---
license: other
license_name: mixed-per-source
task_categories:
- text-generation
- question-answering
language:
- en
size_categories:
- 10M<n<100M
pretty_name: Omni-Frontier Collection — Cybersecurity · Coding · Math · Science · RSI Reasoning SFT Corpus
tags:
- distillation
- knowledge-distillation
- sft
- supervised-fine-tuning
- reasoning
- chain-of-thought
- long-cot
- deepseek-r1
- rsi
- self-improvement
- cybersecurity
- security
- exploit-db
- threat-intelligence
- red-teaming
- vulnerability-analysis
- security-gym
- nosk
- power-grid-security
- mitre-attack
- sql-injection
- ctf
- code
- coding
- math
- science
- instruction-tuning
- conversational
- chatml
- synthetic
- gpt-5.6
- kimi-k3
- qwen3.8-max
- glm-5.2
- omni-frontier
configs:
- config_name: rsi__openmle_sft_traces
  data_files:
  - split: train
    path: Data/RSI/openmle__train-00000-of-00001.parquet
- config_name: rsi__shslab_rsi_selfgen
  data_files:
  - split: train
    path: Data/RSI/shslab_rsi_selfgen__train.parquet
- config_name: coding__omni_frontier
  data_files:
  - split: train
    path: Data/coding/full_clean__v2__coding.parquet
- config_name: coding__omni_frontier__v2
  data_files:
  - split: train
    path: Data/coding/full_clean__data__coding.parquet
- config_name: coding__openmle_sft_traces
  data_files:
  - split: train
    path: Data/coding/openmle__train-00000-of-00001.parquet
- config_name: cyber_security__cyber_god_mode
  data_files:
  - split: train
    path:
    - Data/cyber security/cybergod__train-00000-of-00015.parquet
    - Data/cyber security/cybergod__train-00001-of-00015.parquet
    - Data/cyber security/cybergod__train-00002-of-00015.parquet
    - Data/cyber security/cybergod__train-00003-of-00015.parquet
    - Data/cyber security/cybergod__train-00004-of-00015.parquet
    - Data/cyber security/cybergod__train-00005-of-00015.parquet
    - Data/cyber security/cybergod__train-00006-of-00015.parquet
    - Data/cyber security/cybergod__train-00007-of-00015.parquet
    - Data/cyber security/cybergod__train-00008-of-00015.parquet
    - Data/cyber security/cybergod__train-00009-of-00015.parquet
    - Data/cyber security/cybergod__train-00010-of-00015.parquet
    - Data/cyber security/cybergod__train-00011-of-00015.parquet
    - Data/cyber security/cybergod__train-00012-of-00015.parquet
    - Data/cyber security/cybergod__train-00013-of-00015.parquet
    - Data/cyber security/cybergod__train-00014-of-00015.parquet
- config_name: cyber_security__exploitdb_detoxio
  data_files:
  - split: train
    path: Data/cyber security/exploitdb-t1__train-00000-of-00001.parquet
- config_name: cyber_security__exploitdb_grantabejar
  data_files:
  - split: train
    path: Data/cyber security/exploitdb-g__train-00000-of-00001.parquet
- config_name: cyber_security__exploitdb_texts_viewer
  data_files:
  - split: train
    path: Data/cyber security/exploitdb_texts__viewer.parquet
- config_name: cyber_security__exploitdb_waiper__v3
  data_files:
  - split: train
    path: Data/cyber security/exploitdb-w__exploitdb_dataset.parquet
- config_name: cyber_security__misc
  data_files:
  - split: train
    path: Data/cyber security/cybersec_bench.parquet
- config_name: cyber_security__misc__v2
  data_files:
  - split: train
    path:
    - Data/cyber security/threat_intel_blended.parquet
    - Data/cyber security/threat_intel_eval.parquet
    - Data/cyber security/threat_intel_train.parquet
- config_name: cyber_security__misc__v3
  data_files:
  - split: train
    path: Data/cyber security/blackhat_kali.parquet
- config_name: cyber_security__misc__v4
  data_files:
  - split: train
    path:
    - Data/cyber security/voltvar_train.parquet
    - Data/cyber security/voltvar_validation.parquet
- config_name: cyber_security__misc__v5
  data_files:
  - split: train
    path: Data/cyber security/threat_intel_multiturn.parquet
- config_name: cyber_security__misc__v6
  data_files:
  - split: train
    path: Data/cyber security/cybersecurity_qa_jsonl.parquet
- config_name: cyber_security__omni_frontier__v10
  data_files:
  - split: train
    path:
    - Data/cyber security/cybersecurity_qa_0000_0000.parquet
    - Data/cyber security/cybersecurity_qa_0000_0001.parquet
    - Data/cyber security/cybersecurity_qa_0000_0002.parquet
    - Data/cyber security/cybersecurity_qa_0000_0003.parquet
    - Data/cyber security/data-0001.parquet
