Download benchmarks/perma/README.md from PolarSnowLeopard/RPMem-data: direct link, hf CLI and curl.
- Browser
- Download file 3.37 kB
-
https://huggingface.co/datasets/PolarSnowLeopard/RPMem-data/resolve/main/benchmarks/perma/README.md
- Command line
-
hf download hf://datasets/PolarSnowLeopard/RPMem-data/benchmarks/perma/README.md
-
curl -L -o README.md https://huggingface.co/datasets/PolarSnowLeopard/RPMem-data/resolve/main/benchmarks/perma/README.md
RPMem perma Prepared Inputs
Local publication draft. Not uploaded or approved for distribution.
Prepared inputs for compiler-memory generation, downstream Gate training and held-out evaluation. Original splits, option order and data files are preserved; no predictions, latent caches, trained gates or baseline results are included.
Source: ustclsc/PERMA at 440e64e4fb8baec6f7ad10c1de135505f93e7cb1.
Payload: 868,527,429 bytes in 1,165 files.
Upstream attribution and redistribution terms still apply. The code license
does not grant a new license to these data. Public release approval is pending.
Counts
{
"clean_sd": {
"questions": 705,
"questions_by_user": {
"334": 75,
"354": 75,
"123": 75,
"1377": 75,
"507": 75,
"914": 60,
"112": 60,
"419": 75,
"108": 75,
"109": 60
}
},
"noisy_sd": {
"questions": 705,
"questions_by_user": {
"334": 75,
"354": 75,
"123": 75,
"1377": 75,
"507": 75,
"914": 60,
"112": 60,
"419": 75,
"108": 75,
"109": 60
}
},
"style_sd": {
"questions": 705,
"questions_by_user": {
"334": 75,
"354": 75,
"123": 75,
"1377": 75,
"507": 75,
"914": 60,
"112": 60,
"419": 75,
"108": 75,
"109": 60
}
},
"style_long_sd": {
"questions": 705,
"questions_by_user": {
"334": 75,
"354": 75,
"123": 75,
"1377": 75,
"507": 75,
"914": 60,
"112": 60,
"419": 75,
"108": 75,
"109": 60
}
},
"clean_md": {
"questions": 389,
"questions_by_user": {
"334": 42,
"354": 47,
"123": 41,
"1377": 28,
"507": 50,
"914": 39,
"112": 30,
"419": 46,
"108": 39,
"109": 27
}
},
"noisy_md": {
"questions": 389,
"questions_by_user": {
"334": 42,
"354": 47,
"123": 41,
"1377": 28,
"507": 50,
"914": 39,
"112": 30,
"419": 46,
"108": 39,
"109": 27
}
},
"style_md": {
"questions": 389,
"questions_by_user": {
"334": 42,
"354": 47,
"123": 41,
"1377": 28,
"507": 50,
"914": 39,
"112": 30,
"419": 46,
"108": 39,
"109": 27
}
}
}
Split policy
{
"protocol": "leave_one_user_out",
"held_out_users": [
334,
354,
123,
1377,
507,
914,
112,
419,
108,
109
],
"training_users": "all other users in the same variant"
}
For PERMA, point PERMA_DATA_ROOT at this directory. For PersonaMem-v2 and
PrefEval, pass this directory as --data-root. Keep the directory structure
intact: compilation reads historical conversations, not only question tables.
Use the code repository's docs/benchmarks.md for the training/evaluation steps.
PrefEval materializes each form row at 10/70/300 turns; form-row counts are not
the final question counts. PERMA's folds share input files but not trained gates.
release_manifest.json lists exact files and counts. Historical freeze files
retain their original names and identifiers for loader compatibility. Public
raw-source preparation code remains available; raw upstream snapshots are not
necessarily bundled in every benchmark package.