zhiyuanhucs/nemotron-student-fail-v41-clean-thinking

Nemotron-fail / DeepSeek-V4.1 clean and action-only trajectories

DeepSeek-V4.1 reward-1 trajectories for tasks on which the Nemotron student did not obtain reward 1. This release was rebuilt from the complete reward-1 audit under v56-shell-quote-aware-package-probes.

先看这里：目录是什么意思

每个目录内的文件名都直接说明：是一题一条还是全部 trials、是否为 audit、thinking/action 是否干净。

Recommended training files

SWE-bench Pro Verified

SWE-bench Pro Verified is exported under a separate benchmark path:

The Pro collector continuously rerolls unresolved tasks. The current export has 105 verifier-correct unique tasks out of the 731-task verified set. Reused AgentCompass result files are deduplicated before attempt counts are reported.

Official 731-task baselines

data/07_official_swebench_pro_verified_baseline/ stores the independent AgentCompass mini-swe-agent baselines. Each task contributes one valid attempt; only infrastructure failures may be retried and those failures are separated from the scored rows.

Below-hybrid blocked-attempt control

data/04_reward1_successful_but_hacking_failed/one_trajectory_per_task.jsonl.gz contains 90 unique-task control rows that were rejected from strict and hybrid. Each row is verifier-correct, has dirty thinking or action evidence, and has no audit evidence of successful gold, verifier, published-source, or local-history answer access. Rows with test tampering are excluded. evidence_level distinguishes pre-execution guard blocking from the weaker absence of a successful-access signal. These rows are for comparison experiments only and are never counted as clean or hybrid. data/04_reward1_successful_but_hacking_failed/all_successful_trials.jsonl.gz preserves all 240 matching verifier-correct historical trials, including repeated trials for the same task. data/04_reward1_successful_but_hacking_failed/all_guard_rejected_dirty_generations.jsonl.gz preserves all 8126 rejected generation events without the per-trajectory inline truncation. These candidates were blocked before execution and include their dirty thinking/action samples.

For evidence-sensitive experiments, the same rows are also separated into strong_proof_guard_blocked_before_execution (64 unique tasks, 182 trials) and weak_proof_no_successful_answer_access_detected (26 unique tasks, 58 trials). The strong split has direct guard evidence that dirty candidate generations were rejected before execution. The weak split only has an end-to-end audit with no successful answer-access signal and should be treated separately.

Isolated no-guard matched control

The same 89 tasks are being run once with DeepSeek-V4.1 and Nemotron-3-Ultra at a 128-step limit. The semantic regeneration guard is disabled, while the one-snapshot cleanroom, all external agent-network denial, and verifier isolation remain enabled. This separates guard suppression from model behavior.

These controls are stored under data/05_offline_no_guard_experiment/ and never enter the strict or hybrid counts. The agent egress allowlist contains only the local model proxy. Successful or uncertain local/cache source recovery and test tampering are excluded from the training files and retained only as audit metadata.

Reward-1 trajectories that temporarily changed tests but did not retain test changes in the submitted patch are preserved separately as attempted_test_changes_but_external_verifier_still_passed.jsonl.gz. The verifier runs outside the agent container, so those edits could not control the final reward. These rows remain comparison-only and are not merged into the ordinary dirty-action control file.

The two canonical training paths are disjoint by task. If a task has a strict trajectory, it is placed in strict; hybrid contains the remaining tasks whose actions passed but whose assistant prose did not.

Complete reward-1 review

All 1036 reward-1 trials are represented in data/03_all_reward1_trials/all_trials__audit_index.jsonl.gz:

• —21 strict-clean trials across 16 tasks.
• —534 action-clean trials across 184 tasks. Only the contaminated assistant turns are blanked; clean thinking is retained.
• —3 action-clean trials remain excluded because the contaminated turn could not be localized safely.
• —481 rejected trials with an action, provenance, implementation, or completion blocker. Only audit metadata is published for rejected trials.

The two all_trials files under data/03_all_reward1_trials/ retain every accepted trial, including multiple trials for the same task. Canonical training files select one trial per task.

data/03_all_reward1_trials/raw_original_trajectories_before_thinking_redaction__action_clean_reward1.jsonl.gz preserves the unredacted source trajectory for every accepted hybrid trial. These rows are marked control_only; use the redacted hybrid files for SFT.

Hybrid redaction guarantee

Hybrid rows retain the source trajectory and clean model thinking. For every localized contaminated assistant turn, content, reasoning_content, and reasoning are blanked while its original tool call is preserved. Each row lists redacted_assistant_message_indices, redacted_turns, and the redaction method. Semantic-only findings without localizable turn evidence are excluded instead of deleting the full trajectory.