": "55b04ac7875ad148fecf1c7cdfb5ab48e171abe3754ed9b19809cb9b93cd9ae0"
+    },
+    "training/source_blobs/56543798ef7fcda70687accfb7f6e62a541453752c0fe10acbebb6e7c33563e3.py": {
+      "bytes": 11620,
+      "sha256": "56543798ef7fcda70687accfb7f6e62a541453752c0fe10acbebb6e7c33563e3"
+    },
+    "training/source_blobs/56946ebf45af107f1952d03428437f83905be51bd2ce65605f4b381e55d76c06.py": {
+      "bytes": 3607,
+      "sha256": "56946ebf45af107f1952d03428437f83905be51bd2ce65605f4b381e55d76c06"
+    },
+    "training/source_blobs/56b487defef5b61803857ef671f307ae3109ba96d6e29fce2c3004a888966952.py": {
+      "bytes": 2268,
+      "sha256": "56b487defef5b61803857ef671f307ae3109ba96d6e29fce2c3004a888966952"
+    },
+    "training/source_blobs/57555e09d1c1716f2026001accf8522a9d27fb0685bcbad2904152f980913880.py": {
+      "bytes": 18387,
+      "sha256": "57555e09d1c1716f2026001accf8522a9d27fb0685bcbad2904152f980913880"
+    },
+    "training/source_blobs/58e0bdc793f076f0fa6e7ec5b0cc4df51506a9e7505c09322d157fa080471dbc.py": {
+      "bytes": 1444,
+      "sha256": "58e0bdc793f076f0fa6e7ec5b0cc4df51506a9e7505c09322d157fa080471dbc"
+    },
+    "training/source_blobs/59d1adb31400e3c288914b4c0cc8a21681a73ca3dfa7c722ea4e13d21dc38528.py": {
+      "bytes": 2328,
+      "sha256": "59d1adb31400e3c288914b4c0cc8a21681a73ca3dfa7c722ea4e13d21dc38528"
+    },
+    "training/source_blobs/5a4bda6fc5e174f0d7dd9d602867ce0c746ec2609e243043a3657793b94401fa.py": {
+      "bytes": 5067,
+      "sha256": "5a4bda6fc5e174f0d7dd9d602867ce0c746ec2609e243043a3657793b94401fa"
+    },
+    "training/source_blobs/5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8.py": {
+      "bytes": 4129,
+      "sha256": "5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8"
+    },
+    "training/source_blobs/5cbce166031d4056b6e69c7b9b9efb45528e182d6ba996739c382bc5806bf477.py": {
+      "bytes": 24905,
+      "sha256": "5cbce166031d4056b6e69c7b9b9efb45528e182d6ba996739c382bc5806bf477"
+    },
+    "training/source_blobs/5cda62bae42610986c169ecd7d6cb92fdd7f0f7eb2255fa912a7eb54216a9cbb.py": {
+      "bytes": 10630,
+      "sha256": "5cda62bae42610986c169ecd7d6cb92fdd7f0f7eb2255fa912a7eb54216a9cbb"
+    },
+    "training/source_blobs/5dcaa086355bc0fd73c3ea1504bd596b48edba1dc2d757a148dcad11547cec02.py": {
+      "bytes": 4530,
+      "sha256": "5dcaa086355bc0fd73c3ea1504bd596b48edba1dc2d757a148dcad11547cec02"
+    },
+    "training/source_blobs/5f6474ff3b49542dd553d88d9b78512241cf7a25f8e37c03fe9bb9d90d1dd77f.py": {
+      "bytes": 5273,
+      "sha256": "5f6474ff3b49542dd553d88d9b78512241cf7a25f8e37c03fe9bb9d90d1dd77f"
+    },
+    "training/source_blobs/5fffbc7cb346228589779105aab9c5c0e26a71b6d6e3528a29777f68869b430c.py": {
+      "bytes": 6227,
+      "sha256": "5fffbc7cb346228589779105aab9c5c0e26a71b6d6e3528a29777f68869b430c"
+    },
+    "training/source_blobs/60c792369c4a3d65ff960ef05e5389d8c1571cf553eb7a1eb2673258c59b069e.py": {
+      "bytes": 4867,
+      "sha256": "60c792369c4a3d65ff960ef05e5389d8c1571cf553eb7a1eb2673258c59b069e"
+    },
+    "training/source_blobs/6155114ea8f946e116dec8afd080b91d37275d9565d62004a2303fdcff74cbc7.py": {
+      "bytes": 1678,
+      "sha256": "6155114ea8f946e116dec8afd080b91d37275d9565d62004a2303fdcff74cbc7"
+    },
+    "training/source_blobs/631451a45b020120610b7116e8e63276ba5043011359fbd13e49e63ec8e2b02b.py": {
+      "bytes": 8219,
+      "sha256": "631451a45b020120610b7116e8e63276ba5043011359fbd13e49e63ec8e2b02b"
+    },
+    "training/source_blobs/63dc102460925b5146e5d73dc32dbde53a668ca1b662b49782561910994b9de5.py": {
+      "bytes": 14439,
+      "sha256": "63dc102460925b5146e5d73dc32dbde53a668ca1b662b49782561910994b9de5"
+    },
+    "training/source_blobs/641140ca75f9a64d8afc015898ec0a8c294b11fe7fc7bfd4c06ab1726c59df04.py": {
+      "bytes": 3597,
+      "sha256": "641140ca75f9a64d8afc015898ec0a8c294b11fe7fc7bfd4c06ab1726c59df04"
+    },
+    "training/source_blobs/655888aa035d4939001ebf08ccde1a5a8a78d1ce5189da3ed11c60d2b1dedd97.py": {
+      "bytes": 15584,
+      "sha256": "655888aa035d4939001ebf08ccde1a5a8a78d1ce5189da3ed11c60d2b1dedd97"
+    },
+    "training/source_blobs/669eccb3d19fe369bfa6e04ed47cc2b2a2cdb26f9407ae0de9a993431f1880eb.py": {
+      "bytes": 10748,
+      "sha256": "669eccb3d19fe369bfa6e04ed47cc2b2a2cdb26f9407ae0de9a993431f1880eb"
+    },
+    "training/source_blobs/695895ece361823ce772e651db19d1a26fc372cc8913fb9dc6fcc9231342c6c0.py": {
+      "bytes": 25568,
+      "sha256": "695895ece361823ce772e651db19d1a26fc372cc8913fb9dc6fcc9231342c6c0"
+    },
+    "training/source_blobs/69babc5ed58d0b859c6cd2419ce9c0cdb37823d0b6b174cc6de82bd22816eb35.py": {
+      "bytes": 44249,
+      "sha256": "69babc5ed58d0b859c6cd2419ce9c0cdb37823d0b6b174cc6de82bd22816eb35"
+    },
+    "training/source_blobs/6a7a6c7ee6dee3d4ee81d5f2a627b57c3b4444808a2c9c1f7b280789a9260e23.py": {
+      "bytes": 22389,
+      "sha256": "6a7a6c7ee6dee3d4ee81d5f2a627b57c3b4444808a2c9c1f7b280789a9260e23"
+    },
+    "training/source_blobs/6bf0cf7f1fb395aa90227b1bca52214f3151628b930087b58bc90101828dc467.py": {
+      "bytes": 3428,
+      "sha256": "6bf0cf7f1fb395aa90227b1bca52214f3151628b930087b58bc90101828dc467"
+    },
+    "training/source_blobs/6dac260e47b46258d9d582362055b172ac24bdf1ecc57acae7e2769a0097a129.py": {
+      "bytes": 14151,
+      "sha256": "6dac260e47b46258d9d582362055b172ac24bdf1ecc57acae7e2769a0097a129"
+    },
+    "training/source_blobs/6e854aa53c69baa40955091e7f5318ddfea3aa8569a606fa342f9fbe07c1006b.py": {
+      "bytes": 9893,
+      "sha256": "6e854aa53c69baa40955091e7f5318ddfea3aa8569a606fa342f9fbe07c1006b"
+    },
+    "training/source_blobs/6ebf9b6ed5a5b3d78e179c609950fc0f7788b6af12dde11cd55ec79c9b3e28d7.py": {
+      "bytes": 892,
+      "sha256": "6ebf9b6ed5a5b3d78e179c609950fc0f7788b6af12dde11cd55ec79c9b3e28d7"
+    },
+    "training/source_blobs/70e36e12d33547bfce505bd0302605e3d66c9ee65fc0facb26ccf5341bde8b4b.py": {
+      "bytes": 15293,
+      "sha256": "70e36e12d33547bfce505bd0302605e3d66c9ee65fc0facb26ccf5341bde8b4b"
+    },
+    "training/source_blobs/71ec4fc999f1e87c24e0b2fe2e60e6a59d716a13f76b29ca88a3856c05c246da.py": {
+      "bytes": 1205,
+      "sha256": "71ec4fc999f1e87c24e0b2fe2e60e6a59d716a13f76b29ca88a3856c05c246da"
+    },
+    "training/source_blobs/72cc9ed68cee16a17a5271bcee77870da6fafb3949a7d052bc39abe6e7daccc2.py": {
+      "bytes": 8233,
+      "sha256": "72cc9ed68cee16a17a5271bcee77870da6fafb3949a7d052bc39abe6e7daccc2"
+    },
+    "training/source_blobs/72e7ca777f0e9078d955e13ff958ef2cc43374f924d72d69ad0ee770e9ae8271.py": {
+      "bytes": 7294,
+      "sha256": "72e7ca777f0e9078d955e13ff958ef2cc43374f924d72d69ad0ee770e9ae8271"
+    },
+    "training/source_blobs/73cbf0c17cf74a44bb7c74a8f9e5bb63671057a1f664a6a4af0e0777b2c8f1a4.py": {
+      "bytes": 5469,
+      "sha256": "73cbf0c17cf74a44bb7c74a8f9e5bb63671057a1f664a6a4af0e0777b2c8f1a4"
+    },
+    "training/source_blobs/7493fcc8857dcc01e5e9d5e72a3d47639411b810719a64a263486d5faddd5a04.py": {
+      "bytes": 8071,
+      "sha256": "7493fcc8857dcc01e5e9d5e72a3d47639411b810719a64a263486d5faddd5a04"
+    },
+    "training/source_blobs/74d4e12f420a4e105933a0bcb87626b7d4fb9c68d2a3b2164a6bfefc0eb6a1e0.py": {
+      "bytes": 37237,
+      "sha256": "74d4e12f420a4e105933a0bcb87626b7d4fb9c68d2a3b2164a6bfefc0eb6a1e0"
+    },
+    "training/source_blobs/75f45d5bb98fc6ac62ab00506503e31b7534cad1e7ea8c0fc69963251c8f0c4d.py": {
+      "bytes": 14051,
+      "sha256": "75f45d5bb98fc6ac62ab00506503e31b7534cad1e7ea8c0fc69963251c8f0c4d"
+    },
+    "training/source_blobs/76a82db3ee8da8956413fe559674d94415b9ea21b95d0a1442a8da732f908e9c.py": {
+      "bytes": 3006,
+      "sha256": "76a82db3ee8da8956413fe559674d94415b9ea21b95d0a1442a8da732f908e9c"
+    },
+    "training/source_blobs/78290186ea157bb44399268dd87f4df32feba21d79229cbe988a35de576e800f.py": {
+      "bytes": 1368,
+      "sha256": "78290186ea157bb44399268dd87f4df32feba21d79229cbe988a35de576e800f"
+    },
+    "training/source_blobs/7b761ff4a4023a0f84f9c0efa0e0c92cc7640df30b0608842d723c95c582f1cf.py": {
+      "bytes": 2599,
+      "sha256": "7b761ff4a4023a0f84f9c0efa0e0c92cc7640df30b0608842d723c95c582f1cf"
+    },
+    "training/source_blobs/7c176d4ca3c55cadb8deffc2e6d640cb5d81dee4884a55c6ef15adecb1f3f6b4.py": {
+      "bytes": 17385,
+      "sha256": "7c176d4ca3c55cadb8deffc2e6d640cb5d81dee4884a55c6ef15adecb1f3f6b4"
+    },
+    "training/source_blobs/7c6f6d0c910f4cea15d29d20c2a98add617c08cae5cf545c3d43be220cf44881.py": {
+      "bytes": 2686,
+      "sha256": "7c6f6d0c910f4cea15d29d20c2a98add617c08cae5cf545c3d43be220cf44881"
+    },
+    "training/source_blobs/7cb885d4bb5d9fa5a796304a9bae8ce995ffdf8d450c6fc68c17f93f0986350d.py": {
+      "bytes": 2411,
+      "sha256": "7cb885d4bb5d9fa5a796304a9bae8ce995ffdf8d450c6fc68c17f93f0986350d"
+    },
+    "training/source_blobs/7cdc7d042c65a996bc8867c2e370503b533a1216f9656619ec6fcffd4774f6fb.py": {
+      "bytes": 12499,
+      "sha256": "7cdc7d042c65a996bc8867c2e370503b533a1216f9656619ec6fcffd4774f6fb"
+    },
+    "training/source_blobs/7e7b298f6716c0e3a013d25bfcf20a766fc42aa59821b0a639b60bb7e5c78b75.py": {
+      "bytes": 3802,
+      "sha256": "7e7b298f6716c0e3a013d25bfcf20a766fc42aa59821b0a639b60bb7e5c78b75"
+    },
+    "training/source_blobs/7e9e02294e03e6b8a12fd202ab288f2f9b0073bb11b8cb2a5fa82211b0d1dfa4.py": {
+      "bytes": 67,
+      "sha256": "7e9e02294e03e6b8a12fd202ab288f2f9b0073bb11b8cb2a5fa82211b0d1dfa4"
+    },
+    "training/source_blobs/7ee92cc6f8765187de5cf72b105a8a7e0784cf59fc8cc9869dcb91bc2643aada.py": {
+      "bytes": 13045,
+      "sha256": "7ee92cc6f8765187de5cf72b105a8a7e0784cf59fc8cc9869dcb91bc2643aada"
+    },
+    "training/source_blobs/7f64ce45cd9e8201f8adce7ecb37be6ee29c80b83c344df699814401fc162780.py": {
+      "bytes": 21775,
+      "sha256": "7f64ce45cd9e8201f8adce7ecb37be6ee29c80b83c344df699814401fc162780"
+    },
+    "training/source_blobs/80d4e25a009727ccaf3885612fb3d2879192b923b248f18a00f1318b04a34253.py": {
+      "bytes": 12118,
+      "sha256": "80d4e25a009727ccaf3885612fb3d2879192b923b248f18a00f1318b04a34253"
+    },
+    "training/source_blobs/80de0fe0d43dea84ce734bc10256fc45a931828d117ef57a6a81b83ec40d2827.py": {
+      "bytes": 7066,
+      "sha256": "80de0fe0d43dea84ce734bc10256fc45a931828d117ef57a6a81b83ec40d2827"
+    },
+    "training/source_blobs/81c300349b75def6cc4d151fd39e9d4cec85adae87c02a6290f74f6fabe665b9.py": {
+      "bytes": 2757,
+      "sha256": "81c300349b75def6cc4d151fd39e9d4cec85adae87c02a6290f74f6fabe665b9"
+    },
+    "training/source_blobs/82d4b5333c98bd0483d22b196548670b26b712623fea00f815f25137f01f0891.py": {
+      "bytes": 2167,
+      "sha256": "82d4b5333c98bd0483d22b196548670b26b712623fea00f815f25137f01f0891"
+    },
+    "training/source_blobs/82e6fc9baae2c5fb23b24bde85880f8d6c2911ed07f0911ca1959434650c2256.py": {
+      "bytes": 1348,
+      "sha256": "82e6fc9baae2c5fb23b24bde85880f8d6c2911ed07f0911ca1959434650c2256"
+    },
+    "training/source_blobs/83b46cb227b9d5e6d8410b1df727c1df32975f04121e4efe4caf7f8552664800.py": {
+      "bytes": 21393,
+      "sha256": "83b46cb227b9d5e6d8410b1df727c1df32975f04121e4efe4caf7f8552664800"
+    },
+    "training/source_blobs/88d7ece9842b43fcab0f061a4de81e9dacf3010df3a17fb40c0ba576f8d11834.py": {
+      "bytes": 25598,
+      "sha256": "88d7ece9842b43fcab0f061a4de81e9dacf3010df3a17fb40c0ba576f8d11834"
+    },
+    "training/source_blobs/89006313590f7e77f6db14ca9ddb5c0e92e9952b387b8289dc8b9f662509979a.py": {
+      "bytes": 5696,
+      "sha256": "89006313590f7e77f6db14ca9ddb5c0e92e9952b387b8289dc8b9f662509979a"
+    },
+    "training/source_blobs/890cc0a233f730cc9ca4eab0c5fbc570f8a7a04f816ca4979940014dccf485b5.py": {
+      "bytes": 38755,
+      "sha256": "890cc0a233f730cc9ca4eab0c5fbc570f8a7a04f816ca4979940014dccf485b5"
+    },
+    "training/source_blobs/898c01ca9ab30c5f9e6ae205060f8457a7abd2d740c8b8d04c57a1030a268d17.py": {
+      "bytes": 65869,
+      "sha256": "898c01ca9ab30c5f9e6ae205060f8457a7abd2d740c8b8d04c57a1030a268d17"
+    },
+    "training/source_blobs/8a8c0e212718e0f53d004acad86827b4fab97539fe7ae1ada5c4c855ef7f57e0.py": {
+      "bytes": 2243,
+      "sha256": "8a8c0e212718e0f53d004acad86827b4fab97539fe7ae1ada5c4c855ef7f57e0"
+    },
+    "training/source_blobs/8ab3b8004f0c208bc1c8c74ad2285e4fa067bb29481816342b5ae6eaa0670652.py": {
+      "bytes": 1966,
+      "sha256": "8ab3b8004f0c208bc1c8c74ad2285e4fa067bb29481816342b5ae6eaa0670652"
+    },
+    "training/source_blobs/8b29bdc076c576d3f951e1d02f2eca1a2a6d78e86202d19704c3814a4cb0a84e.py": {
+      "bytes": 10103,
+      "sha256": "8b29bdc076c576d3f951e1d02f2eca1a2a6d78e86202d19704c3814a4cb0a84e"
+    },
+    "training/source_blobs/8bcc4cdb380d68a5db6e221481cde7ed5cc4b5fc9c87d07cd9d47f41ef056aab.py": {
+      "bytes": 10270,
+      "sha256": "8bcc4cdb380d68a5db6e221481cde7ed5cc4b5fc9c87d07cd9d47f41ef056aab"
+    },
+    "training/source_blobs/8bcfaff689b96ba9716ec0255dc0f91c5583d9b696784605dacbd88a6562cc7e.py": {
+      "bytes": 1659,
+      "sha256": "8bcfaff689b96ba9716ec0255dc0f91c5583d9b696784605dacbd88a6562cc7e"
+    },
+    "training/source_blobs/8eed9a85b1c94e84405f563f12504b4930ddbcd2d9f870aff9ce1fee553e8b5f.py": {
+      "bytes": 199242,
+      "sha256": "8eed9a85b1c94e84405f563f12504b4930ddbcd2d9f870aff9ce1fee553e8b5f"
+    },
+    "training/source_blobs/8f74c3f3ef6840811051fb8cfd5ac50249c94b116f2db47a5ba56ca8d8c92534.py": {
+      "bytes": 5064,
+      "sha256": "8f74c3f3ef6840811051fb8cfd5ac50249c94b116f2db47a5ba56ca8d8c92534"
+    },
+    "training/source_blobs/907b357fdd5c8ade00becbfe952a99791855a2afc90ecc3fd026a7860fb13ebf.py": {
+      "bytes": 7274,
+      "sha256": "907b357fdd5c8ade00becbfe952a99791855a2afc90ecc3fd026a7860fb13ebf"
+    },
+    "training/source_blobs/90c58a30a50cb625e36222825768a134aba5c81df87de8655ab95de2bd47402c.py": {
+      "bytes": 8982,
+      "sha256": "90c58a30a50cb625e36222825768a134aba5c81df87de8655ab95de2bd47402c"
+    },
+    "training/source_blobs/90e9b0e31206f0560e7e2aba2a5cf1adb9e9a2e0f23a5ab0fabdacf22ebb1470.py": {
+      "bytes": 532,
+      "sha256": "90e9b0e31206f0560e7e2aba2a5cf1adb9e9a2e0f23a5ab0fabdacf22ebb1470"
+    },
+    "training/source_blobs/92030c2ce302eea434bd81553ddac5c9170db82c491950cdb903f2caf732aee4.py": {
+      "bytes": 4633,
+      "sha256": "92030c2ce302eea434bd81553ddac5c9170db82c491950cdb903f2caf732aee4"
+    },
+    "training/source_blobs/922bba304680cb24aaad23f879faaa13387f15fcbd02bcd1c03734dc3a652c3b.py": {
+      "bytes": 8176,
+      "sha256": "922bba304680cb24aaad23f879faaa13387f15fcbd02bcd1c03734dc3a652c3b"
+    },
+    "training/source_blobs/93294262c9e8f9d0b6082d0c4f56149925969f9020279e8bc50d928c6e6811d7.py": {
+      "bytes": 24297,
+      "sha256": "93294262c9e8f9d0b6082d0c4f56149925969f9020279e8bc50d928c6e6811d7"
+    },
+    "training/source_blobs/93b86664d0aba1e36083877e1cdb1a995766dffb1b207bb4cd691e226629de53.py": {
+      "bytes": 3204,
+      "sha256": "93b86664d0aba1e36083877e1cdb1a995766dffb1b207bb4cd691e226629de53"
+    },
+    "training/source_blobs/93dcc3aa14fba92023f3b9fc44caab09bded50c9ef54a7b9d95281b6ab5afc25.py": {
+      "bytes": 1204,
+      "sha256": "93dcc3aa14fba92023f3b9fc44caab09bded50c9ef54a7b9d95281b6ab5afc25"
+    },
+    "training/source_blobs/943fd969994e896faff14481b94eb121d23b6fef00244874d5d5361df6bd8cce.py": {
+      "bytes": 2333,
+      "sha256": "943fd969994e896faff14481b94eb121d23b6fef00244874d5d5361df6bd8cce"
+    },
+    "training/source_blobs/94b8de5c15570b753850ca1ba84543c858334e5b98ed20d60a1bb0acf1e6c17e.py": {
+      "bytes": 8453,
+      "sha256": "94b8de5c15570b753850ca1ba84543c858334e5b98ed20d60a1bb0acf1e6c17e"
+    },
+    "training/source_blobs/955993d44637e21b043a0b8fc06b48c97589e40f255d90ff6c795394c6a5b569.py": {
+      "bytes": 1871,
+      "sha256": "955993d44637e21b043a0b8fc06b48c97589e40f255d90ff6c795394c6a5b569"
+    },
+    "training/source_blobs/965c794f040d45cf2abefb16ea17c55e14be81d390b6534d36da7a12cecf5a93.py": {
+      "bytes": 192,
+      "sha256": "965c794f040d45cf2abefb16ea17c55e14be81d390b6534d36da7a12cecf5a93"
+    },
+    "training/source_blobs/96795d99ce287b32200c9b56069856324678feba937834904efbc46a6dcc6404.py": {
+      "bytes": 10878,
+      "sha256": "96795d99ce287b32200c9b56069856324678feba937834904efbc46a6dcc6404"
+    },
+    "training/source_blobs/97243740f45c7b79f5e358a4b9378e54a3e9114f150d2780a718a1e62164f2f9.py": {
+      "bytes": 4771,
+      "sha256": "97243740f45c7b79f5e358a4b9378e54a3e9114f150d2780a718a1e62164f2f9"
+    },
+    "training/source_blobs/977a196d5949f9ebc965b253afd45b818290c8ff5ea3a462efa219c368f6f4bd.py": {
+      "bytes": 1312,
+      "sha256": "977a196d5949f9ebc965b253afd45b818290c8ff5ea3a462efa219c368f6f4bd"
+    },
+    "training/source_blobs/9a99bcf27a5895cc4f1ab34103c303675e05cb745944f9b10dfb26e3651a6581.py": {
+      "bytes": 2191,
+      "sha256": "9a99bcf27a5895cc4f1ab34103c303675e05cb745944f9b10dfb26e3651a6581"
+    },
+    "training/source_blobs/9bf5f9f145991764775f032257ccf80190a1cb5d6f45e6ef504c3cebd87433ca.py": {
+      "bytes": 9803,
+      "sha256": "9bf5f9f145991764775f032257ccf80190a1cb5d6f45e6ef504c3cebd87433ca"
+    },
+    "training/source_blobs/9ea68c491cc79452f58a4a93c483af6c26fdff1f6caeb924e004d5298e927a48.py": {
+      "bytes": 13663,
+      "sha256": "9ea68c491cc79452f58a4a93c483af6c26fdff1f6caeb924e004d5298e927a48"
+    },
+    "training/source_blobs/9efa69e4acf6afef250e0d1ab45ecbac4da042e21a7fb55f38082a13b0fe4abf.py": {
+      "bytes": 12146,
+      "sha256": "9efa69e4acf6afef250e0d1ab45ecbac4da042e21a7fb55f38082a13b0fe4abf"
+    },
+    "training/source_blobs/9efaba2abb2f32c9f5778872fd7849609c8045687aa7ac8146a1001bf653f151.py": {
+      "bytes": 15629,
+      "sha256": "9efaba2abb2f32c9f5778872fd7849609c8045687aa7ac8146a1001bf653f151"
+    },
+    "training/source_blobs/9f3a8646f9661e0c31bcb3188d4b3410a0e6454a128a617d18b79ea0cbf660ee.py": {
+      "bytes": 19571,
+      "sha256": "9f3a8646f9661e0c31bcb3188d4b3410a0e6454a128a617d18b79ea0cbf660ee"
+    },
+    "training/source_blobs/9f7f0804eaa9edc6d9c4a5049df230a454171b7f307144212a881af0b3db3971.py": {
+      "bytes": 15910,
+      "sha256": "9f7f0804eaa9edc6d9c4a5049df230a454171b7f307144212a881af0b3db3971"
+    },
+    "training/source_blobs/9ffff2dcc93bcf275f1bc183c8cef37914602a25342f65c1404fd7b398854c82.py": {
+      "bytes": 26934,
+      "sha256": "9ffff2dcc93bcf275f1bc183c8cef37914602a25342f65c1404fd7b398854c82"
+    },
+    "training/source_blobs/a08b55556fe466c10c7d8188066b1a6a23f86d2545378d85f61898cbe300a67d.py": {
+      "bytes": 5523,
+      "sha256": "a08b55556fe466c10c7d8188066b1a6a23f86d2545378d85f61898cbe300a67d"
+    },
+    "training/source_blobs/a1c6f3f901996aac30d7ffcf2b3fdc89873faba7cf5bd1a8ec301ff8076bb719.py": {
+      "bytes": 4389,
+      "sha256": "a1c6f3f901996aac30d7ffcf2b3fdc89873faba7cf5bd1a8ec301ff8076bb719"
+    },
+    "training/source_blobs/a1f77593681bc49b242a042d4819e5da3baa60ac4e6e65e9cd8ea4d5aae8a9e0.py": {
+      "bytes": 226,
+      "sha256": "a1f77593681bc49b242a042d4819e5da3baa60ac4e6e65e9cd8ea4d5aae8a9e0"
+    },
+    "training/source_blobs/a3dfbd61de43fda14daf4bbd921ccb382ea8af2c8342457fe16d1ad0908e88fe.py": {
+      "bytes": 11399,
+      "sha256": "a3dfbd61de43fda14daf4bbd921ccb382ea8af2c8342457fe16d1ad0908e88fe"
+    },
+    "training/source_blobs/a43649482bd7234bdb704bd014031ad69769086b2ac60c8eaa9190bbf428629c.py": {
+      "bytes": 15463,
+      "sha256": "a43649482bd7234bdb704bd014031ad69769086b2ac60c8eaa9190bbf428629c"
+    },
+    "training/source_blobs/a6b6437f5efbf0771af2a70e26d63bbcdaa45a397fa19f6c214c27a8140e4b78.py": {
+      "bytes": 3938,
+      "sha256": "a6b6437f5efbf0771af2a70e26d63bbcdaa45a397fa19f6c214c27a8140e4b78"
+    },
+    "training/source_blobs/a71eabcc526344b6502cae999020da63c950b2d80c13077629bb86c8757a18ed.py": {
+      "bytes": 6060,
+      "sha256": "a71eabcc526344b6502cae999020da63c950b2d80c13077629bb86c8757a18ed"
+    },
+    "training/source_blobs/a7aa3fc6627fd055924ba7da19c280146bd98e2e1257475816326d18fe41eabc.py": {
+      "bytes": 5152,
+      "sha256": "a7aa3fc6627fd055924ba7da19c280146bd98e2e1257475816326d18fe41eabc"
+    },
+    "training/source_blobs/a7e3879ae4bd1506473d9b34e37dc1a94f9c9d2048995e0cba7fb04c61514554.py": {
+      "bytes": 24797,
+      "sha256": "a7e3879ae4bd1506473d9b34e37dc1a94f9c9d2048995e0cba7fb04c61514554"
+    },
+    "training/source_blobs/a8379185dd82972a931e7270b7c26142ad30bc1340af309c8ff087965d79964a.py": {
+      "bytes": 10906,
+      "sha256": "a8379185dd82972a931e7270b7c26142ad30bc1340af309c8ff087965d79964a"
+    },
+    "training/source_blobs/a8d9306095bb1de53cc4350741fe3132269b2a91600e351e6285d8ec1e3be6e3.py": {
+      "bytes": 501,
+      "sha256": "a8d9306095bb1de53cc4350741fe3132269b2a91600e351e6285d8ec1e3be6e3"
+    },
+    "training/source_blobs/a971a2e697a35c34d973d8add104fbcc5d661cb83473ccafceeed4820d280291.py": {
+      "bytes": 11840,
+      "sha256": "a971a2e697a35c34d973d8add104fbcc5d661cb83473ccafceeed4820d280291"
+    },
+    "training/source_blobs/aad8eba4764e40e490856986f1f369cade78635d087e4fda1b9ad8455ffd07f3.py": {
+      "bytes": 5788,
+      "sha256": "aad8eba4764e40e490856986f1f369cade78635d087e4fda1b9ad8455ffd07f3"
+    },
+    "training/source_blobs/ab01e0e88e2edc87a1960666366fa433f976a0c021270cf412dba32b918914d5.py": {
+      "bytes": 6158,
+      "sha256": "ab01e0e88e2edc87a1960666366fa433f976a0c021270cf412dba32b918914d5"
+    },
+    "training/source_blobs/ac95755ba4b994dae0c0902cd052a9200ec09eecac28c547408a4ebabdbc99b0.py": {
+      "bytes": 805,
+      "sha256": "ac95755ba4b994dae0c0902cd052a9200ec09eecac28c547408a4ebabdbc99b0"
+    },
+    "training/source_blobs/acca03eb1f853e16c9166c144b645f44f470cc81cde0c0aa5a3ed0c4b3c0b070.py": {
+      "bytes": 224,
+      "sha256": "acca03eb1f853e16c9166c144b645f44f470cc81cde0c0aa5a3ed0c4b3c0b070"
+    },
+    "training/source_blobs/ad220f0033d3761d67d8048c8b4a730fb1ce47923e88c6bbe236b41067b83a82.py": {
+      "bytes": 2679,
+      "sha256": "ad220f0033d3761d67d8048c8b4a730fb1ce47923e88c6bbe236b41067b83a82"
+    },
+    "training/source_blobs/adc3a0a6f19e7cfb3073f4c1e0432f664e5fb709141ceb4148aca7cda08ddaab.py": {
+      "bytes": 3259,
+      "sha256": "adc3a0a6f19e7cfb3073f4c1e0432f664e5fb709141ceb4148aca7cda08ddaab"
+    },
+    "training/source_blobs/ae59081bcf4ea79fb7c5a2ab5cbac809976074e0f4fb290ae65bbf57f4733ca2.py": {
+      "bytes": 1867,
+      "sha256": "ae59081bcf4ea79fb7c5a2ab5cbac809976074e0f4fb290ae65bbf57f4733ca2"
+    },
+    "training/source_blobs/af7bb33a7b0253c552bbefd46b1b3ad8857dd7740ea1118480948739075b210a.py": {
+      "bytes": 2707,
+      "sha256": "af7bb33a7b0253c552bbefd46b1b3ad8857dd7740ea1118480948739075b210a"
+    },
+    "training/source_blobs/b0095392ef6a92ae2aad4ed20a8b9d668fe02f8f0ed602bd023a1a5d1ff59f4b.py": {
+      "bytes": 12078,
+      "sha256": "b0095392ef6a92ae2aad4ed20a8b9d668fe02f8f0ed602bd023a1a5d1ff59f4b"
+    },
+    "training/source_blobs/b0bf909b472ddf9bd4eab30f6e6fd8b33bf3997ce038b882fe0663f291958d2d.txt": {
+      "bytes": 1065,
+      "sha256": "b0bf909b472ddf9bd4eab30f6e6fd8b33bf3997ce038b882fe0663f291958d2d"
+    },
+    "training/source_blobs/b0bfe4c4afa5b00155a047fde824fb15baffd6b3b493f94a0a8da1cad6fc5f9e.py": {
+      "bytes": 4993,
+      "sha256": "b0bfe4c4afa5b00155a047fde824fb15baffd6b3b493f94a0a8da1cad6fc5f9e"
+    },
+    "training/source_blobs/b222ce0b6eacf8a244328883cb751325c402a1027170829e91ac5fbdfbb8da3f.py": {
+      "bytes": 17122,
+      "sha256": "b222ce0b6eacf8a244328883cb751325c402a1027170829e91ac5fbdfbb8da3f"
+    },
+    "training/source_blobs/b2f98795df373c7f1e91b2c2c2c158b51c7ef65c47522ad20a685c63e1773f99.py": {
+      "bytes": 15332,
+      "sha256": "b2f98795df373c7f1e91b2c2c2c158b51c7ef65c47522ad20a685c63e1773f99"
+    },
+    "training/source_blobs/b30cd4ab4e99eedaf623580e3ff89b4a197534b14ee1d378c36d771fec303d82.py": {
+      "bytes": 18854,
+      "sha256": "b30cd4ab4e99eedaf623580e3ff89b4a197534b14ee1d378c36d771fec303d82"
+    },
+    "training/source_blobs/b3bb31dd1dcf0fec5416d81664998e80b0244669fecd4389669083be09ec8daa.py": {
+      "bytes": 4707,
+      "sha256": "b3bb31dd1dcf0fec5416d81664998e80b0244669fecd4389669083be09ec8daa"
+    },
+    "training/source_blobs/b575e3334fe9a21b3b9fc428d0f25728dfd4523426f4bf7aa34f6ddd7d4769e8.py": {
+      "bytes": 353,
+      "sha256": "b575e3334fe9a21b3b9fc428d0f25728dfd4523426f4bf7aa34f6ddd7d4769e8"
+    },
+    "training/source_blobs/b62637ccddac28f42b6efa19ce0578365497f7823d0df1dd2a9e27df453fea5c.py": {
+      "bytes": 6778,
+      "sha256": "b62637ccddac28f42b6efa19ce0578365497f7823d0df1dd2a9e27df453fea5c"
+    },
+    "training/source_blobs/b64d1908b52856e19510283933d8b8d80f6ba15a8511a97732499a2820a06c76.py": {
+      "bytes": 5421,
+      "sha256": "b64d1908b52856e19510283933d8b8d80f6ba15a8511a97732499a2820a06c76"
+    },
+    "training/source_blobs/b651f986a836c59591c13776ada2b17c1a359adbe6c96dd8306ef6b208d12231.py": {
+      "bytes": 2638,
+      "sha256": "b651f986a836c59591c13776ada2b17c1a359adbe6c96dd8306ef6b208d12231"
+    },
+    "training/source_blobs/b828e68f67e7ee17cef4ff75d5c2b354672fbe6f559607248d621fd5a1e490f4.py": {
+      "bytes": 582,
+      "sha256": "b828e68f67e7ee17cef4ff75d5c2b354672fbe6f559607248d621fd5a1e490f4"
+    },
+    "training/source_blobs/b9f535fca724153703ed57a300eceb5b34a8f8381fc65bf5c4c5a0f7e066303c.py": {
+      "bytes": 1292,
+      "sha256": "b9f535fca724153703ed57a300eceb5b34a8f8381fc65bf5c4c5a0f7e066303c"
+    },
+    "training/source_blobs/babe89bddb186a02e8c3ab4be1fa2cbb64f2d7bb27f8ac370701b4df39e03001.py": {
+      "bytes": 22208,
+      "sha256": "babe89bddb186a02e8c3ab4be1fa2cbb64f2d7bb27f8ac370701b4df39e03001"
+    },
+    "training/source_blobs/bb3ba5db709a7df15a2936db36793939f877e99dd437956c64042d6e6cc25886.py": {
+      "bytes": 11673,
+      "sha256": "bb3ba5db709a7df15a2936db36793939f877e99dd437956c64042d6e6cc25886"
+    },
+    "training/source_blobs/bbe9f97676c183f91308fce8e4b54c8b2ea04c4cc1e96db9f89d24a1e95efc48.py": {
+      "bytes": 4211,
+      "sha256": "bbe9f97676c183f91308fce8e4b54c8b2ea04c4cc1e96db9f89d24a1e95efc48"
+    },
+    "training/source_blobs/bc89ed266b373257a693d1548c7d0d8304830e9d26517002623d7d29fbb467d2.py": {
+      "bytes": 7208,
+      "sha256": "bc89ed266b373257a693d1548c7d0d8304830e9d26517002623d7d29fbb467d2"
+    },
+    "training/source_blobs/bddaf2b1f3691319843da58f4f37eb9ff02bd077502f6c87a082278cdbb970c6.py": {
+      "bytes": 419525,
+      "sha256": "bddaf2b1f3691319843da58f4f37eb9ff02bd077502f6c87a082278cdbb970c6"
+    },
+    "training/source_blobs/be338639aa34512297c2d563fafd7d3b12f00f4ceb43ae8758e386c232c2439b.py": {
+      "bytes": 36355,
+      "sha256": "be338639aa34512297c2d563fafd7d3b12f00f4ceb43ae8758e386c232c2439b"
+    },
+    "training/source_blobs/bffd8f8c5a74a0bf9c3ca070e6511cc1ab8ceb8343c0092a3d711b4fb9ece21c.py": {
+      "bytes": 3709,
+      "sha256": "bffd8f8c5a74a0bf9c3ca070e6511cc1ab8ceb8343c0092a3d711b4fb9ece21c"
+    },
+    "training/source_blobs/c0468e20922d17555f41cbd76255a183917e1bde85d6c621732862fd2cf176de.py": {
+      "bytes": 1470,
+      "sha256": "c0468e20922d17555f41cbd76255a183917e1bde85d6c621732862fd2cf176de"
+    },
+    "training/source_blobs/c065be118b0c6b8948a15d4b81e7fd7371f4e699ea3c49d34fa0abde258e0505.py": {
+      "bytes": 9445,
+      "sha256": "c065be118b0c6b8948a15d4b81e7fd7371f4e699ea3c49d34fa0abde258e0505"
+    },
+    "training/source_blobs/c2c4b70dcee9caf027c49e0d9e3b8071b24d9bc6d197d1fec502618513c0cda2.py": {
+      "bytes": 11456,
+      "sha256": "c2c4b70dcee9caf027c49e0d9e3b8071b24d9bc6d197d1fec502618513c0cda2"
+    },
+    "training/source_blobs/c2f5c1d6dd2de0195d89370359f91c438a48dacf5f157aa83607b24abe363265.py": {
+      "bytes": 1565,
+      "sha256": "c2f5c1d6dd2de0195d89370359f91c438a48dacf5f157aa83607b24abe363265"
+    },
+    "training/source_blobs/c3565fe2c10ad14e2794e2d1a23d2fa5cce1a206d0064e4626a70b50a06c73f1.py": {
+      "bytes": 4124,
+      "sha256": "c3565fe2c10ad14e2794e2d1a23d2fa5cce1a206d0064e4626a70b50a06c73f1"
+    },
+    "training/source_blobs/c745011c90cffc70e751eee207e2dca10ad94e1f2e167b967f0fdc8273afed36.py": {
+      "bytes": 26967,
+      "sha256": "c745011c90cffc70e751eee207e2dca10ad94e1f2e167b967f0fdc8273afed36"
+    },
+    "training/source_blobs/c75ab44c4dbf49aab3e805bd51b5f1c607d7de917197dc10673771adc846592f.py": {
+      "bytes": 7065,
+      "sha256": "c75ab44c4dbf49aab3e805bd51b5f1c607d7de917197dc10673771adc846592f"
+    },
+    "training/source_blobs/c81e7321a9a9fbf9cfe5a70434e4df2c751f33b6d68d93697ba4f09f69b56d88.py": {
+      "bytes": 4552,
+      "sha256": "c81e7321a9a9fbf9cfe5a70434e4df2c751f33b6d68d93697ba4f09f69b56d88"
+    },
+    "training/source_blobs/c836de3dfbc0df4839e6b9e818406f6791b6b2db0a3531fbca8f90251cde9498.py": {
+      "bytes": 3447,
+      "sha256": "c836de3dfbc0df4839e6b9e818406f6791b6b2db0a3531fbca8f90251cde9498"
+    },
+    "training/source_blobs/c9567f40f61e375ae5cea20ec10355ab4bc270af8b64e7b316b30a69947f4e33.py": {
+      "bytes": 18362,
+      "sha256": "c9567f40f61e375ae5cea20ec10355ab4bc270af8b64e7b316b30a69947f4e33"
+    },
+    "training/source_blobs/c964e65593989b7ee6ca3ad7cc482ee926e95b91dcc123d3ccb4999710f2eccf.py": {
+      "bytes": 8255,
+      "sha256": "c964e65593989b7ee6ca3ad7cc482ee926e95b91dcc123d3ccb4999710f2eccf"
+    },
+    "training/source_blobs/c9cfca2802cf083776ddb75c8805920751e763cb943efeea65f0949b732f90a7.py": {
+      "bytes": 22323,
+      "sha256": "c9cfca2802cf083776ddb75c8805920751e763cb943efeea65f0949b732f90a7"
+    },
+    "training/source_blobs/cad10665efe683f5d7cec5ac8988e18e5970c7b354e248e1502eb7f9c242eab5.py": {
+      "bytes": 8063,
+      "sha256": "cad10665efe683f5d7cec5ac8988e18e5970c7b354e248e1502eb7f9c242eab5"
+    },
+    "training/source_blobs/cb6e341043ea2b86c3058832c443a187c3abfb0bbdbca024360f04bc284479aa.py": {
+      "bytes": 8443,
+      "sha256": "cb6e341043ea2b86c3058832c443a187c3abfb0bbdbca024360f04bc284479aa"
+    },
+    "training/source_blobs/cc7c710aaca40ac224da99f5902d4833b16d36b7a12c23817ee07ae3e29dbfda.py": {
+      "bytes": 8069,
+      "sha256": "cc7c710aaca40ac224da99f5902d4833b16d36b7a12c23817ee07ae3e29dbfda"
+    },
+    "training/source_blobs/cc908731e8d2b2b0a848673a83bf094982a06446016df59896cdece491463b0e.py": {
+      "bytes": 3437,
+      "sha256": "cc908731e8d2b2b0a848673a83bf094982a06446016df59896cdece491463b0e"
+    },
+    "training/source_blobs/cf09c5de39997350d25806866a3b3aedfe315ffeb6e4bdf00cf61ae05b24c3a2.py": {
+      "bytes": 285927,
+      "sha256": "cf09c5de39997350d25806866a3b3aedfe315ffeb6e4bdf00cf61ae05b24c3a2"
+    },
+    "training/source_blobs/cf7a8e609115a29f0b97148dc531367e916c1219e60ecbbb3c372906f994453a.py": {
+      "bytes": 11524,
+      "sha256": "cf7a8e609115a29f0b97148dc531367e916c1219e60ecbbb3c372906f994453a"
+    },
+    "training/source_blobs/cf80cb1bbde860cfc08808dac1fdf22cfd8d5900ce1294c7125b6c59def0de80.py": {
+      "bytes": 3331,
+      "sha256": "cf80cb1bbde860cfc08808dac1fdf22cfd8d5900ce1294c7125b6c59def0de80"
+    },
+    "training/source_blobs/d35a801bc3f2c47528f371aa6ad99574ad1230ab25745b189f79a04268b8ee7c.py": {
+      "bytes": 10751,
+      "sha256": "d35a801bc3f2c47528f371aa6ad99574ad1230ab25745b189f79a04268b8ee7c"
+    },
+    "training/source_blobs/d38da245bf7b2f54bfd4c728d2638dd9050888e3f5a2c7ed2360201b960e75a8.py": {
+      "bytes": 323,
+      "sha256": "d38da245bf7b2f54bfd4c728d2638dd9050888e3f5a2c7ed2360201b960e75a8"
+    },
+    "training/source_blobs/d39a65a18405683724659ce1ea4464d5f65fe6c3a05ca81d409fc6982ab052a0.py": {
+      "bytes": 8248,
+      "sha256": "d39a65a18405683724659ce1ea4464d5f65fe6c3a05ca81d409fc6982ab052a0"
+    },
+    "training/source_blobs/d4393b7b221898b9019429346e71610db27bbca91904f6c80ae75c92b182b5cd.py": {
+      "bytes": 4718,
+      "sha256": "d4393b7b221898b9019429346e71610db27bbca91904f6c80ae75c92b182b5cd"
+    },
+    "training/source_blobs/d5c27314990d2453044f4bc7993819a538ded896eb047ba72244e15c4b495f09.py": {
+      "bytes": 69106,
+      "sha256": "d5c27314990d2453044f4bc7993819a538ded896eb047ba72244e15c4b495f09"
+    },
+    "training/source_blobs/d6a488047f6c58822b2881de9d9c363ce67435ec7fe07793b4043b0cd620c4f8.py": {
+      "bytes": 34547,
+      "sha256": "d6a488047f6c58822b2881de9d9c363ce67435ec7fe07793b4043b0cd620c4f8"
+    },
+    "training/source_blobs/d9f529149989771c6bb04f46fb366413b8c8817925999a1cfa46c229c9782414.py": {
+      "bytes": 2414,
+      "sha256": "d9f529149989771c6bb04f46fb366413b8c8817925999a1cfa46c229c9782414"
+    },
+    "training/source_blobs/da0cf71d99bfca88f04c0fb1182c200321aa27f0366980d566748970e0133f92.py": {
+      "bytes": 72,
+      "sha256": "da0cf71d99bfca88f04c0fb1182c200321aa27f0366980d566748970e0133f92"
+    },
+    "training/source_blobs/da5d0b5d3d9762dd806ad0ed9da1e8f0edd3e69dc8d626f20d9e6af27673f375.py": {
+      "bytes": 3691,
+      "sha256": "da5d0b5d3d9762dd806ad0ed9da1e8f0edd3e69dc8d626f20d9e6af27673f375"
+    },
+    "training/source_blobs/db6cb7946e10423bdee30ebcead28ac1bc6b6e670204aaa6488f1b4ccb9d4240.py": {
+      "bytes": 16072,
+      "sha256": "db6cb7946e10423bdee30ebcead28ac1bc6b6e670204aaa6488f1b4ccb9d4240"
+    },
+    "training/source_blobs/dbd685610812235fa72ac6081a1bc16c8ff55eecdec034c0df0c9c2eadde05e8.py": {
+      "bytes": 6826,
+      "sha256": "dbd685610812235fa72ac6081a1bc16c8ff55eecdec034c0df0c9c2eadde05e8"
+    },
+    "training/source_blobs/dc0601b93d20a58245982b6041763212720e7d5386993749ba88d0f79d8e8479.py": {
+      "bytes": 1508,
+      "sha256": "dc0601b93d20a58245982b6041763212720e7d5386993749ba88d0f79d8e8479"
+    },
+    "training/source_blobs/dc06bedad9c254e70109c0206c54eb594f6f5c427f2ffc0fb9f162d8191b4813.py": {
+      "bytes": 31912,
+      "sha256": "dc06bedad9c254e70109c0206c54eb594f6f5c427f2ffc0fb9f162d8191b4813"
+    },
+    "training/source_blobs/de1306af5306ff9af02d9bdecf07f6a99baffb1ac6a052d4690840cc7ad680c6.py": {
+      "bytes": 5968,
+      "sha256": "de1306af5306ff9af02d9bdecf07f6a99baffb1ac6a052d4690840cc7ad680c6"
+    },
+    "training/source_blobs/de9bf2fac7a6b3ed0ca0d23926cf02e19484383a78eefeeba3af3459f6ee46f4.py": {
+      "bytes": 14785,
+      "sha256": "de9bf2fac7a6b3ed0ca0d23926cf02e19484383a78eefeeba3af3459f6ee46f4"
+    },
+    "training/source_blobs/debe595c69ddc5d19bfb189193b75923e0412f4ba00c833dc2d9b7bafb636fad.py": {
+      "bytes": 4891,
+      "sha256": "debe595c69ddc5d19bfb189193b75923e0412f4ba00c833dc2d9b7bafb636fad"
+    },
+    "training/source_blobs/df16803b62d90800c51ad8072123067814d24a23d4a0569671fc44c3684b9f74.py": {
+      "bytes": 7697,
+      "sha256": "df16803b62d90800c51ad8072123067814d24a23d4a0569671fc44c3684b9f74"
+    },
+    "training/source_blobs/df55a101f2bcf31f32d7e201e9b27071ab404f34b0e7df3f805a8ddb322e1f80.py": {
+      "bytes": 4041,
+      "sha256": "df55a101f2bcf31f32d7e201e9b27071ab404f34b0e7df3f805a8ddb322e1f80"
+    },
+    "training/source_blobs/e00e99d07655b7fdf96a3164d182de3683f81f43817e93c5f8b89e6bf7130c91.py": {
+      "bytes": 2228,
+      "sha256": "e00e99d07655b7fdf96a3164d182de3683f81f43817e93c5f8b89e6bf7130c91"
+    },
+    "training/source_blobs/e137daae9e39c642f5b2cabbfc3e2730446e399a61e44800901e5b9c17da84bd.py": {
+      "bytes": 17258,
+      "sha256": "e137daae9e39c642f5b2cabbfc3e2730446e399a61e44800901e5b9c17da84bd"
+    },
+    "training/source_blobs/e141132efbd5f1beab4b187d4eab15efd46fcf9710d6a2ba7c882f62362a77cf.py": {
+      "bytes": 12645,
+      "sha256": "e141132efbd5f1beab4b187d4eab15efd46fcf9710d6a2ba7c882f62362a77cf"
+    },
+    "training/source_blobs/e14e793f20d6b1830defcbb37ca498f2ecbfac7c6c08b02e2dfd3d1e06311286.py": {
+      "bytes": 1035,
+      "sha256": "e14e793f20d6b1830defcbb37ca498f2ecbfac7c6c08b02e2dfd3d1e06311286"
+    },
+    "training/source_blobs/e178e766849e6c96fdea770a2bb307562c27bfb3761a81bd5eec12a3d84a7995.py": {
+      "bytes": 5791,
+      "sha256": "e178e766849e6c96fdea770a2bb307562c27bfb3761a81bd5eec12a3d84a7995"
+    },
+    "training/source_blobs/e31d4cdcadea1fc6b0d8449098b06d946d58f308ad2e6efe7eac6b9c3f72ed98.py": {
+      "bytes": 329,
+      "sha256": "e31d4cdcadea1fc6b0d8449098b06d946d58f308ad2e6efe7eac6b9c3f72ed98"
+    },
+    "training/source_blobs/e3780f86fb01a1e34e44b78a42d0648f259f41c1d6a0afbfbb600c4d54c737c8.py": {
+      "bytes": 30717,
+      "sha256": "e3780f86fb01a1e34e44b78a42d0648f259f41c1d6a0afbfbb600c4d54c737c8"
+    },
+    "training/source_blobs/e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855.py": {
+      "bytes": 0,
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    "training/source_blobs/e3c3977ad23a40b6dda4164996164dadc44b6424a56abd816d12076b81f5b143.py": {
+      "bytes": 5549,
+      "sha256": "e3c3977ad23a40b6dda4164996164dadc44b6424a56abd816d12076b81f5b143"
+    },
+    "training/source_blobs/e4eb3d7799f756c95f8d714e59618c9171239db79ec9750781bcc5cbbcf4416a.py": {
+      "bytes": 27773,
+      "sha256": "e4eb3d7799f756c95f8d714e59618c9171239db79ec9750781bcc5cbbcf4416a"
+    },
+    "training/source_blobs/e5285b4ce241af63d7038d2cb2470b2b397686ec904a9ca83f02b6fe1a1d7ce8.py": {
+      "bytes": 4832,
+      "sha256": "e5285b4ce241af63d7038d2cb2470b2b397686ec904a9ca83f02b6fe1a1d7ce8"
+    },
+    "training/source_blobs/e57dbb7567a2438a497802b66d4a0188c53c3afdd9eb7c313334b8e312ad4999.py": {
+      "bytes": 5569,
+      "sha256": "e57dbb7567a2438a497802b66d4a0188c53c3afdd9eb7c313334b8e312ad4999"
+    },
+    "training/source_blobs/e672d5bddefb2a79eaec3185ad436940522f352e762f265c461764050d323855.py": {
+      "bytes": 1063,
+      "sha256": "e672d5bddefb2a79eaec3185ad436940522f352e762f265c461764050d323855"
+    },
+    "training/source_blobs/e6778ba8a8215f0b3c20386e768ea3e15b5939e7919935046689733eb96eba28.py": {
+      "bytes": 5442,
+      "sha256": "e6778ba8a8215f0b3c20386e768ea3e15b5939e7919935046689733eb96eba28"
+    },
+    "training/source_blobs/e685abc5e4579508bf432632159c64d89fab91a1831235bcc0a0f12e4fa309f7.py": {
+      "bytes": 1808,
+      "sha256": "e685abc5e4579508bf432632159c64d89fab91a1831235bcc0a0f12e4fa309f7"
+    },
+    "training/source_blobs/e73dd7b7ded430562a55380a8e5a5a5f52b8176f53277da1f0e23f48466b0e34.py": {
+      "bytes": 17302,
+      "sha256": "e73dd7b7ded430562a55380a8e5a5a5f52b8176f53277da1f0e23f48466b0e34"
+    },
+    "training/source_blobs/e8dd3c8676b9cf5bc7ecd501f3afe82555c2c0d4f7700bb6e05417278447a995.py": {
+      "bytes": 191,
+      "sha256": "e8dd3c8676b9cf5bc7ecd501f3afe82555c2c0d4f7700bb6e05417278447a995"
+    },
+    "training/source_blobs/e915efcc450bd027272f029e6ff52c9ca9fc8f5b13bc593e4f019c6efaf191ca.py": {
+      "bytes": 14127,
+      "sha256": "e915efcc450bd027272f029e6ff52c9ca9fc8f5b13bc593e4f019c6efaf191ca"
+    },
+    "training/source_blobs/ea92b8b5a884cce7459a1ef2957da8bf8f4a81a9165ffdca461dcc24a2fc8df4.py": {
+      "bytes": 3628,
+      "sha256": "ea92b8b5a884cce7459a1ef2957da8bf8f4a81a9165ffdca461dcc24a2fc8df4"
+    },
+    "training/source_blobs/ebc86535e538107334e76c16ddfbfd2671a5635281f46d613f11ad68b627f866.py": {
+      "bytes": 11689,
+      "sha256": "ebc86535e538107334e76c16ddfbfd2671a5635281f46d613f11ad68b627f866"
+    },
+    "training/source_blobs/ebd8bcd2612f1c023924e2d7ab6afaf3ea215bb1b3c4fdf451b54fa4c561dbce.py": {
+      "bytes": 9611,
+      "sha256": "ebd8bcd2612f1c023924e2d7ab6afaf3ea215bb1b3c4fdf451b54fa4c561dbce"
+    },
+    "training/source_blobs/ec2d2174ac2524e68f2b20d789f52eb31953e4d0a4d1c5fb70f6516680b6a4ea.py": {
+      "bytes": 7061,
+      "sha256": "ec2d2174ac2524e68f2b20d789f52eb31953e4d0a4d1c5fb70f6516680b6a4ea"
+    },
+    "training/source_blobs/ec619a6e34b3b53aa71aed4bb0b9ffe241f151243cd114d69401229b1fadbe72.py": {
+      "bytes": 2126,
+      "sha256": "ec619a6e34b3b53aa71aed4bb0b9ffe241f151243cd114d69401229b1fadbe72"
+    },
+    "training/source_blobs/ed5eed8c69210ed743e47c0c4025c0b69114f02925f2ca8cd946911b311c6913.py": {
+      "bytes": 2278,
+      "sha256": "ed5eed8c69210ed743e47c0c4025c0b69114f02925f2ca8cd946911b311c6913"
+    },
+    "training/source_blobs/eebc3ef49e86023666bdb9ed6068c5d96b68446be4db8f4a25f13712efdb7de0.py": {
+      "bytes": 28786,
+      "sha256": "eebc3ef49e86023666bdb9ed6068c5d96b68446be4db8f4a25f13712efdb7de0"
+    },
+    "training/source_blobs/eff2f9cd32de9fa73563020d7b36c3a2b1c19d92022fed4f5c1d7c4b0cedc59c.py": {
+      "bytes": 2471,
+      "sha256": "eff2f9cd32de9fa73563020d7b36c3a2b1c19d92022fed4f5c1d7c4b0cedc59c"
+    },
+    "training/source_blobs/f086fb95bfc7bc091598f528dc5c6ef0c164e8c85627947e7d1b9241ad429efe.py": {
+      "bytes": 509,
+      "sha256": "f086fb95bfc7bc091598f528dc5c6ef0c164e8c85627947e7d1b9241ad429efe"
+    },
+    "training/source_blobs/f24ee4feacb1484017893933d0ba2484197d8027a51e522b1e28e48f3b82983f.py": {
+      "bytes": 34706,
+      "sha256": "f24ee4feacb1484017893933d0ba2484197d8027a51e522b1e28e48f3b82983f"
+    },
+    "training/source_blobs/f2a5e8599e7eb3a7026148dd57acb3fb5cf4311b3768f07831bcf7c740ead4a9.py": {
+      "bytes": 24076,
+      "sha256": "f2a5e8599e7eb3a7026148dd57acb3fb5cf4311b3768f07831bcf7c740ead4a9"
+    },
+    "training/source_blobs/f6e985a2f49e4412c0685f7632740abf87aebc32d3c359346d9ce0b8af38f50c.py": {
+      "bytes": 1976,
+      "sha256": "f6e985a2f49e4412c0685f7632740abf87aebc32d3c359346d9ce0b8af38f50c"
+    },
+    "training/source_blobs/f7f2f12b6b9056653aed9b2cc8cd928a6b0dbbaf74130421e3dc5cf8e03eaa1c.py": {
+      "bytes": 2888,
+      "sha256": "f7f2f12b6b9056653aed9b2cc8cd928a6b0dbbaf74130421e3dc5cf8e03eaa1c"
+    },
+    "training/source_blobs/f80f065f6892eea951e94a7ff6925d5a1bff9419a7384e0756cf2faeb92e4775.py": {
+      "bytes": 7312,
+      "sha256": "f80f065f6892eea951e94a7ff6925d5a1bff9419a7384e0756cf2faeb92e4775"
+    },
+    "training/source_blobs/f8d014e088704e7c366e91b61dbc6a959380908e4836eb6e3952a80b6a39768e.py": {
+      "bytes": 2219,
+      "sha256": "f8d014e088704e7c366e91b61dbc6a959380908e4836eb6e3952a80b6a39768e"
+    },
+    "training/source_blobs/f99edeedb6cb328ef728d05643f65ea20371437f7e5806009ae3ad0885bafbf1.py": {
+      "bytes": 3705,
+      "sha256": "f99edeedb6cb328ef728d05643f65ea20371437f7e5806009ae3ad0885bafbf1"
+    },
+    "training/source_blobs/f9cb70123441e889d53dc9c20d282826e4b6d4841562f2e3b21cca965427cf3d.py": {
+      "bytes": 3822,
+      "sha256": "f9cb70123441e889d53dc9c20d282826e4b6d4841562f2e3b21cca965427cf3d"
+    },
+    "training/source_blobs/f9e0e6242ac3dbdc151241cbcca9d173765514e2fb7878648cc886143f1d5376.py": {
+      "bytes": 535,
+      "sha256": "f9e0e6242ac3dbdc151241cbcca9d173765514e2fb7878648cc886143f1d5376"
+    },
+    "training/source_blobs/fab59323c6fc8302dfd226b1b7e0cdef9c865231f72e930e9968fba2139565c3.py": {
+      "bytes": 6059,
+      "sha256": "fab59323c6fc8302dfd226b1b7e0cdef9c865231f72e930e9968fba2139565c3"
+    },
+    "training/source_blobs/fb194687b4e39d11d79d33deb9eb4cc7e7d09b5f06424814263d38b9232d2b03.py": {
+      "bytes": 13134,
+      "sha256": "fb194687b4e39d11d79d33deb9eb4cc7e7d09b5f06424814263d38b9232d2b03"
+    },
+    "training/source_blobs/fbe2b197e9f06a2916d7ef950cd2621aba2a8597faa58bc4b10789b4fef02761.py": {
+      "bytes": 18465,
+      "sha256": "fbe2b197e9f06a2916d7ef950cd2621aba2a8597faa58bc4b10789b4fef02761"
+    },
+    "training/source_blobs/fc20fa387d2f794b572397d684d42d2013d517f383142f3033bf91c9ced2ccae.py": {
+      "bytes": 35124,
+      "sha256": "fc20fa387d2f794b572397d684d42d2013d517f383142f3033bf91c9ced2ccae"
+    },
+    "training/source_blobs/fc74fbb0dc95c48b5f68c802efa4b1be36e59b6456c384f30a54f74a379381c6.py": {
+      "bytes": 192,
+      "sha256": "fc74fbb0dc95c48b5f68c802efa4b1be36e59b6456c384f30a54f74a379381c6"
+    },
+    "training/source_blobs/fc9d9862dd1f7f7c155ff62b4023a803db6752d21a43b7af6931041264aee00e.py": {
+      "bytes": 5584,
+      "sha256": "fc9d9862dd1f7f7c155ff62b4023a803db6752d21a43b7af6931041264aee00e"
+    },
+    "training/source_blobs/fe42694806dede2847762153e12eb7b84ff461f47959c5b39189c4893df5863e.py": {
+      "bytes": 5459,
+      "sha256": "fe42694806dede2847762153e12eb7b84ff461f47959c5b39189c4893df5863e"
+    },
+    "training/source_blobs/fef4e145fd16e2ec890b3113687c5cccc5ead04e3add2a65382f724cc911bc88.py": {
+      "bytes": 4340,
+      "sha256": "fef4e145fd16e2ec890b3113687c5cccc5ead04e3add2a65382f724cc911bc88"
+    },
+    "training/source_blobs/ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957.py": {
+      "bytes": 1924,
+      "sha256": "ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957"
+    },
+    "training/source_blobs/ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6.py": {
+      "bytes": 108207,
+      "sha256": "ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6"
+    },
+    "training/source_blobs/ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7.py": {
+      "bytes": 10504,
+      "sha256": "ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7"
+    },
+    "training/source_blobs/ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba.py": {
+      "bytes": 14327,
+      "sha256": "ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba"
+    }
+  },
+  "excluded": [
+    "FILE_MANIFEST.json",
+    "SHA256SUMS",
+    ".git/",
+    "__pycache__/"
+  ]
+}

