The Dataset Viewer has been disabled on this dataset.

capability-removal traces

Generations of DeepSeek-V4.1-Flash-1017 (vLLM v0.33.0, 8 x H100, TP 8; temperature 1.0, top-p 0.95, reasoning effort 100) and their scores, from the capability-removal project. HLE-Diamond is not included: its dataset card forbids redistribution. Its traces are in a separate private repository.

  • screen/<stage1|stage2|smoke>/stage*.jsonl: one row per sample: bench, id, sample, prompt_tokens, completion_tokens, finish (stop or length), text (reasoning, </think>, reply). Items are referenced by (bench, id); the questions themselves are not included, but the reasoning often restates them.
  • screen/*/scores_stage*.jsonl: one row per sample: score (0-1; FrontierCS score / 100), finish, pred, and judge: the LLM judge's replies for that sample (Gemini 4 Carbon, carbon-g; one per rubric item for HealthBench).
  • screen/*/items_stage*.json: per item: samples, mean score, full-credit count. survivors_stage1.json and boundary_stage2.json: the selected (bench, id) pairs.
  • screen/*/frontiercs*.json: FrontierCS 1.1 local judge scores (0-100) per <id>/<sample>.
  • pilot2/<job>/tp.jsonl: pilot-2 samples on TriviaQA, MMLU-Pro and MATH L4-5 (correct, pred, text).

Each benchmark's questions keep their own license.

Downloads last month
7