    - Data/cyber security/data-0002.parquet
    - Data/cyber security/data-0003.parquet
    - Data/cyber security/data-0004.parquet
    - Data/cyber security/data-0005.parquet
- config_name: cyber_security__omni_frontier__v7
  data_files:
  - split: train
    path: Data/cyber security/full_clean__v2__cyber-security.parquet
- config_name: cyber_security__omni_frontier__v8
  data_files:
  - split: train
    path: Data/cyber security/full_clean__data__cyber-security.parquet
- config_name: cyber_security__omni_frontier__v9
  data_files:
  - split: train
    path:
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0001.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0002.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0003.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0004.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0005.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0006.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0007.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0008.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0009.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0010.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0011.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0012.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0013.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0014.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0015.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0016.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0017.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0018.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0019.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0020.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0021.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0022.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0023.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0024.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0025.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0026.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0027.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0028.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0029.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0030.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0031.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0032.parquet
    - Data/cyber security/cyber_unified_train_cybersecurity_analysis_0033.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0001.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0002.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0003.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0004.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0005.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0006.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0007.parquet
    - Data/cyber security/cyber_unified_train_incident_response_0008.parquet
    - Data/cyber security/cyber_unified_train_threat_intel_0001.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0001.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0002.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0003.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0004.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0005.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0006.parquet
    - Data/cyber security/cyber_unified_train_vulnerability_mgmt_0007.parquet
    - Data/cyber security/val_unified.parquet
- config_name: distillation__gpt56_sol_luna_terra_traces
  data_files:
  - split: train
    path: Data/distillation/stl-traces__train.parquet
- config_name: distillation__kimik3_coding_debug_traces__v2
  data_files:
  - split: train
    path:
    - Data/distillation/k3traces__train-00053.parquet
    - Data/distillation/k3traces__train-00055.parquet
    - Data/distillation/k3traces__train-00081.parquet
    - Data/distillation/k3traces__train-00083.parquet
    - Data/distillation/k3traces__train-00087.parquet
    - Data/distillation/k3traces__train-00090.parquet
    - Data/distillation/k3traces__train-00092.parquet
    - Data/distillation/k3traces__train-00094.parquet
    - Data/distillation/k3traces__train-00095.parquet
    - Data/distillation/k3traces__train-00096.parquet
    - Data/distillation/k3traces__train-00098.parquet