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

PACKAGING_VERIFICATION.json

@@ -0,0 +1,33 @@
+{
+  "date": "2026-11-09",
+  "model": "OpenSML/Solstice-1",
+  "parameter_label": "175M",
+  "parameters": 174202432,
+  "status": "passed",
+  "weight_sha256": "7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515",
+  "environment": {
+    "python": "3.11.13",
+    "macos": "27.0.1",
+    "machine": "arm64",
+    "mlx": "0.32.3",
+    "numpy": "2.4.2",
+    "tokenizers": "0.22.2"
+  },
+  "clean_environment": true,
+  "external_working_directory": true,
+  "inference_dependencies_from_training_checkout": false,
+  "saved_responses_matched": 27,
+  "saved_responses_checked": 27,
+  "checks": [
+    "Tokenizer round trips and literal structural-token isolation",
+    "Explicit chat token boundaries",
+    "Invalid roles, empty messages and turn orders rejected",
+    "Output budget bounds and context overflow rejected",
+    "Non-finite and out-of-range temperatures rejected",
+    "Fixed-seed sampling is repeatable in this runtime"
+  ],
+  "full_public_suite_rerun": false,
+  "public_suite_evidence": "Previously completed full public suite on the identical FP32 tensor file.",
+  "details": "results/package_inference_verification.json",
+  "training_reproduction": "Not rerun. Historical sources/configs are an audit archive with documented external prerequisites."
+}

README.md

@@ -1,3 +1,128 @@
 ---
+language:
+- en
 license: apache-2.0
+pipeline_tag: text-generation
+library_name: mlx
+tags:
+- small-language-model
+- from-scratch
+- text-generation
+- conversational
+- apple-silicon
+- research
+datasets:
+- HuggingFaceTB/smollm-corpus
+- HuggingFaceTB/dclm-edu
+- HuggingFaceFW/finewiki
+- hongliu9903/stack_edu_python
+- hongliu9903/stack_edu_sql
+- iNeil77/SynSQL-2.5M
+- HuggingFaceTB/smol-smoltalk
+- HuggingFaceTB/smoltalk
+- rajpurkar/squad_v2
+- allenai/sciq
 ---
+
+# Solstice-1
+
+**An OpenSML 175M-parameter English chat research model trained from scratch on Apple Silicon.**
+
+Author: William Zebrowski · License: Apache-2.0 · Status: research preview
+
+[Model weights](https://huggingface.co/OpenSML/Solstice-1) · [GitHub source](https://github.com/williamzebrowskI/Solstice-1) · [Run guide](RUN_GUIDE.md) · [Technical report](docs/TECHNICAL_REPORT.md) · [Evaluation](docs/EVALUATION.md) · [Training data](docs/DATA.md) · [Results](results/public/summary.json)
+
+Solstice-1 starts a model family under OpenSML. This first release is a 175M chat model with a 2,048-token context. Its selected weights come from **42.001B pretraining tokens and three adapter fine-tuning stages**. The package provides merged FP32 weights, the original tokenizer, a native MLX runner, training-source snapshots, and recorded evaluation evidence.
+
+The model is intended for research into small English language models, short conversations, reading supplied text, and instruction following. It can produce incorrect answers, repeat itself, decline answerable questions, and invent information that was not provided. The measurements below do not establish dependable factual or production behavior.
+
+## Model specification
+
+| Setting | Value |
+| --- | --- |
+| Public size designation | 175M |
+| Exact parameter count | 174,202,432 |
+| Transformer | 44 layers; width 576; SwiGLU intermediate width 1,536 |
+| Attention | 9 query heads / 3 KV heads; head dimension 64 |
+| Components | RoPE 10,000; RMSNorm; Q/K normalization; tied embeddings |
+| Vocabulary / shared context | 32,000 byte-level BPE tokens / 2,048 tokens |
+| Pretraining exposure | 42,001,367,040 processed token positions |
+| Released weights | Unquantized merged FP32 safetensors |
+| Supported inference | Supplied native MLX runner on Apple Silicon |
+
+This is a different architecture and lineage from OpenSML-150M and OpenSML-JSONLite-175M. Those releases are comparison references, not weight ancestors. The frozen tokenizer is reused from the earlier OpenSML project.
+
+## Run locally
+
+Download the entire model repository, including its tokenizer and native source. While the repository is private, run `hf auth login` after installing `huggingface_hub`, using an account with access to OpenSML/Solstice-1:
+
+```bash
+python3.11 -m venv .venv
+source .venv/bin/activate
+python -m pip install huggingface_hub
+hf download OpenSML/Solstice-1 --local-dir Solstice-1
+cd Solstice-1
+python -m pip install -r requirements.txt
+python inference.py --prompt "What is the main function of the heart?" --max-new-tokens 96
+python inference.py --interactive
+```
+
+The runner uses the model's native structural-token chat format. It checks the weight hash and tokenizer before loading, defaults to greedy FP32 inference, and rejects context overflow without truncation. `/reset` clears an interactive conversation; `/quit` exits. See [recorded package verification](PACKAGING_VERIFICATION.json) for the environment and checks actually completed.
+
+**Use the supplied runner.** This package does not provide a verified Transformers AutoModel, MLX-LM auto-loader, GGUF, or CUDA implementation. The context limit includes previous turns, role delimiters, the current prompt, and the requested output budget.
+
+## Measured results
+
+Full-split, zero-shot likelihood evaluations use native MLX FP32 scoring:
+
+| Benchmark | Examples | Accuracy | Normalized accuracy |
+| --- | --- | --- | --- |
+| ARC Easy | 2376 | 62.58% | 62.33% |
+| ARC Challenge | 1172 | 29.35% | 30.89% |
+| PIQA | 1838 | 66.49% | 66.21% |
+| HellaSwag | 10042 | 32.85% | 39.40% |
+
+`acc_norm` divides each candidate's summed log-likelihood by its Unicode-character count. It is distinct from raw accuracy. The evaluator follows pinned task conventions but is not an installed lm-evaluation-harness run.
+
+| IFEval metric | Score | Passes / checks |
+| --- | --- | --- |
+| Strict prompts | 17.93% | 97/541 |
+| Strict instructions | 30.10% | 251/834 |
+| Loose prompts | 19.22% | 104/541 |
+| Loose instructions | 32.49% | 271/834 |
+
+IFEval uses all 541 prompts / 834 instructions, greedy native-chat generation, no system message or repetition penalty, and a 1,024-token output cap. **130/541 responses reach that cap.** All responses are scored as produced. These numbers do not measure general factual accuracy.
+
+On a separate 192-response authored chat check, 107/180 automatically checked answers match the accepted answers; 191/192 responses stop naturally, and three have repetition flags. Automatic answer matching is conservative, and repetition detection is heuristic. These checks are not a broad human preference study.
+
+![Recorded pretraining history](docs/figures/pretraining-overview.png)
+
+See the [evaluation report](docs/EVALUATION.md) for baseline comparisons, dataset pins, protocol differences, and private-test scope.
+
+## Training lineage
+
+```text
+Fresh Deep44 initialization → streamed prose pretraining
+  → add Python/SQL at 36.801B → 40.001B → 2.000B-token cooldown
+  → 42.001B pretrained base
+  → adapter SFT 1, selected update 2000
+  → adapter SFT 2, selected update 256
+  → adapter SFT 3, selected update 128 → Solstice-1
+```
+
+Only updates through 128 of the final 256-update experiment are inherited. Earlier experimental branches and checkpoint 167 are not additional stages in this lineage. [Checkpoint lineage and hashes](provenance/LINEAGE.json) · [Selected data exposure](provenance/SELECTED_PREFIX.json).
+
+## Known limitations
+
+- The model sometimes gives fluent but incorrect factual explanations.
+- On the new authored evaluation, 111/200 answerable extraction questions triggered unnecessary-abstention signals. On a separate set with different wording, it failed all 15 missing-information cases. These are distinct tests, not contradictory measures of the same inputs.
+- Instruction following remains incomplete; only 97/541 IFEval prompts satisfy every strict requirement.
+- Repetition and token-limit stops remain possible, especially on longer requested outputs.
+- JSON extraction, tool use, robust coding, long documents, multilingual use, safety behavior, and prompt-injection resistance are not validated release capabilities.
+- Public and historical benchmarks informed earlier experiments. The newly reserved 1,000-case panel was evaluated after this finalist was selected, but shares authored task families with training. No comprehensive semantic contamination audit or independent human evaluation is claimed.
+
+## Repository contents and licensing
+
+The weights and author-controlled code are released under [Apache-2.0](LICENSE). Upstream datasets and third-party code retain their own terms; the model license does not relicense those materials. [Source notice](SOURCE_NOTICE.md) · [Third-party notices](THIRD_PARTY_NOTICES.md).
+
+Technical details, data provenance, exact selected training prefixes, source snapshots, and reproduction prerequisites are included. Full raw corpora, optimizer binaries, and all ancestor weights are not distributed. The historical training archive is not a verified one-command retraining package. [Training and reproducibility](TRAINING_AND_REPRODUCIBILITY.md).

REPOSITORIES.json

@@ -0,0 +1,21 @@
+{
+  "model": "Solstice-1",
+  "organization": "OpenSML",
+  "parameter_label": "175M",
+  "license": "Apache-2.0",
+  "release_status": "private review",
+  "repositories": {
+    "github": {
+      "url": "https://github.com/williamzebrowskI/Solstice-1",
+      "visibility": "private",
+      "role": "Native inference code, documentation, training provenance, evaluation evidence"
+    },
+    "huggingface": {
+      "url": "https://huggingface.co/OpenSML/Solstice-1",
+      "visibility": "private",
+      "role": "Complete runnable package, including merged FP32 weights and tokenizer"
+    }
+  },
+  "weight_file": "model.safetensors",
+  "weight_sha256": "7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515"
+}

RUN_GUIDE.md

@@ -0,0 +1,51 @@
+# Run Solstice-1
+
+Use Python 3.11 on Apple Silicon macOS with the pinned requirements. The tested public MLX wheel is 0.32.3; it requires a compatible macOS version. See PACKAGING_VERIFICATION.json for the actual tested system. Inference is offline after downloading the package and installing dependencies.
+
+While the model repository is private, run `hf auth login` after installing `huggingface_hub` below. Use a Hugging Face account with read access to OpenSML/Solstice-1. Tokens belong in the terminal sign-in prompt, never in a shared command, notebook, or repository file.
+
+```bash
+python3.11 -m venv .venv
+source .venv/bin/activate
+python -m pip install huggingface_hub
+hf download OpenSML/Solstice-1 --local-dir Solstice-1
+cd Solstice-1
+python -m pip install -r requirements.txt
+python scripts/verify_package.py
+python inference.py --prompt "Name the largest continent." --max-new-tokens 64
+python inference.py --interactive
+```
+
+For the GitHub source checkout, download the matching `model.safetensors` from the Hugging Face repository into its root. That large weight file is deliberately excluded from Git history.
+
+## Previous messages
+
+Interactive mode retains complete previous turns. `/reset` starts a new conversation and `/quit` exits. The program never silently truncates your history. When history plus output budget would exceed 2,048 tokens, reset it or shorten the request.
+
+For scripted conversations, create a JSON file containing message objects:
+
+```json
+[
+  {"role":"user","content":"My name is Rowan."},
+  {"role":"assistant","content":"Hello, Rowan."},
+  {"role":"user","content":"What is my name?"}
+]
+```
+
+```bash
+python inference.py --messages-file conversation.json --max-new-tokens 64 --json
+```
+
+That file describes the input conversation; it does not ask the model to produce JSON. An optional system message is accepted only as the first message. No system message is inserted by default.
+
+## Generation settings and outputs
+
+Greedy temperature zero is the default and matches benchmark decoding. `--temperature 0.6 --seed 2026` enables sampling; sampled outputs are not the reported benchmark protocol. The supplied runner uses no repetition penalty, output repair, external model, retrieval, or constrained decoder. `--json` wraps the unmodified generated answer with token IDs, lengths, timing and stop reason for inspection.
+
+Stop reasons are `turn_end`, `document_end`, `token_limit`, or `unexpected_structure`. A natural stop says the model completed a turn; it does not establish answer correctness. Generation defaults to 128 new tokens and can be changed within the context budget. In the native playground, additional controls and history trimming may differ from this standalone CLI; use the CLI's recorded protocol for reproducibility.
+
+## Verification and troubleshooting
+
+Run `python scripts/verify_package.py` to check the recorded file hashes. The runner independently verifies its weight checksum and tokenizer manifest. A missing or mismatched weight file is rejected. Unknown role sequences and context overflow are rejected before generation.
+
+Use the supplied `inference.py`; Transformers AutoModel, MLX-LM automatic loading and GGUF conversion are not supported by this release. Do not repair a hash mismatch by editing the expected hash. Redownload the intended package instead.

