prismane16/monitor-accuracy-under-rl

Cached artifacts for "What Happens to a Monitor's Accuracy When You Train Against It"

Abraham Yeung and Anagha Ramaswamy, Stanford University. Interpretability as a Science (InterpScience) Workshop, NeurIPS 2026. Code and paper: https://github.com/Abraham-y/reader-writer-probe-rl

These are the 56 files (0.32 GB) that the paper's regeneration script reads. With them, every number, table and figure in the paper regenerates on CPU, and the script exits non-zero if any published value disagrees with its recomputation. The list is not hand-picked: it was built by logging every file the script opened, and it was then checked by running the full script in an empty checkout that held only the repository and these files.

Use

The dataset mirrors the code repository's paths, so the files land where the scripts look for them.

git clone https://github.com/Abraham-y/reader-writer-probe-rl
cd reader-writer-probe-rl
pip install huggingface_hub   # plus the repository's own requirements
python scripts/artifacts.py fetch
bash scripts/check_everything.sh

fetch downloads exactly the paths listed in artifacts/MANIFEST.tsv and then checks each file's size and SHA-256 against it. It deliberately does not download this card, which would otherwise overwrite the code repository's README. To check a copy you already have: python scripts/artifacts.py verify.

Contents

All activations are the layer-16 residual stream (hidden_states[16]) of a Qwen2.5-0.5B policy at the first token of </think>, stored as float32 arrays of shape (answers, 896). "First block" and "last block" grade the first and the last <answer> block an answer contains, with the exact Countdown verifier.

path	what it is	where the paper uses it
followup/acts/phase0_harvest_runA/{step}/16.npy, labels.parquet	activations and labels for 8 fresh answers per prompt, sampled at temperature 1 on the 406 evaluation prompts, at each of the 11 checkpoints of runA (the probe-as-reward run from C_outcome). Labels hold prompt and answer indices, both block labels, answer length and template features	§4.1, Figure 1, Tables 3 and 5
followup/results/fragility/judge_lag/step_{step}.jsonl	an independent draw of 8 answers per prompt at each runA checkpoint: the proposed equation, the verifier's first-block label, and Qwen2.5-7B-Instruct's renormalised P(YES)	§4.2, Table 4
followup/acts/vanilla_rloo_ladder/{step}/16.npy, labels.parquet	the same kind of activations for five checkpoints of ordinary verifier-reward RLOO: the weak control ladder the paper describes and does not rely on	Appendix B
extension/cache/probe_cache_n500_clean406/*	activations (.npz) and first-block labels (.meta.json) for C_outcome's and C_SFT's headline-protocol answers (16 per prompt, temperature 0.6)	§3, the AUROC column of Table 2
extension/cache/steering/probe_pipeline_C_outcome_l16_pre_answer_temp1.pkl	the reward probe: a scikit-learn StandardScaler + LogisticRegression pipeline	the RL reward; scored in §4.1
extension/cache/probe_cache_temp1/C_outcome_temp1_l16_pre_answer.meta.json	the prompts and answers the reward probe was fit on	the seen/unseen split of Appendix F
eval_*.json	sampled answers with verifier scores: runA and runB after training and C_outcome, C_SFT (headline protocol); arms A and B, and runB re-scored as arm C (arm protocol)	§4, §4.3, Table 2
artifacts/MANIFEST.tsv	every file above with its size and SHA-256	scripts/artifacts.py verify

The two probe files are Python pickles (the .pkl, and .npz files loaded with allow_pickle=True). Loading a pickle can execute code, so load these only from this dataset or another source you trust.

Provenance and licenses

• These artifacts are released under CC BY 4.0, the paper's license.
• The policies are fine-tunes of Qwen/Qwen2.5-0.5B (Apache-2.0), and the judge is Qwen/Qwen2.5-7B-Instruct (Apache-2.0).
• C_SFT is asingh15/qwen-sft-countdown-defaultproj, a public checkpoint we did not train; its model card declares no license. This dataset contains its sampled answers and activations, not its weights or training data.
• Countdown problems are procedurally generated, following Gandhi et al. (2024), "Stream of Search".

What is not here

• Model checkpoints. The RL checkpoints are not public; everything the paper reports is computed from the activations and answers above.
• Anything behind the withdrawn steering result, whose checkpoints no longer exist.
• Training variance: every RL configuration in the paper is a single seed.

Citation

@inproceedings{yeung2026monitor,
  title     = {What Happens to a Monitor's Accuracy When You Train Against It},
  author    = {Yeung, Abraham and Ramaswamy, Anagha},
  booktitle = {Interpretability as a Science (InterpScience) Workshop at NeurIPS},
  year      = {2026}
}