    - Data/distillation/k3traces__train-00106.parquet
    - Data/distillation/k3traces__train-00109.parquet
    - Data/distillation/k3traces__train-00110.parquet
    - Data/distillation/k3traces__train-00114.parquet
    - Data/distillation/k3traces__train-00115.parquet
    - Data/distillation/k3traces__train-00117.parquet
    - Data/distillation/k3traces__train-00119.parquet
    - Data/distillation/k3traces__train-00122.parquet
    - Data/distillation/k3traces__train-00126.parquet
    - Data/distillation/k3traces__train-00132.parquet
    - Data/distillation/k3traces__train-00138.parquet
    - Data/distillation/k3traces__train-00145.parquet
    - Data/distillation/k3traces__train-00151.parquet
    - Data/distillation/k3traces__train-00160.parquet
    - Data/distillation/k3traces__train-00167.parquet
    - Data/distillation/k3traces__train-00170.parquet
    - Data/distillation/k3traces__train-00171.parquet
    - Data/distillation/k3traces__train-00175.parquet
    - Data/distillation/k3traces__train-00180.parquet
    - Data/distillation/k3traces__train-00181.parquet
    - Data/distillation/k3traces__train-00182.parquet
    - Data/distillation/k3traces__train-00185.parquet
    - Data/distillation/k3traces__train-00186.parquet
    - Data/distillation/k3traces__train-00187.parquet
    - Data/distillation/k3traces__train-00189.parquet
    - Data/distillation/k3traces__train-00191.parquet
    - Data/distillation/k3traces__train-00192.parquet
    - Data/distillation/k3traces__train-00193.parquet
    - Data/distillation/k3traces__train-00198.parquet
    - Data/distillation/k3traces__train-00200.parquet
    - Data/distillation/k3traces__train-00204.parquet
    - Data/distillation/k3traces__train-00206.parquet
    - Data/distillation/k3traces__train-00207.parquet
    - Data/distillation/k3traces__train-00208.parquet
    - Data/distillation/k3traces__train-00209.parquet
    - Data/distillation/k3traces__train-00210.parquet
    - Data/distillation/k3traces__train-00211.parquet
    - Data/distillation/k3traces__train-00212.parquet
    - Data/distillation/k3traces__train-00213.parquet
- config_name: distillation__omni_frontier__v156
  data_files:
  - split: train
    path: Data/distillation/full_clean__v2__distillation.parquet
- config_name: distillation__omni_frontier__v157
  data_files:
  - split: train
    path: Data/distillation/full_clean__data__distillation.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v10
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00001__e971ad95.parquet
    - Data/distillation/k3dist__train-00000-of-00006.parquet
    - Data/distillation/k3dist__train-00001-of-00006.parquet
    - Data/distillation/k3dist__train-00002-of-00006.parquet
    - Data/distillation/k3dist__train-00003-of-00006.parquet
    - Data/distillation/k3dist__train-00004-of-00006.parquet
    - Data/distillation/k3dist__train-00005-of-00006.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__4022cac9.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v11
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00001.parquet
    - Data/distillation/k3dist__test-00000-of-00001__0d3d4b1f.parquet
    - Data/distillation/k3dist__test-00000-of-00001__3b4dbc66.parquet
    - Data/distillation/k3dist__test-00000-of-00001__47567728.parquet
    - Data/distillation/k3dist__test-00000-of-00001__4ab60c60.parquet
    - Data/distillation/k3dist__test-00000-of-00001__5300a1a6.parquet
    - Data/distillation/k3dist__test-00000-of-00001__85cee7e6.parquet
    - Data/distillation/k3dist__test-00000-of-00001__91982d36.parquet
    - Data/distillation/k3dist__test-00000-of-00001__a46cef53.parquet
    - Data/distillation/k3dist__test-00000-of-00001__b90e0c78.parquet
    - Data/distillation/k3dist__test-00000-of-00001__c0c9844f.parquet
    - Data/distillation/k3dist__test-00000-of-00001__c48b2bfe.parquet
    - Data/distillation/k3dist__test-00000-of-00001__d8ee5e53.parquet
    - Data/distillation/k3dist__test-00000-of-00001__de50bacc.parquet
    - Data/distillation/k3dist__test-00000-of-00001__e143e2a0.parquet
    - Data/distillation/k3dist__test-00000-of-00001__ee5c8cb1.parquet
    - Data/distillation/k3dist__test-00000-of-00001__fdba9be0.parquet
    - Data/distillation/k3dist__train-00000-of-00001.parquet
    - Data/distillation/k3dist__train-00000-of-00001__1d8a4357.parquet
    - Data/distillation/k3dist__train-00000-of-00001__5b9303e2.parquet