SHA256SUMS

@@ -0,0 +1,466 @@
+70a6a34e99e8ec921af10056c133b8035f937a7a5e43131c06f8817ea04222e0  .gitignore
+f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859  BACKUP_AND_CLEANUP.md
+af646a190d135ec25a6ee2c76d1594c3f9ba8005cd3ca6df908606a15023273e  CITATION.cff
+cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  LICENSE
+0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375  PACKAGING_VERIFICATION.json
+2bcb95ff85d71d6f2c9fa6711bcbad62c52d68ab3367a8317db72b6d88bb04fd  README.md
+152a87d0098752350d54fdd3cde3861a77eb497c53dc973ee1a8370d12a13ff6  REPOSITORIES.json
+fd05e1595bcdc52942e9425b6efb2c7e89553e6b0ad7d5f93d9075969cd75dba  RUN_GUIDE.md
+87fa33719fe1182d467a05ab417be2a362f50eea59f0ac6496d40e677544e27c  SOURCE_NOTICE.md
+85525fd4cfd74dbabeb8fe767074cf02a8834dc40a3fc2e648180ef2a3ff35d9  THIRD_PARTY_NOTICES.md
+a4bbaa7901cce3f88cdcac1ffd30ecdf60d4999c198054dce6bb7ef1d72607b2  TRAINING_AND_REPRODUCIBILITY.md
+c98126fd2051388690f7d6ad709f961afbb654f0268e4fd4db07e0ead915ee46  config.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261022_121457_b3d5f398/corpus.json
+ddf14c4cef8a4fb66392a27e87f8c5a807a88da35fe61599f252c9eba91bec29  configs/pretrain_20261022_121457_b3d5f398/recipe.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261024_111140_5f596b7a/corpus.json
+900c630d3a315d63ed332a9e300a2d1a86f9a2dd8e9627a15ad6430c2077162e  configs/pretrain_20261024_111140_5f596b7a/recipe.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261027_222615_4d065da7/corpus.json
+a81a72c7bfe47b4b9e47ce0b63bb7c6b5a139ddba905c5d2a017bd87f249a0d6  configs/pretrain_20261027_222615_4d065da7/recipe.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261028_093017_dd42e713/corpus.json
+0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1  configs/pretrain_20261028_093017_dd42e713/recipe.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261102_234538_0ee851dc/corpus.json
+0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1  configs/pretrain_20261102_234538_0ee851dc/recipe.json
+c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632  configs/pretrain_20261103_000432_8ac2605e/corpus.json
+0c384bc3e0e802a8a749883d9e60bed74e02d55501cefa29eaacbc65f88a8da1  configs/pretrain_20261103_000432_8ac2605e/recipe.json
+f27196edf4eab3270c3ec9a8e232509fcc4435baa68c0d976acdf6b5881a14cb  configs/pretrain_20261104_170916_448e6f64/corpus.json
+218451ccbe74415b0e4b2e810200feaacf1e17bb582c0a9f1f84e381d6bf9ad9  configs/pretrain_20261104_170916_448e6f64/recipe.json
+f27196edf4eab3270c3ec9a8e232509fcc4435baa68c0d976acdf6b5881a14cb  configs/pretrain_20261107_142423_40700422/corpus.json
+c70c4491d8865d0db8af21d555c309dcfcbaa346760f0668e82a8c0ce54c8d6f  configs/pretrain_20261107_142423_40700422/recipe.json
+fb204be1d9e72eb1cae140bfdb03dbcadc1843c47b1894a731396ceeedb9a22f  configs/sft1/plan.json
+51d66108416988bad1015d7b52cf63db0571f87f202b2b18c7e98e67a29ec62b  configs/sft2/plan.json
+ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284  configs/sft3/plan.json
+e0b0c730321360c40f20e53a22890c52cd5294bad2a5777a0a891096557a2f0a  docs/DATA.md
+5522817f8d698f05b43adfc1c5227bab4e917492c898fa4b792d521d42a512d8  docs/EVALUATION.md
+7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3  docs/EXAMPLES.md
+f98d45b39b9595510fa80d495a3758b211d8c770adc739d381763d3cfb2a2a0b  docs/TECHNICAL_REPORT.md
+0fa840653d1a5305297ef6f51ba08dc70a4bbfaf8029c411627e39dfc3cdc153  docs/figures/cooldown-detail.png
+1071d21211f7d32d9f9f5874e62f46a2a17306f88f0f355b24aee5b4af4c4e0f  docs/figures/cooldown-detail.svg
+c46bba5d4fae1c55701def21829de08fd3756dc9659fee2a54a79f42eb1afcb0  docs/figures/pretraining-overview.png
+862a5400ecde1ececb915441766f378849c65024288522c84a4883d0c4b31a4e  docs/figures/pretraining-overview.svg
+9e15b1b93bbc1160f3988a7956eee88b2467e5d826f097df35401562c116b90d  evaluation/README.md
+f3835c8cb0bdf580e4142a57ac3724c474dcc1d44aa01dbbd805c989a587c7c7  evaluation/benchmark_best.py
+10ae9d2c49236c377d6e9952674b14289f5f92c310c29cbe592a40c98a14c752  evaluation/benchmark_ifeval.py
+48dcb288740c2451e6d4218684809a201c078d517bfbac94a3adec61c8d10c40  evaluation/chat_adapter_search_bench.py
+66883e6b6589b81927fdc331b25f38915e59422cdf4b8339c715cd85d74e2b1d  evaluation/chat_benchmark/__init__.py
+70d78840147614727a34a775f669f9316c8d589721d41fdf2dd8c98deb2475bd  evaluation/chat_benchmark/panel.py
+e63a8715c302e2b59b0fadc3598e912d3f695fb9352b541e33dba3b046017c54  evaluation/chat_benchmark/scoring.py
+cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  evaluation/vendor/ifeval/LICENSE
+bc7bead98aea2499b1226aca21021d267a9e926d36711eb9e2230ebb0841abfd  evaluation/vendor/ifeval/NOTICE
+8f77aceef943735bec21b7f42630794cad7ce20d32b72a83889e20274fb6df31  evaluation/vendor/ifeval/instruction_following_eval/__init__.py
+35decc06000718487f44d7deafa6d3f48a8ec0886281edf40162c0265b7d248c  evaluation/vendor/ifeval/instruction_following_eval/evaluation_lib.py
+60e086f5342a03ce8e18b64bbcccf86308f523c08aa826707a562150a52f3edf  evaluation/vendor/ifeval/instruction_following_eval/instructions.py
+ec92d72c264f6d906978613085db262356174300370a3fffe6fefd5969ce9cfc  evaluation/vendor/ifeval/instruction_following_eval/instructions_registry.py
+a73797261eee5bf447e279d82a2b700b1bdd3cb1193412dbab1270a85832bc6b  evaluation/vendor/ifeval/instruction_following_eval/instructions_util.py
+5e3542d6369138730a11db9b7793fe1a722021b66ea8ad32b2e93d49ade91555  evaluation/vendor/ifeval/prepared_runtime.json
+d1ccd2965b9b8752d51b1e814faa5c4879e0c630047fb01c4fa50dca75527d70  evaluation/vendor/ifeval/source.json
+e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410  inference.py
+0628a9b64e3e12824a303a720a8456b2ea4341bf4b45a12aea25cbc71b68deaf  inference_native/__init__.py
+5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8  inference_native/chat_format.py
+d9bdeeb4b8ea60dacb9af1cafa1dfbd295bfb384c2070d282848ec7e5eb42168  inference_native/common.py
+1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe  inference_native/model.py
+254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5  inference_native/tokenization.py
+7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515  model.safetensors
+1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f  provenance/INFERENCE_EXPORT.json
+490082ad7c96b776738e4d6bb7f11507af0c697facd3a7b95d944abd64e25b3a  provenance/LINEAGE.json
+0db456c4bf1a5e46f831195346626b73d953d412b23c47bc6b27eaa230babf42  provenance/PRETRAINING_PHASES.json
+d8aaf1ef62597ac384b6f7c9510b24c2bbb63470150c3835266203ee6a567fd8  provenance/SELECTED_PREFIX.json
+834edc532edfe9a897b60a68046006db6700bdd21de72f055199eb827aef401a  provenance/TENSOR_INVENTORY.json
+f1adf7d2c3a91ee6a4643beb26ce8d9cdfbdb69156e96439d01be657abf4ad20  provenance/checkpoint_metadata.json
+f1f8519ca75a35343cfec7e88c8197bb84aa7d97d809353861b97c1cf3d77513  provenance/pretrained_base_manifest.json
+8bc81bacb9ef74be0b3c0118e4d60f0d00ace0d17139040be8454b873f2cc448  provenance/pretrained_base_metadata.json
+f9273450799963b2fd8e61a5e9cb6892be7ba4a2308849623e288302a687a1f7  provenance/sft1/checkpoint_manifest.json
+df241d47b313c08bd5f5d11436bcd05f8f3b53c0c6c7bdc065c071ee0c1dbf67  provenance/sft1/contract.json
+d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a  provenance/sft1/corpus_manifest.json
+c6e0e82a2b4f80aa16b803cac7149e7984ad66a076d0105331cca08134a4fdd8  provenance/sft1/state.json
+dcab4ad8e4484ad1cc4b32ffbc6352b6a6ff9ea0a4ddccdb8d367d23db098a3f  provenance/sft2/checkpoint_manifest.json
+c7d30352bfe3bae2d52937c257d9ab48fccddd90054e4f8b83ee2f78f7d55d87  provenance/sft2/contract.json
+d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a  provenance/sft2/corpus_manifest.json
+c11c2bcdbd109e60178da47b9ac830734c5a15557f96c33ae7405009a3159265  provenance/sft2/fresh_manifest.json
+904bcd42c0497f4b0e0b99fdb8dbd5f2d6155d2dd9e9cab31d9c5021d0b7a4ec  provenance/sft2/selected_indices.json
+de16e15cfd6687ddc0480dd113f1f01b6900a17737c99b7a4c46863a763df194  provenance/sft2/state.json
+2118c7446c412a4ee7a4b7a356a211ad91574983e213da45980707b5fa1ca124  provenance/sft3/checkpoint_manifest.json
+5776cf85c123d0560a40d4ebd255c10aae8ac1817006d8cf09f405731604ceae  provenance/sft3/contract.json
+d5485bae1be8d662b7fc796cd6ca50e69f288dae828d017c52a4027ad48c3c9a  provenance/sft3/corpus_manifest.json
+741da3f1897cb5b092cdcab91051630b07262c285933efab3cc00a844a6c3d01  provenance/sft3/fresh_manifest.json
+eee29c2b37c883cb6df6387ee1c734bb1a7011d2106fe97d7351a42d174a2531  provenance/sft3/selected_indices.json
+78371afe2868d7e6278f4bf3a1de4c2033ef1673e6c590c3c1c6b671b8b56cc9  provenance/sft3/state.json
+cfeac63c4d14d73a7b247bf9e39fd0b4642754aa4b69e7ea6cf94e79a594e3d8  requirements.txt
+28fdc175a0019e6a34590256c3c1da8fc3a39813699afae8a477d63b00249a7b  results/FINAL_COMPARISON.json
+f76237736d1aab8f04517ad9e3336925ca9773399a9ad888f07c15d3ceaa4f87  results/baselines/167/choices/results.json
+187b395945254a4d9359694d5f9ead19a8b95caffcb42477cb9a349cf6ad625e  results/baselines/167/ifeval/results.json
+a5d49e3dc60a6c4dda4799090881cc2aab033b922ac0b5dafd52c0c6c7e2e6d8  results/baselines/attempt5.json
+2b59f728ebc9bef6deaf378e6908f7b55723b368f60da83f9af5b2b1731d0346  results/baselines/parent2000.json
+fed55313b61008e8f379a7b4bc49535cc501a31a942c7560687de4a4bd1cd47f  results/package_inference_verification.json
+ba2cfb7c8b446dede8e7f33bb78885855e74bd75f861b0c201d6e01e88bc987f  results/pretraining_steps.csv
+c4e3210780814051e3ad1bd99a9c143b9c5dbe910224239a134fddc546e9c9a8  results/pretraining_validation.csv
+06f43fbe7a4f38dbafb7d82e778d25fe45c009b5939ee9c81919122758d02000  results/public/chat/protocol.json
+fea56aceda29f7e78a9f8ce730c050c89b5e856753bb5609e97d3a2cafce159e  results/public/chat/responses.jsonl
+784d4efc1d63d5b35a201f44525144755e586f27ad812085ee563a4ec2396eaf  results/public/chat/results.json
+d467b86959d28256f44af33d3f3dfe877ea5849ac7d103eaa9c3c348cb1af180  results/public/choices/arc_challenge.jsonl
+92bead9bc0361159c46d047ef6ed04cf93f59a36e38bab0d6eb638f89d995388  results/public/choices/arc_easy.jsonl
+b7ed4f70e9be865b1288e9df0f2f10e5d79ac393c0b6887e672b134d72819067  results/public/choices/hellaswag.jsonl
+a095ec405560fbf0649664058b3705a850475981eafe8c82efbba377b284b92e  results/public/choices/piqa.jsonl
+3321b5b1cc771c69c06ec01d7ea62d1aeb78764f455e2f514b23c73d44ab0938  results/public/choices/results.json
+501160470cb0f9f3f06ed853d1107cb819ab8f8295220cb2ccae0f065b267bdb  results/public/ifeval/inputs.sha256.json
+40e0e8f2ded6b9f95893aff33697252752f574f1aac090e01136a3784026ee83  results/public/ifeval/loose.jsonl
+ba051d49356da173c78f5079eb6450acbfe510ca0215625ed32050441c7a2aff  results/public/ifeval/responses.jsonl
+859d5c8359f15179d021fc8c1676483efd7ea5320dd336105f24608077e24668  results/public/ifeval/results.json
+fdc6f6e6962f76f5603670b6dff4f3ea7c28e6cd5a085f83ceddc3f09196065a  results/public/ifeval/strict.jsonl
+b9a155a97a22e09027507737c5a4e9afe2a312fdce39c95df11a3022756e1b73  results/public/summary.json
+8d07aeaecc61f0acbcf4d18d413472e8440561634b15df95c3a3444dc433f376  results/search/trial_01.json
+4d4271ce322e29ed1fcdfb64e4631dfd358b0a43f907200a1f2fa41652dc0770  results/search/trial_02.json
+6dc15f24abbf6b24528178be803d97c2885939835558f353ff1bf34427f524cd  results/search/trial_03.json
+be38e4323b1b120fa92717e2439c3d231675e70d860a4673d296e202e62a2342  results/search/trial_04.json
+da83222ee38d33eea6c7062c28fd310e61da89a9a0315dc62559c4a9181759d9  results/search/trial_05.json
+8ba998638ee42e06e5fb6d5d26e78ec51be5ce5b8565889d5880457cca614c6e  results/selection/batch_order_audit.json
+28fdc175a0019e6a34590256c3c1da8fc3a39813699afae8a477d63b00249a7b  results/selection/comparison.json
+6f4978a0a95ad88e089de1eda6026a4a1f61d3d03f398ac1234a045f4b150ec6  results/selection/data_audit.json
+a478bc790345129fa960060490d4c2148351ccfc17ac433ec06364969787a8fe  results/selection/design.json
+0567b916bdf27de7f6eafe8c61da7fb3686a69e08361ac6cb5645ac567cf8f69  results/selection/selection_frozen.json
+cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a  scripts/materialize_training_snapshot.py
+c47f496b10f813dcb61697a0d49ed14043137ee1e2a96911a68fccb31fa9df26  scripts/plot_training.py
+002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3  scripts/verify_inference.py
+ea7d6388ad1f17fcf244119b2bf15018d981c4e2c127bedf2af73dc74eaa0528  scripts/verify_package.py
+e4c47b7b6aba08f948de7c2c7cb94c58dbd09dde2ded6ac185c60856ee686d27  tokenizer/corpus.json
+1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb  tokenizer/manifest.json
+e871f690eabebf82da173f0e406f38b79a10cd7d2647bb3592c3fc35ae9a2f16  tokenizer/report.json
+e095f0413a11fa91fbb58f37d00e4c6bcda59beb89b28f6f680db1436a7e8891  tokenizer/tokenizer.json
+619f500553367eb76f5be00a98625130ec857b9c367561334984096c029ba2a9  tokenizer/tokenizer_config.json
+1d16aa9630669fc9b5d03eea415c7c2a9b7423d048b3ba0ebe75d62193cb3066  training/SNAPSHOTS.json
+71f9d16d8a05d4e944ef0863ec52c673e1342cb5378ddfec57a3f7cc7962772c  training/launchers/SOURCE.json
+27544f082d142bb4ee436f3a527667bd3c68559d1f70c67114a2ac16dccca2bc  training/launchers/five_mac_cluster.py
+f289117941679701cb4e29dc79c3dc859f5739df416fcbc2447441d607d00695  training/launchers/launch_adapter_trial_09_replica_20261108.py
+5f74006eae3c6b34cd1901b634dda1f6b8712ab740ef118bd203214767124e76  training/launchers/launch_code_pretraining.py
+7887737805af333126e2755fc04b6e16b1ced7e1e7c8091e358867d3d6176a21  training/launchers/launch_deep44_cooldown.py
+b70ea382d073a0a2ef37c1acb4b5925ea1fe4ef509e0cfdb31282ab054bf5247  training/launchers/launch_research.py
+770dc6f9bb9c60114838bfda93582fb917cd93754a3af33757580f75395031bf  training/launchers/launch_stable_pretraining.py
+58f99d20ee7d37bde6e0ee260f6a31c39ac76296771fa3afcf466a311dfd2327  training/launchers/preflight_chat_repair_search.py
+6ba5201968f384b1823e878ff0b3b173e0977916508891ee33e6065dfd2bc98b  training/launchers/reproduce_parent2000_attempt5.py
+4e4c405eb89b9de3e3d1eecec682ab225bf4288768537ba766bad1d5e4ddfbbf  training/launchers/run_chat_repair_search.py
+94ffce3491a40e8f4c70adfad0a564c7cfa0cf77a4eba6348378a2adf7b9512a  training/snapshots/pretrain_20261022_121457_b3d5f398.json
+d5dd508491a9dd6151c972353ecfe898aa5fafbe9056f152884daa41995b04ee  training/snapshots/pretrain_20261024_111140_5f596b7a.json
+d2c70ebe2b052eafab538244460084263f13f3b4d455eb8cd76a2e92836e47f7  training/snapshots/pretrain_20261027_222615_4d065da7.json
+cd7a2ad3988fd6ba53c0f291f39239625ab390eff0a890a6915ced6bfb145493  training/snapshots/pretrain_20261028_093017_dd42e713.json
+70f212b330a9e25d245629ebc9d1909537fae8c8639bfb65dda17a0f17e4e39c  training/snapshots/pretrain_20261102_234538_0ee851dc.json
+d47efd77e209ad2ca131064c5e943519d63564bee55a2b866f8c6a8ea74676cb  training/snapshots/pretrain_20261103_000432_8ac2605e.json
+45adf703c6004b8269a652775d04328af7811783a5211e9d151bdc2885b0d6c5  training/snapshots/pretrain_20261104_170916_448e6f64.json
+227ca4a97a3581ce39781551ba30558178c1d7b0358a54b923d23dd5d6f9aee3  training/snapshots/pretrain_20261107_142423_40700422.json
+95b573e73247cb8628179dc010c34f73527500e702c7b5e042fe883a7be702cd  training/snapshots/sft1.json
+ffda31c2dc6d0afe2112ffaf8dd3d28f6dab40119eb4360203103c00658e85f1  training/snapshots/sft2.json
+dfbd41d4b60ae84bfc5fe32a0051a244e3489b7c2f24cea333d362aba98f79cc  training/snapshots/sft3.json
+0132642f0e4338c1f38a72a74407430afbd0d269c02e458f83df2366baeb85d6  training/source_blobs/0132642f0e4338c1f38a72a74407430afbd0d269c02e458f83df2366baeb85d6.py
+01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b  training/source_blobs/01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b.py
+01e7b0b2f162a9cdc242aecf6dbb7c22806365308973b4faf2c90d86a796b9e9  training/source_blobs/01e7b0b2f162a9cdc242aecf6dbb7c22806365308973b4faf2c90d86a796b9e9.py
+023fda52d696e3b8349afb663e085a467f10f3d3ba869e4a85ac2c4b9efe8c16  training/source_blobs/023fda52d696e3b8349afb663e085a467f10f3d3ba869e4a85ac2c4b9efe8c16.py
+0348ad7d8aeb3e6749be7ce4df4a5b3029d9db869de9a334f2c5c6058bedb711  training/source_blobs/0348ad7d8aeb3e6749be7ce4df4a5b3029d9db869de9a334f2c5c6058bedb711.py
+03bceb2e4c8a3ca7c63ea11d31b5e89dece07cbb326b23e53ad8d6972e1214fd  training/source_blobs/03bceb2e4c8a3ca7c63ea11d31b5e89dece07cbb326b23e53ad8d6972e1214fd.py
+04945e1ff569c5cd580028d77db3d8f35d6fce9817f6aa52e13768c26e978d62  training/source_blobs/04945e1ff569c5cd580028d77db3d8f35d6fce9817f6aa52e13768c26e978d62.py
+04c8c482003ed8ffa1f2238337f351ec2748e2ef5ae60b608e6d6acdeda5f516  training/source_blobs/04c8c482003ed8ffa1f2238337f351ec2748e2ef5ae60b608e6d6acdeda5f516.py
+0548da651db528075ca56db9570e19812f7824da024d7f6350e585d1058d166a  training/source_blobs/0548da651db528075ca56db9570e19812f7824da024d7f6350e585d1058d166a.py
+064b0370ae2b6c0f6a6fa7c684ded54bd97b4e39958aebb0c2aff7dac1f9b9e4  training/source_blobs/064b0370ae2b6c0f6a6fa7c684ded54bd97b4e39958aebb0c2aff7dac1f9b9e4.py
+06c96307cb6749a4c9f169de9a4e049fe07ec8f753aff30344a0aa00921e5530  training/source_blobs/06c96307cb6749a4c9f169de9a4e049fe07ec8f753aff30344a0aa00921e5530.py
+071b6416f3f0785dde7c5a06e90afe405fc0885a39853c2f2027d670f060556e  training/source_blobs/071b6416f3f0785dde7c5a06e90afe405fc0885a39853c2f2027d670f060556e.py
+072851f20850e4548c0b3a48e0a2e3ceab961d7b956e76a8944ed409cd460898  training/source_blobs/072851f20850e4548c0b3a48e0a2e3ceab961d7b956e76a8944ed409cd460898.py
+072fed811b8a3d13878ab1283991246877a12e7755fbca5a0c88bf2ed95bd382  training/source_blobs/072fed811b8a3d13878ab1283991246877a12e7755fbca5a0c88bf2ed95bd382.py
+0769ac3b1bec8cacf5a24c63b3049271d0907e20bba94a2475bbcfdeac329ae9  training/source_blobs/0769ac3b1bec8cacf5a24c63b3049271d0907e20bba94a2475bbcfdeac329ae9.py
+07751cc81cf60a44c571c0a5ac8194690a6805211eaab10236590849b7c3b90f  training/source_blobs/07751cc81cf60a44c571c0a5ac8194690a6805211eaab10236590849b7c3b90f.py
+077ab6db2082770a4683b37912c0b43e34330101fec97aab1a1cac4bb4d7b2ed  training/source_blobs/077ab6db2082770a4683b37912c0b43e34330101fec97aab1a1cac4bb4d7b2ed.py
+0aaa7019ab53fd8b5e6b0e397f970026b4e8d38e12bf5d5f0fcb7f686de9b82c  training/source_blobs/0aaa7019ab53fd8b5e6b0e397f970026b4e8d38e12bf5d5f0fcb7f686de9b82c.py
+0bb4999f8a50190afebd71df014e9ea0e1db8bea44144d93f47c34b63f646c60  training/source_blobs/0bb4999f8a50190afebd71df014e9ea0e1db8bea44144d93f47c34b63f646c60.py
+0bd916e6d5eafb4be3d5478a4a551cf1b41e9f6dbe11d86c3928dac1fe83acc5  training/source_blobs/0bd916e6d5eafb4be3d5478a4a551cf1b41e9f6dbe11d86c3928dac1fe83acc5.py
+0ca74fa61e1748e3d689826ef2c134d82c5d4cf7f47ca761f15963d586e8c942  training/source_blobs/0ca74fa61e1748e3d689826ef2c134d82c5d4cf7f47ca761f15963d586e8c942.py
+0d0a8d68c9278f292c5e8c59ce8f3e25ea585ae18d07dfe7e6c7e2fac6449f46  training/source_blobs/0d0a8d68c9278f292c5e8c59ce8f3e25ea585ae18d07dfe7e6c7e2fac6449f46.py
+0eafba0490c4db294eb44057a200a7cffaed51146b102845d03095973544bf5c  training/source_blobs/0eafba0490c4db294eb44057a200a7cffaed51146b102845d03095973544bf5c.py
+0f1d2108827df13fc4420c9ab05f2fc22ce33c85d2fc54008e955c79fe7ddbc4  training/source_blobs/0f1d2108827df13fc4420c9ab05f2fc22ce33c85d2fc54008e955c79fe7ddbc4.py
+0f452657e4e5a32f6df03d696874c3a959b9e6935ccc6e0ff86c329207fda7db  training/source_blobs/0f452657e4e5a32f6df03d696874c3a959b9e6935ccc6e0ff86c329207fda7db.py
+1012b31fa62d5b56c959b69cb2d63dd9996e7518541842bdb119699ac00ed2c4  training/source_blobs/1012b31fa62d5b56c959b69cb2d63dd9996e7518541842bdb119699ac00ed2c4.py
+10b5f6451f238387ba7b0342b380453808236f63824ff0ead4e22df3083662c5  training/source_blobs/10b5f6451f238387ba7b0342b380453808236f63824ff0ead4e22df3083662c5.py
+10ca8b5430b7f97d53ea00e50ed9dab53e62208247d310105e3e7671b0198455  training/source_blobs/10ca8b5430b7f97d53ea00e50ed9dab53e62208247d310105e3e7671b0198455.py
+112ccdbbd682d661902d606be4bfc9d6fa41ede284fe48855e704688ca6cc88d  training/source_blobs/112ccdbbd682d661902d606be4bfc9d6fa41ede284fe48855e704688ca6cc88d.py
+116bf40b02262743bb5ddac629c8d223537c0c12ce4631d32347a459844bf4aa  training/source_blobs/116bf40b02262743bb5ddac629c8d223537c0c12ce4631d32347a459844bf4aa.py
+1173c8c94dc11c6b8e446c26804a543b68e366986f01d249384082629dca369b  training/source_blobs/1173c8c94dc11c6b8e446c26804a543b68e366986f01d249384082629dca369b.py
+11766ff21cd6deda9fc4f1ad4ba45007222ebf7534a1c5639d6ad26df4359c03  training/source_blobs/11766ff21cd6deda9fc4f1ad4ba45007222ebf7534a1c5639d6ad26df4359c03.py
+14c0be0a65ff87803a56e1ec031308cdd99a96f6c2732fdf28055fab5fa0a4cc  training/source_blobs/14c0be0a65ff87803a56e1ec031308cdd99a96f6c2732fdf28055fab5fa0a4cc.py
+1829527deaeb7c2ccd8c8a7b2a05f7a7bb26cf159576981d9a52be54a6995901  training/source_blobs/1829527deaeb7c2ccd8c8a7b2a05f7a7bb26cf159576981d9a52be54a6995901.py
+197c0f0f54ac691a52a244425f96dc82d250a2373bb7612fccfc5188d2dd8cad  training/source_blobs/197c0f0f54ac691a52a244425f96dc82d250a2373bb7612fccfc5188d2dd8cad.py
+1ac442f888499a13fab376be1be819b12688ae5f9eac90b67f8800a6454ed08c  training/source_blobs/1ac442f888499a13fab376be1be819b12688ae5f9eac90b67f8800a6454ed08c.py
+1c032c459b232a7b426a05fca18bb6007fe7d8d56b73b291630f570b5575b23e  training/source_blobs/1c032c459b232a7b426a05fca18bb6007fe7d8d56b73b291630f570b5575b23e.py
+1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe  training/source_blobs/1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe.py
+1feb34473bfa8cf3e4b47eddbd767cccf2b3f1775e04f194d2b4556abe730e84  training/source_blobs/1feb34473bfa8cf3e4b47eddbd767cccf2b3f1775e04f194d2b4556abe730e84.py
+210e8f84fce6a32e60cc535296ea4a05d98456a47af977898f02017561a73eec  training/source_blobs/210e8f84fce6a32e60cc535296ea4a05d98456a47af977898f02017561a73eec.py
+24fb7b1b6ab2adb5575308c7d386687389627a0018b4eb730254dc4f9e708219  training/source_blobs/24fb7b1b6ab2adb5575308c7d386687389627a0018b4eb730254dc4f9e708219.py
+254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5  training/source_blobs/254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5.py
+25536570e721a16c61795744882e7868b2aad63f35f9b88c8ea4ba845d92cae1  training/source_blobs/25536570e721a16c61795744882e7868b2aad63f35f9b88c8ea4ba845d92cae1.py
+26d3ff514c47d34363982d6a2b6e02c0c28d53e0df7bc589281149effae859ea  training/source_blobs/26d3ff514c47d34363982d6a2b6e02c0c28d53e0df7bc589281149effae859ea.py
+27392bfdaa565c8150f217ac2d8835135a56d26da3056427545f6a2efd898f4d  training/source_blobs/27392bfdaa565c8150f217ac2d8835135a56d26da3056427545f6a2efd898f4d.py
+274c2760db9514c13e91ea3c78df9ff3717054dd0997b35260238cb83dbefbe8  training/source_blobs/274c2760db9514c13e91ea3c78df9ff3717054dd0997b35260238cb83dbefbe8.py
+28968ef3e5de86c1b523ec5c3cfffdbcfe3c0f523a8ca18bd463dfdba3e709ee  training/source_blobs/28968ef3e5de86c1b523ec5c3cfffdbcfe3c0f523a8ca18bd463dfdba3e709ee.py
+292a9338f1169410e4450138740beea1588be4f1883bd164b8cd83a56d5fc5c0  training/source_blobs/292a9338f1169410e4450138740beea1588be4f1883bd164b8cd83a56d5fc5c0.py
+29310a52ea67aac363214dd8f3d3f26663d1371fa0d10172861df51ada623bd1  training/source_blobs/29310a52ea67aac363214dd8f3d3f26663d1371fa0d10172861df51ada623bd1.py
+2abf5926a22a33b4f29c699855b24f7e459b6e631aa6ef780876ddc205f17758  training/source_blobs/2abf5926a22a33b4f29c699855b24f7e459b6e631aa6ef780876ddc205f17758.py
+2ac80ce082cb5fca838b98c027f352f21dc666a9270005bec79d9c055ca3b99b  training/source_blobs/2ac80ce082cb5fca838b98c027f352f21dc666a9270005bec79d9c055ca3b99b.py
+2b0a53ffa8f1100a9c24ec5991ffbe2c5ed79e884de9486376fd4dd8eac6da93  training/source_blobs/2b0a53ffa8f1100a9c24ec5991ffbe2c5ed79e884de9486376fd4dd8eac6da93.py
+2b6aa65bd43786a1b815bc21e73229b4304d0f4275aed56ccdfbce4c6ea08f68  training/source_blobs/2b6aa65bd43786a1b815bc21e73229b4304d0f4275aed56ccdfbce4c6ea08f68.py
+2beddc756f979309b46b115101191fa17a3c2c8e12579b672154a45a513b4e6d  training/source_blobs/2beddc756f979309b46b115101191fa17a3c2c8e12579b672154a45a513b4e6d.py
+2cca8d901c9fe4f49016a01b1e89ba2b519fe37b42b265bcf9333db64d8ba2d9  training/source_blobs/2cca8d901c9fe4f49016a01b1e89ba2b519fe37b42b265bcf9333db64d8ba2d9.py
+2cd4fe99866d6c85fa1a2481a2058009e96824c8622e915caca56fbbf544e829  training/source_blobs/2cd4fe99866d6c85fa1a2481a2058009e96824c8622e915caca56fbbf544e829.py
+2d901af050a73117f6f2c76f16060377912b765b0f2f4bafe24f3092d01f19c4  training/source_blobs/2d901af050a73117f6f2c76f16060377912b765b0f2f4bafe24f3092d01f19c4.py
+2e41c214043ae68cc2fc533ecf1e7c3358e8390c6edc9ff8f20a962acbf0d5b2  training/source_blobs/2e41c214043ae68cc2fc533ecf1e7c3358e8390c6edc9ff8f20a962acbf0d5b2.py
+2ef091c51f28ea44bce60988693903abae0dd973aee6892e26f69126b9d68db0  training/source_blobs/2ef091c51f28ea44bce60988693903abae0dd973aee6892e26f69126b9d68db0.py
+2f643abaa6ac3f86e47b37632df4d25fa34d0b25e77b59ffecc0290b2385e7dc  training/source_blobs/2f643abaa6ac3f86e47b37632df4d25fa34d0b25e77b59ffecc0290b2385e7dc.py
+2fbd333bb25f0dd02907141f0a16d59ad49d4db08cc5453ea8655f61cadb65f1  training/source_blobs/2fbd333bb25f0dd02907141f0a16d59ad49d4db08cc5453ea8655f61cadb65f1.py
+30b8ecb4b97d2520addb522f77c2e955d01ebc62fb9a2ef9a6d8a3f41dc3d0b3  training/source_blobs/30b8ecb4b97d2520addb522f77c2e955d01ebc62fb9a2ef9a6d8a3f41dc3d0b3.py
+317024d64eca0650b791f08ef92001d05037c6dd81b0f31ad8eac15c509bd8bb  training/source_blobs/317024d64eca0650b791f08ef92001d05037c6dd81b0f31ad8eac15c509bd8bb.py
+317c05d4d17239e62d1316797892288f5e4eae467b3e66a4ca4f15d189960a9b  training/source_blobs/317c05d4d17239e62d1316797892288f5e4eae467b3e66a4ca4f15d189960a9b.py
+321c85734de3ce593cd7401b0082722d155ab71344196a73216c02d04ff0ba6a  training/source_blobs/321c85734de3ce593cd7401b0082722d155ab71344196a73216c02d04ff0ba6a.py
+32527d75848cdc0e90f1ca93cd4ea54aeb2368374a3d0edd8653296b359d09f7  training/source_blobs/32527d75848cdc0e90f1ca93cd4ea54aeb2368374a3d0edd8653296b359d09f7.py
+345331fc34634f7ea7d9ebfb1cf8c91a17123033c52c7d0bbfbe2aa3ed6ab49d  training/source_blobs/345331fc34634f7ea7d9ebfb1cf8c91a17123033c52c7d0bbfbe2aa3ed6ab49d.py
+35c8ad532a20f77fc92edd49aaf877db0b2a4750c7212687e21b5d475b73534e  training/source_blobs/35c8ad532a20f77fc92edd49aaf877db0b2a4750c7212687e21b5d475b73534e.py
+392f3cbeb2c4ad10a8b7173ed4331f63368a5cedda6fa84d23eb0d8dbd434657  training/source_blobs/392f3cbeb2c4ad10a8b7173ed4331f63368a5cedda6fa84d23eb0d8dbd434657.py
+3972d7d2ef83ddbbd883b08f30204c870c4f91fdca47326b4a9e969bb0283f20  training/source_blobs/3972d7d2ef83ddbbd883b08f30204c870c4f91fdca47326b4a9e969bb0283f20.py
+3ad40e26678dbe103cbf895a9f885bf9a06734807afa341336f48725178dc0ad  training/source_blobs/3ad40e26678dbe103cbf895a9f885bf9a06734807afa341336f48725178dc0ad.py
+3b8f68cfd6fce0828c205723e98a31902455a71c27413d8a7bd18b1aec42d11b  training/source_blobs/3b8f68cfd6fce0828c205723e98a31902455a71c27413d8a7bd18b1aec42d11b.py
+3d78ed45dbab52a671d889fe937a210a832e55426bbdc4f905ddbd202ccf2336  training/source_blobs/3d78ed45dbab52a671d889fe937a210a832e55426bbdc4f905ddbd202ccf2336.py
+3e1c35e75a1f27dff42c6e90246972d811df32c6a3666dfe5f2a58e8daae5ea4  training/source_blobs/3e1c35e75a1f27dff42c6e90246972d811df32c6a3666dfe5f2a58e8daae5ea4.py
+3e82b28d76197292a35d4d479b77fe14e5482bff0f7fd04d2a34e8d7cec4c9c4  training/source_blobs/3e82b28d76197292a35d4d479b77fe14e5482bff0f7fd04d2a34e8d7cec4c9c4.py
+3eed2836d8c4111cf409815d9e749806c0f54fb8164e704bbcc9297b19e829ae  training/source_blobs/3eed2836d8c4111cf409815d9e749806c0f54fb8164e704bbcc9297b19e829ae.py
+3f6164c3f80f5fc8794b8c1776a6ceb7bc80af9cd36e03f6dcf3c0442ca662b5  training/source_blobs/3f6164c3f80f5fc8794b8c1776a6ceb7bc80af9cd36e03f6dcf3c0442ca662b5.py
+40dbfada2856c91b0a7ffe5df7b9584175dbd479a1c6b581d1935d38dd6228aa  training/source_blobs/40dbfada2856c91b0a7ffe5df7b9584175dbd479a1c6b581d1935d38dd6228aa.py
+41618bc3978700ba85bdae4a179ac27ac4faafee2f81c67f6887a72722e91f99  training/source_blobs/41618bc3978700ba85bdae4a179ac27ac4faafee2f81c67f6887a72722e91f99.py
+41c77ec50fa0cfae3de7ccffa0ba792a5eaf754f0342159adad1316c4d8c8fe0  training/source_blobs/41c77ec50fa0cfae3de7ccffa0ba792a5eaf754f0342159adad1316c4d8c8fe0.py
+41e6a98d9b41dc844063fc28ce81a08d83b919306b9daf0716b76a0bc21365fc  training/source_blobs/41e6a98d9b41dc844063fc28ce81a08d83b919306b9daf0716b76a0bc21365fc.py
+4255b93bbc18d120318d83fb26b2fda1fe16d19ca1423250c77188940cc4f008  training/source_blobs/4255b93bbc18d120318d83fb26b2fda1fe16d19ca1423250c77188940cc4f008.py
+42d96626fffbee32722c4f685a4aa121ec32a046283d51a9ac325fcea68e7996  training/source_blobs/42d96626fffbee32722c4f685a4aa121ec32a046283d51a9ac325fcea68e7996.py
+441089f82a4a8e5037cd8f3f02f24f11852b6722ba8d32fff619618c6efba211  training/source_blobs/441089f82a4a8e5037cd8f3f02f24f11852b6722ba8d32fff619618c6efba211.py
+44dfd6367ab2cd890d9630c47e3a49fe36b7f56107657196946343033fe2bebf  training/source_blobs/44dfd6367ab2cd890d9630c47e3a49fe36b7f56107657196946343033fe2bebf.py
+464357ccafaa6674fad3980792773e1c2369b3fcaef2bd41fe2407eb888d6a26  training/source_blobs/464357ccafaa6674fad3980792773e1c2369b3fcaef2bd41fe2407eb888d6a26.py
+467c462e44925f2a1810cb54778ad496b194a8a8ba4e788521e9bbc09164c000  training/source_blobs/467c462e44925f2a1810cb54778ad496b194a8a8ba4e788521e9bbc09164c000.py
+4883847e90e5db21bf4305ad9cab06d031b486d11483492cfc2691c559b503bc  training/source_blobs/4883847e90e5db21bf4305ad9cab06d031b486d11483492cfc2691c559b503bc.py
+489371cd03eec40cc26b9ad872177d022f6844e84b982b77bc35b4692decc511  training/source_blobs/489371cd03eec40cc26b9ad872177d022f6844e84b982b77bc35b4692decc511.py
+4a177d4a54761cd46cf044f6a6cf9c30831665a9e0aa2b84bd52f9e563f4ebfc  training/source_blobs/4a177d4a54761cd46cf044f6a6cf9c30831665a9e0aa2b84bd52f9e563f4ebfc.py
+4a7da603e70e60435b6bece218bc2b7a8ec06a54483a37e3088e51f92d0d0058  training/source_blobs/4a7da603e70e60435b6bece218bc2b7a8ec06a54483a37e3088e51f92d0d0058.py
+4b3b6862e0bc364f6d27adfdfbd187742173bbc45163109f6af0be8a81564c28  training/source_blobs/4b3b6862e0bc364f6d27adfdfbd187742173bbc45163109f6af0be8a81564c28.py
+4c0c1486aeeaa2c2640f3fce552568a71c6c32c7bf194fc40247eda71ae2ae16  training/source_blobs/4c0c1486aeeaa2c2640f3fce552568a71c6c32c7bf194fc40247eda71ae2ae16.py
+4ca1d4825452fa5ff01024f07d3bb718d1bd1919695f4450101a33fd1bb6d312  training/source_blobs/4ca1d4825452fa5ff01024f07d3bb718d1bd1919695f4450101a33fd1bb6d312.py
+50087489251785aaf70b00a66fcad27689f154fb08089ccf54f959e0abf2700b  training/source_blobs/50087489251785aaf70b00a66fcad27689f154fb08089ccf54f959e0abf2700b.py
+5041b972c2b7e3c80a5483329c5b52512e566391959f74c2aa4c4ea0fa1f83d7  training/source_blobs/5041b972c2b7e3c80a5483329c5b52512e566391959f74c2aa4c4ea0fa1f83d7.py
+50b208a8b67a2fc7f5626ab42ef597a0844f70a11a42e048faa35ffdeadaddf0  training/source_blobs/50b208a8b67a2fc7f5626ab42ef597a0844f70a11a42e048faa35ffdeadaddf0.py
+5105b1b526a8f8d734634231243e054649b0cc0abbfd28e8783d95891e29fcb2  training/source_blobs/5105b1b526a8f8d734634231243e054649b0cc0abbfd28e8783d95891e29fcb2.py
+51060d603ab35f695729e585233b9e2d3fe4c11e27145251eb0a6f15a4862fbf  training/source_blobs/51060d603ab35f695729e585233b9e2d3fe4c11e27145251eb0a6f15a4862fbf.py
+51411b86701c4b0d1c38691094232c43ffefea05c38dc8629ca1fc70316efaf4  training/source_blobs/51411b86701c4b0d1c38691094232c43ffefea05c38dc8629ca1fc70316efaf4.py
+5234bc6b824e4b07f63dda48e5353bbdb5f5abac4c98c299b642d1815ce85d94  training/source_blobs/5234bc6b824e4b07f63dda48e5353bbdb5f5abac4c98c299b642d1815ce85d94.py
+53d795e591f3e67f26ecd1f9eac289b0e3a66c4d83183cf47859168bb93c3817  training/source_blobs/53d795e591f3e67f26ecd1f9eac289b0e3a66c4d83183cf47859168bb93c3817.py
+548f1142be10f46b302886abae76790f5802c0f159412994eee66a7a5a3366a9  training/source_blobs/548f1142be10f46b302886abae76790f5802c0f159412994eee66a7a5a3366a9.py
+554977e377d8a247404479de2909a6d544ffe98c00f3f0cacfcea741dce0a3d2  training/source_blobs/554977e377d8a247404479de2909a6d544ffe98c00f3f0cacfcea741dce0a3d2.py
+55b04ac7875ad148fecf1c7cdfb5ab48e171abe3754ed9b19809cb9b93cd9ae0  training/source_blobs/55b04ac7875ad148fecf1c7cdfb5ab48e171abe3754ed9b19809cb9b93cd9ae0.py
+56543798ef7fcda70687accfb7f6e62a541453752c0fe10acbebb6e7c33563e3  training/source_blobs/56543798ef7fcda70687accfb7f6e62a541453752c0fe10acbebb6e7c33563e3.py
+56946ebf45af107f1952d03428437f83905be51bd2ce65605f4b381e55d76c06  training/source_blobs/56946ebf45af107f1952d03428437f83905be51bd2ce65605f4b381e55d76c06.py
+56b487defef5b61803857ef671f307ae3109ba96d6e29fce2c3004a888966952  training/source_blobs/56b487defef5b61803857ef671f307ae3109ba96d6e29fce2c3004a888966952.py
+57555e09d1c1716f2026001accf8522a9d27fb0685bcbad2904152f980913880  training/source_blobs/57555e09d1c1716f2026001accf8522a9d27fb0685bcbad2904152f980913880.py
