CoolFace
Datasetpublic

JingweiNi/iclr-2action-controller-chains

ICLR 2-action controller — rollout chains Chain-of-thought rollouts with counterfactual stopping probes, for a single-chain {continue, terminate} controller on Qwen3-1.7B (native thinking). Each JSONL line is one chain: field meaning qkey, rep, gold question id, chain index, gold answer outcome natural | length (budget cut) answer, correct the chain's own answer and whether it is right blocks, thinking_tokens length of the reasoning completion_ids the full… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/iclr-2action-controller-chains.

sourceHugging Faceotherupdated 23d agoView on Hugging Face
0likes123downloads
Dataset Card

ICLR 2-action controller — rollout chains

Chain-of-thought rollouts with counterfactual stopping probes, for a single-chain {continue, terminate} controller on Qwen3-1.7B (native thinking).

Each JSONL line is one chain:

fieldmeaning
qkey, rep, goldquestion id, chain index, gold answer
outcomenatural \length (budget cut)
answer, correctthe chain's own answer and whether it is right
blocks, thinking_tokenslength of the reasoning
completion_idsthe full token ids — the CoT itself
block_ends, query_pointsblock boundaries and the probed prefixes
probes[]{block, prefix_len, forced, forced_tokens} — the answer the model gives if forced to stop there

Probes are placed at every 4th block. Boxed pools force </think>\n\n\boxed{ (≤ 32 tokens); code pools force a ``` `python ``` program (≤ 768 tokens) graded by unit tests. That makes the value of stopping at any point observable, which is what the controller is trained and evaluated on.

Contents

filequestions × chainsdomain
chains/math_union_fit.jsonl4 241 × 10hard math (DAPO, Numina-hard, DeepScaleR) — training
chains/math_union_select.jsonl1 065 × 10same — model selection
chains/math_t1_report.jsonl, chains/math_t2_report.jsonl841 / 853 × 10sealed math test splits
chains/math_ood_report.jsonl951 × 8math OOD (MATH-500, AIME 24/25, OlympiadBench, GaoKao)
chains/code_pool_{fit,select,report}.jsonl891 / 223 / 386 × 8–16SYNTHETIC-2-RL code pool
chains/code_ood_pooled.jsonl958 × 8HumanEval+, MBPP+, LiveCodeBench ≥ 2024-08
chains/gpqa_diamond_report.jsonl198 × 8GPQA-Diamond as numbered-option MCQ
chains/mixed_{fit,select}.jsonlconcatenations of the math and code splits
manifests/, fire_tables/question manifests and cross-fitted π₀ tables
ladders/ladders_json.tar.gzfull decision-rule sweeps behind every reported table
ASSET_INVENTORY.jsonpath → bytes → lines → sha256 for every file here

Companion checkpoints: `JingweiNi/iclr-2action-controller-heads`.

Provenance and reproduction

Code, frozen question manifests, all 70 result tables and the full reproducibility record live on the branch archive/iclr-2action-controller-2026-08-30 of the project repo; see docs/reproducibility/iclr-2action-controller-assets.md there. Generation is seeded by (qkey, rep, seed), so any shard regenerates byte-identically from (manifest, commit, seed).

GPQA-Diamond questions are derived from the gated Idavidrein/gpqa (revision 633f5ee8…); this repo redistributes only the model's own rollouts over them, not the source CSV.