    - Data/distillation/k3dist__train-00000-of-00001__721fd95f.parquet
    - Data/distillation/k3dist__train-00000-of-00001__77fb4d1b.parquet
    - Data/distillation/k3dist__train-00000-of-00001__8e42a2c6.parquet
    - Data/distillation/k3dist__train-00000-of-00001__b517b6e7.parquet
    - Data/distillation/k3dist__train-00000-of-00001__cc454e67.parquet
    - Data/distillation/k3dist__train-00000-of-00002.parquet
    - Data/distillation/k3dist__train-00000-of-00002__17d7bb41.parquet
    - Data/distillation/k3dist__train-00000-of-00002__f7193c04.parquet
    - Data/distillation/k3dist__train-00000-of-00004.parquet
    - Data/distillation/k3dist__train-00000-of-00005.parquet
    - Data/distillation/k3dist__train-00000-of-00005__14478d89.parquet
    - Data/distillation/k3dist__train-00000-of-00005__34d979a3.parquet
    - Data/distillation/k3dist__train-00000-of-00005__540d1843.parquet
    - Data/distillation/k3dist__train-00000-of-00005__f65669a9.parquet
    - Data/distillation/k3dist__train-00000-of-00006__275fea53.parquet
    - Data/distillation/k3dist__train-00000-of-00006__b37bc37d.parquet
    - Data/distillation/k3dist__train-00000-of-00006__dde40623.parquet
    - Data/distillation/k3dist__train-00000-of-00006__e3bad213.parquet
    - Data/distillation/k3dist__train-00001-of-00002.parquet
    - Data/distillation/k3dist__train-00001-of-00002__152d6b06.parquet
    - Data/distillation/k3dist__train-00001-of-00002__9c47c4bc.parquet
    - Data/distillation/k3dist__train-00001-of-00004.parquet
    - Data/distillation/k3dist__train-00001-of-00005.parquet
    - Data/distillation/k3dist__train-00001-of-00005__2e1e938b.parquet
    - Data/distillation/k3dist__train-00001-of-00005__ba6ca043.parquet
    - Data/distillation/k3dist__train-00001-of-00005__efa47d99.parquet
    - Data/distillation/k3dist__train-00001-of-00005__ff388ab7.parquet
    - Data/distillation/k3dist__train-00001-of-00006__5d534f52.parquet
    - Data/distillation/k3dist__train-00001-of-00006__7ef16925.parquet
    - Data/distillation/k3dist__train-00001-of-00006__bbf7d600.parquet
    - Data/distillation/k3dist__train-00001-of-00006__fcb7aca0.parquet
    - Data/distillation/k3dist__train-00002-of-00004.parquet
    - Data/distillation/k3dist__train-00002-of-00005.parquet
    - Data/distillation/k3dist__train-00002-of-00005__70cdb84b.parquet
    - Data/distillation/k3dist__train-00002-of-00005__74aba895.parquet
    - Data/distillation/k3dist__train-00002-of-00005__c7f460d2.parquet
    - Data/distillation/k3dist__train-00002-of-00005__efea8ea5.parquet
    - Data/distillation/k3dist__train-00002-of-00006__06517817.parquet
    - Data/distillation/k3dist__train-00002-of-00006__2f0ae26f.parquet
    - Data/distillation/k3dist__train-00002-of-00006__54191dbe.parquet
    - Data/distillation/k3dist__train-00002-of-00006__81aa6f1b.parquet
    - Data/distillation/k3dist__train-00003-of-00004.parquet
    - Data/distillation/k3dist__train-00003-of-00005.parquet
    - Data/distillation/k3dist__train-00003-of-00005__19d65f2c.parquet
    - Data/distillation/k3dist__train-00003-of-00005__82a804c6.parquet
    - Data/distillation/k3dist__train-00003-of-00005__9949017c.parquet
    - Data/distillation/k3dist__train-00003-of-00005__a3e3575b.parquet
    - Data/distillation/k3dist__train-00003-of-00006__2475ad17.parquet
    - Data/distillation/k3dist__train-00003-of-00006__832625c5.parquet
    - Data/distillation/k3dist__train-00003-of-00006__e50b151e.parquet
    - Data/distillation/k3dist__train-00003-of-00006__ff001afe.parquet
    - Data/distillation/k3dist__train-00004-of-00005.parquet
    - Data/distillation/k3dist__train-00004-of-00005__1b9431cc.parquet
    - Data/distillation/k3dist__train-00004-of-00005__5a6e84dc.parquet
    - Data/distillation/k3dist__train-00004-of-00005__9bd6a829.parquet
    - Data/distillation/k3dist__train-00004-of-00005__e9607ade.parquet
    - Data/distillation/k3dist__train-00004-of-00006__2a009aab.parquet
    - Data/distillation/k3dist__train-00004-of-00006__47141bff.parquet
    - Data/distillation/k3dist__train-00004-of-00006__806fd502.parquet
    - Data/distillation/k3dist__train-00004-of-00006__c36bee8a.parquet
    - Data/distillation/k3dist__train-00005-of-00006__49cf1561.parquet
    - Data/distillation/k3dist__train-00005-of-00006__56e745d1.parquet
    - Data/distillation/k3dist__train-00005-of-00006__9c6702ac.parquet