+58e0bdc793f076f0fa6e7ec5b0cc4df51506a9e7505c09322d157fa080471dbc  training/source_blobs/58e0bdc793f076f0fa6e7ec5b0cc4df51506a9e7505c09322d157fa080471dbc.py
+59d1adb31400e3c288914b4c0cc8a21681a73ca3dfa7c722ea4e13d21dc38528  training/source_blobs/59d1adb31400e3c288914b4c0cc8a21681a73ca3dfa7c722ea4e13d21dc38528.py
+5a4bda6fc5e174f0d7dd9d602867ce0c746ec2609e243043a3657793b94401fa  training/source_blobs/5a4bda6fc5e174f0d7dd9d602867ce0c746ec2609e243043a3657793b94401fa.py
+5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8  training/source_blobs/5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8.py
+5cbce166031d4056b6e69c7b9b9efb45528e182d6ba996739c382bc5806bf477  training/source_blobs/5cbce166031d4056b6e69c7b9b9efb45528e182d6ba996739c382bc5806bf477.py
+5cda62bae42610986c169ecd7d6cb92fdd7f0f7eb2255fa912a7eb54216a9cbb  training/source_blobs/5cda62bae42610986c169ecd7d6cb92fdd7f0f7eb2255fa912a7eb54216a9cbb.py
+5dcaa086355bc0fd73c3ea1504bd596b48edba1dc2d757a148dcad11547cec02  training/source_blobs/5dcaa086355bc0fd73c3ea1504bd596b48edba1dc2d757a148dcad11547cec02.py
+5f6474ff3b49542dd553d88d9b78512241cf7a25f8e37c03fe9bb9d90d1dd77f  training/source_blobs/5f6474ff3b49542dd553d88d9b78512241cf7a25f8e37c03fe9bb9d90d1dd77f.py
+5fffbc7cb346228589779105aab9c5c0e26a71b6d6e3528a29777f68869b430c  training/source_blobs/5fffbc7cb346228589779105aab9c5c0e26a71b6d6e3528a29777f68869b430c.py
+60c792369c4a3d65ff960ef05e5389d8c1571cf553eb7a1eb2673258c59b069e  training/source_blobs/60c792369c4a3d65ff960ef05e5389d8c1571cf553eb7a1eb2673258c59b069e.py
+6155114ea8f946e116dec8afd080b91d37275d9565d62004a2303fdcff74cbc7  training/source_blobs/6155114ea8f946e116dec8afd080b91d37275d9565d62004a2303fdcff74cbc7.py
+631451a45b020120610b7116e8e63276ba5043011359fbd13e49e63ec8e2b02b  training/source_blobs/631451a45b020120610b7116e8e63276ba5043011359fbd13e49e63ec8e2b02b.py
+63dc102460925b5146e5d73dc32dbde53a668ca1b662b49782561910994b9de5  training/source_blobs/63dc102460925b5146e5d73dc32dbde53a668ca1b662b49782561910994b9de5.py
+641140ca75f9a64d8afc015898ec0a8c294b11fe7fc7bfd4c06ab1726c59df04  training/source_blobs/641140ca75f9a64d8afc015898ec0a8c294b11fe7fc7bfd4c06ab1726c59df04.py
+655888aa035d4939001ebf08ccde1a5a8a78d1ce5189da3ed11c60d2b1dedd97  training/source_blobs/655888aa035d4939001ebf08ccde1a5a8a78d1ce5189da3ed11c60d2b1dedd97.py
+669eccb3d19fe369bfa6e04ed47cc2b2a2cdb26f9407ae0de9a993431f1880eb  training/source_blobs/669eccb3d19fe369bfa6e04ed47cc2b2a2cdb26f9407ae0de9a993431f1880eb.py
+695895ece361823ce772e651db19d1a26fc372cc8913fb9dc6fcc9231342c6c0  training/source_blobs/695895ece361823ce772e651db19d1a26fc372cc8913fb9dc6fcc9231342c6c0.py
+69babc5ed58d0b859c6cd2419ce9c0cdb37823d0b6b174cc6de82bd22816eb35  training/source_blobs/69babc5ed58d0b859c6cd2419ce9c0cdb37823d0b6b174cc6de82bd22816eb35.py
+6a7a6c7ee6dee3d4ee81d5f2a627b57c3b4444808a2c9c1f7b280789a9260e23  training/source_blobs/6a7a6c7ee6dee3d4ee81d5f2a627b57c3b4444808a2c9c1f7b280789a9260e23.py
+6bf0cf7f1fb395aa90227b1bca52214f3151628b930087b58bc90101828dc467  training/source_blobs/6bf0cf7f1fb395aa90227b1bca52214f3151628b930087b58bc90101828dc467.py
+6dac260e47b46258d9d582362055b172ac24bdf1ecc57acae7e2769a0097a129  training/source_blobs/6dac260e47b46258d9d582362055b172ac24bdf1ecc57acae7e2769a0097a129.py
+6e854aa53c69baa40955091e7f5318ddfea3aa8569a606fa342f9fbe07c1006b  training/source_blobs/6e854aa53c69baa40955091e7f5318ddfea3aa8569a606fa342f9fbe07c1006b.py
+6ebf9b6ed5a5b3d78e179c609950fc0f7788b6af12dde11cd55ec79c9b3e28d7  training/source_blobs/6ebf9b6ed5a5b3d78e179c609950fc0f7788b6af12dde11cd55ec79c9b3e28d7.py
+70e36e12d33547bfce505bd0302605e3d66c9ee65fc0facb26ccf5341bde8b4b  training/source_blobs/70e36e12d33547bfce505bd0302605e3d66c9ee65fc0facb26ccf5341bde8b4b.py
+71ec4fc999f1e87c24e0b2fe2e60e6a59d716a13f76b29ca88a3856c05c246da  training/source_blobs/71ec4fc999f1e87c24e0b2fe2e60e6a59d716a13f76b29ca88a3856c05c246da.py
+72cc9ed68cee16a17a5271bcee77870da6fafb3949a7d052bc39abe6e7daccc2  training/source_blobs/72cc9ed68cee16a17a5271bcee77870da6fafb3949a7d052bc39abe6e7daccc2.py
+72e7ca777f0e9078d955e13ff958ef2cc43374f924d72d69ad0ee770e9ae8271  training/source_blobs/72e7ca777f0e9078d955e13ff958ef2cc43374f924d72d69ad0ee770e9ae8271.py
+73cbf0c17cf74a44bb7c74a8f9e5bb63671057a1f664a6a4af0e0777b2c8f1a4  training/source_blobs/73cbf0c17cf74a44bb7c74a8f9e5bb63671057a1f664a6a4af0e0777b2c8f1a4.py
+7493fcc8857dcc01e5e9d5e72a3d47639411b810719a64a263486d5faddd5a04  training/source_blobs/7493fcc8857dcc01e5e9d5e72a3d47639411b810719a64a263486d5faddd5a04.py
+74d4e12f420a4e105933a0bcb87626b7d4fb9c68d2a3b2164a6bfefc0eb6a1e0  training/source_blobs/74d4e12f420a4e105933a0bcb87626b7d4fb9c68d2a3b2164a6bfefc0eb6a1e0.py
+75f45d5bb98fc6ac62ab00506503e31b7534cad1e7ea8c0fc69963251c8f0c4d  training/source_blobs/75f45d5bb98fc6ac62ab00506503e31b7534cad1e7ea8c0fc69963251c8f0c4d.py
+76a82db3ee8da8956413fe559674d94415b9ea21b95d0a1442a8da732f908e9c  training/source_blobs/76a82db3ee8da8956413fe559674d94415b9ea21b95d0a1442a8da732f908e9c.py
+78290186ea157bb44399268dd87f4df32feba21d79229cbe988a35de576e800f  training/source_blobs/78290186ea157bb44399268dd87f4df32feba21d79229cbe988a35de576e800f.py
+7b761ff4a4023a0f84f9c0efa0e0c92cc7640df30b0608842d723c95c582f1cf  training/source_blobs/7b761ff4a4023a0f84f9c0efa0e0c92cc7640df30b0608842d723c95c582f1cf.py
+7c176d4ca3c55cadb8deffc2e6d640cb5d81dee4884a55c6ef15adecb1f3f6b4  training/source_blobs/7c176d4ca3c55cadb8deffc2e6d640cb5d81dee4884a55c6ef15adecb1f3f6b4.py
+7c6f6d0c910f4cea15d29d20c2a98add617c08cae5cf545c3d43be220cf44881  training/source_blobs/7c6f6d0c910f4cea15d29d20c2a98add617c08cae5cf545c3d43be220cf44881.py
+7cb885d4bb5d9fa5a796304a9bae8ce995ffdf8d450c6fc68c17f93f0986350d  training/source_blobs/7cb885d4bb5d9fa5a796304a9bae8ce995ffdf8d450c6fc68c17f93f0986350d.py
+7cdc7d042c65a996bc8867c2e370503b533a1216f9656619ec6fcffd4774f6fb  training/source_blobs/7cdc7d042c65a996bc8867c2e370503b533a1216f9656619ec6fcffd4774f6fb.py
+7e7b298f6716c0e3a013d25bfcf20a766fc42aa59821b0a639b60bb7e5c78b75  training/source_blobs/7e7b298f6716c0e3a013d25bfcf20a766fc42aa59821b0a639b60bb7e5c78b75.py
+7e9e02294e03e6b8a12fd202ab288f2f9b0073bb11b8cb2a5fa82211b0d1dfa4  training/source_blobs/7e9e02294e03e6b8a12fd202ab288f2f9b0073bb11b8cb2a5fa82211b0d1dfa4.py
+7ee92cc6f8765187de5cf72b105a8a7e0784cf59fc8cc9869dcb91bc2643aada  training/source_blobs/7ee92cc6f8765187de5cf72b105a8a7e0784cf59fc8cc9869dcb91bc2643aada.py
+7f64ce45cd9e8201f8adce7ecb37be6ee29c80b83c344df699814401fc162780  training/source_blobs/7f64ce45cd9e8201f8adce7ecb37be6ee29c80b83c344df699814401fc162780.py
+80d4e25a009727ccaf3885612fb3d2879192b923b248f18a00f1318b04a34253  training/source_blobs/80d4e25a009727ccaf3885612fb3d2879192b923b248f18a00f1318b04a34253.py
+80de0fe0d43dea84ce734bc10256fc45a931828d117ef57a6a81b83ec40d2827  training/source_blobs/80de0fe0d43dea84ce734bc10256fc45a931828d117ef57a6a81b83ec40d2827.py
+81c300349b75def6cc4d151fd39e9d4cec85adae87c02a6290f74f6fabe665b9  training/source_blobs/81c300349b75def6cc4d151fd39e9d4cec85adae87c02a6290f74f6fabe665b9.py
+82d4b5333c98bd0483d22b196548670b26b712623fea00f815f25137f01f0891  training/source_blobs/82d4b5333c98bd0483d22b196548670b26b712623fea00f815f25137f01f0891.py
+82e6fc9baae2c5fb23b24bde85880f8d6c2911ed07f0911ca1959434650c2256  training/source_blobs/82e6fc9baae2c5fb23b24bde85880f8d6c2911ed07f0911ca1959434650c2256.py
+83b46cb227b9d5e6d8410b1df727c1df32975f04121e4efe4caf7f8552664800  training/source_blobs/83b46cb227b9d5e6d8410b1df727c1df32975f04121e4efe4caf7f8552664800.py
+88d7ece9842b43fcab0f061a4de81e9dacf3010df3a17fb40c0ba576f8d11834  training/source_blobs/88d7ece9842b43fcab0f061a4de81e9dacf3010df3a17fb40c0ba576f8d11834.py
+89006313590f7e77f6db14ca9ddb5c0e92e9952b387b8289dc8b9f662509979a  training/source_blobs/89006313590f7e77f6db14ca9ddb5c0e92e9952b387b8289dc8b9f662509979a.py
+890cc0a233f730cc9ca4eab0c5fbc570f8a7a04f816ca4979940014dccf485b5  training/source_blobs/890cc0a233f730cc9ca4eab0c5fbc570f8a7a04f816ca4979940014dccf485b5.py
+898c01ca9ab30c5f9e6ae205060f8457a7abd2d740c8b8d04c57a1030a268d17  training/source_blobs/898c01ca9ab30c5f9e6ae205060f8457a7abd2d740c8b8d04c57a1030a268d17.py
+8a8c0e212718e0f53d004acad86827b4fab97539fe7ae1ada5c4c855ef7f57e0  training/source_blobs/8a8c0e212718e0f53d004acad86827b4fab97539fe7ae1ada5c4c855ef7f57e0.py
+8ab3b8004f0c208bc1c8c74ad2285e4fa067bb29481816342b5ae6eaa0670652  training/source_blobs/8ab3b8004f0c208bc1c8c74ad2285e4fa067bb29481816342b5ae6eaa0670652.py
+8b29bdc076c576d3f951e1d02f2eca1a2a6d78e86202d19704c3814a4cb0a84e  training/source_blobs/8b29bdc076c576d3f951e1d02f2eca1a2a6d78e86202d19704c3814a4cb0a84e.py
+8bcc4cdb380d68a5db6e221481cde7ed5cc4b5fc9c87d07cd9d47f41ef056aab  training/source_blobs/8bcc4cdb380d68a5db6e221481cde7ed5cc4b5fc9c87d07cd9d47f41ef056aab.py
+8bcfaff689b96ba9716ec0255dc0f91c5583d9b696784605dacbd88a6562cc7e  training/source_blobs/8bcfaff689b96ba9716ec0255dc0f91c5583d9b696784605dacbd88a6562cc7e.py
+8eed9a85b1c94e84405f563f12504b4930ddbcd2d9f870aff9ce1fee553e8b5f  training/source_blobs/8eed9a85b1c94e84405f563f12504b4930ddbcd2d9f870aff9ce1fee553e8b5f.py
+8f74c3f3ef6840811051fb8cfd5ac50249c94b116f2db47a5ba56ca8d8c92534  training/source_blobs/8f74c3f3ef6840811051fb8cfd5ac50249c94b116f2db47a5ba56ca8d8c92534.py
+907b357fdd5c8ade00becbfe952a99791855a2afc90ecc3fd026a7860fb13ebf  training/source_blobs/907b357fdd5c8ade00becbfe952a99791855a2afc90ecc3fd026a7860fb13ebf.py
+90c58a30a50cb625e36222825768a134aba5c81df87de8655ab95de2bd47402c  training/source_blobs/90c58a30a50cb625e36222825768a134aba5c81df87de8655ab95de2bd47402c.py
+90e9b0e31206f0560e7e2aba2a5cf1adb9e9a2e0f23a5ab0fabdacf22ebb1470  training/source_blobs/90e9b0e31206f0560e7e2aba2a5cf1adb9e9a2e0f23a5ab0fabdacf22ebb1470.py
+92030c2ce302eea434bd81553ddac5c9170db82c491950cdb903f2caf732aee4  training/source_blobs/92030c2ce302eea434bd81553ddac5c9170db82c491950cdb903f2caf732aee4.py
+922bba304680cb24aaad23f879faaa13387f15fcbd02bcd1c03734dc3a652c3b  training/source_blobs/922bba304680cb24aaad23f879faaa13387f15fcbd02bcd1c03734dc3a652c3b.py
+93294262c9e8f9d0b6082d0c4f56149925969f9020279e8bc50d928c6e6811d7  training/source_blobs/93294262c9e8f9d0b6082d0c4f56149925969f9020279e8bc50d928c6e6811d7.py
+93b86664d0aba1e36083877e1cdb1a995766dffb1b207bb4cd691e226629de53  training/source_blobs/93b86664d0aba1e36083877e1cdb1a995766dffb1b207bb4cd691e226629de53.py
+93dcc3aa14fba92023f3b9fc44caab09bded50c9ef54a7b9d95281b6ab5afc25  training/source_blobs/93dcc3aa14fba92023f3b9fc44caab09bded50c9ef54a7b9d95281b6ab5afc25.py
+943fd969994e896faff14481b94eb121d23b6fef00244874d5d5361df6bd8cce  training/source_blobs/943fd969994e896faff14481b94eb121d23b6fef00244874d5d5361df6bd8cce.py
+94b8de5c15570b753850ca1ba84543c858334e5b98ed20d60a1bb0acf1e6c17e  training/source_blobs/94b8de5c15570b753850ca1ba84543c858334e5b98ed20d60a1bb0acf1e6c17e.py
+955993d44637e21b043a0b8fc06b48c97589e40f255d90ff6c795394c6a5b569  training/source_blobs/955993d44637e21b043a0b8fc06b48c97589e40f255d90ff6c795394c6a5b569.py
+965c794f040d45cf2abefb16ea17c55e14be81d390b6534d36da7a12cecf5a93  training/source_blobs/965c794f040d45cf2abefb16ea17c55e14be81d390b6534d36da7a12cecf5a93.py
+96795d99ce287b32200c9b56069856324678feba937834904efbc46a6dcc6404  training/source_blobs/96795d99ce287b32200c9b56069856324678feba937834904efbc46a6dcc6404.py
+97243740f45c7b79f5e358a4b9378e54a3e9114f150d2780a718a1e62164f2f9  training/source_blobs/97243740f45c7b79f5e358a4b9378e54a3e9114f150d2780a718a1e62164f2f9.py
+977a196d5949f9ebc965b253afd45b818290c8ff5ea3a462efa219c368f6f4bd  training/source_blobs/977a196d5949f9ebc965b253afd45b818290c8ff5ea3a462efa219c368f6f4bd.py
+9a99bcf27a5895cc4f1ab34103c303675e05cb745944f9b10dfb26e3651a6581  training/source_blobs/9a99bcf27a5895cc4f1ab34103c303675e05cb745944f9b10dfb26e3651a6581.py
+9bf5f9f145991764775f032257ccf80190a1cb5d6f45e6ef504c3cebd87433ca  training/source_blobs/9bf5f9f145991764775f032257ccf80190a1cb5d6f45e6ef504c3cebd87433ca.py
+9ea68c491cc79452f58a4a93c483af6c26fdff1f6caeb924e004d5298e927a48  training/source_blobs/9ea68c491cc79452f58a4a93c483af6c26fdff1f6caeb924e004d5298e927a48.py
+9efa69e4acf6afef250e0d1ab45ecbac4da042e21a7fb55f38082a13b0fe4abf  training/source_blobs/9efa69e4acf6afef250e0d1ab45ecbac4da042e21a7fb55f38082a13b0fe4abf.py
+9efaba2abb2f32c9f5778872fd7849609c8045687aa7ac8146a1001bf653f151  training/source_blobs/9efaba2abb2f32c9f5778872fd7849609c8045687aa7ac8146a1001bf653f151.py
+9f3a8646f9661e0c31bcb3188d4b3410a0e6454a128a617d18b79ea0cbf660ee  training/source_blobs/9f3a8646f9661e0c31bcb3188d4b3410a0e6454a128a617d18b79ea0cbf660ee.py
+9f7f0804eaa9edc6d9c4a5049df230a454171b7f307144212a881af0b3db3971  training/source_blobs/9f7f0804eaa9edc6d9c4a5049df230a454171b7f307144212a881af0b3db3971.py
+9ffff2dcc93bcf275f1bc183c8cef37914602a25342f65c1404fd7b398854c82  training/source_blobs/9ffff2dcc93bcf275f1bc183c8cef37914602a25342f65c1404fd7b398854c82.py
+a08b55556fe466c10c7d8188066b1a6a23f86d2545378d85f61898cbe300a67d  training/source_blobs/a08b55556fe466c10c7d8188066b1a6a23f86d2545378d85f61898cbe300a67d.py
+a1c6f3f901996aac30d7ffcf2b3fdc89873faba7cf5bd1a8ec301ff8076bb719  training/source_blobs/a1c6f3f901996aac30d7ffcf2b3fdc89873faba7cf5bd1a8ec301ff8076bb719.py
+a1f77593681bc49b242a042d4819e5da3baa60ac4e6e65e9cd8ea4d5aae8a9e0  training/source_blobs/a1f77593681bc49b242a042d4819e5da3baa60ac4e6e65e9cd8ea4d5aae8a9e0.py
+a3dfbd61de43fda14daf4bbd921ccb382ea8af2c8342457fe16d1ad0908e88fe  training/source_blobs/a3dfbd61de43fda14daf4bbd921ccb382ea8af2c8342457fe16d1ad0908e88fe.py
+a43649482bd7234bdb704bd014031ad69769086b2ac60c8eaa9190bbf428629c  training/source_blobs/a43649482bd7234bdb704bd014031ad69769086b2ac60c8eaa9190bbf428629c.py
+a6b6437f5efbf0771af2a70e26d63bbcdaa45a397fa19f6c214c27a8140e4b78  training/source_blobs/a6b6437f5efbf0771af2a70e26d63bbcdaa45a397fa19f6c214c27a8140e4b78.py
+a71eabcc526344b6502cae999020da63c950b2d80c13077629bb86c8757a18ed  training/source_blobs/a71eabcc526344b6502cae999020da63c950b2d80c13077629bb86c8757a18ed.py
+a7aa3fc6627fd055924ba7da19c280146bd98e2e1257475816326d18fe41eabc  training/source_blobs/a7aa3fc6627fd055924ba7da19c280146bd98e2e1257475816326d18fe41eabc.py
+a7e3879ae4bd1506473d9b34e37dc1a94f9c9d2048995e0cba7fb04c61514554  training/source_blobs/a7e3879ae4bd1506473d9b34e37dc1a94f9c9d2048995e0cba7fb04c61514554.py
+a8379185dd82972a931e7270b7c26142ad30bc1340af309c8ff087965d79964a  training/source_blobs/a8379185dd82972a931e7270b7c26142ad30bc1340af309c8ff087965d79964a.py
+a8d9306095bb1de53cc4350741fe3132269b2a91600e351e6285d8ec1e3be6e3  training/source_blobs/a8d9306095bb1de53cc4350741fe3132269b2a91600e351e6285d8ec1e3be6e3.py
+a971a2e697a35c34d973d8add104fbcc5d661cb83473ccafceeed4820d280291  training/source_blobs/a971a2e697a35c34d973d8add104fbcc5d661cb83473ccafceeed4820d280291.py
+aad8eba4764e40e490856986f1f369cade78635d087e4fda1b9ad8455ffd07f3  training/source_blobs/aad8eba4764e40e490856986f1f369cade78635d087e4fda1b9ad8455ffd07f3.py
+ab01e0e88e2edc87a1960666366fa433f976a0c021270cf412dba32b918914d5  training/source_blobs/ab01e0e88e2edc87a1960666366fa433f976a0c021270cf412dba32b918914d5.py
+ac95755ba4b994dae0c0902cd052a9200ec09eecac28c547408a4ebabdbc99b0  training/source_blobs/ac95755ba4b994dae0c0902cd052a9200ec09eecac28c547408a4ebabdbc99b0.py
+acca03eb1f853e16c9166c144b645f44f470cc81cde0c0aa5a3ed0c4b3c0b070  training/source_blobs/acca03eb1f853e16c9166c144b645f44f470cc81cde0c0aa5a3ed0c4b3c0b070.py
+ad220f0033d3761d67d8048c8b4a730fb1ce47923e88c6bbe236b41067b83a82  training/source_blobs/ad220f0033d3761d67d8048c8b4a730fb1ce47923e88c6bbe236b41067b83a82.py
+adc3a0a6f19e7cfb3073f4c1e0432f664e5fb709141ceb4148aca7cda08ddaab  training/source_blobs/adc3a0a6f19e7cfb3073f4c1e0432f664e5fb709141ceb4148aca7cda08ddaab.py
+ae59081bcf4ea79fb7c5a2ab5cbac809976074e0f4fb290ae65bbf57f4733ca2  training/source_blobs/ae59081bcf4ea79fb7c5a2ab5cbac809976074e0f4fb290ae65bbf57f4733ca2.py
+af7bb33a7b0253c552bbefd46b1b3ad8857dd7740ea1118480948739075b210a  training/source_blobs/af7bb33a7b0253c552bbefd46b1b3ad8857dd7740ea1118480948739075b210a.py
+b0095392ef6a92ae2aad4ed20a8b9d668fe02f8f0ed602bd023a1a5d1ff59f4b  training/source_blobs/b0095392ef6a92ae2aad4ed20a8b9d668fe02f8f0ed602bd023a1a5d1ff59f4b.py
+b0bf909b472ddf9bd4eab30f6e6fd8b33bf3997ce038b882fe0663f291958d2d  training/source_blobs/b0bf909b472ddf9bd4eab30f6e6fd8b33bf3997ce038b882fe0663f291958d2d.txt
+b0bfe4c4afa5b00155a047fde824fb15baffd6b3b493f94a0a8da1cad6fc5f9e  training/source_blobs/b0bfe4c4afa5b00155a047fde824fb15baffd6b3b493f94a0a8da1cad6fc5f9e.py
+b222ce0b6eacf8a244328883cb751325c402a1027170829e91ac5fbdfbb8da3f  training/source_blobs/b222ce0b6eacf8a244328883cb751325c402a1027170829e91ac5fbdfbb8da3f.py
+b2f98795df373c7f1e91b2c2c2c158b51c7ef65c47522ad20a685c63e1773f99  training/source_blobs/b2f98795df373c7f1e91b2c2c2c158b51c7ef65c47522ad20a685c63e1773f99.py
+b30cd4ab4e99eedaf623580e3ff89b4a197534b14ee1d378c36d771fec303d82  training/source_blobs/b30cd4ab4e99eedaf623580e3ff89b4a197534b14ee1d378c36d771fec303d82.py
+b3bb31dd1dcf0fec5416d81664998e80b0244669fecd4389669083be09ec8daa  training/source_blobs/b3bb31dd1dcf0fec5416d81664998e80b0244669fecd4389669083be09ec8daa.py
+b575e3334fe9a21b3b9fc428d0f25728dfd4523426f4bf7aa34f6ddd7d4769e8  training/source_blobs/b575e3334fe9a21b3b9fc428d0f25728dfd4523426f4bf7aa34f6ddd7d4769e8.py
+b62637ccddac28f42b6efa19ce0578365497f7823d0df1dd2a9e27df453fea5c  training/source_blobs/b62637ccddac28f42b6efa19ce0578365497f7823d0df1dd2a9e27df453fea5c.py
+b64d1908b52856e19510283933d8b8d80f6ba15a8511a97732499a2820a06c76  training/source_blobs/b64d1908b52856e19510283933d8b8d80f6ba15a8511a97732499a2820a06c76.py
+b651f986a836c59591c13776ada2b17c1a359adbe6c96dd8306ef6b208d12231  training/source_blobs/b651f986a836c59591c13776ada2b17c1a359adbe6c96dd8306ef6b208d12231.py
+b828e68f67e7ee17cef4ff75d5c2b354672fbe6f559607248d621fd5a1e490f4  training/source_blobs/b828e68f67e7ee17cef4ff75d5c2b354672fbe6f559607248d621fd5a1e490f4.py
+b9f535fca724153703ed57a300eceb5b34a8f8381fc65bf5c4c5a0f7e066303c  training/source_blobs/b9f535fca724153703ed57a300eceb5b34a8f8381fc65bf5c4c5a0f7e066303c.py
+babe89bddb186a02e8c3ab4be1fa2cbb64f2d7bb27f8ac370701b4df39e03001  training/source_blobs/babe89bddb186a02e8c3ab4be1fa2cbb64f2d7bb27f8ac370701b4df39e03001.py
+bb3ba5db709a7df15a2936db36793939f877e99dd437956c64042d6e6cc25886  training/source_blobs/bb3ba5db709a7df15a2936db36793939f877e99dd437956c64042d6e6cc25886.py
+bbe9f97676c183f91308fce8e4b54c8b2ea04c4cc1e96db9f89d24a1e95efc48  training/source_blobs/bbe9f97676c183f91308fce8e4b54c8b2ea04c4cc1e96db9f89d24a1e95efc48.py
+bc89ed266b373257a693d1548c7d0d8304830e9d26517002623d7d29fbb467d2  training/source_blobs/bc89ed266b373257a693d1548c7d0d8304830e9d26517002623d7d29fbb467d2.py
+bddaf2b1f3691319843da58f4f37eb9ff02bd077502f6c87a082278cdbb970c6  training/source_blobs/bddaf2b1f3691319843da58f4f37eb9ff02bd077502f6c87a082278cdbb970c6.py
+be338639aa34512297c2d563fafd7d3b12f00f4ceb43ae8758e386c232c2439b  training/source_blobs/be338639aa34512297c2d563fafd7d3b12f00f4ceb43ae8758e386c232c2439b.py
+bffd8f8c5a74a0bf9c3ca070e6511cc1ab8ceb8343c0092a3d711b4fb9ece21c  training/source_blobs/bffd8f8c5a74a0bf9c3ca070e6511cc1ab8ceb8343c0092a3d711b4fb9ece21c.py
+c0468e20922d17555f41cbd76255a183917e1bde85d6c621732862fd2cf176de  training/source_blobs/c0468e20922d17555f41cbd76255a183917e1bde85d6c621732862fd2cf176de.py
+c065be118b0c6b8948a15d4b81e7fd7371f4e699ea3c49d34fa0abde258e0505  training/source_blobs/c065be118b0c6b8948a15d4b81e7fd7371f4e699ea3c49d34fa0abde258e0505.py
+c2c4b70dcee9caf027c49e0d9e3b8071b24d9bc6d197d1fec502618513c0cda2  training/source_blobs/c2c4b70dcee9caf027c49e0d9e3b8071b24d9bc6d197d1fec502618513c0cda2.py
+c2f5c1d6dd2de0195d89370359f91c438a48dacf5f157aa83607b24abe363265  training/source_blobs/c2f5c1d6dd2de0195d89370359f91c438a48dacf5f157aa83607b24abe363265.py
+c3565fe2c10ad14e2794e2d1a23d2fa5cce1a206d0064e4626a70b50a06c73f1  training/source_blobs/c3565fe2c10ad14e2794e2d1a23d2fa5cce1a206d0064e4626a70b50a06c73f1.py
+c745011c90cffc70e751eee207e2dca10ad94e1f2e167b967f0fdc8273afed36  training/source_blobs/c745011c90cffc70e751eee207e2dca10ad94e1f2e167b967f0fdc8273afed36.py
+c75ab44c4dbf49aab3e805bd51b5f1c607d7de917197dc10673771adc846592f  training/source_blobs/c75ab44c4dbf49aab3e805bd51b5f1c607d7de917197dc10673771adc846592f.py
+c81e7321a9a9fbf9cfe5a70434e4df2c751f33b6d68d93697ba4f09f69b56d88  training/source_blobs/c81e7321a9a9fbf9cfe5a70434e4df2c751f33b6d68d93697ba4f09f69b56d88.py
+c836de3dfbc0df4839e6b9e818406f6791b6b2db0a3531fbca8f90251cde9498  training/source_blobs/c836de3dfbc0df4839e6b9e818406f6791b6b2db0a3531fbca8f90251cde9498.py
+c9567f40f61e375ae5cea20ec10355ab4bc270af8b64e7b316b30a69947f4e33  training/source_blobs/c9567f40f61e375ae5cea20ec10355ab4bc270af8b64e7b316b30a69947f4e33.py
+c964e65593989b7ee6ca3ad7cc482ee926e95b91dcc123d3ccb4999710f2eccf  training/source_blobs/c964e65593989b7ee6ca3ad7cc482ee926e95b91dcc123d3ccb4999710f2eccf.py
+c9cfca2802cf083776ddb75c8805920751e763cb943efeea65f0949b732f90a7  training/source_blobs/c9cfca2802cf083776ddb75c8805920751e763cb943efeea65f0949b732f90a7.py
+cad10665efe683f5d7cec5ac8988e18e5970c7b354e248e1502eb7f9c242eab5  training/source_blobs/cad10665efe683f5d7cec5ac8988e18e5970c7b354e248e1502eb7f9c242eab5.py
+cb6e341043ea2b86c3058832c443a187c3abfb0bbdbca024360f04bc284479aa  training/source_blobs/cb6e341043ea2b86c3058832c443a187c3abfb0bbdbca024360f04bc284479aa.py
+cc7c710aaca40ac224da99f5902d4833b16d36b7a12c23817ee07ae3e29dbfda  training/source_blobs/cc7c710aaca40ac224da99f5902d4833b16d36b7a12c23817ee07ae3e29dbfda.py
+cc908731e8d2b2b0a848673a83bf094982a06446016df59896cdece491463b0e  training/source_blobs/cc908731e8d2b2b0a848673a83bf094982a06446016df59896cdece491463b0e.py
+cf09c5de39997350d25806866a3b3aedfe315ffeb6e4bdf00cf61ae05b24c3a2  training/source_blobs/cf09c5de39997350d25806866a3b3aedfe315ffeb6e4bdf00cf61ae05b24c3a2.py
+cf7a8e609115a29f0b97148dc531367e916c1219e60ecbbb3c372906f994453a  training/source_blobs/cf7a8e609115a29f0b97148dc531367e916c1219e60ecbbb3c372906f994453a.py
+cf80cb1bbde860cfc08808dac1fdf22cfd8d5900ce1294c7125b6c59def0de80  training/source_blobs/cf80cb1bbde860cfc08808dac1fdf22cfd8d5900ce1294c7125b6c59def0de80.py
+d35a801bc3f2c47528f371aa6ad99574ad1230ab25745b189f79a04268b8ee7c  training/source_blobs/d35a801bc3f2c47528f371aa6ad99574ad1230ab25745b189f79a04268b8ee7c.py
+d38da245bf7b2f54bfd4c728d2638dd9050888e3f5a2c7ed2360201b960e75a8  training/source_blobs/d38da245bf7b2f54bfd4c728d2638dd9050888e3f5a2c7ed2360201b960e75a8.py
+d39a65a18405683724659ce1ea4464d5f65fe6c3a05ca81d409fc6982ab052a0  training/source_blobs/d39a65a18405683724659ce1ea4464d5f65fe6c3a05ca81d409fc6982ab052a0.py
+d4393b7b221898b9019429346e71610db27bbca91904f6c80ae75c92b182b5cd  training/source_blobs/d4393b7b221898b9019429346e71610db27bbca91904f6c80ae75c92b182b5cd.py
+d5c27314990d2453044f4bc7993819a538ded896eb047ba72244e15c4b495f09  training/source_blobs/d5c27314990d2453044f4bc7993819a538ded896eb047ba72244e15c4b495f09.py
+d6a488047f6c58822b2881de9d9c363ce67435ec7fe07793b4043b0cd620c4f8  training/source_blobs/d6a488047f6c58822b2881de9d9c363ce67435ec7fe07793b4043b0cd620c4f8.py
+d9f529149989771c6bb04f46fb366413b8c8817925999a1cfa46c229c9782414  training/source_blobs/d9f529149989771c6bb04f46fb366413b8c8817925999a1cfa46c229c9782414.py
+da0cf71d99bfca88f04c0fb1182c200321aa27f0366980d566748970e0133f92  training/source_blobs/da0cf71d99bfca88f04c0fb1182c200321aa27f0366980d566748970e0133f92.py
+da5d0b5d3d9762dd806ad0ed9da1e8f0edd3e69dc8d626f20d9e6af27673f375  training/source_blobs/da5d0b5d3d9762dd806ad0ed9da1e8f0edd3e69dc8d626f20d9e6af27673f375.py
+db6cb7946e10423bdee30ebcead28ac1bc6b6e670204aaa6488f1b4ccb9d4240  training/source_blobs/db6cb7946e10423bdee30ebcead28ac1bc6b6e670204aaa6488f1b4ccb9d4240.py
+dbd685610812235fa72ac6081a1bc16c8ff55eecdec034c0df0c9c2eadde05e8  training/source_blobs/dbd685610812235fa72ac6081a1bc16c8ff55eecdec034c0df0c9c2eadde05e8.py
+dc0601b93d20a58245982b6041763212720e7d5386993749ba88d0f79d8e8479  training/source_blobs/dc0601b93d20a58245982b6041763212720e7d5386993749ba88d0f79d8e8479.py
+dc06bedad9c254e70109c0206c54eb594f6f5c427f2ffc0fb9f162d8191b4813  training/source_blobs/dc06bedad9c254e70109c0206c54eb594f6f5c427f2ffc0fb9f162d8191b4813.py
+de1306af5306ff9af02d9bdecf07f6a99baffb1ac6a052d4690840cc7ad680c6  training/source_blobs/de1306af5306ff9af02d9bdecf07f6a99baffb1ac6a052d4690840cc7ad680c6.py
+de9bf2fac7a6b3ed0ca0d23926cf02e19484383a78eefeeba3af3459f6ee46f4  training/source_blobs/de9bf2fac7a6b3ed0ca0d23926cf02e19484383a78eefeeba3af3459f6ee46f4.py
+debe595c69ddc5d19bfb189193b75923e0412f4ba00c833dc2d9b7bafb636fad  training/source_blobs/debe595c69ddc5d19bfb189193b75923e0412f4ba00c833dc2d9b7bafb636fad.py
+df16803b62d90800c51ad8072123067814d24a23d4a0569671fc44c3684b9f74  training/source_blobs/df16803b62d90800c51ad8072123067814d24a23d4a0569671fc44c3684b9f74.py
+df55a101f2bcf31f32d7e201e9b27071ab404f34b0e7df3f805a8ddb322e1f80  training/source_blobs/df55a101f2bcf31f32d7e201e9b27071ab404f34b0e7df3f805a8ddb322e1f80.py
+e00e99d07655b7fdf96a3164d182de3683f81f43817e93c5f8b89e6bf7130c91  training/source_blobs/e00e99d07655b7fdf96a3164d182de3683f81f43817e93c5f8b89e6bf7130c91.py
+e137daae9e39c642f5b2cabbfc3e2730446e399a61e44800901e5b9c17da84bd  training/source_blobs/e137daae9e39c642f5b2cabbfc3e2730446e399a61e44800901e5b9c17da84bd.py
+e141132efbd5f1beab4b187d4eab15efd46fcf9710d6a2ba7c882f62362a77cf  training/source_blobs/e141132efbd5f1beab4b187d4eab15efd46fcf9710d6a2ba7c882f62362a77cf.py
+e14e793f20d6b1830defcbb37ca498f2ecbfac7c6c08b02e2dfd3d1e06311286  training/source_blobs/e14e793f20d6b1830defcbb37ca498f2ecbfac7c6c08b02e2dfd3d1e06311286.py
+e178e766849e6c96fdea770a2bb307562c27bfb3761a81bd5eec12a3d84a7995  training/source_blobs/e178e766849e6c96fdea770a2bb307562c27bfb3761a81bd5eec12a3d84a7995.py
+e31d4cdcadea1fc6b0d8449098b06d946d58f308ad2e6efe7eac6b9c3f72ed98  training/source_blobs/e31d4cdcadea1fc6b0d8449098b06d946d58f308ad2e6efe7eac6b9c3f72ed98.py
+e3780f86fb01a1e34e44b78a42d0648f259f41c1d6a0afbfbb600c4d54c737c8  training/source_blobs/e3780f86fb01a1e34e44b78a42d0648f259f41c1d6a0afbfbb600c4d54c737c8.py
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  training/source_blobs/e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855.py
+e3c3977ad23a40b6dda4164996164dadc44b6424a56abd816d12076b81f5b143  training/source_blobs/e3c3977ad23a40b6dda4164996164dadc44b6424a56abd816d12076b81f5b143.py
+e4eb3d7799f756c95f8d714e59618c9171239db79ec9750781bcc5cbbcf4416a  training/source_blobs/e4eb3d7799f756c95f8d714e59618c9171239db79ec9750781bcc5cbbcf4416a.py
+e5285b4ce241af63d7038d2cb2470b2b397686ec904a9ca83f02b6fe1a1d7ce8  training/source_blobs/e5285b4ce241af63d7038d2cb2470b2b397686ec904a9ca83f02b6fe1a1d7ce8.py
+e57dbb7567a2438a497802b66d4a0188c53c3afdd9eb7c313334b8e312ad4999  training/source_blobs/e57dbb7567a2438a497802b66d4a0188c53c3afdd9eb7c313334b8e312ad4999.py
+e672d5bddefb2a79eaec3185ad436940522f352e762f265c461764050d323855  training/source_blobs/e672d5bddefb2a79eaec3185ad436940522f352e762f265c461764050d323855.py
+e6778ba8a8215f0b3c20386e768ea3e15b5939e7919935046689733eb96eba28  training/source_blobs/e6778ba8a8215f0b3c20386e768ea3e15b5939e7919935046689733eb96eba28.py
+e685abc5e4579508bf432632159c64d89fab91a1831235bcc0a0f12e4fa309f7  training/source_blobs/e685abc5e4579508bf432632159c64d89fab91a1831235bcc0a0f12e4fa309f7.py
+e73dd7b7ded430562a55380a8e5a5a5f52b8176f53277da1f0e23f48466b0e34  training/source_blobs/e73dd7b7ded430562a55380a8e5a5a5f52b8176f53277da1f0e23f48466b0e34.py
+e8dd3c8676b9cf5bc7ecd501f3afe82555c2c0d4f7700bb6e05417278447a995  training/source_blobs/e8dd3c8676b9cf5bc7ecd501f3afe82555c2c0d4f7700bb6e05417278447a995.py
+e915efcc450bd027272f029e6ff52c9ca9fc8f5b13bc593e4f019c6efaf191ca  training/source_blobs/e915efcc450bd027272f029e6ff52c9ca9fc8f5b13bc593e4f019c6efaf191ca.py
+ea92b8b5a884cce7459a1ef2957da8bf8f4a81a9165ffdca461dcc24a2fc8df4  training/source_blobs/ea92b8b5a884cce7459a1ef2957da8bf8f4a81a9165ffdca461dcc24a2fc8df4.py
+ebc86535e538107334e76c16ddfbfd2671a5635281f46d613f11ad68b627f866  training/source_blobs/ebc86535e538107334e76c16ddfbfd2671a5635281f46d613f11ad68b627f866.py
+ebd8bcd2612f1c023924e2d7ab6afaf3ea215bb1b3c4fdf451b54fa4c561dbce  training/source_blobs/ebd8bcd2612f1c023924e2d7ab6afaf3ea215bb1b3c4fdf451b54fa4c561dbce.py
+ec2d2174ac2524e68f2b20d789f52eb31953e4d0a4d1c5fb70f6516680b6a4ea  training/source_blobs/ec2d2174ac2524e68f2b20d789f52eb31953e4d0a4d1c5fb70f6516680b6a4ea.py
+ec619a6e34b3b53aa71aed4bb0b9ffe241f151243cd114d69401229b1fadbe72  training/source_blobs/ec619a6e34b3b53aa71aed4bb0b9ffe241f151243cd114d69401229b1fadbe72.py
+ed5eed8c69210ed743e47c0c4025c0b69114f02925f2ca8cd946911b311c6913  training/source_blobs/ed5eed8c69210ed743e47c0c4025c0b69114f02925f2ca8cd946911b311c6913.py
+eebc3ef49e86023666bdb9ed6068c5d96b68446be4db8f4a25f13712efdb7de0  training/source_blobs/eebc3ef49e86023666bdb9ed6068c5d96b68446be4db8f4a25f13712efdb7de0.py
+eff2f9cd32de9fa73563020d7b36c3a2b1c19d92022fed4f5c1d7c4b0cedc59c  training/source_blobs/eff2f9cd32de9fa73563020d7b36c3a2b1c19d92022fed4f5c1d7c4b0cedc59c.py
+f086fb95bfc7bc091598f528dc5c6ef0c164e8c85627947e7d1b9241ad429efe  training/source_blobs/f086fb95bfc7bc091598f528dc5c6ef0c164e8c85627947e7d1b9241ad429efe.py
+f24ee4feacb1484017893933d0ba2484197d8027a51e522b1e28e48f3b82983f  training/source_blobs/f24ee4feacb1484017893933d0ba2484197d8027a51e522b1e28e48f3b82983f.py
+f2a5e8599e7eb3a7026148dd57acb3fb5cf4311b3768f07831bcf7c740ead4a9  training/source_blobs/f2a5e8599e7eb3a7026148dd57acb3fb5cf4311b3768f07831bcf7c740ead4a9.py
+f6e985a2f49e4412c0685f7632740abf87aebc32d3c359346d9ce0b8af38f50c  training/source_blobs/f6e985a2f49e4412c0685f7632740abf87aebc32d3c359346d9ce0b8af38f50c.py
+f7f2f12b6b9056653aed9b2cc8cd928a6b0dbbaf74130421e3dc5cf8e03eaa1c  training/source_blobs/f7f2f12b6b9056653aed9b2cc8cd928a6b0dbbaf74130421e3dc5cf8e03eaa1c.py
+f80f065f6892eea951e94a7ff6925d5a1bff9419a7384e0756cf2faeb92e4775  training/source_blobs/f80f065f6892eea951e94a7ff6925d5a1bff9419a7384e0756cf2faeb92e4775.py
+f8d014e088704e7c366e91b61dbc6a959380908e4836eb6e3952a80b6a39768e  training/source_blobs/f8d014e088704e7c366e91b61dbc6a959380908e4836eb6e3952a80b6a39768e.py
+f99edeedb6cb328ef728d05643f65ea20371437f7e5806009ae3ad0885bafbf1  training/source_blobs/f99edeedb6cb328ef728d05643f65ea20371437f7e5806009ae3ad0885bafbf1.py
+f9cb70123441e889d53dc9c20d282826e4b6d4841562f2e3b21cca965427cf3d  training/source_blobs/f9cb70123441e889d53dc9c20d282826e4b6d4841562f2e3b21cca965427cf3d.py
+f9e0e6242ac3dbdc151241cbcca9d173765514e2fb7878648cc886143f1d5376  training/source_blobs/f9e0e6242ac3dbdc151241cbcca9d173765514e2fb7878648cc886143f1d5376.py
+fab59323c6fc8302dfd226b1b7e0cdef9c865231f72e930e9968fba2139565c3  training/source_blobs/fab59323c6fc8302dfd226b1b7e0cdef9c865231f72e930e9968fba2139565c3.py
+fb194687b4e39d11d79d33deb9eb4cc7e7d09b5f06424814263d38b9232d2b03  training/source_blobs/fb194687b4e39d11d79d33deb9eb4cc7e7d09b5f06424814263d38b9232d2b03.py
+fbe2b197e9f06a2916d7ef950cd2621aba2a8597faa58bc4b10789b4fef02761  training/source_blobs/fbe2b197e9f06a2916d7ef950cd2621aba2a8597faa58bc4b10789b4fef02761.py
+fc20fa387d2f794b572397d684d42d2013d517f383142f3033bf91c9ced2ccae  training/source_blobs/fc20fa387d2f794b572397d684d42d2013d517f383142f3033bf91c9ced2ccae.py
+fc74fbb0dc95c48b5f68c802efa4b1be36e59b6456c384f30a54f74a379381c6  training/source_blobs/fc74fbb0dc95c48b5f68c802efa4b1be36e59b6456c384f30a54f74a379381c6.py
+fc9d9862dd1f7f7c155ff62b4023a803db6752d21a43b7af6931041264aee00e  training/source_blobs/fc9d9862dd1f7f7c155ff62b4023a803db6752d21a43b7af6931041264aee00e.py
+fe42694806dede2847762153e12eb7b84ff461f47959c5b39189c4893df5863e  training/source_blobs/fe42694806dede2847762153e12eb7b84ff461f47959c5b39189c4893df5863e.py
+fef4e145fd16e2ec890b3113687c5cccc5ead04e3add2a65382f724cc911bc88  training/source_blobs/fef4e145fd16e2ec890b3113687c5cccc5ead04e3add2a65382f724cc911bc88.py
+ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957  training/source_blobs/ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957.py
+ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6  training/source_blobs/ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6.py
+ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7  training/source_blobs/ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7.py
+ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba  training/source_blobs/ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba.py

