name stringclasses 2
values | calibration dict | ppl dict | status stringclasses 1
value | env dict | eval_dtype stringclasses 1
value | finished timestamp[s]date 2026-11-07 14:02:54 2026-11-07 15:30:10 | spec stringclasses 1
value |
|---|---|---|---|---|---|---|---|
s16_q128_d1e-3 | {
"corpus": "kit/data/calibration_mixed_en_ja_dedup.txt",
"samples": 16,
"max_tokens": 128,
"output_block": "small",
"damping": 0.001,
"calib_dtype": null,
"raw_output": false,
"ja_ratio": null,
"overlap_checked": true
} | {
"en": {
"chunks": 100,
"tokens": 12700,
"bf16_ppl": 35.380284412245416,
"candidate_ppl": 41.79434488873292,
"delta_nll_per_token": 0.16660631314976015,
"config_hash": "1d36ba24ec58bc2d"
},
"ja": {
"chunks": 100,
"tokens": 12700,
"bf16_ppl": 43.41654459182924,
"candidate_p... | complete | {
"JAX_PLATFORM_NAME": "cpu",
"jax_version": "0.11.1",
"lalamo_version": "0.17.0",
"lalamo_source": "https://github.com/trymirai/lalamo",
"lalamo_commit": "e401e17761f2fcd337b0cf20218d67d656e6ecba"
} | float32 | 2026-11-07T14:02:54 | /kaggle/working/kit/lanes/robustness-08b-shard2of5.json |
s32_q128_d1e-2 | {
"corpus": "kit/data/calibration_mixed_en_ja_dedup.txt",
"samples": 32,
"max_tokens": 128,
"output_block": "small",
"damping": 0.01,
"calib_dtype": null,
"raw_output": false,
"ja_ratio": null,
"overlap_checked": true
} | {
"en": {
"chunks": 100,
"tokens": 12700,
"bf16_ppl": 35.380284412245416,
"candidate_ppl": 39.084481911707016,
"delta_nll_per_token": 0.0995707786740283,
"config_hash": "1d36ba24ec58bc2d"
},
"ja": {
"chunks": 100,
"tokens": 12700,
"bf16_ppl": 43.41654459182924,
"candidate_p... | complete | {
"JAX_PLATFORM_NAME": "cpu",
"jax_version": "0.11.1",
"lalamo_version": "0.17.0",
"lalamo_source": "https://github.com/trymirai/lalamo",
"lalamo_commit": "e401e17761f2fcd337b0cf20218d67d656e6ecba"
} | float32 | 2026-11-07T15:30:10 | /kaggle/working/kit/lanes/robustness-08b-shard2of5.json |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
uzu results
Kaggle 実行担当が run_lane.py の出力を上げる場所。
レイアウト
<lane>/
manifest.json # 条件一覧・問題・モデル・eval 設定
<condition>/
result.json # calibration 要約 + ppl(en/ja) + config_hash
ppl_en.json
ppl_ja.json
calibration.json
lane.log
アップロード
実行担当のアカウントで上げる。owner 側のトークンは不要。
python scripts/run_lane.py --spec lanes/clean-suite-08b.json \
--hf-results <自分のユーザー名>/uzu-results
取得（owner 側）
hf download <ユーザー名>/uzu-results --repo-type dataset \
--local-dir results/<ユーザー名>
ルール
manifest.jsonにproblemsが入っていたら未完了。上げ直すか owner に報告calibration.npzは上げない（1.4GB）。result.jsonとppl_*.jsonで十分- 条件を追加したら lane 名を変える。同じ lane に別条件を混ぜない
- 旧 eval / 旧 calibration の結果は上げない
- Downloads last month
- 70