    - Data/distillation/k3dist__train-00005-of-00006__bf852d0c.parquet
    - Data/distillation/k3dist__validation-00000-of-00001.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__08d958c6.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__172c758d.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__2c51653b.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__3c37a761.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__46787e9c.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__6207cf53.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__74aea4c7.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__78b5379d.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__7d4e0e13.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__7fc84027.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__89fc1be8.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__d3792574.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__d38d3c2c.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__dcfb1ae7.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__ec1c5e5b.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__f42096d9.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v12
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00001__2863a780.parquet
    - Data/distillation/k3dist__train-00000-of-00001__b74bda7f.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__b7e7bda6.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v13
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__prompt_completion_text.parquet
    - Data/distillation/k3dist__sft_reasoning.parquet
    - Data/distillation/k3dist__sft_tools.parquet
    - Data/distillation/k3dist__smoke.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v5
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00002.parquet
    - Data/distillation/k3dist__test-00001-of-00002.parquet
    - Data/distillation/k3dist__train-00000-of-00019.parquet
    - Data/distillation/k3dist__train-00000-of-00021.parquet
    - Data/distillation/k3dist__train-00001-of-00019.parquet
    - Data/distillation/k3dist__train-00001-of-00021.parquet
    - Data/distillation/k3dist__train-00002-of-00019.parquet
    - Data/distillation/k3dist__train-00002-of-00021.parquet
    - Data/distillation/k3dist__train-00003-of-00019.parquet
    - Data/distillation/k3dist__train-00003-of-00021.parquet
    - Data/distillation/k3dist__train-00004-of-00019.parquet
    - Data/distillation/k3dist__train-00004-of-00021.parquet
    - Data/distillation/k3dist__train-00005-of-00019.parquet
    - Data/distillation/k3dist__train-00005-of-00021.parquet
    - Data/distillation/k3dist__train-00006-of-00019.parquet
    - Data/distillation/k3dist__train-00006-of-00021.parquet
    - Data/distillation/k3dist__train-00007-of-00019.parquet
    - Data/distillation/k3dist__train-00007-of-00021.parquet
    - Data/distillation/k3dist__train-00008-of-00019.parquet
    - Data/distillation/k3dist__train-00008-of-00021.parquet
    - Data/distillation/k3dist__train-00009-of-00019.parquet
    - Data/distillation/k3dist__train-00009-of-00021.parquet
    - Data/distillation/k3dist__train-00010-of-00019.parquet
    - Data/distillation/k3dist__train-00010-of-00021.parquet
    - Data/distillation/k3dist__train-00011-of-00019.parquet
    - Data/distillation/k3dist__train-00011-of-00021.parquet
    - Data/distillation/k3dist__train-00012-of-00019.parquet
    - Data/distillation/k3dist__train-00012-of-00021.parquet
    - Data/distillation/k3dist__train-00013-of-00019.parquet
    - Data/distillation/k3dist__train-00013-of-00021.parquet
    - Data/distillation/k3dist__train-00014-of-00019.parquet
    - Data/distillation/k3dist__train-00014-of-00021.parquet
    - Data/distillation/k3dist__train-00015-of-00019.parquet
    - Data/distillation/k3dist__train-00015-of-00021.parquet
    - Data/distillation/k3dist__train-00016-of-00019.parquet
    - Data/distillation/k3dist__train-00016-of-00021.parquet
    - Data/distillation/k3dist__train-00017-of-00019.parquet
    - Data/distillation/k3dist__train-00017-of-00021.parquet
    - Data/distillation/k3dist__train-00018-of-00019.parquet
    - Data/distillation/k3dist__train-00018-of-00021.parquet
    - Data/distillation/k3dist__train-00019-of-00021.parquet
    - Data/distillation/k3dist__train-00020-of-00021.parquet
    - Data/distillation/k3dist__validation-00000-of-00002.parquet