SOURCE_NOTICE.md

@@ -0,0 +1,9 @@
+# Source and model notice
+
+Copyright 2026 William Zebrowski.
+
+The selected Solstice-1 model weights, frozen tokenizer artifacts to the extent controlled by the author, and author-controlled project code/documentation are released under Apache License 2.0. See LICENSE. This is the author's explicit release decision for Solstice-1, rather than an inferred license from earlier OpenSML releases.
+
+Third-party software and training materials retain their own licenses and attribution. Apache-2.0 does not relicense upstream datasets, their underlying documents/code repositories, or separately licensed vendored components. Raw training datasets are not redistributed here. See THIRD_PARTY_NOTICES.md and docs/DATA.md.
+
+The native model implementation and tokenizer originate in the owner's OpenSML MLX research workspace. Original source hashes, exact inference weight identity and sanitized archival hashes are included. Examples and verification outputs describe this specific selected checkpoint; they are not guarantees of general accuracy or safety.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,32 @@
+# Third-party notices
+
+## Runtime and source
+
+MLX is an Apple open-source project: [ml-explore/mlx](https://github.com/ml-explore/mlx). NumPy and Hugging Face Tokenizers are dependencies distributed under their own terms. Their packages are installed separately and are not relicensed by Solstice-1.
+
+Archived training code includes SQLGlot source and its retained license where required by the historical code filter. Snapshot manifests preserve the license file mapping. Google IFEval evaluator source and its Apache-2.0 attribution are preserved with the evaluation archive. Dataset content is not made Apache-2.0 by being used in training or evaluation.
+
+## Training sources
+
+| Source | Pinned revision | Recorded terms / role |
+| --- | --- | --- |
+| [HuggingFaceTB/smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus/tree/3ba9d605774198c5868892d7a8deda78031a781f) | `3ba9d605774198c5868892d7a8deda78031a781f` | ODC-BY; preserve underlying source notices |
+| [HuggingFaceTB/dclm-edu](https://huggingface.co/datasets/HuggingFaceTB/dclm-edu/tree/dbad8ad71224482740cd9c9d353591adbf62fe04) | `dbad8ad71224482740cd9c9d353591adbf62fe04` | CC-BY-4.0 |
+| [HuggingFaceFW/finewiki](https://huggingface.co/datasets/HuggingFaceFW/finewiki/tree/8bd13e72e6a002407649b3e898535f42ceb1aeb9) | `8bd13e72e6a002407649b3e898535f42ceb1aeb9` | CC-BY-SA-4.0 / GFDL |
+| [hongliu9903/stack_edu_python](https://huggingface.co/datasets/hongliu9903/stack_edu_python/tree/76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376) | `76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376` | Original source repository licenses; preserve provenance and notices |
+| [hongliu9903/stack_edu_sql](https://huggingface.co/datasets/hongliu9903/stack_edu_sql/tree/079ce8b93c0e83a767ee61af20b9de9931a7e07c) | `079ce8b93c0e83a767ee61af20b9de9931a7e07c` | Original source repository licenses; preserve provenance and notices |
+| [iNeil77/SynSQL-2.5M](https://huggingface.co/datasets/iNeil77/SynSQL-2.5M/tree/56fed46215366defbbc3d7d13064349bbd990a8f) | `56fed46215366defbbc3d7d13064349bbd990a8f` | See upstream source terms |
+| [HuggingFaceTB/smol-smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/tree/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc) | `f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc` | apache-2.0 |
+| [HuggingFaceTB/smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk/tree/5feaf2fd3ffca7c237fc38d1861bc30365d48ffa) | `5feaf2fd3ffca7c237fc38d1861bc30365d48ffa` | see upstream component license |
+| [rajpurkar/squad_v2](https://huggingface.co/datasets/rajpurkar/squad_v2/tree/3ffb306f725f7d2ce8394bc1873b24868140c412) | `3ffb306f725f7d2ce8394bc1873b24868140c412` | ['cc-by-sa-4.0'] |
+| [allenai/sciq](https://huggingface.co/datasets/allenai/sciq/tree/2c94ad3e1aafab77146f384e23536f97a4849815) | `2c94ad3e1aafab77146f384e23536f97a4849815` | cc-by-nc-3.0 |
+
+SciQ is recorded as CC-BY-NC-3.0 and SQuAD v2 as CC-BY-SA-4.0. SmolTalk-family synthetic responses retain their upstream provenance. FineWeb-related database terms and underlying webpage rights are separate. The Python/SQL mirrors carry heterogeneous original repository notices and did not require an explicit license on every retained record. These statements describe the frozen records, not a newly completed item-by-item rights audit.
+
+## Evaluation sources
+
+[ARC](https://huggingface.co/datasets/allenai/ai2_arc), [PIQA](https://huggingface.co/datasets/baber/piqa), [HellaSwag](https://huggingface.co/datasets/Rowan/hellaswag), and [IFEval](https://huggingface.co/datasets/google/IFEval) are external evaluation datasets. Their versions and source code hashes are in the public result records. The repository distributes aggregate scores, authored checks and model-generated answers as documented; full public benchmark datasets are obtained separately under upstream terms.
+
+## Prior OpenSML references
+
+[OpenSML-150M](https://huggingface.co/OpenSML/OpenSML-150M) and [OpenSML-JSONLite-175M](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) are earlier releases by the same author. Their weights are not ancestors of Solstice-1. Prior documentation layouts informed this package; their task-specific results are not silently assigned to this model.

TRAINING_AND_REPRODUCIBILITY.md

@@ -0,0 +1,40 @@
+# Training and reproducibility
+
+## What is reproducible from this package
+
+The selected FP32 inference weights, tokenizer, architecture, native serialization and runner are supplied. The release verification records loading, example output parity, token/control boundaries and package integrity checks. The existing complete public evaluations identify these exact weight bytes.
+
+The package includes frozen source snapshots for every recorded pretraining job in the inherited lineage and for the three selected SFT stages. Identical files are stored once under `training/source_blobs`; snapshot manifests map their original module paths to these blobs and record original versus sanitized hashes. Use:
+
+```bash
+python scripts/materialize_training_snapshot.py --list
+python scripts/materialize_training_snapshot.py --snapshot sft3 --output /tmp/solstice-sft3-source
+```
+
+The materializer creates a new directory and never overwrites an existing one. Read the resulting source and the corresponding `configs/` plan before attempting a training launch. Historical entry points include `sml_v1.pretrain`, `sft.lora_train`, `sft.lora_search_train`, and `sft.lora_repair_train`; they are archival training programs with dependencies on original run assets, rather than a verified one-command retraining interface.
+
+## Required external inputs for exact retraining
+
+- Pinned upstream datasets and the same filtering, splits, row identities and deterministic order.
+- The original pretrained initialization and stream history for a full from-scratch reproduction.
+- Exact ancestor base masters and adapters for a stage-only reproduction.
+- Original prepared corpora, prose-replay pools, data receipts, comparison assets and optimizer/resume state where the selected stage imports them.
+- A compatible distributed MLX/JACCL environment for pretraining. The public inference wheel does not itself establish five-worker training parity.
+
+Raw source corpora, prepared training pools, optimizer binaries and all intermediate checkpoint weights are not redistributed. Full selected-pool file hashes, dataset revisions, plans, counters and later-stage selected batch indices are included. These establish provenance and permit auditing when the original inputs are available; they do not make unavailable corpora reconstructible from a hash.
+
+## Stage boundaries and counters
+
+The selected base contains 42,001,367,040 pretraining positions. SFT stages add 2,000, 256 and 128 updates respectively. The final stage was planned for 256 updates, but its later updates are not inherited. `provenance/SELECTED_PREFIX.json` distinguishes complete pools from the actually selected prefix and distinguishes unique rows from repeated presentations.
+
+The original training runtime used MLX `0.32.3.dev20260920+7241f12`. Packaged inference uses the pinned stock runtime listed in requirements and verification. A complete fresh training rerun was not performed to validate numerical equivalence between these runtimes.
+
+## Source sanitization and integrity
+
+Absolute workspace/user paths become `${WORKSPACE}` / `${USER_HOME}` in archived configuration and source. Local network addresses are removed where present. Snapshot manifests retain original hashes and packaged hashes separately. Replacing placeholders and satisfying historical file guards is an explicit reproduction prerequisite. Private SSH configurations, network inventories, credentials and raw operational logs are not included.
+
+The release source does not pretend that every imported helper recipe was a model ancestor. Only the phase/configuration and checkpoint lineage records identify selected training exposure. In particular, 167, 1200, 417, previous full-weight SFT attempts and DPO experiments are not additional ancestors of Solstice-1.
+
+## Evaluations
+
+`evaluation/` contains the recorded native scorer entry points and their dependencies/provenance. Frozen source hashes identify the historical implementation. Public metric JSON files include dataset revisions, count checks, precision and inference settings. New runs should use a fresh output directory and preserve generated responses. Package verification is deliberately reported separately from a full benchmark rerun.

config.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "step_0000128_5937a6def881",
+  "compute_precision": "float32",
+  "context_limit": 2048,
+  "format": "opensml-native-chat-v1",
+  "license": "apache-2.0",
+  "model": {
+    "attention_impl": "vanilla",
+    "bias": false,
+    "ce_impl": "reference",
+    "d_model": 576,
+    "ffn_impl": "reference",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "model_id": "OpenSML/Solstice-1",
+  "model_name": "Solstice-1",
+  "organization": "OpenSML",
+  "parameter_label": "175M",
+  "parameters": 174202432,
+  "serialization": "explicit-ids-assistant-mask-v1",
+  "tokenizer_fingerprint": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb",
+  "weight_sha256": "7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515"
+}

configs/pretrain_20261022_121457_b3d5f398/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261022_121457_b3d5f398/recipe.json

@@ -0,0 +1,65 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "seed": 7337,
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261024_111140_5f596b7a/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261024_111140_5f596b7a/recipe.json

@@ -0,0 +1,87 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261027_222615_4d065da7/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261027_222615_4d065da7/recipe.json

@@ -0,0 +1,124 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261028_093017_dd42e713/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261028_093017_dd42e713/recipe.json

@@ -0,0 +1,141 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "gradual_decay_transition": {
+    "decay_end_tokens": 100000000000,
+    "floor_lr": 3e-05,
+    "format": "sml-v2-open-ended-gradual-v1",
+    "parent_contract": {
+      "code": "1bc7d03c972319212ea775d4667fc0d3f5c0d2d738f3bfff4967c0093913df00",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "19905de3f7b0d4309d2a3341ad393ebb7934e8ef53fe7822afc97f5197acdfff",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "f6238b6fd8a3447dd4cf669f83ab0705bc352e8b5b58bc5b46533246f97fd5f4",
+    "parent_step": 129579,
+    "shape": "linear",
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 16984178688,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261102_234538_0ee851dc/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261102_234538_0ee851dc/recipe.json

@@ -0,0 +1,141 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "gradual_decay_transition": {
+    "decay_end_tokens": 100000000000,
+    "floor_lr": 3e-05,
+    "format": "sml-v2-open-ended-gradual-v1",
+    "parent_contract": {
+      "code": "1bc7d03c972319212ea775d4667fc0d3f5c0d2d738f3bfff4967c0093913df00",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "19905de3f7b0d4309d2a3341ad393ebb7934e8ef53fe7822afc97f5197acdfff",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "f6238b6fd8a3447dd4cf669f83ab0705bc352e8b5b58bc5b46533246f97fd5f4",
+    "parent_step": 129579,
+    "shape": "linear",
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 16984178688,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261103_000432_8ac2605e/corpus.json

@@ -0,0 +1,58 @@
+{
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "research-20b-live-english-prose-v1",
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 70
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 20
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 10
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261103_000432_8ac2605e/recipe.json

@@ -0,0 +1,141 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "gradual_decay_transition": {
+    "decay_end_tokens": 100000000000,
+    "floor_lr": 3e-05,
+    "format": "sml-v2-open-ended-gradual-v1",
+    "parent_contract": {
+      "code": "1bc7d03c972319212ea775d4667fc0d3f5c0d2d738f3bfff4967c0093913df00",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "19905de3f7b0d4309d2a3341ad393ebb7934e8ef53fe7822afc97f5197acdfff",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "f6238b6fd8a3447dd4cf669f83ab0705bc352e8b5b58bc5b46533246f97fd5f4",
+    "parent_step": 129579,
+    "shape": "linear",
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 16984178688,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261104_170916_448e6f64/corpus.json

@@ -0,0 +1,294 @@
+{
+  "code_validation": {
+    "excluded_content_keys": [
+      "08b5b1ba858dcc4735bb66d304552d51c7b4cb9e09e0c4a5f76ef9847bb6d280",
+      "08cf74e1f4e2c0cd0a6da7e3967c6078df34d77a3f821b46e50c4f7f3fa105cd",
+      "09e9c4d3e6531b853abfc336eb1108890938bfaef18c130dcb34e755fe2c4152",
+      "0d39776a10139d74bcf4be8dd4d5995e79250ddaf0f3aa774ec04fb6ddcb345e",
+      "0d5657b52456eb5c86def02acf3e48318a02f84156ecce6a93b5b163237a54fa",
+      "122be194cc01c08ece5ca2963889fa0f832a9dea93df4b2167df2fb5c52edde9",
+      "1d551c6746418a80059b8a91d59edda84702237c34e91164595eb71b0f407bd4",
+      "226a6abf548fecd027726c22c1955000cb2564cd4cdd59671718b437d1a72539",
+      "2272dc1f261d64ba4a5408767a0c4a4657df2e529896a66914be92dc5895f2ea",
+      "237d921824da9070732b02963c62f472f193b337de7e6b3b71b45f6e326054b2",
+      "2629b2da8684627610d62d8c4d3d85200fec971f88a19300a20495f8dc74a512",
+      "2c38a41f5e6feb03fc2247a084b91916402619870843e22fb53fbc2a8fd330ef",
+      "31d69258a6797aa6563792d0782586ab688e118bed55fad6aa046320c2a70f35",
+      "3632860aeca3ea4b5d684f9c033ed5cc8d3d37cb85eab5526f44e0a578f468b4",
+      "36dde4bc7d0745c15ff12932b830183b79f5a6302bf800d38b58cfdac37f7660",
+      "370e8a14410af597be9dbca64b2f810227ea9ab0d6ff8e25861aad41762cd8f0",
+      "37a28d12641f1f861c6714fcb39d6b384d087ccc8882cdcaeb493041279b1b7a",
+      "389e360ee797751c34139fb80e438c09bc2e633c6eccf84207d4474e7904fa51",
+      "38b715099c0c8ff8a29da8ee6aab0de70511644ecd3536eaaf1daa3c9272a438",
+      "3c4fd1ec175f0d607bc76f0fc403b09bd89f00f39686f4267fbfd579138a699d",
+      "3d4e11b3cadc6694af7692270c0e5fc948dd8a4facf8c541237e6deaf95ce249",
+      "3e552d387eeadfcd6eee143a6448580edc77c3922598e05f49b5a02b13067f4d",
+      "3fcd7df015ad72401934b00142c6b376fc8224f8da40228aa281fde483755340",
+      "418e27e85222a65b6d12a5410b18004231c0a272df29ce7ed12d9f791b16249d",
+      "44a070c90eafff5a7a75a3100acf6625d8b53c988e9eae25f46938eaf606c77e",
+      "47551c7fdcc40c723ffbfb208fef3a50dca21429a3a919a571edbc2449da5ec3",
+      "4ba039726827e69cab1d99be769d9557f9afde6a324a0afbfb8bc2e13f3108e0",
+      "4e82c21a44381ffdd7267a55cb62bd895e67becb654859dcab77927a260787f0",
+      "518c0cb14dc7a6209c53d126f063084e43868bba49e7ad6209bbf7f83e717547",
+      "5329ebfcb0b7909c99fd9263cd32fab1d4c1127214370d0ef7eca45265280454",
+      "535f61e917bc83dd829c1e5879e617d3eb2a10adf2c24202a6428ce8a1f19564",
+      "56a15f625a58b7bfe2ef7f5c8a1e3f997cc655da612f92fba410ba0b40a5c88b",
+      "5903890e66f7fbcba23d7b8a660c854ab38b1032ec93b9e37000ef68c49187dd",
+      "5d2d9776d96b3b3fb7442d05fd6a19884f7c15885adc2a9f6f63beb801b653c3",
+      "66200637354d0023e5bfc1267129d30cc741207a56485da56d507a5027dd0f1e",
+      "6919decb44a2962f4016a5a80ad5b3e1418589547ee4e44eed44e8ae49c5d9d5",
+      "69296bb650fc3edb51fc22a87570c8fe74ae5470822f5b8a8d7601846f882510",
+      "69b1d10f8738d815c86ca6dbf4b25a4fb7e026dc82ebbc10d77a0705c33558be",
+      "6def147f96b57191c11b3e5837244b9cb2782d4fcbac2cab5ec117e9e057dcd2",
+      "6f2fd18b4416df36e733e50560f857ba928d8f5a969dbd8b4f90a9ad1893d664",
+      "6fa61db6aa2c6c57642d868a0c8de647a5d589feafa7113d3fdbea54ed00564d",
+      "70def7eb6110256e84436e057be4257187df903616a924e55567517b922aa9dc",
+      "70f0b4ca4af111a506c62323277135a95cbc850584adbbd2b328855450f093f1",
+      "776e9930bc4fb4e3ae4027dc5d911a0ae03aa5cdb1ec55e6e67cd701cd04735c",
+      "806a4ecae59173c6f899197db153877670d14c8ced2cffca1f7230c40a4138ce",
+      "890b111b6bf7e6cb62598a049dafa5fbf0d1b20cd00d6816b0bd74363139b8d1",
+      "8ca2ff38d8d25fe96382d157d660a95e64581f194c907ecf2bfef8dc091f21ec",
+      "a7ebe97f86560ac023f5c7032f2ae5e0b156235ffebaef667012d72c090dcf78",
+      "b12d9d5665a7a7c1276d2c66a1e0b248f8efedb132c42d7058f4faf86538abd1",
+      "b31ba3a54899565789422821d9069be1a08adc6f1933934a7522f32f77ee62d9",
+      "b41b2b21b7df3ae441ef4b127ce894b74ab929d3f2100d02726fdb0dac61fe06",
+      "b6919fef0a28f857fb5659a6919dd8e7a4d46477bffb18fe33653bcf58fc669b",
+      "bb57476f00cf4dd574bd4364cec4da646706c778f6570e5c62c4b4da58289dad",
+      "bb9f2e9774508071947a739285a72b2e75ad7e33fe3942936c5adeb9ad741851",
+      "c261eaf172509ed43f1e733dcc08e79e55dc5f7b6e6c63d834071fa6823ec800",
+      "c32e278f6c6b67fa5de35e8a3df3ae90f7b1e285590da5c7932f5286ee7c0439",
+      "c7d120103d1cf9fc3142cb36645bd54543072f579fd0cd4af64158fd797341af",
+      "c81cf714da2e5cf627222a910a602c42ff2b445e9646e1bf1ce5ea4a480763af",
+      "c93274b966ce10113e9559fdd00e758a8dcd1b4fba5dc49d2f50b2a9b5126e4b",
+      "ca60dfcd82dc85c7f1aac743c9ebc799721b59a404f25717e8d45ca053802160",
+      "cb8e3691b4513a063f738f9b0e2adaf7ad6f54e033b700c2f3eae0928c356df2",
+      "ce298eec02cdda9b80999d6d4d5f83e98d8011c742d3af26c4f92a794bc2cc85",
+      "d126c3c9fd2ddf8f774e605e50ebfb82da972cd17b77b3ed7061c2910cdbb09b",
+      "d160f7f0150624a438a259ac94c4fa5fc72ae656678b764af153a38c4cb4a280",
+      "d5f7d6a12cd4b124fbbcb8ab893bb20f030ecce4def4d21c51079599c01542fa",
+      "d6f8e71eb189de2073f985bf055252ba8ba0b6853ca10562e9db45ab09461750",
+      "d98e902d44585100466d5e5991280decb93b938f74d96c0a6d3ab642e6f05407",
+      "dacf411bc176fa1c64f5656606848ac167f63a52e7c235f2a21453046ecb66db",
+      "dec4288d321e694d1a15dadabbd882f75fbfd151d487bbfa7eecd7905a049c6d",
+      "df04a03f7ebe2bdb6b65c40178342f3abf3655442b5ee9ace187093a4b25b4e1",
+      "e2ea5f987fe3ddad84ea91169baa6931a69e82c2827ba1c88989267a83a12e47",
+      "e5083ed8625be5f1b504ea1e6505d0f7390cdd3364a292bbfaa00efbe45333eb",
+      "e5b14324d302b7849c871328b2e65fbe957fefa941e10314bdc1b693b9f449c3",
+      "e9c0bef83bed2a0077d649defd6b96fb572eb9e5feefe943748938e612176fbc",
+      "ebeacc07897d24a98c18af4307545e213e5d234b8bb358cf3f4b868d97ce8f2c",
+      "ec9af0ec16d18cf655ba0ae197488c076e0a5b6dfddc21a567a573b525ab5183",
+      "ed0252613ae99ddda19bf87323b80e370cb0f0bbae5c2263655a3ce4aaad4464",
+      "f05dac8db278593a240548d207e047a61dab771c54903933c9cad0b28045905f",
+      "f51376acd8a20affee0e785e25c86d081aa74253ee29d6327263a59a5bdfe26c",
+      "fc7812e727a9b2a19481d9f8633fb62bd7ff347964246e5fe5d987aa48ef8f17",
+      "fd1fa0111249a7806808497f416d423641435652f373bb9e4cfa7c3650d68c1e",
+      "fd9734ecdefbb34e470ca31c77bd71b49317e0ab7797200f170152ab1cf63c20"
+    ],
+    "format": "sml-python-sql-heldout-v1",
+    "seq_len": 2048,
+    "sequences": 16,
+    "sha256": "641529f48090540b75bb444deca63680f73ffd7b542639398a2d6c47f4e9e170",
+    "sources": "e9ce1db0add5eb8b8ba38e403c6efdf30ef0106a97c298fc0ef1eb8593f2b39e",
+    "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+  },
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "deep44-prose80-python15-sql5-public-v1",
+  "reference_corpus": {
+    "max_bytes": 200000,
+    "max_rows_per_source": 50000000,
+    "min_chars": 200,
+    "name": "research-20b-live-english-prose-v1",
+    "seed": 7337,
+    "shuffle_buffer": 10000,
+    "sources": [
+      {
+        "columns": [
+          "text",
+          "id"
+        ],
+        "config": "fineweb-edu-dedup",
+        "label": "web",
+        "license": "ODC-BY; preserve underlying source notices",
+        "repo": "HuggingFaceTB/smollm-corpus",
+        "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+        "split": "train",
+        "text_field": "text",
+        "weight": 70
+      },
+      {
+        "columns": [
+          "text",
+          "id",
+          "edu_int_score"
+        ],
+        "label": "dclm",
+        "license": "CC-BY-4.0",
+        "min_score": {
+          "field": "edu_int_score",
+          "value": 3
+        },
+        "repo": "HuggingFaceTB/dclm-edu",
+        "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+        "split": "train",
+        "text_field": "text",
+        "weight": 20
+      },
+      {
+        "columns": [
+          "text",
+          "id"
+        ],
+        "config": "en",
+        "label": "wiki",
+        "license": "CC-BY-SA-4.0 / GFDL",
+        "repo": "HuggingFaceFW/finewiki",
+        "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+        "split": "train",
+        "text_field": "text",
+        "weight": 10
+      }
+    ],
+    "split_buckets": 10000,
+    "text_filter": "english-prose-v1"
+  },
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 56
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 16
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 8
+    },
+    {
+      "code_filter": {
+        "language": "Python",
+        "require_explicit_license": false
+      },
+      "columns": [
+        "blob_id",
+        "repo_name",
+        "path",
+        "src_encoding",
+        "length_bytes",
+        "int_score",
+        "detected_licenses",
+        "license_type",
+        "language",
+        "text",
+        "download_success"
+      ],
+      "config": "default",
+      "label": "python",
+      "license": "Original source repository licenses; preserve provenance and notices",
+      "min_score": {
+        "field": "int_score",
+        "value": 3
+      },
+      "repo": "hongliu9903/stack_edu_python",
+      "revision": "76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376",
+      "split": "train",
+      "text_field": "text",
+      "upstream_config": "Python",
+      "upstream_repo": "HuggingFaceTB/stack-edu",
+      "upstream_revision": "eeec5caac5cc3758a18f1d3ba4416837a9ba814c",
+      "weight": 15
+    },
+    {
+      "additional_sources": [
+        {
+          "columns": [
+            "db_id",
+            "schema",
+            "sql",
+            "question",
+            "external_knowledge"
+          ],
+          "config": "default",
+          "renderer": "synsql-schema-query-v1",
+          "repo": "iNeil77/SynSQL-2.5M",
+          "revision": "56fed46215366defbbc3d7d13064349bbd990a8f",
+          "split": "train",
+          "upstream_repo": "seeklhy/SynSQL-2.5M",
+          "upstream_revision": "cca2c84cc3b41afa6b51534762a6a3a4a420baca"
+        }
+      ],
+      "code_filter": {
+        "language": "SQL",
+        "require_explicit_license": false
+      },
+      "columns": [
+        "blob_id",
+        "repo_name",
+        "path",
+        "src_encoding",
+        "length_bytes",
+        "int_score",
+        "detected_licenses",
+        "license_type",
+        "language",
+        "text",
+        "download_success"
+      ],
+      "config": "default",
+      "label": "sql",
+      "license": "Original source repository licenses; preserve provenance and notices",
+      "min_score": {
+        "field": "int_score",
+        "value": 3
+      },
+      "repo": "hongliu9903/stack_edu_sql",
+      "revision": "079ce8b93c0e83a767ee61af20b9de9931a7e07c",
+      "split": "train",
+      "text_field": "text",
+      "upstream_config": "SQL",
+      "upstream_repo": "HuggingFaceTB/stack-edu",
+      "upstream_revision": "eeec5caac5cc3758a18f1d3ba4416837a9ba814c",
+      "weight": 5
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261104_170916_448e6f64/recipe.json

@@ -0,0 +1,155 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "code_data_transition": {
+    "format": "sml-prose-python-sql-phase-v1",
+    "manifest": "42ba571cb16dccbf79e65781e82a21881c78c244bb16fac46c40a6e0f123b1e2",
+    "parent_contract": {
+      "code": "b7ff0ae2c1f095bd36300267f5dc2fd60d58be71d828a1b85a796ddf176e942e",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "4f6def65dbd6e1f236be703c2ab977928a55f6bb144ef33688a6e3b0488aacb3",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "051bad7730581c758ce7f469b0214d56cf2e8a9b2df6791c1066b1290881361f",
+    "parent_step": 280771,
+    "start_tokens": 36801216512,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "gradual_decay_transition": {
+    "decay_end_tokens": 100000000000,
+    "floor_lr": 3e-05,
+    "format": "sml-v2-open-ended-gradual-v1",
+    "parent_contract": {
+      "code": "1bc7d03c972319212ea775d4667fc0d3f5c0d2d738f3bfff4967c0093913df00",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "19905de3f7b0d4309d2a3341ad393ebb7934e8ef53fe7822afc97f5197acdfff",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "f6238b6fd8a3447dd4cf669f83ab0705bc352e8b5b58bc5b46533246f97fd5f4",
+    "parent_step": 129579,
+    "shape": "linear",
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 16984178688,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/pretrain_20261107_142423_40700422/corpus.json

