Dataset Viewer

The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.

Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('text', {}), NamedSplit('test'): ('json', {})}

Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

JudgeJack Pilot Data -- Repo Map

NOTE: Files are not yet physically reorganized into folders due to time constraints at the end of the compute window. This README maps the EXISTING paths to their INPUT / OUTPUT / NOTEBOOK role. Physical reorganization into inputs/ outputs/ notebooks/ folders is a follow-up task, not yet done.

INPUTS (cleaned data others need before running anything)

prm800k/prm800k_train.json, prm800k/prm800k_holdout.json
math_shepherd/math_shepherd_train.json, math_shepherd/math_shepherd_holdout.json
math_shepherd/math_shepherd_sampled_pool.json, math_shepherd/math_shepherd_parsed_full.json

OUTPUTS (results produced by a run -- includes both checkpoints and eval results)

Checkpoints (in benjaminrtoney/judgejack-judge-checkpoints, separate repo):

prm800k/clean_judge_full/checkpoint-2500, prm800k/poisoned_judge_full/checkpoint-2500 -- Qwen, ORIGINAL RUN, USE THESE (not _incremental or _mid_v* variants)
prm800k/deepseek_clean_judge_full/checkpoint-10000 -- DeepSeek, USE THIS (not checkpoint-35082, which underperformed per probe history)
prm800k/deepseek_poisoned_judge_full/checkpoint-7500 -- DeepSeek, USE THIS (not checkpoint-35082)
See prm800k/deepseek_judges_README.md in the checkpoint repo for the full explanation.

Eval results and analysis outputs (this repo):

prm800k/deepseek_eval_results_sample.json -- DeepSeek x PRM800K real result (clean +0.0657, poisoned +0.5763)
prm800k/ablation_n50_trajectories_incremental_FINAL_SNAPSHOT.jsonl -- Qwen token-cap ablation, 191/300 trajectories (run interrupted by server crash; 10K and 32K budgets have real data, 64K has none)
math_shepherd/deepseek_stress_test_WEAK_INCONCLUSIVE_signal.json + math_shepherd/deepseek_stress_test_README_WEAK_RESULT.md -- Math-Shepherd result is WEAK/INCONCLUSIVE, read the README before using this

NOTEBOOKS

notebooks/*.ipynb -- all notebooks used tonight, pushed with full cell outputs/logs intact

TODO (follow-up, not done tonight)

Physically reorganize into inputs/, outputs/, notebooks/ folders
Resume the ablation to complete 64K budget (see chat log / notebooks for exact resume plan: missing problem_ids per budget/judge)

Downloads last month: 190

Total file size:

4.89 GB