    - Data/distillation/k3dist__validation-00001-of-00002.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v6
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__glm47-le-16384.parquet
    - Data/distillation/k3dist__glm47-le-4096.parquet
    - Data/distillation/k3dist__glm47-le-8192.parquet
    - Data/distillation/k3dist__llama31_final-le-32768.parquet
    - Data/distillation/k3dist__llama31_final-le-4096.parquet
    - Data/distillation/k3dist__llama31_final-le-8192.parquet
    - Data/distillation/k3dist__llama31_final-long-outlier.parquet
    - Data/distillation/k3dist__qwen3-le-32768.parquet
    - Data/distillation/k3dist__qwen3-le-4096.parquet
    - Data/distillation/k3dist__qwen3-le-8192.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v7
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00001__aef6279c.parquet
    - Data/distillation/k3dist__train-00000-of-00005__f2c0cfec.parquet
    - Data/distillation/k3dist__train-00001-of-00005__10235724.parquet
    - Data/distillation/k3dist__train-00002-of-00005__d62ad9c9.parquet
    - Data/distillation/k3dist__train-00003-of-00005__bf06bd57.parquet
    - Data/distillation/k3dist__train-00004-of-00005__a88ee91c.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__c33338d9.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v8
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__test-00000-of-00001__7dd01e4d.parquet
    - Data/distillation/k3dist__test-00000-of-00001__b72a9a9c.parquet
    - Data/distillation/k3dist__train-00000-of-00001__43ddd9fc.parquet
    - Data/distillation/k3dist__train-00000-of-00005__2a3e7711.parquet
    - Data/distillation/k3dist__train-00000-of-00006__7970e2a3.parquet
    - Data/distillation/k3dist__train-00001-of-00005__7ea40ecd.parquet
    - Data/distillation/k3dist__train-00001-of-00006__6c9d0c25.parquet
    - Data/distillation/k3dist__train-00002-of-00005__25a62cae.parquet
    - Data/distillation/k3dist__train-00002-of-00006__6c88e07e.parquet
    - Data/distillation/k3dist__train-00003-of-00005__061f8df6.parquet
    - Data/distillation/k3dist__train-00003-of-00006__924cc29f.parquet
    - Data/distillation/k3dist__train-00004-of-00005__08a348f0.parquet
    - Data/distillation/k3dist__train-00004-of-00006__9f03ae5b.parquet
    - Data/distillation/k3dist__train-00005-of-00006__e7d22d54.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__6d5615d7.parquet
    - Data/distillation/k3dist__validation-00000-of-00001__bb0be95f.parquet
- config_name: distillation__qwen38_glm52_kimik3_distill__v9
  data_files:
  - split: train
    path:
    - Data/distillation/k3dist__train-00000-of-00005__8be01a0d.parquet
    - Data/distillation/k3dist__train-00001-of-00005__f3c52c60.parquet
    - Data/distillation/k3dist__train-00002-of-00005__c90038bf.parquet
    - Data/distillation/k3dist__train-00003-of-00005__3821c8a6.parquet
    - Data/distillation/k3dist__train-00004-of-00005__83a92678.parquet
- config_name: science__omni_frontier
  data_files:
  - split: train
    path: Data/science/train-00000-of-00001.parquet
- config_name: science__omni_frontier__v2
  data_files:
  - split: train
    path: Data/science/full_clean__v2__science.parquet
- config_name: science__omni_frontier__v3
  data_files:
  - split: train
    path: Data/science/full_clean__data__science.parquet
- config_name: applied__omni_frontier
  data_files:
  - split: train
    path: Data/applied/full_clean__v2__applied.parquet
- config_name: applied__omni_frontier__v2
  data_files:
  - split: train
    path: Data/applied/full_clean__data__applied.parquet
- config_name: humanities__omni_frontier
  data_files:
  - split: train
    path: Data/humanities/full_clean__v2__humanities.parquet
- config_name: humanities__omni_frontier__v2
  data_files:
  - split: train
    path: Data/humanities/full_clean__data__humanities.parquet
- config_name: instructions__omni_frontier
  data_files:
  - split: train
    path: Data/instructions/full_clean__v2__instructions.parquet
- config_name: instructions__omni_frontier__v2
  data_files:
  - split: train
    path: Data/instructions/full_clean__data__instructions.parquet
- config_name: index__omni_frontier
  data_files:
  - split: train
    path: Data/index/full_clean__v2__index.parquet
- config_name: index__omni_frontier__v2
  data_files:
  - split: train
    path: Data/index/full_clean__data__index.parquet
---