@@ -0,0 +1,294 @@
+{
+  "code_validation": {
+    "excluded_content_keys": [
+      "08b5b1ba858dcc4735bb66d304552d51c7b4cb9e09e0c4a5f76ef9847bb6d280",
+      "08cf74e1f4e2c0cd0a6da7e3967c6078df34d77a3f821b46e50c4f7f3fa105cd",
+      "09e9c4d3e6531b853abfc336eb1108890938bfaef18c130dcb34e755fe2c4152",
+      "0d39776a10139d74bcf4be8dd4d5995e79250ddaf0f3aa774ec04fb6ddcb345e",
+      "0d5657b52456eb5c86def02acf3e48318a02f84156ecce6a93b5b163237a54fa",
+      "122be194cc01c08ece5ca2963889fa0f832a9dea93df4b2167df2fb5c52edde9",
+      "1d551c6746418a80059b8a91d59edda84702237c34e91164595eb71b0f407bd4",
+      "226a6abf548fecd027726c22c1955000cb2564cd4cdd59671718b437d1a72539",
+      "2272dc1f261d64ba4a5408767a0c4a4657df2e529896a66914be92dc5895f2ea",
+      "237d921824da9070732b02963c62f472f193b337de7e6b3b71b45f6e326054b2",
+      "2629b2da8684627610d62d8c4d3d85200fec971f88a19300a20495f8dc74a512",
+      "2c38a41f5e6feb03fc2247a084b91916402619870843e22fb53fbc2a8fd330ef",
+      "31d69258a6797aa6563792d0782586ab688e118bed55fad6aa046320c2a70f35",
+      "3632860aeca3ea4b5d684f9c033ed5cc8d3d37cb85eab5526f44e0a578f468b4",
+      "36dde4bc7d0745c15ff12932b830183b79f5a6302bf800d38b58cfdac37f7660",
+      "370e8a14410af597be9dbca64b2f810227ea9ab0d6ff8e25861aad41762cd8f0",
+      "37a28d12641f1f861c6714fcb39d6b384d087ccc8882cdcaeb493041279b1b7a",
+      "389e360ee797751c34139fb80e438c09bc2e633c6eccf84207d4474e7904fa51",
+      "38b715099c0c8ff8a29da8ee6aab0de70511644ecd3536eaaf1daa3c9272a438",
+      "3c4fd1ec175f0d607bc76f0fc403b09bd89f00f39686f4267fbfd579138a699d",
+      "3d4e11b3cadc6694af7692270c0e5fc948dd8a4facf8c541237e6deaf95ce249",
+      "3e552d387eeadfcd6eee143a6448580edc77c3922598e05f49b5a02b13067f4d",
+      "3fcd7df015ad72401934b00142c6b376fc8224f8da40228aa281fde483755340",
+      "418e27e85222a65b6d12a5410b18004231c0a272df29ce7ed12d9f791b16249d",
+      "44a070c90eafff5a7a75a3100acf6625d8b53c988e9eae25f46938eaf606c77e",
+      "47551c7fdcc40c723ffbfb208fef3a50dca21429a3a919a571edbc2449da5ec3",
+      "4ba039726827e69cab1d99be769d9557f9afde6a324a0afbfb8bc2e13f3108e0",
+      "4e82c21a44381ffdd7267a55cb62bd895e67becb654859dcab77927a260787f0",
+      "518c0cb14dc7a6209c53d126f063084e43868bba49e7ad6209bbf7f83e717547",
+      "5329ebfcb0b7909c99fd9263cd32fab1d4c1127214370d0ef7eca45265280454",
+      "535f61e917bc83dd829c1e5879e617d3eb2a10adf2c24202a6428ce8a1f19564",
+      "56a15f625a58b7bfe2ef7f5c8a1e3f997cc655da612f92fba410ba0b40a5c88b",
+      "5903890e66f7fbcba23d7b8a660c854ab38b1032ec93b9e37000ef68c49187dd",
+      "5d2d9776d96b3b3fb7442d05fd6a19884f7c15885adc2a9f6f63beb801b653c3",
+      "66200637354d0023e5bfc1267129d30cc741207a56485da56d507a5027dd0f1e",
+      "6919decb44a2962f4016a5a80ad5b3e1418589547ee4e44eed44e8ae49c5d9d5",
+      "69296bb650fc3edb51fc22a87570c8fe74ae5470822f5b8a8d7601846f882510",
+      "69b1d10f8738d815c86ca6dbf4b25a4fb7e026dc82ebbc10d77a0705c33558be",
+      "6def147f96b57191c11b3e5837244b9cb2782d4fcbac2cab5ec117e9e057dcd2",
+      "6f2fd18b4416df36e733e50560f857ba928d8f5a969dbd8b4f90a9ad1893d664",
+      "6fa61db6aa2c6c57642d868a0c8de647a5d589feafa7113d3fdbea54ed00564d",
+      "70def7eb6110256e84436e057be4257187df903616a924e55567517b922aa9dc",
+      "70f0b4ca4af111a506c62323277135a95cbc850584adbbd2b328855450f093f1",
+      "776e9930bc4fb4e3ae4027dc5d911a0ae03aa5cdb1ec55e6e67cd701cd04735c",
+      "806a4ecae59173c6f899197db153877670d14c8ced2cffca1f7230c40a4138ce",
+      "890b111b6bf7e6cb62598a049dafa5fbf0d1b20cd00d6816b0bd74363139b8d1",
+      "8ca2ff38d8d25fe96382d157d660a95e64581f194c907ecf2bfef8dc091f21ec",
+      "a7ebe97f86560ac023f5c7032f2ae5e0b156235ffebaef667012d72c090dcf78",
+      "b12d9d5665a7a7c1276d2c66a1e0b248f8efedb132c42d7058f4faf86538abd1",
+      "b31ba3a54899565789422821d9069be1a08adc6f1933934a7522f32f77ee62d9",
+      "b41b2b21b7df3ae441ef4b127ce894b74ab929d3f2100d02726fdb0dac61fe06",
+      "b6919fef0a28f857fb5659a6919dd8e7a4d46477bffb18fe33653bcf58fc669b",
+      "bb57476f00cf4dd574bd4364cec4da646706c778f6570e5c62c4b4da58289dad",
+      "bb9f2e9774508071947a739285a72b2e75ad7e33fe3942936c5adeb9ad741851",
+      "c261eaf172509ed43f1e733dcc08e79e55dc5f7b6e6c63d834071fa6823ec800",
+      "c32e278f6c6b67fa5de35e8a3df3ae90f7b1e285590da5c7932f5286ee7c0439",
+      "c7d120103d1cf9fc3142cb36645bd54543072f579fd0cd4af64158fd797341af",
+      "c81cf714da2e5cf627222a910a602c42ff2b445e9646e1bf1ce5ea4a480763af",
+      "c93274b966ce10113e9559fdd00e758a8dcd1b4fba5dc49d2f50b2a9b5126e4b",
+      "ca60dfcd82dc85c7f1aac743c9ebc799721b59a404f25717e8d45ca053802160",
+      "cb8e3691b4513a063f738f9b0e2adaf7ad6f54e033b700c2f3eae0928c356df2",
+      "ce298eec02cdda9b80999d6d4d5f83e98d8011c742d3af26c4f92a794bc2cc85",
+      "d126c3c9fd2ddf8f774e605e50ebfb82da972cd17b77b3ed7061c2910cdbb09b",
+      "d160f7f0150624a438a259ac94c4fa5fc72ae656678b764af153a38c4cb4a280",
+      "d5f7d6a12cd4b124fbbcb8ab893bb20f030ecce4def4d21c51079599c01542fa",
+      "d6f8e71eb189de2073f985bf055252ba8ba0b6853ca10562e9db45ab09461750",
+      "d98e902d44585100466d5e5991280decb93b938f74d96c0a6d3ab642e6f05407",
+      "dacf411bc176fa1c64f5656606848ac167f63a52e7c235f2a21453046ecb66db",
+      "dec4288d321e694d1a15dadabbd882f75fbfd151d487bbfa7eecd7905a049c6d",
+      "df04a03f7ebe2bdb6b65c40178342f3abf3655442b5ee9ace187093a4b25b4e1",
+      "e2ea5f987fe3ddad84ea91169baa6931a69e82c2827ba1c88989267a83a12e47",
+      "e5083ed8625be5f1b504ea1e6505d0f7390cdd3364a292bbfaa00efbe45333eb",
+      "e5b14324d302b7849c871328b2e65fbe957fefa941e10314bdc1b693b9f449c3",
+      "e9c0bef83bed2a0077d649defd6b96fb572eb9e5feefe943748938e612176fbc",
+      "ebeacc07897d24a98c18af4307545e213e5d234b8bb358cf3f4b868d97ce8f2c",
+      "ec9af0ec16d18cf655ba0ae197488c076e0a5b6dfddc21a567a573b525ab5183",
+      "ed0252613ae99ddda19bf87323b80e370cb0f0bbae5c2263655a3ce4aaad4464",
+      "f05dac8db278593a240548d207e047a61dab771c54903933c9cad0b28045905f",
+      "f51376acd8a20affee0e785e25c86d081aa74253ee29d6327263a59a5bdfe26c",
+      "fc7812e727a9b2a19481d9f8633fb62bd7ff347964246e5fe5d987aa48ef8f17",
+      "fd1fa0111249a7806808497f416d423641435652f373bb9e4cfa7c3650d68c1e",
+      "fd9734ecdefbb34e470ca31c77bd71b49317e0ab7797200f170152ab1cf63c20"
+    ],
+    "format": "sml-python-sql-heldout-v1",
+    "seq_len": 2048,
+    "sequences": 16,
+    "sha256": "641529f48090540b75bb444deca63680f73ffd7b542639398a2d6c47f4e9e170",
+    "sources": "e9ce1db0add5eb8b8ba38e403c6efdf30ef0106a97c298fc0ef1eb8593f2b39e",
+    "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+  },
+  "max_bytes": 200000,
+  "max_rows_per_source": 50000000,
+  "min_chars": 200,
+  "name": "deep44-prose80-python15-sql5-public-v1",
+  "reference_corpus": {
+    "max_bytes": 200000,
+    "max_rows_per_source": 50000000,
+    "min_chars": 200,
+    "name": "research-20b-live-english-prose-v1",
+    "seed": 7337,
+    "shuffle_buffer": 10000,
+    "sources": [
+      {
+        "columns": [
+          "text",
+          "id"
+        ],
+        "config": "fineweb-edu-dedup",
+        "label": "web",
+        "license": "ODC-BY; preserve underlying source notices",
+        "repo": "HuggingFaceTB/smollm-corpus",
+        "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+        "split": "train",
+        "text_field": "text",
+        "weight": 70
+      },
+      {
+        "columns": [
+          "text",
+          "id",
+          "edu_int_score"
+        ],
+        "label": "dclm",
+        "license": "CC-BY-4.0",
+        "min_score": {
+          "field": "edu_int_score",
+          "value": 3
+        },
+        "repo": "HuggingFaceTB/dclm-edu",
+        "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+        "split": "train",
+        "text_field": "text",
+        "weight": 20
+      },
+      {
+        "columns": [
+          "text",
+          "id"
+        ],
+        "config": "en",
+        "label": "wiki",
+        "license": "CC-BY-SA-4.0 / GFDL",
+        "repo": "HuggingFaceFW/finewiki",
+        "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+        "split": "train",
+        "text_field": "text",
+        "weight": 10
+      }
+    ],
+    "split_buckets": 10000,
+    "text_filter": "english-prose-v1"
+  },
+  "seed": 7337,
+  "shuffle_buffer": 10000,
+  "sources": [
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "fineweb-edu-dedup",
+      "label": "web",
+      "license": "ODC-BY; preserve underlying source notices",
+      "repo": "HuggingFaceTB/smollm-corpus",
+      "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
+      "split": "train",
+      "text_field": "text",
+      "weight": 56
+    },
+    {
+      "columns": [
+        "text",
+        "id",
+        "edu_int_score"
+      ],
+      "label": "dclm",
+      "license": "CC-BY-4.0",
+      "min_score": {
+        "field": "edu_int_score",
+        "value": 3
+      },
+      "repo": "HuggingFaceTB/dclm-edu",
+      "revision": "dbad8ad71224482740cd9c9d353591adbf62fe04",
+      "split": "train",
+      "text_field": "text",
+      "weight": 16
+    },
+    {
+      "columns": [
+        "text",
+        "id"
+      ],
+      "config": "en",
+      "label": "wiki",
+      "license": "CC-BY-SA-4.0 / GFDL",
+      "repo": "HuggingFaceFW/finewiki",
+      "revision": "8bd13e72e6a002407649b3e898535f42ceb1aeb9",
+      "split": "train",
+      "text_field": "text",
+      "weight": 8
+    },
+    {
+      "code_filter": {
+        "language": "Python",
+        "require_explicit_license": false
+      },
+      "columns": [
+        "blob_id",
+        "repo_name",
+        "path",
+        "src_encoding",
+        "length_bytes",
+        "int_score",
+        "detected_licenses",
+        "license_type",
+        "language",
+        "text",
+        "download_success"
+      ],
+      "config": "default",
+      "label": "python",
+      "license": "Original source repository licenses; preserve provenance and notices",
+      "min_score": {
+        "field": "int_score",
+        "value": 3
+      },
+      "repo": "hongliu9903/stack_edu_python",
+      "revision": "76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376",
+      "split": "train",
+      "text_field": "text",
+      "upstream_config": "Python",
+      "upstream_repo": "HuggingFaceTB/stack-edu",
+      "upstream_revision": "eeec5caac5cc3758a18f1d3ba4416837a9ba814c",
+      "weight": 15
+    },
+    {
+      "additional_sources": [
+        {
+          "columns": [
+            "db_id",
+            "schema",
+            "sql",
+            "question",
+            "external_knowledge"
+          ],
+          "config": "default",
+          "renderer": "synsql-schema-query-v1",
+          "repo": "iNeil77/SynSQL-2.5M",
+          "revision": "56fed46215366defbbc3d7d13064349bbd990a8f",
+          "split": "train",
+          "upstream_repo": "seeklhy/SynSQL-2.5M",
+          "upstream_revision": "cca2c84cc3b41afa6b51534762a6a3a4a420baca"
+        }
+      ],
+      "code_filter": {
+        "language": "SQL",
+        "require_explicit_license": false
+      },
+      "columns": [
+        "blob_id",
+        "repo_name",
+        "path",
+        "src_encoding",
+        "length_bytes",
+        "int_score",
+        "detected_licenses",
+        "license_type",
+        "language",
+        "text",
+        "download_success"
+      ],
+      "config": "default",
+      "label": "sql",
+      "license": "Original source repository licenses; preserve provenance and notices",
+      "min_score": {
+        "field": "int_score",
+        "value": 3
+      },
+      "repo": "hongliu9903/stack_edu_sql",
+      "revision": "079ce8b93c0e83a767ee61af20b9de9931a7e07c",
+      "split": "train",
+      "text_field": "text",
+      "upstream_config": "SQL",
+      "upstream_repo": "HuggingFaceTB/stack-edu",
+      "upstream_revision": "eeec5caac5cc3758a18f1d3ba4416837a9ba814c",
+      "weight": 5
+    }
+  ],
+  "split_buckets": 10000,
+  "text_filter": "english-prose-v1"
+}

configs/pretrain_20261107_142423_40700422/recipe.json

@@ -0,0 +1,172 @@
+{
+  "batches": [
+    7,
+    3,
+    2,
+    2,
+    2
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "cache_gib_by_mac": [
+    24,
+    24,
+    8,
+    8,
+    8
+  ],
+  "code_data_transition": {
+    "format": "sml-prose-python-sql-phase-v1",
+    "manifest": "42ba571cb16dccbf79e65781e82a21881c78c244bb16fac46c40a6e0f123b1e2",
+    "parent_contract": {
+      "code": "b7ff0ae2c1f095bd36300267f5dc2fd60d58be71d828a1b85a796ddf176e942e",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "4f6def65dbd6e1f236be703c2ab977928a55f6bb144ef33688a6e3b0488aacb3",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "051bad7730581c758ce7f469b0214d56cf2e8a9b2df6791c1066b1290881361f",
+    "parent_step": 280771,
+    "start_tokens": 36801216512,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "data_mode": "hf_stream",
+  "decay_start_fraction": 0.8,
+  "eps": 1e-08,
+  "eval_batch_size": 2,
+  "eval_batches_per_source": 32,
+  "eval_every_tokens": 100000000,
+  "grad_accum": 4,
+  "grad_clip": 1.0,
+  "gradual_decay_transition": {
+    "decay_end_tokens": 100000000000,
+    "floor_lr": 3e-05,
+    "format": "sml-v2-open-ended-gradual-v1",
+    "parent_contract": {
+      "code": "1bc7d03c972319212ea775d4667fc0d3f5c0d2d738f3bfff4967c0093913df00",
+      "data": "4412630e242b60857c2ad4a15afb494338ce9686247e6a9e073867d378f53237",
+      "recipe": "19905de3f7b0d4309d2a3341ad393ebb7934e8ef53fe7822afc97f5197acdfff",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "f6238b6fd8a3447dd4cf669f83ab0705bc352e8b5b58bc5b46533246f97fd5f4",
+    "parent_step": 129579,
+    "shape": "linear",
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 16984178688,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "keep_checkpoints": 3,
+  "log_every": 10,
+  "min_lr_ratio": 0.1,
+  "model": {
+    "attention_impl": "fast",
+    "bias": false,
+    "ce_impl": "metal",
+    "d_model": 576,
+    "ffn_impl": "packed-metal",
+    "loss_dtype": "float32",
+    "max_seq_len": 2048,
+    "mlp_multiple_of": 256,
+    "mlp_ratio": 4.0,
+    "n_heads": 9,
+    "n_kv_heads": 3,
+    "n_layers": 44,
+    "qk_norm": true,
+    "vocab_size": 32000
+  },
+  "name": "opensml-deep44-174m-fresh-prose-stream-20b-20261022",
+  "peak_lr": 0.0003,
+  "plateau_scheduler": {
+    "cooldown": 0,
+    "factor": 0.8,
+    "format": "sml-v2-plateau-v1",
+    "initial_lr": 0.0003,
+    "min_delta": 0.0,
+    "min_lr": 3e-05,
+    "patience": 10
+  },
+  "playground_publish_best": false,
+  "reserve_gib": 16,
+  "sample_every_tokens": 100000000,
+  "sample_prompts_file": "sample_prompts.json",
+  "save_every_tokens": 100000000,
+  "scheduler_transition": {
+    "baseline_best": 3.1120742715895178,
+    "format": "sml-v2-scheduler-transition-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "89d8b84a2fe6ff50ef274ccb70b01e4d5394546ea8371876dccf032f817a2f14",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "52bfaaef0938ea869a6ab16890e6d311df863e56156eafcab58a7e2a694d9ec4",
+    "parent_step": 43145,
+    "start_tokens": 5655101440
+  },
+  "seed": 7337,
+  "stable_cooldown_transition": {
+    "duration_tokens": 2000027648,
+    "final_lr": 1e-05,
+    "format": "sml-v2-stable-cooldown-v1",
+    "parent_contract": {
+      "code": "ae36174cc1f5b20f6bf7cf4bc56c7c5f59136dd22104d3dff775d3628600271f",
+      "data": "42ba571cb16dccbf79e65781e82a21881c78c244bb16fac46c40a6e0f123b1e2",
+      "recipe": "d25763cd3f2a452c0cbcb4adec3503644d5bbdccea1deab56412a4843c3d8ec4",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b6d6568e89215fb219747c3426536ac21e9bcebb9a97bc264e3481d091e37641",
+    "parent_step": 305186,
+    "shape": "1-sqrt",
+    "start_lr": 7.93658732685024e-05,
+    "start_tokens": 40001339392,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa"
+  },
+  "stable_schedule_transition": {
+    "format": "sml-v2-open-ended-stable-v1",
+    "parent_contract": {
+      "code": "680fdd3a84c6cae314e9d9038ca00e1f2712284e255d5d8d66599984f57ebb6c",
+      "data": "2f863c538203e9f286b549c7f56598fb9d9b51cef81af77f9627ba7a405e5ec7",
+      "recipe": "99caa9bb0f8182f591c4b466b09a6c9aad4b2ff27ac9314bc44fe1ca821cd9ca",
+      "tokenizer": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb"
+    },
+    "parent_replica_hash": "b86edb7b0448bc58c6f8e1266ba194eebe4442fc860d3ba888901671ef04663e",
+    "parent_scheduler": {
+      "bad_evals": 7,
+      "best": 2.914223026111722,
+      "cap": 9.830400000000001e-05,
+      "config_fingerprint": "ba037a7d13990f03806dfb7dab1a1107b9fccd2373da86870deebf3457f2caa1",
+      "cooldown_remaining": 0,
+      "evaluations": 100,
+      "format": "sml-v2-plateau-v1",
+      "last_loss": 2.9307962145656345,
+      "last_step": 119019,
+      "last_tokens": 15600058368,
+      "reductions": 5
+    },
+    "parent_step": 119722,
+    "parent_stream": {
+      "dedup_window": 100000,
+      "prefetch_batches": 4,
+      "read_attempts": 5,
+      "shuffle_buffer": 256,
+      "tokenizer_corpus_policy": "reuse-frozen"
+    },
+    "stable_lr": 9.830400000000001e-05,
+    "start_lr": 9.830400000000001e-05,
+    "start_tokens": 15692201984,
+    "validation_fingerprint": "2e9444ddc7de3cf4685911d967c1ad605d9b8cfd9244f7039621e828bf088ffa",
+    "warmup_tokens": 0
+  },
+  "stream": {
+    "dedup_window": 100000,
+    "prefetch_batches": 4,
+    "read_attempts": 5,
+    "shuffle_buffer": 256,
+    "source_policy": "all-available-v1",
+    "tokenizer_corpus_policy": "reuse-frozen"
+  },
+  "target_tokens": 20000000000,
+  "warmup_fraction": 0.01,
+  "weight_decay": 0.1
+}

configs/sft1/plan.json

@@ -0,0 +1,189 @@
+{
+  "adaptation": {
+    "adapter_compute": "float32",
+    "alpha": 64,
+    "base_compute": "bfloat16",
+    "dropout": 0.0,
+    "expected_trainable_parameters": 3785856,
+    "layers": "all",
+    "method": "lora",
+    "rank": 32,
+    "targets": [
+      "attn.qkv_proj",
+      "attn.proj"
+    ],
+    "trainable_token_ids": [
+      2,
+      3
+    ]
+  },
+  "answer_max_tokens": {
+    "constraints": 320,
+    "dialogue": 256,
+    "general": 256,
+    "grounded": 160,
+    "knowledge": 80,
+    "rewrite": 320,
+    "skills": 160,
+    "social": 128
+  },
+  "authored_holdouts": {
+    "constraints": 128,
+    "dialogue": 128,
+    "skills": 192
+  },
+  "authored_quotas": {
+    "constraints": 4200,
+    "dialogue": 6800,
+    "skills": 12000
+  },
+  "baseline_cache": "${WORKSPACE}/pretraining/prepared/chat_adapter_search_20261108/baseline_cache",
+  "batches": [
+    8
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "checked_components": [
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue"
+  ],
+  "comparison_checkpoint": "${WORKSPACE}/reports/benchmarks/deep44-reliable-chat-167-fp32-20261106/choices/checkpoint",
+  "component_quotas": {
+    "constraints": 6200,
+    "dialogue": 8000,
+    "general": 4800,
+    "grounded": 11000,
+    "knowledge": 4000,
+    "rewrite": 1000,
+    "skills": 12000,
+    "social": 1000
+  },
+  "components": [
+    "general",
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue",
+    "social",
+    "rewrite"
+  ],
+  "epochs": 2,
+  "eps": 1e-08,
+  "eval_every": 200,
+  "eval_examples_per_component": 64,
+  "execution": "single-mac",
+  "final_lr": 2e-05,
+  "format": "opensml-verified-chat-v2",
+  "generated_eval_per_component": 64,
+  "grad_accum": 8,
+  "grad_clip": 1.0,
+  "keep_checkpoints": 6,
+  "log_every": 10,
+  "max_messages": 8,
+  "max_repetition_rate": 0.05,
+  "max_retention_loss_increase": 0.03,
+  "max_sequence_length": 2048,
+  "min_domain_pass_cases": {
+    "constraints": 24,
+    "dialogue": 32,
+    "grounded": 20,
+    "knowledge": 24,
+    "skills": 16
+  },
+  "min_natural_stop": 0.95,
+  "min_task_gain_cases": 8,
+  "minimum_source_tokens": 42000000000,
+  "output": "${WORKSPACE}/pretraining/runs/deep44_adapter_trial_09_replica_20261108",
+  "peak_lr": 0.0002,
+  "prepared": "${WORKSPACE}/pretraining/prepared/adapter_trial_09_replica_20261108/data",
+  "public_holdouts": {
+    "constraints_smol": 32,
+    "dialogue_smol": 64,
+    "general_smol": 64,
+    "grounded_no": 32,
+    "grounded_yes": 96,
+    "knowledge": 128,
+    "rewrite_smol": 32,
+    "social": 32
+  },
+  "public_quotas": {
+    "constraints_smol": 2000,
+    "dialogue_smol": 1200,
+    "general_smol": 4800,
+    "grounded_no": 2750,
+    "grounded_yes": 8250,
+    "knowledge": 4000,
+    "rewrite_smol": 1000,
+    "social": 1000
+  },
+  "qualitative_eval_per_component": 8,
+  "regression_patience": 2,
+  "replay_length": 512,
+  "replay_objective_weight": 0.2,
+  "replay_policy": "prose-only",
+  "replay_rows_per_update": 16,
+  "retention": "${WORKSPACE}/pretraining/data/deep44_pretraining_validation_20261102",
+  "seed": 202611053,
+  "selection_reference": "Retained checkpoint 167 for behavior; exact FP32 cooled 42B base for prose retention",
+  "sft_objective_weight": 0.8,
+  "sft_rows_per_update": 48,
+  "source_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107/step_0320445_08b689895ed2",
+  "source_run": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107",
+  "sources": [
+    {
+      "label": "smol",
+      "license": "apache-2.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00000-of-00004.parquet",
+        "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00001-of-00004.parquet",
+        "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00002-of-00004.parquet",
+        "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00003-of-00004.parquet"
+      ],
+      "repo": "HuggingFaceTB/smol-smoltalk",
+      "revision": "f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc",
+      "split": "train"
+    },
+    {
+      "label": "social",
+      "license": "see upstream component license",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/HuggingFaceTB/smoltalk/resolve/5feaf2fd3ffca7c237fc38d1861bc30365d48ffa/data/everyday-conversations/train-00000-of-00001.parquet"
+      ],
+      "repo": "HuggingFaceTB/smoltalk",
+      "revision": "5feaf2fd3ffca7c237fc38d1861bc30365d48ffa",
+      "split": "train"
+    },
+    {
+      "label": "squad",
+      "license": [
+        "cc-by-sa-4.0"
+      ],
+      "parquet_urls": [
+        "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/train-00000-of-00001.parquet"
+      ],
+      "repo": "rajpurkar/squad_v2",
+      "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
+      "split": "train"
+    },
+    {
+      "label": "sciq",
+      "license": "cc-by-nc-3.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/allenai/sciq/resolve/2c94ad3e1aafab77146f384e23536f97a4849815/data/train-00000-of-00001.parquet"
+      ],
+      "repo": "allenai/sciq",
+      "revision": "2c94ad3e1aafab77146f384e23536f97a4849815",
+      "split": "train"
+    }
+  ],
+  "strategy": "Fresh 42B base, tied chat tokens, rank 32 attention adapters; 2 complete passes. Fixed data, checks and benchmark protocol.",
+  "tokenizer": "${WORKSPACE}/reference/opensml-150m/tokenizer/bytebpe32k_v1",
+  "warmup_fraction": 0.05,
+  "weight_decay": 0.0
+}

configs/sft2/plan.json

@@ -0,0 +1,209 @@
+{
+  "adaptation": {
+    "adapter_compute": "float32",
+    "alpha": 64,
+    "base_compute": "bfloat16",
+    "dropout": 0.0,
+    "expected_trainable_parameters": 3785856,
+    "layers": "all",
+    "method": "lora",
+    "rank": 32,
+    "targets": [
+      "attn.qkv_proj",
+      "attn.proj"
+    ],
+    "trainable_token_ids": [
+      2,
+      3
+    ]
+  },
+  "answer_max_tokens": {
+    "constraints": 320,
+    "dialogue": 256,
+    "general": 256,
+    "grounded": 160,
+    "knowledge": 80,
+    "rewrite": 320,
+    "skills": 160,
+    "social": 128
+  },
+  "authored_holdouts": {
+    "constraints": 128,
+    "dialogue": 128,
+    "skills": 192
+  },
+  "authored_quotas": {
+    "constraints": 4200,
+    "dialogue": 6800,
+    "skills": 12000
+  },
+  "baseline_cache": "${WORKSPACE}/pretraining/prepared/chat_adapter_search_20261108/baseline_cache",
+  "batches": [
+    8
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "checked_components": [
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue"
+  ],
+  "comparison_checkpoint": "${WORKSPACE}/reports/benchmarks/deep44-reliable-chat-167-fp32-20261106/choices/checkpoint",
+  "component_quotas": {
+    "constraints": 800,
+    "dialogue": 2000,
+    "general": 2400,
+    "grounded": 2000,
+    "skills": 800
+  },
+  "components": [
+    "general",
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue",
+    "social",
+    "rewrite"
+  ],
+  "epochs": 1,
+  "eps": 1e-08,
+  "eval_every": 64,
+  "eval_examples_per_component": 64,
+  "execution": "single-mac",
+  "final_lr": 5e-06,
+  "format": "opensml-verified-chat-v2",
+  "fresh_mixture": {
+    "constraints": 0.4,
+    "dialogue": 0.1,
+    "grounded": 0.1,
+    "skills": 0.4
+  },
+  "fresh_weight": 0.5,
+  "generated_eval_per_component": 64,
+  "grad_accum": 8,
+  "grad_clip": 1.0,
+  "initial_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_adapter_trial_09_replica_20261108/step_0002000_f3d1ffd185f9",
+  "keep_checkpoints": 6,
+  "kl_weight": 0.5,
+  "log_every": 10,
+  "max_cumulative_retention_loss_increase": 0.03,
+  "max_messages": 8,
+  "max_repetition_rate": 0.03333333333333333,
+  "max_retention_loss_increase": 0.003,
+  "max_sequence_length": 2048,
+  "max_steps": 256,
+  "method": "Editing and instruction skills with preservation",
+  "microbatch_rows": 4,
+  "min_domain_pass_cases": {},
+  "min_natural_stop": 0.9833333333333333,
+  "min_task_gain_cases": 4,
+  "minimum_source_tokens": 42000000000,
+  "original_pretraining_retention_loss": 2.8353193145245315,
+  "output": "${WORKSPACE}/pretraining/runs/deep44_parent2000_generalize_05_replica_20261109",
+  "peak_lr": 5e-05,
+  "prepared": "${WORKSPACE}/pretraining/runs/deep44_parent2000_generalize_05_replica_20261109/input/corpus",
+  "public_holdouts": {
+    "constraints_smol": 32,
+    "dialogue_smol": 64,
+    "general_smol": 64,
+    "grounded_no": 32,
+    "grounded_yes": 96,
+    "knowledge": 128,
+    "rewrite_smol": 32,
+    "social": 32
+  },
+  "public_quotas": {
+    "constraints_smol": 2000,
+    "dialogue_smol": 1200,
+    "general_smol": 4800,
+    "grounded_no": 2750,
+    "grounded_yes": 8250,
+    "knowledge": 4000,
+    "rewrite_smol": 1000,
+    "social": 1000
+  },
+  "qualitative_eval_per_component": 8,
+  "regression_patience": 2,
+  "replay_length": 512,
+  "replay_mixture": {
+    "constraints": 0.125,
+    "dialogue": 0.125,
+    "general": 0.1875,
+    "grounded": 0.125,
+    "knowledge": 0.1875,
+    "rewrite": 0.0625,
+    "skills": 0.125,
+    "social": 0.0625
+  },
+  "replay_objective_weight": 0.2,
+  "replay_policy": "prose-only",
+  "replay_rows_per_update": 16,
+  "retention": "${WORKSPACE}/pretraining/data/deep44_pretraining_validation_20261102",
+  "rows_per_group": 16,
+  "seed": 202611097,
+  "selection_reference": "Parent 2000: +4/320 development, +8/704 final, no domain/control regressions; fresh 256-case check must not regress; full-suite guard after private pass",
+  "sft_objective_weight": 0.8,
+  "sft_rows_per_update": 48,
+  "sft_stage": 2,
+  "source_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107/step_0320445_08b689895ed2",
+  "source_run": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107",
+  "sources": [
+    {
+      "file": "train.jsonl.gz",
+      "label": "helpsteer2",
+      "license": "cc-by-4.0",
+      "repo": "nvidia/HelpSteer2",
+      "revision": "990b2711a36180dd19d9c94b8627844866f8982a"
+    },
+    {
+      "file": "data/train-00000-of-00001-b42a775f407cee45.parquet",
+      "label": "oasst1",
+      "license": "apache-2.0",
+      "repo": "OpenAssistant/oasst1",
+      "revision": "fdf72ae0827c1cda404aff25b6603abec9e3399b"
+    },
+    {
+      "files": [
+        "data/smol-magpie-ultra/train-00000-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00001-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00002-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00003-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00004-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00005-of-00006.parquet"
+      ],
+      "label": "smol_magpie_ultra",
+      "license": "apache-2.0",
+      "repo": "HuggingFaceTB/smoltalk",
+      "revision": "5feaf2fd3ffca7c237fc38d1861bc30365d48ffa"
+    },
+    {
+      "label": "squad",
+      "license": "cc-by-sa-4.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/train-00000-of-00001.parquet"
+      ],
+      "repo": "rajpurkar/squad_v2",
+      "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
+      "split": "train"
+    },
+    {
+      "label": "squad_validation",
+      "license": "cc-by-sa-4.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/validation-00000-of-00001.parquet"
+      ],
+      "repo": "rajpurkar/squad_v2",
+      "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
+      "split": "validation"
+    }
+  ],
+  "strategy": "Fresh computed instructions and context tasks plus broad original-training replay; optional forward KL to the frozen parent on replay assistant tokens; all attempts restart exactly from parent2000 with fresh moments",
+  "tokenizer": "${WORKSPACE}/reference/opensml-150m/tokenizer/bytebpe32k_v1",
+  "warmup_fraction": 0.1,
+  "weight_decay": 0.0
+}

configs/sft3/plan.json

@@ -0,0 +1,210 @@
+{
+  "adaptation": {
+    "adapter_compute": "float32",
+    "alpha": 64,
+    "base_compute": "bfloat16",
+    "dropout": 0.0,
+    "expected_trainable_parameters": 3785856,
+    "layers": "all",
+    "method": "lora",
+    "rank": 32,
+    "targets": [
+      "attn.qkv_proj",
+      "attn.proj"
+    ],
+    "trainable_token_ids": [
+      2,
+      3
+    ]
+  },
+  "answer_max_tokens": {
+    "constraints": 320,
+    "dialogue": 256,
+    "general": 256,
+    "grounded": 160,
+    "knowledge": 80,
+    "rewrite": 320,
+    "skills": 160,
+    "social": 128
+  },
+  "authored_holdouts": {
+    "constraints": 128,
+    "dialogue": 128,
+    "skills": 192
+  },
+  "authored_quotas": {
+    "constraints": 4200,
+    "dialogue": 6800,
+    "skills": 12000
+  },
+  "baseline_cache": "${WORKSPACE}/pretraining/prepared/chat_adapter_search_20261108/baseline_cache",
+  "batches": [
+    8
+  ],
+  "betas": [
+    0.9,
+    0.95
+  ],
+  "checked_components": [
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue"
+  ],
+  "comparison_checkpoint": "${WORKSPACE}/reports/benchmarks/deep44-reliable-chat-167-fp32-20261106/choices/checkpoint",
+  "component_quotas": {
+    "constraints": 800,
+    "dialogue": 2000,
+    "general": 2400,
+    "grounded": 2000,
+    "skills": 800
+  },
+  "components": [
+    "general",
+    "knowledge",
+    "grounded",
+    "skills",
+    "constraints",
+    "dialogue",
+    "social",
+    "rewrite"
+  ],
+  "epochs": 1,
+  "eps": 1e-08,
+  "eval_every": 64,
+  "eval_examples_per_component": 64,
+  "execution": "single-mac",
+  "final_lr": 2.0000000000000003e-06,
+  "format": "opensml-verified-chat-v2",
+  "fresh_mixture": {
+    "abstention": 0.1,
+    "dialogue": 0.15,
+    "grounded": 0.15,
+    "reasoning": 0.6
+  },
+  "fresh_weight": 0.4,
+  "generated_eval_per_component": 64,
+  "grad_accum": 8,
+  "grad_clip": 1.0,
+  "initial_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_parent2000_generalize_05_replica_20261109/step_0000256_6074b98fe66b",
+  "keep_checkpoints": 6,
+  "kl_weight": 0.25,
+  "log_every": 10,
+  "max_cumulative_retention_loss_increase": 0.03,
+  "max_messages": 8,
+  "max_repetition_rate": 0.025,
+  "max_retention_loss_increase": 0.003,
+  "max_sequence_length": 2048,
+  "max_steps": 256,
+  "method": "Short checked reasoning",
+  "microbatch_rows": 4,
+  "min_domain_pass_cases": {},
+  "min_natural_stop": 0.9861111111111112,
+  "min_task_gain_cases": 4,
+  "minimum_source_tokens": 42000000000,
+  "original_pretraining_retention_loss": 2.8353193145245315,
+  "output": "${WORKSPACE}/pretraining/runs/deep44_chat_repair_04_20261109",
+  "peak_lr": 2e-05,
+  "prepared": "${WORKSPACE}/pretraining/runs/deep44_chat_repair_04_20261109/input/corpus",
+  "public_holdouts": {
+    "constraints_smol": 32,
+    "dialogue_smol": 64,
+    "general_smol": 64,
+    "grounded_no": 32,
+    "grounded_yes": 96,
+    "knowledge": 128,
+    "rewrite_smol": 32,
+    "social": 32
+  },
+  "public_quotas": {
+    "constraints_smol": 2000,
+    "dialogue_smol": 1200,
+    "general_smol": 4800,
+    "grounded_no": 2750,
+    "grounded_yes": 8250,
+    "knowledge": 4000,
+    "rewrite_smol": 1000,
+    "social": 1000
+  },
+  "qualitative_eval_per_component": 8,
+  "regression_patience": 2,
+  "repair_parent": "attempt5",
+  "replay_length": 512,
+  "replay_mixture": {
+    "constraints": 0.125,
+    "dialogue": 0.125,
+    "general": 0.1875,
+    "grounded": 0.125,
+    "knowledge": 0.1875,
+    "rewrite": 0.0625,
+    "skills": 0.125,
+    "social": 0.0625
+  },
+  "replay_objective_weight": 0.2,
+  "replay_policy": "prose-only",
+  "replay_rows_per_update": 16,
+  "retention": "${WORKSPACE}/pretraining/data/deep44_pretraining_validation_20261102",
+  "rows_per_group": 16,
+  "seed": 2026110915,
+  "selection_reference": "New 300-case development test, with historical 320-case and prose retention guards. Final 1000-case test is unopened until a single finalist is frozen.",
+  "sft_objective_weight": 0.8,
+  "sft_rows_per_update": 48,
+  "sft_stage": 2,
+  "source_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107/step_0320445_08b689895ed2",
+  "source_run": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107",
+  "sources": [
+    {
+      "file": "train.jsonl.gz",
+      "label": "helpsteer2",
+      "license": "cc-by-4.0",
+      "repo": "nvidia/HelpSteer2",
+      "revision": "990b2711a36180dd19d9c94b8627844866f8982a"
+    },
+    {
+      "file": "data/train-00000-of-00001-b42a775f407cee45.parquet",
+      "label": "oasst1",
+      "license": "apache-2.0",
+      "repo": "OpenAssistant/oasst1",
+      "revision": "fdf72ae0827c1cda404aff25b6603abec9e3399b"
+    },
+    {
+      "files": [
+        "data/smol-magpie-ultra/train-00000-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00001-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00002-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00003-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00004-of-00006.parquet",
+        "data/smol-magpie-ultra/train-00005-of-00006.parquet"
+      ],
+      "label": "smol_magpie_ultra",
+      "license": "apache-2.0",
+      "repo": "HuggingFaceTB/smoltalk",
+      "revision": "5feaf2fd3ffca7c237fc38d1861bc30365d48ffa"
+    },
+    {
+      "label": "squad",
+      "license": "cc-by-sa-4.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/train-00000-of-00001.parquet"
+      ],
+      "repo": "rajpurkar/squad_v2",
+      "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
+      "split": "train"
+    },
+    {
+      "label": "squad_validation",
+      "license": "cc-by-sa-4.0",
+      "parquet_urls": [
+        "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/validation-00000-of-00001.parquet"
+      ],
+      "repo": "rajpurkar/squad_v2",
+      "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
+      "split": "validation"
+    }
+  ],
+  "strategy": "Independent small adapter updates from parent2000 or preserved attempt5; verified text tasks and 60% replay objective with frozen-source probability preservation.",
+  "tokenizer": "${WORKSPACE}/reference/opensml-150m/tokenizer/bytebpe32k_v1",
+  "warmup_fraction": 0.1,
+  "weight_decay": 0.0
+}

docs/DATA.md

@@ -0,0 +1,48 @@
+# Solstice-1 data provenance
+
+## Pretraining
+
+Configured token shares are 70/20/10 web/DCLM/wiki until 36,801,216,512 positions, then 56/16/8/15/5 web/DCLM/wiki/Python/SQL through 42,001,367,040. These shares describe the mixture controller; they are not a separately audited count of all unique content delivered. The SQL configuration includes a SynSQL mirror continuation source. Configuration establishes eligibility, not proof that every configured downstream shard was consumed. Streaming cursor/validation records are not represented as a global deduplicated data manifest.
+
+The frozen stream uses bounded accepted-document shuffling and a 100,000-document recent-deduplication window. Content identity controls the fixed train/validation partition. English prose and code filters are archived in each source snapshot. DCLM and code sources require their configured educational score at least 3. Text size bounds are 200 characters through 200,000 UTF-8 bytes where applicable. Null characters, corruption/replacement patterns and repetitive content are filtered by the source-specific code. These rules are heuristic; they do not guarantee every retained example is correct, licensed identically, or uncontaminated.
+
+Python and SQL full-text mirrors retain upstream repository/license fields. The recorded filter sets `require_explicit_license=false`; it must not be described as an explicit-license-only corpus. No raw training corpus is redistributed with this release.
+
+## Dataset identities
+
+The following pins and terms come from the frozen configurations and corpus manifests. They are source provenance, not a blanket grant of rights to all underlying content.
+
+| Source | Pinned revision | Recorded terms / role |
+| --- | --- | --- |
+| [HuggingFaceTB/smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus/tree/3ba9d605774198c5868892d7a8deda78031a781f) | `3ba9d605774198c5868892d7a8deda78031a781f` | ODC-BY; preserve underlying source notices |
+| [HuggingFaceTB/dclm-edu](https://huggingface.co/datasets/HuggingFaceTB/dclm-edu/tree/dbad8ad71224482740cd9c9d353591adbf62fe04) | `dbad8ad71224482740cd9c9d353591adbf62fe04` | CC-BY-4.0 |
+| [HuggingFaceFW/finewiki](https://huggingface.co/datasets/HuggingFaceFW/finewiki/tree/8bd13e72e6a002407649b3e898535f42ceb1aeb9) | `8bd13e72e6a002407649b3e898535f42ceb1aeb9` | CC-BY-SA-4.0 / GFDL |
+| [hongliu9903/stack_edu_python](https://huggingface.co/datasets/hongliu9903/stack_edu_python/tree/76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376) | `76eeb3eb59bbad948e4fbfae9f4e6159d6f8f376` | Original source repository licenses; preserve provenance and notices |
+| [hongliu9903/stack_edu_sql](https://huggingface.co/datasets/hongliu9903/stack_edu_sql/tree/079ce8b93c0e83a767ee61af20b9de9931a7e07c) | `079ce8b93c0e83a767ee61af20b9de9931a7e07c` | Original source repository licenses; preserve provenance and notices |
+| [iNeil77/SynSQL-2.5M](https://huggingface.co/datasets/iNeil77/SynSQL-2.5M/tree/56fed46215366defbbc3d7d13064349bbd990a8f) | `56fed46215366defbbc3d7d13064349bbd990a8f` | See upstream source terms |
+| [HuggingFaceTB/smol-smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/tree/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc) | `f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc` | apache-2.0 |
+| [HuggingFaceTB/smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk/tree/5feaf2fd3ffca7c237fc38d1861bc30365d48ffa) | `5feaf2fd3ffca7c237fc38d1861bc30365d48ffa` | see upstream component license |
+| [rajpurkar/squad_v2](https://huggingface.co/datasets/rajpurkar/squad_v2/tree/3ffb306f725f7d2ce8394bc1873b24868140c412) | `3ffb306f725f7d2ce8394bc1873b24868140c412` | ['cc-by-sa-4.0'] |
+| [allenai/sciq](https://huggingface.co/datasets/allenai/sciq/tree/2c94ad3e1aafab77146f384e23536f97a4849815) | `2c94ad3e1aafab77146f384e23536f97a4849815` | cc-by-nc-3.0 |
+
+## Stage-1 conversations
+
+The 48,000-conversation pool contains 6,200 constraint examples, 8,000 dialogue examples, 4,800 general responses, 11,000 grounded examples, 4,000 knowledge examples, 1,000 rewrites, 12,000 skills examples and 1,000 social examples. Public sources are Smol-SmolTalk, SmolTalk everyday conversations, SQuAD v2 TRAIN and SciQ TRAIN. Computed/authored tasks supply much of the skills, correction/dialogue and constraint coverage. Two complete passes produce 96,000 conversation presentations; prose replay contributes another 32,000 512-token sequences.
+
+Public answers were filtered and sample-reviewed; not every statement received independent factual verification. SQuAD span checks verify annotation consistency, not every annotation's truth. The frozen data manifests preserve rejection rules, token statistics and file hashes.
+
+## Later selected prefixes
+
+Stage 2 draws 4,096 distinct new tasks from an 8,000-task authored pool, plus 4,096 distinct replay conversations. Stage 3 draws from a new 20,480-task pool; the released prefix inherits 2,048 fresh tasks and 2,048 replay conversations. Stage-3 fresh counts are reasoning 1,230; grounding 324; dialogue 306; abstention 188. The full five-experiment search did not all become training exposure for this model.
+
+The replay pool contains prior stage-1 material, including SciQ and SQuAD. Later plan files contain inherited source declarations such as HelpSteer2/OASST/Smol-Magpie; those declarations are **not evidence those sources were fetched for these stages**. Actual selected rows, subsets and source-file hashes in [SELECTED_PREFIX.json](../provenance/SELECTED_PREFIX.json) are authoritative. The release's dataset metadata lists the actual inherited sources rather than those stale plan declarations.
+
+## Filtering and evaluation separation
+
+Conversations are retained whole and serialized using the frozen tokenizer. The selected text-focused SFT excludes dedicated Python, SQL, JSON and tool tasks. Exact prompt exclusions and grouped splits were checked. The new authored panel separates entity values and wording between train/development/test, but shares task families. All 21,780 newly authored train/dev/test rows were independently parsed and target-checked during preparation. Its 1,000-case test was unopened until one finalist was frozen.
+
+Historical 320/704 panels and public benchmarks had prior research exposure. Their use does not provide a fresh blind evaluation of the entire research program. Hash-based decontamination does not establish absence of paraphrases or source-document overlap in pretraining.
+
+## Redistribution
+
+Author-controlled code and selected model weights use Apache-2.0 by the author's release decision. Dataset terms remain distinct. In particular, recorded SQuAD terms are CC-BY-SA-4.0 and SciQ terms are CC-BY-NC-3.0; Python/SQL repositories can have heterogeneous terms. The release does not relicense or redistribute the raw source datasets. See [THIRD_PARTY_NOTICES.md](../THIRD_PARTY_NOTICES.md) for attribution and the exact upstream links.

