Skip to page content
SAVRN
Search Plan a project

Dataset · Text generation

nemotron-student-fail-v41-clean-thinking

by Zhiyuan Hu zhiyuanhucs/nemotron-student-fail-v41-clean-thinking

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+v57-target-source-recheck.

Rows378
Configurations2
Size1.1 GB
Licenseother
AccessPublicly accessible
Monthly Downloads10.3k

Dataset Card

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+v57-target-source-recheck. 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。 目录是按轨迹分类,同一道题在不同 trial 中可能属于不同目录,不能把各目录的任务数直接相加。 其中 raw thinking 与 action 均干净的唯一任务,连同 05 中已复核的干净轨迹去重后为 语言分布为 {'go': 4, 'js': 6, 'python': 8, 'ts': 1}。 05 的干净轨迹保持单独文件,便于识别其无在线 guard 的采集设置。 05 的 V4.1 full-pool「reward=1 且未观察到成功获取答案」有 其中可以直接用于 clean/hybrid 训练的仍只有 02、04 按轨迹质量分组:02 的 action 审计通过,删除可定位的污染 thinking turn 后可作 hybrid 训练; 04 的轨迹低于 hybrid 门槛,仅作对照。05 按采集设置分组,含 reward=1、reward=0 和未评分试跑; 同一道题可以在不同…

Excerpt from the card by Zhiyuan Hu, licensed other.

Structure

clean_thinking_and_actions 16 rows

SplitRowsSize
train166.3 MB
taskstringtask_namestringlanguagestringrepostringmodelstringstudent_modelstringstudent_failedboolteacher_rewardfloat64audit_policy_versionstringrunstringtrialstringfinished_atstringinstructionstringaction_trajectorylistfinal_patchstringfinal_patch_capturestringauditvalueintegrity_guard_sentence_removedboolsource_links_sanitizedbooltask_variantstringsource_sanitization_policynullstrict_cleanboolhybridbooltrajectoryvalueraw_thinking_cleanbooloriginal_thinking_removedbooltraining_viewstring

reward1_successful_but_hacking_failed 362 rows

SplitRowsSize
one_per_task101213.2 MB
all_trials261339.2 MB
taskstringtask_namestringlanguagestringrepostringmodelstringstudent_modelstringstudent_failedboolteacher_rewardfloat64runstringtrialstringsource_pathstringinstructionstringtrajectoryvalueaction_trajectorylistfinal_patchstringaudit_reasonslistcontrol_onlyboolclean_or_hybrid_eligibleboolcontrol_categorystringevidence_levelstringcandidate_generation_rejected_before_executionboolexternal_answer_observedboolexternal_answer_observation_basisstringblocked_event_countint64blocked_events_truncatedboolblocked_eventslisttraining_viewstring

Details

Repository
zhiyuanhucs/nemotron-student-fail-v41-clean-thinking
Publisher
Zhiyuan Hu
Task category
Text generation
Tags
Not stated by the source
Size category
Not stated by the source
Languages
en
Revision
7963b135d444f92b762ec610391fbd487e61f9f6
Last updated
2026-09-30

Files

91 files, 1.1 GB in total.