<div align="center">

<img src="assets/shslab_avatar.webp" width="120" alt="SHS-Lab"/>

# 🧬 Omni-Frontier Collection
### Cybersecurity · Coding · Math · Science · RSI Reasoning — one unified SFT package

<img src="assets/banner.jpg" width="100%" alt="Omni-Frontier banner"/><br><br>

![Parquet Rows](https://img.shields.io/badge/parquet_rows-19%2C096%2C153-3B82F6?style=for-the-badge)
![Size](https://img.shields.io/badge/data_size-14.56_GB-06B6D4?style=for-the-badge)
![Files](https://img.shields.io/badge/files-2%2C462-8B5CF6?style=for-the-badge)<br>
![Categories](https://img.shields.io/badge/taxonomy-12_categories-F59E0B?style=for-the-badge)
![Viewer](https://img.shields.io/badge/viewer_configs-44-10B981?style=for-the-badge)
![License](https://img.shields.io/badge/license-mixed_per_source-lightgrey?style=for-the-badge)

**A unified, deduplicated, fully-browsable distillation & SFT corpus — every row real, every row visible.**

</div>

---

## 📖 Jump to

[What's inside](#-whats-inside-the-real-numbers) · [🔁 Aggregation audit](#-full-aggregation-from-manusagentsomni-frontier--the-feature-manifest) · [🛡 Cybersecurity](#-cybersecurity-45m-rows) · [💻 Coding](#-coding-126m-rows) · [🏭 Distillation deep-dive](#-distillation-the-honest-deep-dive-16m-rows) · [🔁 RSI](#-rsi--recursive-self-improvement) · [🧮 Math/Science/More](#-math--science--humanities--applied--instructions) · [🎓 Training guide](#-training-guide-how-to-actually-use-this) · [🔎 Browsing](#-browsing-the-dataset-viewer) · [🧹 Quality](#-quality--dedup--qa) · [🗺 Roadmap](#-roadmap) · [📄 License](#-licensing--attribution)

---

## 📌 What's inside — the REAL numbers

> <table><tr><td>
>
> ⚠️ **Read this first:** the **19,096,153** figure is the **total parquet row count of the whole collection** — it is *not* all distillation data. Distillation is only **8.6%** of it. The bulk is **coding (66.3%)** and **cybersecurity (23.7%)**. Full breakdown below, measured file-by-file from actual parquet footers.
>
> </td></tr></table>

| Category | Parquet rows | Share | Size (GB) | Files |
|---|---:|---:|---:|---:|
| 💻 `coding` | **12,663,252** | 66.3% | 1.28 | 4 |
| 🛡 `cyber security` | **4,519,197** | 23.7% | **9.99** | 1,145 |
| 🏭 `distillation` | **1,642,673** | 8.6% | 2.97 | 1,212 |
| 📝 `instructions` | 114,857 | 0.6% | 0.04 | 2 |
| 🗂 `index` | 51,458 | 0.3% | 0.01 | 76 |
| 🔁 `RSI` | **26,334** | 0.1% | 0.22 | 2 |
| 🌌 `humanities` | 36,901 | 0.2% | <0.01 | 2 |
| 🔬 `science` | 22,024 | 0.1% | <0.01 | 3 |
| ⚙️ `applied` | 19,457 | 0.1% | <0.01 | 2 |
| 🧮 `math` | jsonl only | — | 0.01 | 1 |
| 🎭 `emotions` | reserved | — | — | 0 |
| 🧭 `State` | reserved | — | — | 0 |

On top of the parquet: **~6.8 GB of JSONL/JSON/CSV + raw archives** — including a **1.05 GB GPT-5.6 agent-trace log**, **~1.95 GB of MITRE/CVE security records**, and **3.0 GB of security-gym v4 zstd SQLite experiment DBs**.

<img src="assets/chart_categories.png" width="100%" alt="Category breakdown"/>

---

## 🔁 Full aggregation from Manusagents/Omni-Frontier — the feature manifest

**Every one of the 2,214 files of the original `Manusagents/Omni-Frontier-Distillation-SFT-Cyber-Coding-Med-dataset-collection` is accounted for here** — verified by sha256 content hashes and parquet row counts, not by eyeballing. This is not a fork: it is the **upgraded continuation** of Omni-Frontier, with every original feature carried over, upgraded, or transparently flagged.

<table><tr><td>

✅ **Audit result: 2,214 / 2,214 original files reconciled** — 2,075 mapped (incl. 5 security-gym files re-added by this audit) · 78 exact duplicates (sha256-verified) · 16 row-verified subsumed copies · 9 derived `_viewable` copies (row-verified) · 30 VCS files · 5 root assets replaced. Full evidence: **`Data/index/aggregation_report.json`**