docs/EVALUATION.md

@@ -0,0 +1,69 @@
+# Solstice-1 evaluation
+
+All results identify the same release weight hash, `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`. Public summary: [results/public/summary.json](../results/public/summary.json). The package verification is a separate check of loading/output parity, not a new training result.
+
+## Likelihood scoring
+
+| Benchmark | Examples | Accuracy | Normalized accuracy |
+| --- | --- | --- | --- |
+| ARC Easy | 2376 | 62.58% | 62.33% |
+| ARC Challenge | 1172 | 29.35% | 30.89% |
+| PIQA | 1838 | 66.49% | 66.21% |
+| HellaSwag | 10042 | 32.85% | 39.40% |
+
+Native MLX evaluates complete ARC Easy/Challenge test splits, PIQA validation and HellaSwag validation. It uses zero-shot raw completion prompts, no chat template or inserted BOS/EOS, FP32 weights, vanilla attention, batch size four and chunked scoring. Raw accuracy ranks summed continuation log-likelihood; normalized accuracy divides by choice Unicode-character count. No generated explanation or LLM judge determines the answer. No choices were context-truncated. Dataset revisions and scorer settings are preserved in [the result record](../results/public/choices/results.json).
+
+## Instruction following
+
+| IFEval metric | Score | Passes / checks |
+| --- | --- | --- |
+| Strict prompts | 17.93% | 97/541 |
+| Strict instructions | 30.10% | 251/834 |
+| Loose prompts | 19.22% | 104/541 |
+| Loose instructions | 32.49% | 271/834 |
+
+Google IFEval is pinned to dataset revision `966cd89545d6b6acfd7638bc708b98261ca58e84` and google-research commit `83f1e7b59e762648c953de751d31f75720683622`. All 541 prompts and 834 instructions use the recorded strict/loose programmatic verifiers. Generation is greedy, native-chat formatted, FP32, without a system message or repetition penalty. Output cap: 1,024 tokens. Token-cap hits: **130/541**; context-cap hits: **0**. All generated responses are scored, including cap hits. The official dataset's split is named `train`; it is used as evaluation input here, not SFT training material.
+
+## Historical comparison
+
+| Metric | Published 150M | 167 | Parent 2000 | Prior attempt 5 | Solstice-1 |
+| --- | --- | --- | --- | --- | --- |
+| ARC Easy acc_norm | 55.43% | 61.49% | 62.33% | 62.33% | 62.33% |
+| ARC Challenge acc_norm | 29.52% | 30.46% | 30.63% | 31.23% | 30.89% |
+| PIQA acc_norm | 64.09% | 65.67% | 66.43% | 66.43% | 66.21% |
+| HellaSwag acc_norm | 33.94% | 38.60% | 39.51% | 39.44% | 39.40% |
+| IFEval strict prompt | 15.16% | 16.45% | 14.97% | 15.53% | 17.93% |
+| IFEval strict instruction | 25.30% | 29.14% | 26.86% | 28.18% | 30.10% |
+| IFEval loose prompt | 15.71% | 17.93% | 16.82% | 16.45% | 19.22% |
+| IFEval loose instruction | 25.78% | 31.29% | 29.62% | 30.34% | 32.49% |
+
+Solstice, parent 2000, prior attempt 5 and 167 share the recorded native FP32 protocol. The 150M numbers are from its [published model card](https://huggingface.co/OpenSML/OpenSML-150M), not a fresh release-bundle rerun. Its IFEval uses plain `User:/Assistant:` formatting, EOS ID 1 and a 1,280-token cap; newer models use structural chat IDs and 1,024 tokens. The earlier choice run uses batch size one. These differences limit controlled head-to-head claims. No external-model leaderboard rank or statistical significance of every small benchmark difference is asserted.
+
+JSONLite has different extraction panels, so its 89.54% extraction correctness and 99.93% schema validity are not comparable with these scores. No JSONLite ARC/IFEval score is inferred from its pretrained ancestor.
+
+## Authored checks: keep denominators separate
+
+| Panel | Parent 2000 | Prior attempt 5 | Solstice-1 |
+| --- | ---: | ---: | ---: |
+| Historical development answers | 197/320 | 204/320 | 200/320 |
+| Historical final answers | 367/704 | 384/704 | 379/704 |
+| New reserved authored panel | 200/1,000 | 259/1,000 | 519/1,000 |
+| Separate chat answer matches | 102/180 | 101/180 | 107/180 |
+| Separate chat natural stopping | 189/192 | 191/192 | 191/192 |
+| Separate chat repetition flags | 4/192 | 2/192 | 3/192 |
+
+The 320 development and 704 final counts arise from different selected evaluation subsets. They are not a single test that changed denominators. The 192-response chat panel contains 180 objectively checked cases plus 12 open-ended qualitative cases. Automatic answer matching is conservative; it is not comprehensive semantic review. The new panel contains five domains with 200 cases each. Solstice passes 172 abstention, 25 constraints, 194 dialogue, 81 grounded and 47 reasoning cases. Strict formatting reduces its total from 519 to 518.
+
+Against parent 2000, the new-panel gain is 31.9 percentage points, with a 97.5% scenario-cluster paired bootstrap interval of 28.69–35.04 points. Against attempt 5, it is 26.0 points, interval 22.93–29.00. The method uses 10,000 draws and keeps paired present/missing-fact scenarios together. These intervals describe sampling variation within this authored task distribution; they do not measure all conversational ability or correct for the entire historical experiment program.
+
+## Remaining failure modes
+
+The new panel flags unnecessary abstention on 111/200 answerable extraction cases. Separately, all 15 differently worded missing-information chat cases fail appropriate abstention. Saved examples include invented values and copied irrelevant text, so this second issue is not explained solely by strict answer matching. Wording sensitivity remains a limitation even though the final selection guards passed.
+
+Manual inspection of the three chat repetition flags found two clear loops and one milder repeated sentence/idea. Keep the original three-flag result for consistent comparisons; it does not mean all three are equally severe. All 180 objective chat responses stop naturally; the qualitative cases account for the single token-cap hit and the repetition flags.
+
+## Selection and test exposure
+
+The final five-trial search selected one checkpoint using the 300-case new development panel plus historical development, stopping, repetition and prose-retention guards. Finalist identity was frozen before the new 1,000-case test. Small tolerated regressions were part of the declared guards; passing them does not mean every metric improved. The complete comparison and selection record includes the thresholds and reasons.
+
+Historical and public evaluations were used in earlier research, and some training tasks resemble evaluation task families. Exact exclusions do not establish semantic decontamination. No broad human study, safety certification, calibrated uncertainty test or verified real-world factuality score is claimed.

docs/EXAMPLES.md

@@ -0,0 +1,47 @@
+# Recorded inference examples
+
+Greedy FP32, no system message, no repetition penalty. Reproduced in the clean release environment. These include a factual error so that successful loading is not confused with correctness.
+
+## knowledge-00
+
+**User:** What is the capital of Japan? Give a short answer.
+
+**Generated answer:** Tokyo
+
+Stop: `turn_end`. Saved answer matcher: `True`.
+
+## knowledge-01
+
+**User:** Which planet is closest to the Sun? Give a short answer.
+
+**Generated answer:** venus
+
+Stop: `turn_end`. Saved answer matcher: `False`.
+
+## reading-00
+
+**User:** Use only the information below. Give a short answer.
+
+Anika brought a teal notebook. Greta brought a black notebook.
+What color was Anika’s notebook?
+
+**Generated answer:** teal
+
+Stop: `turn_end`. Saved answer matcher: `True`.
+
+## conversation-00
+
+**User:** My reservation number is 217.
+
+**Assistant:** Understood.
+
+**User:** Correction: my reservation number is 317.
+
+**Assistant:** Understood.
+
+**User:** What is my current reservation number? Give only the number.
+
+**Generated answer:** 317
+
+Stop: `turn_end`. Saved answer matcher: `True`.
+

docs/TECHNICAL_REPORT.md

@@ -0,0 +1,97 @@
+# Solstice-1 technical report
+
+Author: William Zebrowski · Release candidate prepared 2026-11-09 · Public size designation: **175M**
+
+## 1. Scope and identity
+
+Solstice-1 is the merged FP32 export of final-stage checkpoint `step_0000128_5937a6def881`. It contains exactly **174,202,432 parameters**. The 175M name is a rounded public designation. This report describes the selected checkpoint rather than all experiments in the workspace.
+
+Weight SHA-256: `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`.
+
+The complete tensor shape inventory is in [TENSOR_INVENTORY.json](../provenance/TENSOR_INVENTORY.json). The inference configuration is [config.json](../config.json). The published parameter count describes the merged dense model, not the number of trainable adapter parameters.
+
+## 2. Architecture
+
+The model is a decoder-only causal transformer with 44 residual blocks and hidden width 576. Attention uses a combined QKV projection, nine query heads, three KV heads, and head dimension 64. Each query/key head is RMS-normalized. Rotary position embeddings use base 10,000. Blocks use pre-normalization and SwiGLU; the actual FFN intermediate width is 1,536. The packed gate/up matrix has shape 3,072 × 576. Input embeddings and output projection are tied. There are no linear biases or dropout.
+
+The model has a deeper, narrower shape than the earlier 24-layer, width-768 JSONLite ancestor or 20-layer OpenSML-150M. These runs also differ in data, optimization, and training exposure, so their results do not isolate an architectural advantage.
+
+Training uses MLX fast attention and packed Metal FFN/cross-entropy paths where specified by each frozen recipe. Release inference and the reported FP32 benchmarks use vanilla attention and reference FFN/cross-entropy paths. No quantization is applied to the release weights.
+
+## 3. Tokenizer and serialization
+
+The existing 32,000-token byte-level BPE tokenizer was reused unchanged. It was fitted for the earlier OpenSML project, not retrained on the entire Solstice corpus. Its manifest fingerprint is `1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb`. Token IDs are pad=0, document-end=1, turn-start=2, turn-end=3. The runner verifies the manifest, vocabulary and IDs. The tokenizer files and original corpus metadata are included.
+
+Each chat message is serialized as ID 2, tokenized `role + "\n"`, tokenized content, and ID 3. Generation ends with an open assistant header. A training conversation additionally ends with document-end ID 1. Ordinary text that resembles a control token is encoded as ordinary text rather than injected structural IDs. Role headers and user/system text are masked from SFT loss. Assistant content and end markers are supervised according to the stage's loss mask.
+
+The maximum context is 2,048 tokens. SFT preserves complete conversations and rejects oversized examples rather than truncating them. The largest stage-1 serialized training conversation is 1,385 tokens including its final boundary; the shifted input length recorded in its manifest is 1,384. Stage-2 authored tasks reach 115 serialized tokens; stage-3 authored tasks reach 137. Selected conversational replay is at most 512 serialized tokens. These maxima are data checks, not a claim that long-dialogue behavior is reliable.
+
+## 4. Pretraining data and exposure
+
+The model starts from random initialization with seed 7337. Initial token-mixture shares are 70% FineWeb-Edu, 20% DCLM-Edu, and 10% English FineWiki. At 36,801,216,512 lifetime tokens, training switches to 80% prose, 15% Python, and 5% SQL: web/DCLM/wiki become 56%/16%/8%. Python/SQL use public Stack-Edu text mirrors and the configured SQL continuation source. The final base contains **42,001,367,040 processed positions** at step 320,445.
+
+These are processed token counts, not a measured inventory of globally unique training tokens. Sources are streamed with bounded buffering and deduplication. The all-available policy introduced during continuation bypasses the earlier 50-million-row ceiling; an inherited corpus field containing that number does not describe the active policy. The frozen recipes record the actual source policy.
+
+The active stream records a 256-document shuffle buffer, 100,000-document deduplication window, four prefetched batches, and five read attempts. Exact source pins, phase changes, text/code filters and provenance are documented in [DATA.md](DATA.md). Identity partitioning and bounded duplicate checks do not establish semantic decontamination of internet-derived training data.
+
+## 5. Distributed optimization and schedules
+
+Pretraining uses five Apple Silicon workers connected for distributed communication. Per-worker microbatches are [7,3,2,2,2], with four gradient-accumulation microsteps and sequence length 2,048: **131,072 positions per optimizer update**. Gradient aggregation is weighted for unequal worker token counts. AdamW uses betas (0.9,0.95), epsilon 1e-8, weight decay 0.1 and gradient clipping at 1. BF16 compute is backed by FP32 master weights, optimizer state and gradient accumulation. Historical hardware/runtime configuration is archived; no new energy or cross-hardware throughput benchmark is asserted here.
+
+| Boundary | Effective change |
+| --- | --- |
+| Start | Warm up over approximately 200M tokens to 3e-4; original 20B target |
+| 5,655,101,440 tokens | Plateau controller: factor 0.8 after 10 non-improving evaluations; floor 3e-5 |
+| 15,692,201,984 tokens | All-available streaming; hold inherited LR 9.8304e-5 |
+| 16,984,178,688 tokens | Begin gradual linear decay toward 3e-5 at 100B lifetime tokens |
+| 36,801,216,512 tokens | Introduce the 80/15/5 prose/Python/SQL mixture |
+| 40,001,339,392 tokens | Begin final cooldown at 7.93658732685024e-5 |
+| 42,001,367,040 tokens | Finish cooldown at 1e-5; selected pretrained base |
+
+Cooldown lasts 2,000,027,648 additional positions. With progress p from 0 to 1, LR is `1e-5 + (start_lr - 1e-5) * (1 - sqrt(p))`. This is a substantial late reduction, unlike extending the previous 100B schedule unchanged. Earlier plateau state remains in metadata for provenance; it is not an active second controller during cooldown.
+
+Nominal validation and checkpoint intervals are 100M tokens. Recorded training and validation curves are in [pretraining_steps.csv](../results/pretraining_steps.csv) and [pretraining_validation.csv](../results/pretraining_validation.csv); displayed log losses have the precision of the source logs. CSV token positions for validation are reconstructed from the constant 131,072-position batch and step counter.
+
+![Pretraining loss and learning rate](figures/pretraining-overview.png)
+
+![Late training and cooldown](figures/cooldown-detail.png)
+
+The base checkpoint's saved BF16-path best prose validation loss is 2.8356336173. The FP32 base measurement used by SFT retention checks is 2.8353193145. Solstice-1's FP32 retention loss is 2.8633367460. These precision-specific measurements must not be silently combined. Python and SQL validation losses are separate from prose-best selection. Losses on these datasets cannot be compared directly with unrelated models' losses.
+
+## 6. Selected fine-tuning lineage
+
+| Stage | Inherited updates | Fresh/supervised presentations | Replay presentations | Peak → planned final LR |
+| --- | --- | --- | --- | --- |
+| SFT 1 | 2000 | 96,000 (48,000 unique conversations; two passes) | 32,000 prose sequences | 2e-4 → 2e-5 |
+| SFT 2 | 256 | 4,096 authored tasks | 4,096 prior conversations | 5e-5 → 5e-6 |
+| SFT 3 | 128 | 2,048 authored tasks | 2,048 prior conversations | 2e-5 → 2e-6 over 256 planned updates |
+
+All three stages train locally on one Mac. Rank-32 LoRA adapters with alpha 64 and no dropout target attention QKV and output projections in all 44 blocks. The two tied chat-token embedding rows, IDs 2 and 3, are also trainable. There are **3,785,856 trainable parameters**. The original base is frozen, computed in BF16 during training; adapters use FP32. Each subsequent stage inherits the preceding adapters/token deltas and starts fresh AdamW moments. Stage metadata, source hashes, actual batch indices, and selected-prefix counts are in [provenance](../provenance/LINEAGE.json).
+
+SFT 1 uses 48 conversation rows plus 16 prose-replay sequences per update, microbatch eight. Its loss combines 0.8 assistant-masked conversation loss with 0.2 prose replay. It makes two complete passes over 48,000 curated conversations. Warmup is 5% of 2,000 updates, followed by cosine decay. The selected checkpoint is final update 2,000, reproduced from the earlier attempt-9 recipe. It did not satisfy every earlier quality gate; preservation and later continuation are documented rather than recast as universal success.
+
+SFT 2 uses 16 fresh tasks and 16 conversational-replay rows per update, microbatch four. The fresh mixture is 40% editing/skills, 40% constraints, 10% dialogue and 10% grounding, with integer rounding recorded in the selected-prefix file. Its objective is `0.5 * fresh_NLL + 0.5 * (replay_NLL + 0.5 * KL(parent || policy))`. The reference is a frozen copy of its immediate parent. Fresh tasks supervise the final assistant answer; replay supervises all assistant turns. This is supervised fine-tuning with parent-distribution preservation, not DPO or reinforcement learning. The preserved attempt-5 reproduction is selected at update 256.
+
+SFT 3 plans 256 updates with the same 16+16 batching, 10% warmup and cosine schedule. Its objective is `0.4 * fresh_NLL + 0.6 * (replay_NLL + 0.25 * KL(parent || policy))`. Thus the effective outer KL coefficient is 0.15. The configured fresh draw is 60% short reasoning, 15% grounding, 15% dialogue and 10% abstention. Solstice inherits only the first 128 updates: **1,230 reasoning, 324 grounding, 306 dialogue and 188 abstention presentations**, plus 2,048 replay presentations. The selected prefix differs slightly from the full planned proportions because batches are deterministically shuffled and grouped by length.
+
+All SFT stages use AdamW betas (0.9,0.95), epsilon 1e-8, no weight decay, gradient clipping at 1, and no dropout. Stage seeds are 202611053, 202611097, and 2026110915. Stage-3's selected step is halfway through a 256-update schedule; the planned final 2e-6 rate is not falsely attributed to update 128.
+
+Selected recorded assistant-token totals are 4,440,608, 279,606 and 120,629 by stage. Recorded input-position totals are 29,271,514, 816,428 and 411,448; stage 1 includes its prose replay. Counts describe training presentations and processed positions, not unique documents across all stages.
+
+No new external teacher model was called to label these stages. Public SmolTalk-family sources include upstream synthetic responses. Parent KL preservation uses a frozen local reference during training. Inference requires only the merged model.
+
+## 7. Export precision and selection
+
+The inference export combines original FP32 pretrained masters with FP32 adapter and tied-token deltas. It does not merely cast rounded BF16 base weights back to FP32. The merged tensor file is exactly the one evaluated in the saved public suite; no optimizer or training corpus is required for inference. Hash checks and native-package verification are recorded separately from training metrics.
+
+Five bounded repair experiments were compared using a new 300-case development panel and historical 320-case guards. Experiment 4, update 128, was frozen as the finalist before opening the reserved 1,000-case test. The selection rule prioritized new-panel correctness among candidates meeting the predeclared retention/behavior tolerances. It did not maximize every historical metric: another checkpoint could score higher on a particular reused test.
+
+A scoring correction during setup separated content correctness from exact formatting. The initial trial-1 prefix was restarted under the corrected evaluator; its reproduced step-64 adapter/optimizer hashes matched. This was not an extra training recipe. The evidence records the five experiments, selection rules and scoring scope.
+
+The selected model scores 519/1,000 on the reserved authored panel versus 200 for parent 2000 and 259 for the prior attempt 5. Scenario-cluster bootstrap intervals support improvement within that panel, not universal chat superiority. Historical 704-case scores are 379, 367 and 384 respectively: the finalist improves over parent 2000 but trails attempt 5 on this separate check. See [EVALUATION.md](EVALUATION.md).
+
+## 8. Limitations and reproducibility
+
+The strongest new-panel gains do not establish broad factuality. The model still over-abstains on some answerable inputs and invents missing facts under different wording. Longer instructions frequently hit the output cap. Repetition flags and low strict IFEval prompt accuracy remain. Coding, SQL and JSON capabilities are not release claims merely because code appeared in pretraining.
+
+The report includes exact checkpoint identities, source/dataset revisions, batch-selection indices, recorded metrics and sanitized training-source snapshots. Original raw corpora, resumable optimizer state and all intermediate weights are not bundled. Source paths and local network addresses were sanitized and changes are recorded. Reconstructing a historical run requires the external assets and compatible runtime described in [TRAINING_AND_REPRODUCIBILITY.md](../TRAINING_AND_REPRODUCIBILITY.md); exact end-to-end retraining has not been independently rerun.

evaluation/README.md

@@ -0,0 +1,15 @@
+# Evaluation source and recorded evidence
+
+The complete saved choice and IFEval results are in `../results/public/`. Choice files retain example indices, summed likelihoods, predictions, gold labels and truncation counts. IFEval files retain generated responses and strict/loose verifier decisions; `dataset_row_index` maps back to the pinned Google dataset. Upstream prompt fields are omitted from these copies. None of these files represents a newly rerun suite during packaging.
+
+`benchmark_best.py`, `benchmark_ifeval.py` and `chat_adapter_search_bench.py` are historical entry points. Their paths and imports expect the original research workspace. Materialize the matching `sft3` source snapshot to audit their imported training modules. Running the historical suite additionally requires prepared dataset caches, original checkpoint metadata and the pinned IFEval runtime. They are not advertised as standalone commands for this smaller inference package.
+
+The Google evaluator source, original headers and license are retained under `vendor/ifeval`; dependency versions and the source commit are recorded there. ARC/PIQA/HellaSwag dataset pins and evaluation settings are in the saved choice result JSON. The authored chat panel/scorer source is in `chat_benchmark/`.
+
+For a standalone package check that does work directly in this release, run:
+
+```bash
+python scripts/verify_inference.py --output /tmp/solstice-inference-check.json
+```
+
+It runs 27 position-selected examples across seven chat categories, compares exact text/stop/token counts against saved FP32 answers, checks context/serialization guards, and checks fixed-seed sampling. Both correct and incorrect saved answers are included. This is package parity verification, not a substitute for the full benchmark suite or a new accuracy estimate.

evaluation/benchmark_best.py

@@ -0,0 +1,467 @@
+#!/usr/bin/env python3
+"""Prepare/run a frozen zero-shot multiple-choice benchmark (CPU by default).
+
+Task prompts and character-normalized accuracy follow EleutherAI's task
+definitions. This is a native MLX runner, not an lm-eval-harness execution.
+No training imports, optimizer loads or distributed initialization. An explicit
+device='gpu' configuration runs locally on the GPU after training has stopped.
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+import os
+from pathlib import Path
+import re
+import resource
+import shutil
+import statistics
+import sys
+import time
+import uuid
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT / 'pretraining'))
+DEFAULT_CONFIG = ROOT / 'configs/benchmark_best_low_impact.json'
+from sft.metadata import sft_info
+
+
+def read_json(path):
+    return json.loads(Path(path).read_text())
+
+
+def checkpoint_tokenizer(metadata):
+    info=sft_info(metadata)
+    if info is not None:
+        return info['contract']['tokenizer']
+    return metadata['v2_contract']['tokenizer']
+
+
+def atomic_json(path, data):
+    path = Path(path)
+    temp = path.with_name(path.name + '.tmp')
+    temp.write_text(json.dumps(data, indent=2, allow_nan=False) + '\n')
+    temp.replace(path)
+
+
+def sha256(path):
+    digest = hashlib.sha256()
+    with Path(path).open('rb') as file:
+        for block in iter(lambda: file.read(4 * 1024**2), b''):
+            digest.update(block)
+    return digest.hexdigest()
+
+
+def peak_rss_gib():
+    peak = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+    return peak / 1024**3 if sys.platform == 'darwin' else peak / 1024**2
+
+
+def local_path(value):
+    path = Path(value)
+    return path if path.is_absolute() else ROOT / path
+
+
+def configure_process(cache):
+    # Process-local settings; do not change the training runtime or its cache.
+    for name in ('VECLIB_MAXIMUM_THREADS', 'OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS',
+                 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+        os.environ[name] = '1'
+    os.environ['TOKENIZERS_PARALLELISM'] = 'false'
+    os.environ['HF_HUB_CACHE'] = str(cache / 'hub')
+    os.environ['HF_DATASETS_CACHE'] = str(cache / 'datasets')
+    if hasattr(os, 'setpriority'):
+        os.setpriority(os.PRIO_PROCESS, 0, 19)
+
+
+def prepare(config, output):
+    output.mkdir(parents=True, exist_ok=True)
+    frozen = output / 'checkpoint'
+    if not frozen.exists():
+        staging = output / ('.checkpoint-' + uuid.uuid4().hex)
+        staging.mkdir()
+        try:
+            source_run = local_path(config['run_dir'])
+            name = read_json(source_run / 'best.json')['bundle']
+            if not re.fullmatch(r'step_\d+_[a-f0-9]+', name):
+                raise ValueError('Invalid best checkpoint name')
+            source = source_run / name
+            if source.is_symlink():
+                raise ValueError('Checkpoint cannot be a symlink')
+            manifest = read_json(source / 'manifest.json')
+            records = {}
+            for filename in ('model.safetensors', 'model.safetensors.json'):
+                origin = source / filename
+                record = manifest['files'][filename]
+                if origin.is_symlink():
+                    raise ValueError('Checkpoint files cannot be symlinks')
+                target = staging / filename
+                shutil.copyfile(origin, target)
+                if target.stat().st_size != record['bytes'] or sha256(target) != record['sha256']:
+                    raise ValueError('Checkpoint changed or failed its checksum; retry preparation')
+                records[filename] = record
+            shutil.copytree(local_path(config['tokenizer_dir']), staging / 'tokenizer')
+            from sml_v1.tokenization import Tokenizer
+            tokenizer = Tokenizer(staging / 'tokenizer')
+            metadata = read_json(staging / 'model.safetensors.json')
+            info=sft_info(metadata)
+            if tokenizer.fingerprint != checkpoint_tokenizer(metadata):
+                raise ValueError('Checkpoint/tokenizer fingerprints do not match')
+            if tokenizer.vocab_size != metadata['recipe']['model']['vocab_size']:
+                raise ValueError('Checkpoint/tokenizer vocabularies do not match')
+            atomic_json(staging / 'manifest.json', {
+                'format': 'sml-benchmark-snapshot-v1', 'source_bundle': name,
+                'step': metadata['step'], 'tokens': metadata['tokens'],
+                'best_val_loss': metadata.get('best_val_loss'), 'files': records,
+                **({'phase': 'SFT', 'parent_step': info['parent_step'],
+                    'loss_scope': 'SFT uses assistant NLL; pre-training val_loss is evaluated separately.'} if info is not None else {}),
+            })
+            staging.rename(frozen)
+        finally:
+            if staging.exists():
+                shutil.rmtree(staging)
+    saved_config = output / 'config.json'
+    if saved_config.exists() and read_json(saved_config) != config:
+        raise ValueError('Prepared configuration differs; use a fresh output directory')
+    atomic_json(saved_config, config)
+    # Load/download public data before loading any model; no extra packages needed.
+    import pyarrow as pa
+    pa.set_cpu_count(1)
+    pa.set_io_thread_count(1)
+    from datasets import DownloadConfig, load_dataset, load_from_disk
+    for task, spec in config['tasks'].items():
+        target = output / 'datasets' / task
+        if target.exists():
+            dataset = load_from_disk(str(target))
+        else:
+            print(f'[prepare] {task} · {spec["split"]} · {spec["revision"]}', flush=True)
+            dataset = load_dataset(spec['path'], spec['name'], split=spec['split'],
+                                   revision=spec['revision'],
+                                   download_config=DownloadConfig(num_proc=1, max_retries=3))
+            if len(dataset) != spec['expected_examples']:
+                raise ValueError(f'{task}: unexpected example count {len(dataset)}')
+            staging = target.with_name('.' + task + '-' + uuid.uuid4().hex)
+            dataset.save_to_disk(str(staging))
+            staging.rename(target)
+        if len(dataset) != spec['expected_examples']:
+            raise ValueError(f'{task}: prepared dataset is incomplete')
+    print('[prepared] ' + str(frozen), flush=True)
+    print(json.dumps(read_json(frozen / 'manifest.json'), indent=2), flush=True)
+
+
+def example(task, doc):
+    if task.startswith('arc_'):
+        return ('Question: ' + doc['question'] + '\nAnswer:',
+                doc['choices']['text'], doc['choices']['label'].index(doc['answerKey']))
+    if task == 'piqa':
+        return ('Question: ' + doc['goal'] + '\nAnswer:',
+                [doc['sol1'], doc['sol2']], int(doc['label']))
+    if task == 'hellaswag':
+        def clean(text):
+            text = text.strip().replace(' [title]', '. ')
+            return re.sub(r'\[.*?\]', '', text).replace('  ', ' ')
+        context = doc['activity_label'] + ': ' + doc['ctx_a'] + ' ' + doc['ctx_b'].capitalize()
+        return clean(context), [clean(text) for text in doc['endings']], int(doc['label'])
+    raise ValueError('Unknown task: ' + task)
+
+
+def pair_tokens(tokenizer, context, choice, max_length):
+    trailing = len(context) - len(context.rstrip())
+    if trailing:
+        continuation = context[-trailing:] + ' ' + choice
+        context = context[:-trailing]
+    else:
+        continuation = ' ' + choice
+    prefix = tokenizer.encode(context)
+    joint = tokenizer.encode(context + continuation)
+    if not prefix or joint[:len(prefix)] != prefix:
+        raise ValueError('Tokenizer boundary mismatch; do not silently mis-score an answer')
+    answer = joint[len(prefix):]
+    if not answer or len(answer) > max_length:
+        raise ValueError('Answer cannot be scored within the checkpoint context window')
+    # At most max_length input tokens + the final token being predicted.
+    removed = max(0, len(joint) - (max_length + 1))
+    return joint[removed:], len(prefix) - removed, removed
+
+
+def predictions(scores, choices):
+    if any(not math.isfinite(score) for score in scores) or any(not text for text in choices):
+        raise ValueError('Non-finite scores or empty choices')
+    raw = max(range(len(scores)), key=lambda i: scores[i])
+    normalized = max(range(len(scores)), key=lambda i: scores[i] / len(choices[i]))
+    return raw, normalized
+
+
+def training_records(run):
+    paths = list((run / 'jobs').glob('*/train.log'))
+    if not paths:
+        return []
+    path = max(paths, key=lambda p: p.stat().st_mtime_ns)
+    with path.open('rb') as file:
+        file.seek(max(0, path.stat().st_size - 128 * 1024))
+        text = file.read().decode('utf-8', errors='replace')
+    return [(int(step), float(rate.replace(',', '')))
+            for step, rate in re.findall(r'\[step\s+(\d+)\].*?\stok/s=([\d,\.]+)', text)
+            if float(rate.replace(',', '')) > 0]
+
+
+class TrainingGuard:
+    def __init__(self, config):
+        self.run = local_path(config['run_dir'])
+        self.ratio = config['throughput_floor_ratio']
+        records = training_records(self.run)
+        self.baseline = statistics.median(rate for _, rate in records[-20:]) if records else None
+        self.initial_step = self.last_step = records[-1][0] if records else None
+        self.last_advance = time.monotonic()
+        self.last_check = 0.0
+        self.low_checks = 0
+        self.checked_step = None
+
+    def check(self):
+        now = time.monotonic()
+        if now - self.last_check < 10:
+            return
+        self.last_check = now
+        records = training_records(self.run)
+        if not records:
+            return
+        step = records[-1][0]
+        if self.last_step is None or step > self.last_step:
+            self.last_step = step
+            self.last_advance = now
+        elif now - self.last_advance > 180:
+            raise RuntimeError('Benchmark stopped: training has not advanced for 3 minutes; inspect its logs')
+        fresh = [rate for s, rate in records if self.initial_step is not None and s > self.initial_step]
+        if self.baseline is not None and len(fresh) >= 6 and step != self.checked_step:
+            self.checked_step = step
+            recent = statistics.median(fresh[-6:])
+            self.low_checks = self.low_checks + 1 if recent < self.baseline * self.ratio else 0
+            if self.low_checks >= 2:
+                raise RuntimeError(f'Benchmark stopped: recent training throughput {recent:,.0f} '
+                                   f'is below {self.ratio:.0%} of baseline {self.baseline:,.0f}')
+
+
+class Scorer:
+    def __init__(self, output, config, guard):
+        import mlx.core as mx
+        device=config.get('device','cpu')
+        if device not in ('cpu','gpu'):raise ValueError('Benchmark device must be cpu or gpu')
+        mx.set_default_device(mx.gpu if device=='gpu' else mx.cpu)
+        mx.set_cache_limit(32 * 1024**2)
+        mx.set_memory_limit(int(config['rss_stop_gib'] * 1024**3))  # MLX guideline, not an OS cap.
+        from sml_v1.model import TransformerConfig, TransformerLM
+        from sml_v1.inference import _cast_model_floats
+        from sml_v1.tokenization import Tokenizer
+        frozen = output / 'checkpoint'
+        manifest = read_json(frozen / 'manifest.json')
+        for name, record in manifest['files'].items():
+            path = frozen / name
+            if path.stat().st_size != record['bytes'] or sha256(path) != record['sha256']:
+                raise ValueError('Frozen checkpoint integrity check failed: ' + name)
+        metadata = read_json(frozen / 'model.safetensors.json')
+        model_config = dict(metadata['recipe']['model'])
+        # Same weights and equations, CPU-supported reference operations.
+        model_config.update(attention_impl='vanilla', ce_impl='reference', ffn_impl='reference')
+        self.model = TransformerLM(TransformerConfig(**model_config))
+        precision = mx.float32 if metadata.get('compute_precision') == 'float32' else mx.bfloat16
+        _cast_model_floats(self.model, precision)
+        self.model.load_weights(str(frozen / 'model.safetensors'), strict=True)
+        from mlx.utils import tree_flatten
+        if any(value.dtype != precision for _, value in tree_flatten(self.model.parameters())):
+            raise ValueError('Saved weight dtype differs from declared inference precision')
+        self.model.eval()
+        mx.eval(self.model.parameters())
+        self.tokenizer = Tokenizer(frozen / 'tokenizer')
+        if self.tokenizer.fingerprint != checkpoint_tokenizer(metadata):
+            raise ValueError('Frozen tokenizer integrity mismatch')
+        self.mx, self.config, self.guard = mx, config, guard
+        self.max_length = model_config['max_seq_len']
+        self.check_memory()
+        print(f'[model] step={metadata["step"]} tokens={metadata["tokens"]:,} '
+              f'device={mx.default_device()} batch={config.get("batch_size", 1)} '
+              f'chunk={config["chunk_tokens"]}', flush=True)
+
+    def check_memory(self):
+        if peak_rss_gib() > self.config['rss_stop_gib']:
+            raise RuntimeError('Benchmark stopped: process peak RSS exceeded its configured budget')
+
+    def score(self, context, choice):
+        return self.score_choices(context, [choice])[0]
+
+    def score_choices(self, context, choices):
+        """Score right-padded answer sequences together, excluding all padding.
+
+        Causal attention prevents trailing padding from affecting real tokens.
+        Each row retains its own continuation boundary and truncation count.
+        """
+        if not choices:
+            return []
+        mx = self.mx
+        pairs = [pair_tokens(self.tokenizer, context, choice, self.max_length) for choice in choices]
+        totals, caches = [0.0] * len(pairs), None
+        input_length = max(len(tokens) - 1 for tokens, _, _ in pairs)
+        chunk = self.config['chunk_tokens']
+        try:
+            for start in range(0, input_length, chunk):
+                self.guard.check()
+                stop = min(start + chunk, input_length)
+                rows = []
+                for tokens, _, _ in pairs:
+                    row = tokens[:-1][start:stop]
+                    rows.append(row + [0] * (stop - start - len(row)))
+                ids = mx.array(rows, dtype=mx.int32)
+                logits, next_caches = self.model.logits(ids, caches=caches)
+                scores = []
+                for index, (tokens, prefix_length, _) in enumerate(pairs):
+                    first = max(start, prefix_length - 1)
+                    end = min(stop, len(tokens) - 1)
+                    if first >= end:
+                        continue
+                    selected = logits[index, first - start:end - start, :].astype(mx.float32)
+                    targets = mx.array(tokens[first + 1:end + 1], dtype=mx.int32)
+                    target_logits = mx.take_along_axis(selected, targets[:, None], axis=-1)[:, 0]
+                    score = (target_logits - mx.logsumexp(selected, axis=-1)).sum()
+                    scores.append((index, score))
+                    del selected, targets, target_logits, score
+                mx.eval(next_caches, [score for _, score in scores])
+                for index, score in scores:
+                    totals[index] += float(score.item())
+                del scores
+                caches = next_caches
+                del logits, ids, next_caches
+                self.check_memory()
+            if any(not math.isfinite(total) for total in totals):
+                raise ValueError('Non-finite likelihood')
+            return [(total, pair[2]) for total, pair in zip(totals, pairs)]
+        finally:
+            caches = None
+            mx.clear_cache()
+            time.sleep(self.config['pause_seconds'])
+
+
+def summarize(output, config, limit, status):
+    results = {}
+    for task, spec in config['tasks'].items():
+        path = output / f'{task}.jsonl'
+        rows = [json.loads(line) for line in path.read_text().splitlines()] if path.exists() else []
+        selected = [row for row in rows if limit == 0 or row['index'] < limit]
+        count = len(selected)
+        intended = min(limit or spec['expected_examples'], spec['expected_examples'])
+        results[task] = {
+            'evaluated': count, 'expected': intended, 'complete': count == intended,
+            'acc': sum(row['correct'] for row in selected) / count if count else None,
+            'acc_norm': sum(row['correct_norm'] for row in selected) / count if count else None,
+            'context_truncated_choices': sum(sum(n > 0 for n in row['truncated']) for row in selected),
+        }
+    report = {
+        'format': 'sml-choice-results-v1', 'status': status,
+        'checkpoint': read_json(output / 'checkpoint/manifest.json'),
+        'settings': {'device': config.get('device','cpu'), 'batch_size': config.get('batch_size', 1),
+                     'chunk_tokens': config['chunk_tokens'],
+                     'limit_per_task': limit, 'num_fewshot': 0, 'chat_template': False,
+                     'scoring': 'sum continuation log-likelihood; acc_norm divides by choice characters',
+                     'implementation': 'native MLX; lm-eval-style task prompts; not harness-validated'},
+        'datasets': config['tasks'], 'results': results,
+    }
+    metadata_path = output / 'checkpoint/model.safetensors.json'
+    if metadata_path.exists():
+        metadata = read_json(metadata_path)
+        report['settings'].update(
+            compute_precision=metadata.get('compute_precision', 'bfloat16'),
+            weight_provenance=metadata.get('inference_export'),
+            mlx_enable_tf32=os.environ.get('MLX_ENABLE_TF32', 'default'))
+    atomic_json(output / 'results.json', report)
+    return report
+
+
+def run(config, output, limit):
+    if not (output / 'config.json').exists():
+        raise ValueError('Run --prepare first; it freezes the best checkpoint and downloads data')
+    if read_json(output / 'config.json') != config:
+        raise ValueError('Settings differ from the frozen run configuration')
+    import pyarrow as pa
+    pa.set_cpu_count(1)
+    pa.set_io_thread_count(1)
+    from datasets import load_from_disk
+    guard = TrainingGuard(config)
+    print(f'[guard] training baseline={guard.baseline}; stop below '
+          f'{config["throughput_floor_ratio"]:.0%}; RSS stop={config["rss_stop_gib"]} GiB', flush=True)
+    started = time.monotonic()
+    status = 'running'
+    try:
+        scorer = Scorer(output, config, guard)
+        for task in config['tasks']:
+            dataset = load_from_disk(str(output / 'datasets' / task))
+            count = min(limit or len(dataset), len(dataset))
+            path = output / f'{task}.jsonl'
+            done = {json.loads(line)['index'] for line in path.read_text().splitlines()} if path.exists() else set()
+            with path.open('a') as log:
+                for index in range(count):
+                    if index in done:
+                        continue
+                    context, choices, gold = example(task, dataset[index])
+                    scores, truncated = [], []
+                    batch_size = config.get('batch_size', 1)
+                    for start in range(0, len(choices), batch_size):
+                        for score, removed in scorer.score_choices(context, choices[start:start + batch_size]):
+                            scores.append(score)
+                            truncated.append(removed)
+                    raw, normalized = predictions(scores, choices)
+                    log.write(json.dumps({'index': index, 'scores': scores, 'gold': gold,
+                                          'prediction': raw, 'prediction_norm': normalized,
+                                          'correct': raw == gold, 'correct_norm': normalized == gold,
+                                          'truncated': truncated}, allow_nan=False) + '\n')
+                    log.flush()
+                    if (index + 1) % 10 == 0 or index + 1 == count:
+                        summarize(output, config, limit, 'running')
+                        print(f'[{task}] {index + 1}/{count} elapsed={time.monotonic() - started:.1f}s', flush=True)
+            summarize(output, config, limit, 'running')
+        status = 'complete'
+    except (KeyboardInterrupt, Exception) as error:
+        status = 'interrupted' if isinstance(error, KeyboardInterrupt) else 'stopped'
+        print(f'[{status}] {error or "Ctrl+C"}; completed examples are saved. Re-run to resume.', flush=True)
+        if not isinstance(error, KeyboardInterrupt):
+            raise
+    finally:
+        report = summarize(output, config, limit, status)
+        report['runtime'] = {'elapsed_seconds': time.monotonic() - started,
+                             'peak_rss_gib': peak_rss_gib(),
+                             'training_baseline_tokens_per_second': guard.baseline,
+                             'last_observed_training_step': guard.last_step}
+        atomic_json(output / 'results.json', report)
+        print(json.dumps(report['results'], indent=2), flush=True)
+        print(f'[memory] peak process RSS={peak_rss_gib():.3f} GiB', flush=True)
+        print('[results] ' + str(output / 'results.json'), flush=True)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--config', type=Path, default=DEFAULT_CONFIG)
+    parser.add_argument('--prepare', action='store_true')
+    parser.add_argument('--limit', type=int, default=0, help='First N examples per task; 0 runs full splits')
+    args = parser.parse_args()
+    if args.limit < 0:
+        parser.error('--limit must be >= 0')
+    config = read_json(args.config)
+    if config.get('device','cpu') not in ('cpu','gpu'):parser.error('device must be cpu or gpu')
+    output = local_path(config['output_dir'])
+    if not 1 <= config['chunk_tokens'] <= 128:
+        parser.error('Chunk size must be between 1 and 128 tokens')
+    if type(config.get('batch_size', 1)) is not int or not 1 <= config.get('batch_size', 1) <= 8:
+        parser.error('Batch size must be an integer between 1 and 8 answer sequences')
+    configure_process(output.parent / 'cache')
+    if args.prepare:
+        prepare(config, output)
+    else:
+        # Prevent two evaluators appending to the same result logs.
+        import fcntl
+        output.mkdir(parents=True, exist_ok=True)
+        with (output / '.run.lock').open('w') as lock:
+            fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
+            run(config, output, args.limit)
+
+
+if __name__ == '__main__':
+    main()