Data89 files · 1.1 GB
Documentation1 file · 17.5 KB
Repository1 file · 2.5 KB
Every file
FileTypeSizeSHA-256
blocked_attempt_control_manifest.jsonlData33.4 KB—
data/01_clean_thinking_and_actions/all_successful_trials.jsonl.gzData1.1 MB0691fef02958
data/01_clean_thinking_and_actions/one_trajectory_per_task.jsonl.gzData894.0 KBff299e60f66c
data/02_clean_actions_dirty_thinking_removed/all_successful_trials.jsonl.gzData39.8 MB81fb9759153b
data/02_clean_actions_dirty_thinking_removed/one_trajectory_per_task.jsonl.gzData14.3 MB4183cc5fc661
data/03_all_reward1_trials/all_trials__audit_index.jsonl.gzData50.5 KB58c441511964
data/03_all_reward1_trials/clean_actions_dirty_thinking_removed__all_trials.jsonl.gzData39.8 MB81fb9759153b
data/03_all_reward1_trials/clean_thinking_and_actions__all_trials.jsonl.gzData1.1 MB0691fef02958
data/03_all_reward1_trials/raw_original_trajectories_before_thinking_redaction__action_clean_reward1.jsonl.gzData40.5 MBc0daf205b057
data/03_all_reward1_trials/rejected_trials__audit_only.jsonl.gzData35.8 KB87a2df45b56c
data/03_all_reward1_trials/target_published_source_lookup__quarantine_audit.jsonl.gzData44.9 KB8b4a80ab33ee
data/04_reward1_successful_but_hacking_failed/all_guard_rejected_dirty_generations.jsonl.gzData1.3 MB93e548f73d3a
data/04_reward1_successful_but_hacking_failed/all_successful_trials.jsonl.gzData32.2 MB3ce15a80b566
data/04_reward1_successful_but_hacking_failed/all_trials__audit_index.jsonl.gzData16.6 KB75d2caa3b4e0
data/04_reward1_successful_but_hacking_failed/one_trajectory_per_task.jsonl.gzData15.6 MB4546fd73b43d
data/04_reward1_successful_but_hacking_failed/strong_proof_guard_blocked_before_execution/all_successful_trials.jsonl.gzData14.2 MB3f25605fb72a
data/04_reward1_successful_but_hacking_failed/strong_proof_guard_blocked_before_execution/one_trajectory_per_task.jsonl.gzData5.5 MB7e6fdbe8386f
data/04_reward1_successful_but_hacking_failed/target_published_source_lookup__quarantine_audit.jsonl.gzData7.7 KB856972f20c53
data/04_reward1_successful_but_hacking_failed/weak_proof_no_successful_answer_access_detected/all_successful_trials.jsonl.gzData18.0 MB5789cb2be81e
data/04_reward1_successful_but_hacking_failed/weak_proof_no_successful_answer_access_detected/one_trajectory_per_task.jsonl.gzData10.1 MBea8e8672478b
data/05_offline_no_guard_experiment/deepseek_v4_1/all_reward1_without_successful_hack.jsonl.gzData53 B27fe8834bfd5
data/05_offline_no_guard_experiment/deepseek_v4_1/all_scored_tasks__audit.jsonl.gzData6.2 KB5efd4322a8e3
data/05_offline_no_guard_experiment/deepseek_v4_1/attempted_test_changes_but_external_verifier_still_passed.jsonl.gzData62 B80937dea99ce
data/05_offline_no_guard_experiment/deepseek_v4_1/clean_thinking_and_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1/dirty_thinking_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1/legacy_public_egress_untrusted_reward1.jsonl.gzData4.9 MBd75b0c86a005
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/all_attempts_including_unscored__audit_index.jsonl.gzData181.3 KB6ac5f9809209
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/all_reward1_without_successful_hack.jsonl.gzData19.1 MBe5153c4205cb
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/all_scored_tasks__audit.jsonl.gzData181.3 KB6ac5f9809209
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/attempted_test_changes_but_external_verifier_still_passed.jsonl.gzData8.2 MB3b34e24d7a98
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/clean_thinking_and_actions__reward1.jsonl.gzData323.8 KB993e424be722
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData11.9 MBb7d9dae125e2
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/dirty_thinking_only__reward1.jsonl.gzData6.9 MB649476d662dd
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/legacy_public_egress_untrusted_reward1.jsonl.gzData7.4 MB62bc9f1e3f0e
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/network_isolated_reward0_all_raw_trajectories__audit_only.jsonl.gzData549.3 MB2082b11ad651
data/05_offline_no_guard_experiment/deepseek_v4_1_full_pool/network_isolated_reward1_all_raw_trajectories__audit_only.jsonl.gzData37.1 MB6170551de2d6
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/all_reward1_without_successful_hack.jsonl.gzData53 B27fe8834bfd5
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/all_scored_tasks__audit.jsonl.gzData291 B3f5c77b44394
