zml5418's picture
Update README.md
132dbe7 verified
|
Raw History Blame Contribute Delete
5.1 kB
metadata
license: apache-2.0
language:
  - en
pretty_name: Thinking Leakage Rollouts
tags:
  - reasoning
  - math
  - hybrid-reasoning
  - nothink
  - activation-steering
size_categories:
  - 100K<n<1M
configs:
  - config_name: base
    data_files:
      - split: test
        path: data/base/*.parquet
  - config_name: GRPO
    data_files:
      - split: test
        path: data/GRPO/*.parquet
  - config_name: OPSD
    data_files:
      - split: test
        path: data/OPSD/*.parquet
  - config_name: OPSD-LoRA
    data_files:
      - split: test
        path: data/OPSD-LoRA/*.parquet
  - config_name: SFT
    data_files:
      - split: test
        path: data/SFT/*.parquet
  - config_name: steer
    data_files:
      - split: test
        path: data/steer/*.parquet
  - config_name: counter_steer
    data_files:
      - split: test
        path: data/counter_steer/*.parquet

Thinking Leakage: rollouts

All model generations behind the paper Thinking Leakage: A Causal Audit of NoThink Post-Training in Hybrid Reasoning Models: 199,680 rollouts, 16 samples for each of 120 competition math problems, from Qwen3-8B, Qwen3-4B and MiniCPM4.1-8B before and after NoThink post-training (GRPO, OPSD, OPSD-LoRA, SFT), plus the activation-steering runs reported in the main text. The trained checkpoints are released as zml5418/{Method}-{Model} (for example zml5418/GRPO-Qwen3-8B, one branch per training step). Generation and evaluation details are given in the paper.

Configs

config contents files
base base models in Think and NoThink mode 6
GRPO GRPO checkpoint (steps 50 to 500) in NoThink mode, plus GRPO-Qwen3-8B steps 200 and 500 in Think mode 32
OPSD OPSD checkpoints, steps 25 to 100, NoThink 12
OPSD-LoRA OPSD-LoRA checkpoints, steps 25 to 100, NoThink 12
SFT SFT checkpoints, steps 25 to 100, NoThink 12
steer base model in NoThink mode with alpha * v_leak added at layer 20, five doses per model 15
counter_steer trained checkpoints in NoThink mode with their own training drift along v_leak removed at layer 20 (dose 1) 15

Every (config, model, method, step, thinking_mode, dose) combination holds 120 x 16 = 1,920 rollouts.

from datasets import load_dataset
ds = load_dataset("zml5418/thinking-leakage-rollouts", "GRPO", split="test")
ds = ds.filter(lambda r: r["model"] == "Qwen3-8B" and r["step"] == 500 and r["thinking_mode"] == "nothink")

Fields

field meaning
model Qwen3-8B, Qwen3-4B or MiniCPM4.1-8B
method base, GRPO, OPSD, OPSD-LoRA or SFT
step training step of the checkpoint (0 for the base model). For counter_steer, method + step name the checkpoint that was intervened on; its un-intervened rollouts are in that method's config
thinking_mode think or nothink (chat template enable_thinking true / false)
condition bare (no intervention), steer or counter_steer
dose steer: alpha, in units of the base model's Think-NoThink gap. counter_steer: gamma (always 1). Null for bare
qid our row index (0-119) into the combined 120-problem test set; internal, not part of any original release
source_dataset original release the problem comes from (see below)
source_id the problem's id in that release
problem problem statement as released
ground_truth reference answer
rollout_idx sample index, 0-15
response the generated text (prompt excluded)
pred content of the last complete \boxed{} in response (empty if there is none)
is_correct pred equals ground_truth, or the two are equivalent under Math-Verify (both parsed as LaTeX)

Problems

The test set concatenates four releases, 30 problems each. To match a row to its original problem, use source_dataset + source_id:

qid source_dataset source_id
0-29 math-ai/aime24 id (60-89)
30-59 math-ai/aime25 id (0-29)
60-89 MathArena/hmmt_feb_2025 problem_idx (1-30)
90-119 MathArena/hmmt_nov_2025 problem_idx (1-30)

The user message sent to every model was problem followed by "\n\nPlease reason step by step, and put your final answer within \\boxed{}.", rendered with each model's own chat template.

Citation

If you find this dataset useful, please cite:

@misc{liu2026thinkingleakagecausalaudit,
      title={Thinking Leakage: A Causal Audit of NoThink Post-Training in Hybrid Reasoning Models},
      author={Zehao Liu and Vasant G. Honavar},
      year={2026},
      eprint={2609.28682},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2609.28682},
}

The problems belong to their original releases; see those for their terms.