evaluation/benchmark_ifeval.py

@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+"""Generate SFT answers and score with the existing pinned Google IFEval code."""
+import argparse
+import dataclasses
+import fcntl
+import hashlib
+import json
+import os
+from pathlib import Path
+import random
+import sys
+import time
+
+ROOT = Path(__file__).resolve().parents[2]
+RUNTIME = ROOT / 'reports/benchmarks/ifeval-runtime-20261008'
+sys.path[:0] = [str(ROOT/'pretraining'), str(RUNTIME/'site-packages'), str(RUNTIME)]
+
+
+def digest(path):
+    with Path(path).open('rb') as stream:
+        return hashlib.file_digest(stream, 'sha256').hexdigest()
+
+
+def read(path):
+    return json.loads(Path(path).read_text())
+
+
+def write(path, value):
+    path = Path(path)
+    temp = path.with_suffix(path.suffix+'.tmp')
+    temp.write_text(json.dumps(value, indent=2, ensure_ascii=False, allow_nan=False)+'\n')
+    temp.replace(path)
+
+
+def persist_rows(path, rows):
+    temp = path.with_suffix('.jsonl.tmp')
+    temp.write_text(''.join(json.dumps(row, ensure_ascii=False)+'\n' for row in rows))
+    temp.replace(path)
+
+
+def resolve_weights(path):
+    """Accept a verified inference snapshot as well as a training bundle."""
+    directory = Path(path).absolute()
+    if directory.name == 'model.safetensors':
+        directory = directory.parent
+    manifest_path = directory/'manifest.json'
+    if manifest_path.exists():
+        manifest = read(manifest_path)
+        if manifest.get('format') == 'sml-benchmark-snapshot-v1':
+            required = {'model.safetensors', 'model.safetensors.json'}
+            if directory.is_symlink() or not required <= manifest['files'].keys():
+                raise ValueError('Incomplete inference snapshot')
+            for name, record in manifest['files'].items():
+                if Path(name).name != name:
+                    raise ValueError('Unsafe inference snapshot filename')
+                file = directory/name
+                if file.is_symlink() or file.stat().st_size != record['bytes'] or digest(file) != record['sha256']:
+                    raise ValueError('Inference snapshot integrity check failed: '+name)
+            return str(directory/'model.safetensors')
+    from sml_v1.checkpoint_bundle import resolve_bundle
+    return resolve_bundle(path)
+
+
+def score(inputs, records):
+    from instruction_following_eval import evaluation_lib as ev
+    outputs = {'strict': [], 'loose': []}
+    for inp, record in zip(inputs, records):
+        example = ev.InputExample(**inp)
+        response = {inp['prompt']: record['response']}
+        for kind in outputs:
+            random.seed(2026)
+            check = getattr(ev, 'test_instruction_following_'+kind)
+            outputs[kind].append(dataclasses.asdict(check(example, response)))
+    metrics = {}
+    for kind, rows in outputs.items():
+        count = sum(len(row['follow_instruction_list']) for row in rows)
+        metrics[kind] = {
+            'prompt_accuracy': sum(row['follow_all_instructions'] for row in rows)/len(rows) if rows else None,
+            'instruction_accuracy': sum(sum(row['follow_instruction_list']) for row in rows)/count if count else None,
+            'prompts': len(rows), 'instructions': count,
+        }
+    return metrics, outputs
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--checkpoint', type=Path, required=True)
+    parser.add_argument('--tokenizer', type=Path, default=ROOT/'reference/opensml-150m/tokenizer/bytebpe32k_v1')
+    parser.add_argument('--output', type=Path, required=True)
+    parser.add_argument('--max-tokens', type=int, default=1024)
+    parser.add_argument('--limit', type=int, default=0, help='Smoke/subset only; 0 runs all 541 prompts')
+    parser.add_argument('--prepare-only', action='store_true')
+    args = parser.parse_args()
+    if args.limit < 0 or not 1 <= args.max_tokens <= 2048:
+        parser.error('limit must be nonnegative; max-tokens must be 1–2048')
+    args.output.mkdir(parents=True, exist_ok=True)
+    os.environ['TOKENIZERS_PARALLELISM'] = 'false'
+    os.environ['NLTK_DATA'] = str(RUNTIME/'nltk_data')
+    with (args.output/'.run.lock').open('w') as lock:
+        fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
+        from sml_v1.tokenization import Tokenizer
+        weights = resolve_weights(args.checkpoint)
+        metadata = read(weights+'.json')
+        precision = metadata.get('compute_precision', 'bfloat16')
+        if precision not in ('float32', 'bfloat16'):
+            raise ValueError('Unsupported inference precision: '+str(precision))
+        model_config = dict(metadata['recipe']['model'])
+        if precision == 'float32':
+            model_config.update(attention_impl='vanilla', ce_impl='reference', ffn_impl='reference')
+        from sft.metadata import sft_info
+        info=sft_info(metadata)
+        if info is None:
+            raise ValueError('This runner requires an SFT chat checkpoint')
+        tokenizer = Tokenizer(args.tokenizer)
+        if tokenizer.fingerprint != info['contract']['tokenizer']:
+            raise ValueError('Tokenizer mismatch')
+        source = read(RUNTIME/'source.json')
+        for name,expected in source.get('evaluator_blob_sha1',{}).items():
+            raw=(RUNTIME/'instruction_following_eval'/name).read_bytes()
+            if hashlib.sha1(b'blob '+str(len(raw)).encode()+b'\0'+raw).hexdigest()!=expected:
+                raise ValueError('Pinned IFEval evaluator changed: '+name)
+        identity = {
+            'format': 'opensml-ifeval-v1', 'checkpoint': str(Path(weights).parent),
+            'step': metadata['step'], 'model_sha256': digest(weights),
+            'metadata_sha256': digest(weights+'.json'), 'tokenizer': tokenizer.fingerprint,
+            'dataset': source['dataset'], 'dataset_revision': source['dataset_revision'],
+            'google_research_commit': source['google_research_commit'],
+            'evaluator_files': {p.name: digest(p) for p in sorted((RUNTIME/'instruction_following_eval').glob('*.py'))},
+            'runner_sha256': digest(__file__), 'temperature': 0,
+            'repetition_penalty': 1, 'max_tokens': args.max_tokens,
+            'system_prompt': None, 'serialization': 'native SFT chat template',
+            'prompt_truncation': False, 'context_limit': metadata['recipe']['model']['max_seq_len'],
+            'compute_precision': precision, 'model_implementation': model_config,
+            'weight_provenance': metadata.get('inference_export'),
+            'mlx_enable_tf32': os.environ.get('MLX_ENABLE_TF32', 'default'),
+        }
+        identity_path = args.output/'protocol.json'
+        if identity_path.exists() and read(identity_path) != identity:
+            raise ValueError('Protocol differs; choose a fresh output directory')
+        write(identity_path, identity)
+        inputs_path = args.output/'inputs.json'
+        receipt_path = args.output/'inputs.sha256.json'
+        if inputs_path.exists():
+            if read(receipt_path)['sha256'] != digest(inputs_path):
+                raise ValueError('Frozen IFEval inputs changed')
+            inputs = read(inputs_path)
+        else:
+            from datasets import load_dataset
+            dataset = load_dataset(source['dataset'], 'default', split='train', revision=source['dataset_revision'],
+                                   cache_dir=str(ROOT/'pretraining/data/.sft_hf_cache'))
+            inputs = [dict(key=int(row['key']), prompt=row['prompt'],
+                           instruction_id_list=list(row['instruction_id_list']),
+                           kwargs=[{k: v for k, v in item.items() if v is not None} for item in row['kwargs']])
+                      for row in dataset]
+            if len(inputs) != 541 or len({row['key'] for row in inputs}) != 541:
+                raise ValueError('Unexpected IFEval prompt count or duplicate keys')
+            write(inputs_path, inputs)
+            write(receipt_path, {'sha256': digest(inputs_path), 'expected': 541})
+        if len(inputs) != 541:
+            raise ValueError('Frozen prompt count differs')
+        import nltk
+        from langdetect import DetectorFactory
+        DetectorFactory.seed = 2026
+        nltk.data.path.insert(0, str(RUNTIME/'nltk_data'))
+        try:
+            nltk.data.load('nltk:tokenizers/punkt/english.pickle')
+        except LookupError:
+            for resource in ('punkt', 'punkt_tab'):
+                if not nltk.download(resource, download_dir=str(RUNTIME/'nltk_data'), quiet=True):
+                    raise RuntimeError('Could not prepare NLTK '+resource)
+            nltk.data.load('nltk:tokenizers/punkt/english.pickle')
+        # Exercise the scorer on every prompt before loading the model.
+        score(inputs, [{'response': ''} for _ in inputs])
+        if args.prepare_only:
+            print(f'[prepared] IFEval 541 prompts; SFT step {metadata["step"]}; {args.output}', flush=True)
+            return
+        responses_path = args.output/'responses.jsonl'
+        records = [json.loads(line) for line in responses_path.read_text().splitlines()] if responses_path.exists() else []
+        for inp, record in zip(inputs, records):
+            if record['key'] != inp['key'] or record['prompt'] != inp['prompt']:
+                raise ValueError('Saved responses differ from frozen prompt order')
+        if len(records) > len(inputs):
+            raise ValueError('Too many saved responses')
+        selected = inputs[:args.limit] if args.limit else inputs
+        if len(records) < len(selected):
+            import mlx.core as mx
+            from sml_v1.model import TransformerConfig, TransformerLM
+            from sft.data import serialize
+            from sft.evaluate import generate
+            model = TransformerLM(TransformerConfig(**model_config))
+            model.load_weights(weights)
+            dtype = mx.float32 if precision == 'float32' else mx.bfloat16
+            from mlx.utils import tree_flatten
+            if any(value.dtype != dtype for _, value in tree_flatten(model.parameters())):
+                raise ValueError('Saved weight dtype differs from declared inference precision')
+            model.eval()
+            mx.eval(model.parameters())
+            print(f'[model] compute_precision={precision} attention={model_config["attention_impl"]}', flush=True)
+            started = time.monotonic()
+            for inp in selected[len(records):]:
+                messages = [{'role': 'user', 'content': inp['prompt']}]
+                ids, _ = serialize(tokenizer, messages, generation=True)
+                available = model.cfg.max_seq_len-len(ids)
+                if available < 1:
+                    raise ValueError(f'Prompt {inp["key"]} exceeds context; no silent truncation')
+                result = generate(model, tokenizer, messages, limit=args.max_tokens, temperature=0)
+                records.append(dict(key=inp['key'], prompt=inp['prompt'], response=result['text'],
+                                    tokens=result['tokens'], stop_reason=result['stop_reason'],
+                                    prompt_tokens=len(ids), effective_max_tokens=min(args.max_tokens, available)))
+                persist_rows(responses_path, records)
+                if len(records)%10 == 0 or len(records)==len(selected):
+                    print(f'[ifeval] {len(records)}/{len(selected)} elapsed={time.monotonic()-started:.1f}s', flush=True)
+        metrics, outputs = score(inputs[:len(records)], records)
+        for kind, rows in outputs.items():
+            persist_rows(args.output/f'{kind}.jsonl', rows)
+        result = dict(protocol=identity, evaluated=len(records), expected=541,
+                      complete=len(records)==541, metrics=metrics,
+                      token_limited=sum(row['stop_reason']=='token_limit' for row in records),
+                      context_limited=sum(row['effective_max_tokens']<args.max_tokens for row in records),
+                      scope='Official pinned IFEval strict/loose checks; native SFT generation, no harness parity claim.')
+        write(args.output/'results.json', result)
+        print(json.dumps(result, indent=2))
+        print('[results] '+str(args.output/'results.json'), flush=True)
+
+
+if __name__ == '__main__':
+    main()

evaluation/chat_adapter_search_bench.py

@@ -0,0 +1,125 @@
+"""Matched full FP32 choice/IFEval/chat evaluation for a merged adapter export."""
+from pathlib import Path
+import json
+import os
+import shutil
+import subprocess
+import sys
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT/'pretraining'))
+from sml_v1.common import read_json, atomic_json, file_sha256
+
+BASELINE = ROOT/'reports/benchmarks/deep44-reliable-chat-167-fp32-20261106'
+
+
+def verify_scoring():
+    receipt = read_json(BASELINE/'readiness.json')
+    names = ['pretraining/scripts/benchmark_best.py', 'pretraining/scripts/benchmark_ifeval.py',
+             'pretraining/sml_v1/model.py', 'pretraining/sml_v1/tokenization.py',
+             'pretraining/sft/data.py', 'pretraining/sft/evaluate.py',
+             'pretraining/chat_benchmark/panel.py', 'pretraining/chat_benchmark/scoring.py']
+    for name in names:
+        if file_sha256(ROOT/name) != receipt['files'][name]:
+            raise ValueError('Scoring no longer matches 167: '+name)
+    return {n:receipt['files'][n] for n in names}
+
+
+def chat(snapshot, output, training_data):
+    import mlx.core as mx
+    from mlx.utils import tree_flatten
+    from sml_v1.model import TransformerLM, TransformerConfig
+    from sml_v1.tokenization import Tokenizer
+    from sft.evaluate import generate
+    from chat_benchmark.panel import build
+    from chat_benchmark.scoring import score, summarize, normalize
+    cases = build()
+    expected = [json.loads(l) for l in (BASELINE/'chat/cases.jsonl').read_text().splitlines()]
+    if cases != expected: raise ValueError('Private chat cases differ from 167')
+    prompts = {normalize(m['content']) for l in Path(training_data).read_text().splitlines()
+               for m in json.loads(l)['messages'] if m['role']=='user'}
+    if any(normalize(c['messages'][-1]['content']) in prompts for c in cases):
+        raise ValueError('Private chat prompt overlaps training')
+    meta = read_json(snapshot/'model.safetensors.json')
+    model = TransformerLM(TransformerConfig(**dict(meta['recipe']['model'], attention_impl='vanilla',
+                                                    ffn_impl='reference', ce_impl='reference')))
+    model.load_weights(str(snapshot/'model.safetensors'), strict=True)
+    mx.eval(model.parameters()); model.eval()
+    if any(v.dtype != mx.float32 for _,v in tree_flatten(model.parameters())):
+        raise ValueError('Chat evaluation requires FP32 weights')
+    tok = Tokenizer(snapshot/'tokenizer')
+    output.mkdir(parents=True, exist_ok=True)
+    path = output/'responses.jsonl'
+    rows = [json.loads(l) for l in path.read_text().splitlines()] if path.exists() else []
+    if len(rows)>len(cases) or any(r['case']!=c for r,c in zip(rows,cases)):
+        raise ValueError('Saved chat responses differ')
+    with path.open('a') as stream:
+        for case in cases[len(rows):]:
+            value=generate(model,tok,case['messages'],limit=case['max_tokens'],temperature=0)
+            row=dict(case=case,response=value,scores=score(case,value));rows.append(row)
+            stream.write(json.dumps(row,ensure_ascii=False)+'\n');stream.flush()
+            atomic_json(output/'progress.json',dict(completed=len(rows),total=len(cases)))
+    result=summarize(rows)
+    atomic_json(output/'results.json',result)
+    atomic_json(output/'protocol.json',dict(scoring=verify_scoring(),cases=file_sha256(BASELINE/'chat/cases.jsonl'),
+        precision='FP32 base masters plus FP32 trained deltas',model_sha256=file_sha256(snapshot/'model.safetensors'),
+        weight_provenance=meta['inference_export'],temperature=0,repetition_penalty=1,
+        attention='vanilla',ffn='reference',system_prompt=None,prompt_truncation=False))
+    return result
+
+
+def run(bundle, staging, output):
+    verify_scoring()
+    bundle,staging,output=map(Path,(bundle,staging,output))
+    output.mkdir(parents=True,exist_ok=True)
+    snapshot=output/'choices/checkpoint'
+    if not snapshot.exists():
+        snapshot.mkdir(parents=True)
+        source=bundle/'inference'
+        manifest=read_json(source/'manifest.json')
+        for name in ('model.safetensors','model.safetensors.json'):
+            if file_sha256(source/name)!=manifest['files'][name]['sha256']:
+                raise ValueError('Merged export checksum mismatch')
+            shutil.copyfile(source/name,snapshot/name)
+        shutil.copytree(staging/'tokenizer',snapshot/'tokenizer')
+        meta=read_json(snapshot/'model.safetensors.json')
+        atomic_json(snapshot/'manifest.json',dict(format='sml-benchmark-snapshot-v1',source_bundle=bundle.name,
+            step=meta['step'],tokens=meta['tokens'],phase='SFT',files=manifest['files']))
+    cfg=read_json(ROOT/'configs/benchmark_reliable_chat_167_fp32_20261106.json')
+    cfg.update(run_dir=str(bundle.parent),output_dir=str(output/'choices'),explicit_checkpoint=str(bundle/'inference'))
+    config=output/'config.json';atomic_json(config,cfg)
+    if not (output/'choices/datasets').exists():
+        shutil.copytree(BASELINE/'choices/datasets',output/'choices/datasets')
+    (output/'ifeval').mkdir(exist_ok=True)
+    for name in ('inputs.json','inputs.sha256.json'):
+        target=output/'ifeval'/name
+        if not target.exists():shutil.copyfile(BASELINE/'ifeval'/name,target)
+    env=dict(os.environ,HF_HUB_OFFLINE='1',HF_DATASETS_OFFLINE='1',HF_HUB_DISABLE_IMPLICIT_TOKEN='1',TOKENIZERS_PARALLELISM='false')
+    commands=[('choices-prepare',[sys.executable,str(ROOT/'pretraining/scripts/benchmark_best.py'),'--config',str(config),'--prepare']),
+              ('choices',[sys.executable,'-u',str(ROOT/'pretraining/scripts/benchmark_best.py'),'--config',str(config)]),
+              ('ifeval',[sys.executable,'-u',str(ROOT/'pretraining/scripts/benchmark_ifeval.py'),'--checkpoint',str(snapshot),
+                         '--tokenizer',str(snapshot/'tokenizer'),'--output',str(output/'ifeval'),'--max-tokens','1024'])]
+    for label,command in commands:
+        with (output/(label+'.log')).open('a') as log:
+            subprocess.run(command,cwd=ROOT,env=env,stdout=log,stderr=subprocess.STDOUT,check=True)
+    choices=read_json(output/'choices/results.json');ifeval=read_json(output/'ifeval/results.json')
+    if choices['status']!='complete' or not ifeval['complete'] or ifeval['evaluated']!=541:
+        raise ValueError('Full benchmark did not complete')
+    for name,spec in cfg['tasks'].items():
+        if not choices['results'][name]['complete'] or choices['results'][name]['evaluated']!=spec['expected_examples']:
+            raise ValueError('Incomplete choice task')
+    old=read_json(BASELINE/'ifeval/protocol.json');new=ifeval['protocol']
+    for key in ('dataset','dataset_revision','google_research_commit','evaluator_files','runner_sha256',
+                'temperature','repetition_penalty','max_tokens','system_prompt','serialization',
+                'prompt_truncation','context_limit','compute_precision','model_implementation','mlx_enable_tf32'):
+        if old[key]!=new[key]:raise ValueError('IFEval differs from 167: '+key)
+    result=dict(choices=choices['results'],ifeval=ifeval['metrics'],
+                chat=chat(snapshot,output/'chat',staging/'corpus/train.jsonl'),
+                complete=True,model_sha256=file_sha256(snapshot/'model.safetensors'))
+    atomic_json(output/'summary.json',result)
+    return result
+
+
+if __name__=='__main__':
+    if len(sys.argv)!=4:raise SystemExit('bundle staging output required')
+    run(*sys.argv[1:])

evaluation/chat_benchmark/__init__.py

@@ -0,0 +1 @@
+"""Frozen, text-only chatbot evaluations; never used to select training weights."""

evaluation/chat_benchmark/panel.py

@@ -0,0 +1,136 @@
+"""Authored English chatbot diagnostic v1, frozen before seeing model outputs.
+
+30 cases per objective category, plus 12 qualitative conversations. Variants
+within a family are correlated; this is not a representative population sample.
+"""
+
+
+def build():
+    rows = []
+
+    def add(category, family, prompt, answers, *, messages=None, case_sensitive=False,
+            answerable=True, qualitative=False):
+        number = sum(r['category'] == category for r in rows)
+        rows.append(dict(id=f'{category}-{number:02d}', category=category, family=family,
+                         messages=messages or [dict(role='user', content=prompt)],
+                         answers=answers, case_sensitive=case_sensitive,
+                         answerable=answerable, qualitative=qualitative,
+                         max_tokens=160 if qualitative else 128))
+
+    facts = [
+        ('What is the capital of Japan?', ['Tokyo']),
+        ('Which planet is closest to the Sun?', ['Mercury']),
+        ('What is the chemical symbol for gold?', ['Au']),
+        ('How many sides does a hexagon have?', ['six', '6']),
+        ('Which ocean is the largest on Earth?', ['Pacific', 'Pacific Ocean', 'the Pacific Ocean']),
+        ('Who wrote Pride and Prejudice?', ['Jane Austen']),
+        ('On which continent is Egypt?', ['Africa']),
+        ('What force pulls objects toward Earth?', ['gravity', 'gravitational force']),
+        ('What gas do humans need to breathe to survive?', ['oxygen']),
+        ('What is the process by which plants use sunlight to make food called?', ['photosynthesis']),
+        ('How many minutes are in one hour?', ['60', 'sixty']),
+        ('At standard atmospheric pressure, at what Celsius temperature does pure water freeze?', ['0', 'zero', '0 degrees Celsius', '0°C', 'zero degrees Celsius']),
+        ('Which instrument is used to measure temperature?', ['thermometer', 'a thermometer']),
+        ('Which organ pumps blood around the human body?', ['heart', 'the heart']),
+        ('What is the largest living land animal?', ['African elephant', 'African bush elephant', 'African savanna elephant', 'elephant', 'the African elephant']),
+        ('What is the opposite of north?', ['south']),
+        ('What is the name of the natural satellite that orbits Earth?', ['Moon', 'the Moon']),
+        ('Who painted the Mona Lisa?', ['Leonardo da Vinci', 'da Vinci']),
+        ('In which country is the city of Barcelona?', ['Spain']),
+        ('How many days does a leap year have?', ['366', 'three hundred sixty-six', 'three hundred and sixty-six']),
+        ('What is the freezing of a liquid into a solid called?', ['freezing', 'solidification']),
+        ('What is the square root of 81?', ['9', 'nine']),
+        ('Which language is primarily spoken in Brazil?', ['Portuguese', 'Brazilian Portuguese']),
+        ('Which animal is known for building dams?', ['beaver', 'beavers', 'a beaver']),
+        ('What part of a plant typically absorbs water from the soil?', ['roots', 'the roots', 'root']),
+        ('What is the main gas in Earth’s atmosphere?', ['nitrogen']),
+        ('How many degrees are in a right angle?', ['90', 'ninety', '90 degrees']),
+        ('Which metal is liquid at ordinary room temperature, around 20 degrees Celsius?', ['mercury']),
+        ('What is the name of the galaxy that contains our solar system?', ['Milky Way', 'the Milky Way']),
+        ('Which sense uses the ears?', ['hearing']),
+    ]
+    for q, answers in facts:
+        add('knowledge', 'stable_fact', q + ' Give a short answer.', answers)
+
+    names = ['Anika', 'Bram', 'Celia', 'Dante', 'Esme', 'Farid']
+    other = ['Greta', 'Hector', 'Ines', 'Jules', 'Kira', 'Lucio']
+    colors = ['teal', 'maroon', 'ivory', 'indigo', 'ochre', 'turquoise']
+    items = ['notebook', 'umbrella', 'camera', 'scarf', 'compass', 'lantern']
+    cities = ['Kyoto', 'Porto', 'Bergen', 'Ghent', 'Bari', 'Lyon']
+    days = ['Tuesday', 'Friday', 'Sunday', 'Monday', 'Thursday', 'Wednesday']
+    foods = ['rice', 'pasta', 'lentils', 'potatoes', 'noodles', 'beans']
+    for i, name in enumerate(names):
+        lead = 'Use only the information below. Give a short answer.\n\n'
+        add('reading', 'contrast', lead + f'{name} brought a {colors[i]} {items[i]}. {other[i]} brought a black {items[i]}.\nWhat color was {name}’s {items[i]}?', [colors[i]])
+        add('reading', 'schedule_exception', lead + f'The club normally meets on Saturday. This week its meeting has been moved to {days[i]}.\nOn which day is this week’s meeting?', [days[i]])
+        add('reading', 'relation', lead + f'{name} visited {cities[i]}. {other[i]} stayed in Naples.\nWhich city did {name} visit?', [cities[i]])
+        add('reading', 'negation', lead + f'For dinner, {name} chose {foods[i]}, not bread.\nWhat did {name} choose for dinner?', [foods[i]])
+        add('reading', 'quantity', lead + f'The green shelf has {11+i} books. The orange shelf has {23+i} books.\nHow many books are on the orange shelf?', [str(23+i)])
+
+        a, b = 7+i, 3+i
+        add('reasoning', 'addition', f'{name} has {a} postcards and receives {b} more. How many postcards does {name} have now? Give only the number.', [str(a+b)])
+        add('reasoning', 'subtraction', f'A tray contains {20+2*i} cups. Someone removes {b} cups. How many remain? Give only the number.', [str(20+2*i-b)])
+        add('reasoning', 'multiplication', f'There are {b} packets, each containing 4 stickers. How many stickers are there altogether? Give only the number.', [str(b*4)])
+        add('reasoning', 'ordering', f'{name} is taller than {other[i]}. {other[i]} is taller than Zola. Who is tallest? Give only the name.', [name])
+        add('reasoning', 'state_change', f'{name} puts a coin in a drawer, then moves it into a cup. The drawer is closed. Where is the coin now? Give only the container.', ['cup', 'the cup'])
+
+        add('instructions', 'case', f'Convert this word to uppercase and return only the result: {items[i]}', [items[i].upper()], case_sensitive=True)
+        words = [colors[i], items[i], cities[i].lower()]
+        add('instructions', 'reverse_list', 'Reverse the order of these three items. Return the items separated by commas, with no extra words: ' + ', '.join(words), [', '.join(reversed(words))])
+        values = [10+i*2, 3+i*2, 8+i*2]
+        add('instructions', 'sort_numbers', 'Put these numbers in ascending order. Return only the comma-separated numbers: ' + ', '.join(map(str, values)), [', '.join(map(str, sorted(values)))])
+        add('instructions', 'replacement', f'Replace "small" with "large" in this sentence and return only the revised sentence: The small {items[i]} is here.', [f'The large {items[i]} is here.'], case_sensitive=True)
+        add('instructions', 'copy', f'Repeat this label exactly, without a prefix or explanation: QZ-{417+i}-R', [f'QZ-{417+i}-R'], case_sensitive=True)
+
+        def dialogue(first, last, middle=None):
+            turns = [dict(role='user', content=first), dict(role='assistant', content='Understood.')]
+            if middle:
+                turns += [dict(role='user', content=middle), dict(role='assistant', content='Understood.')]
+            return turns + [dict(role='user', content=last)]
+        add('conversation', 'correction', '', [str(317+i)], messages=dialogue(f'My reservation number is {217+i}.', 'What is my current reservation number? Give only the number.', f'Correction: my reservation number is {317+i}.'))
+        add('conversation', 'preference', '', [foods[i]], messages=dialogue(f'My favorite dinner is {foods[i]}.', 'What did I say my favorite dinner was? Give only the food.'))
+        add('conversation', 'entity_tracking', '', [colors[i]], messages=dialogue(f'My name is {name}. My bag is {colors[i]}. {other[i]} has a black bag.', 'What color is my bag? Give only the color.'))
+        add('conversation', 'distraction', '', [cities[i]], messages=dialogue(f'My next trip is to {cities[i]}.', 'Which city did I say I am visiting? Give only the city.', 'I am packing a coat and two books.'))
+        add('conversation', 'updated_plan', '', [days[i]], messages=dialogue('I planned to visit the museum on Saturday.', 'On which day do I now plan to visit the museum? Give only the day.', f'I changed my plans. I will visit the museum on {days[i]} instead.'))
+
+        available = i % 2 == 0
+        suffix = ' Use only these facts. If the answer is missing, reply "Not provided." Otherwise give a short answer.'
+        info = [
+            ('missing_time', f'{name} booked a train ticket.' + (f' Departure is at {8+i}:30.' if available else ''), 'What time does the train leave?', [f'{8+i}:30']),
+            ('missing_quantity', f'{name} bought apples.' + (f' The bag contains {8+i} apples.' if available else ''), 'How many apples did they buy?', [str(8+i)]),
+            ('missing_identity', f'A {items[i]} was found in the park.' + (f' It belongs to {name}.' if available else ''), 'Who owns it?', [name]),
+            ('missing_location', f'{name} mailed a parcel.' + (f' The destination is {cities[i]}.' if available else ''), 'To which city was the parcel sent?', [cities[i]]),
+            ('missing_attribute', f'{name} has a bicycle.' + (f' The bicycle is {colors[i]}.' if available else ''), 'What color is the bicycle?', [colors[i]]),
+        ]
+        for family, passage, question, answer in info:
+            add('answerability', family, passage + '\n' + question + suffix,
+                answer if available else ['Not provided.', 'Not stated.', 'Not specified.', 'Unknown.', 'I don’t know.', 'I do not know.', 'Not enough information.', 'The information is not provided.'],
+                answerable=available)
+
+    qualitative = [
+        ('greeting', 'Hi! How are you?', 'A brief friendly greeting; no invented personal history.'),
+        ('thanks', 'Thanks, that helped.', 'A brief relevant acknowledgement.'),
+        ('explanation', 'Why does an ice cube melt in a warm room? Explain simply in one or two sentences.', 'Heat from the surroundings makes solid ice turn into liquid water.'),
+        ('explanation', 'Why is a bicycle easier to pedal in a lower gear when going uphill?', 'Lower gearing reduces the pedal force needed at the cost of more pedal turns or lower speed.'),
+        ('practical', 'Give me two simple ways to keep my desk organized.', 'Two distinct useful organization suggestions, without irrelevant filler.'),
+        ('practical', 'I keep forgetting to bring my reusable bag to the store. Suggest one practical habit.', 'One relevant concrete reminder or preparation habit.'),
+        ('writing', 'Make this request more polite without changing its meaning: Send me the notes today.', 'A polite request for the notes today; preserve the deadline.'),
+        ('comparison', 'What is the difference between borrowing and buying a book?', 'Borrowing is temporary access with return; buying confers ownership.'),
+        ('explanation', 'What does the word "fragile" mean? Give a simple example.', 'Easily broken or damaged, with a suitable example.'),
+        ('planning', 'I have ten minutes to tidy my room. What should I do first?', 'A feasible brief starting action, not an elaborate or unrelated plan.'),
+        ('uncertainty', 'What color is the mug on my desk right now?', 'Acknowledge lack of visual/contextual information; do not invent a color.'),
+        ('followup', None, 'Use the previously stated no-spicy-food preference and rice/vegetables ingredients.'),
+    ]
+    for family, prompt, rubric in qualitative:
+        messages = None
+        if prompt is None:
+            messages = [dict(role='user',content='I do not like spicy food. I have rice and vegetables.'),
+                        dict(role='assistant',content='Understood.'),
+                        dict(role='user',content='Suggest a simple dinner using those ingredients.')]
+        add('open_chat', family, prompt, [], messages=messages, qualitative=True)
+        rows[-1]['review_rubric'] = rubric
+
+    assert len(rows) == 192
+    assert len({r['id'] for r in rows}) == len(rows)
+    assert len({str(r['messages']) for r in rows}) == len(rows)
+    return rows

evaluation/chat_benchmark/scoring.py

@@ -0,0 +1,57 @@
+"""Conservative answer matching, separate from response control measures."""
+from collections import Counter
+import re
+
+
+def normalize(text, case_sensitive=False):
+    text = text.replace('’', "'").strip().strip('"').strip("'")
+    text = re.sub(r'\s+', ' ', text).rstrip('.!?')
+    text = re.sub(r'\s*,\s*', ', ', text)
+    return text if case_sensitive else text.casefold()
+
+
+def looping(text):
+    words = re.findall(r'\w+', text.casefold())
+    if len(words) >= 24:
+        grams = [tuple(words[i:i+3]) for i in range(len(words)-2)]
+        if len(set(grams)) / len(grams) < .68:
+            return True
+    for chunks in (text.splitlines(), re.split(r'[.!?]+', text)):
+        normalized = [normalize(s) for s in chunks if len(s.split()) >= 4]
+        if normalized and max(Counter(normalized).values()) >= 3:
+            return True
+    return False
+
+
+def score(case, response):
+    text = response['text']
+    normal = response['stop_reason'] in ('turn_end', 'document_end')
+    repeat = looping(text)
+    empty = not bool(text.strip())
+    exact = None
+    wrapped = None
+    if not case['qualitative']:
+        aliases = {normalize(a, case['case_sensitive']) for a in case['answers']}
+        exact = normalize(text, case['case_sensitive']) in aliases
+        # Only strip these neutral wrappers; substring matches are never correctness.
+        unwrapped = re.sub(r'^(?:answer\s*:\s*|the answer is\s+|the result is\s+|it is\s+|it\x27s\s+)', '', text, flags=re.I)
+        wrapped = normalize(unwrapped, case['case_sensitive']) in aliases
+    return dict(answer_match=exact, answer_match_with_neutral_wrapper=wrapped,
+                natural_stop=normal, repeated=repeat, empty=empty,
+                token_limit=response['stop_reason'] == 'token_limit',
+                unexpected_structure=response['stop_reason'] == 'unexpected_structure',
+                clean_response=normal and not repeat and not empty,
+                strict_task_success=None if exact is None else exact and normal and not repeat and not empty)
+
+
+def summarize(rows):
+    objective = [r for r in rows if not r['case']['qualitative']]
+    def metrics(selected):
+        n = len(selected)
+        return dict(count=n, **{k:sum(bool(r['scores'][k]) for r in selected) for k in
+                    ('answer_match','answer_match_with_neutral_wrapper','natural_stop','repeated','empty','token_limit','unexpected_structure','clean_response','strict_task_success')})
+    categories = sorted({r['case']['category'] for r in objective})
+    return dict(objective=metrics(objective), categories={c:metrics([r for r in objective if r['case']['category']==c]) for c in categories},
+                response_control_all=metrics(rows),
+                answerability={str(a):metrics([r for r in objective if r['case']['category']=='answerability' and r['case']['answerable']==a]) for a in (True,False)},
+                caveat='Answer match is conservative automatic scoring, not a complete semantic correctness judgment. Non-matching answers and open chat require review. Repetition detection is heuristic.')

evaluation/vendor/ifeval/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.