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/clean_thinking_and_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_low_reasoning_pilot/dirty_thinking_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/all_reward1_without_successful_hack.jsonl.gzData91.6 KBe644ddb413e6
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/all_scored_tasks__audit.jsonl.gzData535 B85c24d7f27d6
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/clean_thinking_and_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData91.6 KB7dc769126688
data/05_offline_no_guard_experiment/deepseek_v4_1_prompt_only_pilot/dirty_thinking_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/deepseek_v4_1_vs_nemotron_3_ultra/all_paired_tasks__audit.jsonl.gzData742 Bcf5c058d8d6c
data/05_offline_no_guard_experiment/deepseek_v4_1_vs_nemotron_3_ultra/v4_1_success_nemotron_failed.jsonl.gzData57 Ba47124ddadc5
data/05_offline_no_guard_experiment/nemotron_3_ultra/all_reward1_without_successful_hack.jsonl.gzData53 B27fe8834bfd5
data/05_offline_no_guard_experiment/nemotron_3_ultra/all_scored_tasks__audit.jsonl.gzData6.3 KB1707f00287ac
data/05_offline_no_guard_experiment/nemotron_3_ultra/attempted_test_changes_but_external_verifier_still_passed.jsonl.gzData62 B80937dea99ce
data/05_offline_no_guard_experiment/nemotron_3_ultra/clean_thinking_and_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra/dirty_thinking_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra/legacy_public_egress_untrusted_reward1.jsonl.gzData1.0 MBe5c9496478dd
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/all_reward1_without_successful_hack.jsonl.gzData53 B27fe8834bfd5
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/all_scored_tasks__audit.jsonl.gzData288 Bb80f0d3624c7
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/attempted_test_changes_but_external_verifier_still_passed.jsonl.gzData62 B80937dea99ce
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/clean_thinking_and_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/dirty_actions_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/dirty_thinking_and_dirty_actions__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/dirty_thinking_only__reward1.jsonl.gzData32 B9022aad581c6
data/05_offline_no_guard_experiment/nemotron_3_ultra_step256_retry/legacy_public_egress_untrusted_reward1.jsonl.gzData65 Ba064e420ef57
data/06_swebench_pro_verified/all_reward1_trials__audit_index.jsonl.gzData207.7 KBfb4a2bbb6a38
data/06_swebench_pro_verified/clean_actions_dirty_thinking_removed__all_trials.jsonl.gzData35.5 MB19ce29dea954
data/06_swebench_pro_verified/clean_actions_dirty_thinking_removed__one_per_task.jsonl.gzData31.8 MBc21f8dad168e
data/06_swebench_pro_verified/clean_thinking_and_actions__all_trials.jsonl.gzData40 Beec3ad3a60e1
data/06_swebench_pro_verified/clean_thinking_and_actions__one_per_task.jsonl.gzData32 B9022aad581c6
data/07_official_swebench_pro_verified_baseline/deepseek_v4_1/all_valid_once_per_task.jsonl.gzData83.3 MB018c4b34513d
data/07_official_swebench_pro_verified_baseline/deepseek_v4_1/infrastructure_failures_excluded.jsonl.gzData13.7 KB51b6bbe0e1ce
data/07_official_swebench_pro_verified_baseline/deepseek_v4_1/resolved_once_per_task.jsonl.gzData32.9 MB9952da8a1688
data/07_official_swebench_pro_verified_baseline/nemotron_3_ultra/all_valid_once_per_task.jsonl.gzData38.0 MBb93e7617dd33
data/07_official_swebench_pro_verified_baseline/nemotron_3_ultra/infrastructure_failures_excluded.jsonl.gzData29.1 KBa4a3702e8ddd
data/07_official_swebench_pro_verified_baseline/nemotron_3_ultra/resolved_once_per_task.jsonl.gzData14.1 MB4e6f6ce9a341
data/08_source_link_sanitized_recovery/clean_actions_dirty_thinking_removed__one_per_task.jsonl.gzData191.0 KB5b1217b8aab5
data/08_source_link_sanitized_recovery/clean_thinking_and_actions__one_per_task.jsonl.gzData67 B1c2d16c635ea
isolated_no_guard_manifest.jsonlData52.1 KB—
manifest.jsonlData97.6 KB—
official_swebench_pro_verified_baseline_manifest.jsonlData358.9 KB—
rejected-summary.jsonData583 B—
rejected.jsonlData59.9 KB—
summary.jsonData20.7 KB—
swebench_pro_verified_manifest.jsonlData75.2 KB—
README.mdDocumentation17.5 KB—
.gitattributesRepository2.5 KB—

License and Download

License
other
Access
No access gate
Download from Zhiyuan Hu

Released by Zhiyuan Hu through its official repository on Hugging Face.