</td></tr></table>

| Original feature (Manusagents card) | Where it lives in this repo | Status |
|---|---|---|
| **Subset A — cleaned SFT core** (`full_clean`, quality-scored) | `coding/` `distillation/` `cyber security/` … — `full_clean` row-split by category (v2 **and** data variants both kept) | ✅ carried + re-organized |
| **Subset B — human-crafted specialist corpora** | cybersecurity-qa-v2 (**2,121,468 rows**), NOSK (**242,307 rows**), threat-intel ×4, VoltVAR (**100,000 rows**), BlackHat-Kali, CyberSec-Bench, ChemicalData | ✅ carried, per-source mapped |
| **12 domains** | 12-category fixed taxonomy, stable paths (`Data/<category>/`) | ✅ upgraded |
| **80+ source datasets** | 15 source repos consolidated; **per-file sha256 provenance** in `source_map.json` (original had no per-file provenance) | ✅ upgraded |
| **11 viewer configs** | **44 schema-homogeneous viewer configs covering every one of the 356 parquet files — all 19,096,153 rows browsable**; 283 raw jsonl/csv/text files kept with documented schemas | ✅ upgraded |
| **SHA-256 dedup (65% reduction)** | original cleaning stats reproduced below **+** our global file-level dedup (166 dupes removed, 78 more verified & skipped by content hash) | ✅ upgraded |
| **Quality scores 0–100** (`quality_score` column) | columns preserved untouched in all Omni splits | ✅ carried |
| **Native `messages` chat format** | kept 1:1 across RSI / k3dist / mitre-harmony / CYBER-GOD chat data | ✅ carried |
| **Medical & robotics subsets** | ⚠️ **no medical/robotics files were ever shipped in the original repo** — flagged honestly in `aggregation_report.json` instead of silently dropped | 🧾 honest flag |
| **Power-grid security (VoltVAR IEEE-123)** | `applied/` — 80,000 train + 20,000 validation attack-command rows | ✅ carried |
| **Chemical science (Jaafer ChemicalData)** | `science/` + `humanities/` | ✅ carried |
| **NOSK pentesting corpus (~255K claimed)** | `cyber security/` — 49 shards, **242,307 footer-measured rows** | ✅ carried, measured |
| **Threat-intel train/eval/multiturn/blended** | `cyber security/` — row-verified 9,172 / 798 / 1,500 / 12,229 | ✅ carried |
| **Croissant metadata** | security-gym `croissant.json` shipped + HF auto-croissant on all 44 configs | ✅ upgraded |
| **Charts (distribution + radar)** | 5 matplotlib charts + 2 AI art images + banner + org avatar | ✅ upgraded |
| **Apache-2.0 single license** | honest **per-source licensing table** (Apache/MIT/CC-BY/CC-BY-NC) | 📄 upgraded |
| **en/fr/fa multilingual claim** | card declares `en` (dominant); original shipped no per-row language field to verify fr/fa | 🧾 honest flag |
| **Streaming / Axolotl / Unsloth / TRL** | streaming works on all parquet configs; TRL + LoRA recipes below | ✅ carried |

### 📈 The original cleaning pipeline, reproduced

Verbatim from the original `cleaning_stats.json` — this is what produced the 6.94M-row core:

| category | raw | exact dups removed | final |
|---|---:|---:|---:|
| coding | 16,853,976 | 7,233,388 | 6,322,417 |
| cybersecurity | 1,869,568 | 20,717 | 401,511 |
| applied | 200,000 | 165,645 | 18,794 |
| humanities | 200,000 | 151,329 | 30,653 |
| instruction | 454,100 | 15 | 62,619 |
| science | 184,896 | 104,756 | 12,676 |
| distilled | 179,588 | 69,703 | 65,297 |
| index | 50,000 | 17,631 | 26,109 |
| **total** | **19,992,128** | **7,813,184** | **6,940,076** |

### 🆕 What this repo adds (not in the original)

- 🔁 **RSI category** — OpenMLE traces (26,259, dual-listed in coding) + **76 DeepSeek-R1-style self-generated reasoning traces** built by SHS-Lab
- 🗄 **Security-gym v4 raw SQLite archives** — 3.0 GB of zstd RL-experiment DBs (`exp01_90d`, `exp_30d_heavy`, `exp_7d_brute`) re-added after the audit
- 🧾 **`aggregation_report.json`** — the full 2,214-file reconciliation with evidence
- 🥷 **Exploit-DB corpus ×5 sources consolidated** — 1,030 raw exploit `.txt` files also packed into one browsable parquet
- 📊 **Footer-measured honesty** — every row count on this card comes from real parquet footers

---

## 📏 Quality — measured, not claimed

Every number below was computed by sampling real rows from the actual parquet files (footer row counts are exact, content depth is sampled over ~120 rows per config and weighted by rows):

| Category | Rows (exact) | Avg content depth* | What "deep" means here |
|---|---:|---:|---|
| 🛡 `cyber security` | **4,519,197** | **~11,684 chars** | CYBER-GOD-MODE deep ChatML analyses avg **25,768 chars**; exploit texts avg **71,584 chars**; cybersecurity-QA avg 2,593 |
| 💻 `coding` | **12,663,252** | **~2,319 chars** | OpenMLE traces avg **37,553 chars with 100% `<think>`**; Omni coding splits avg 1,400–3,092 |
| 🏭 `distillation` | **1,642,673** | **~4,775 chars** | kimi-k3 debug traces avg 20,768; GPT-5.6 sessions avg 15,376; think-shards avg 10,548 @ 100% `<think>` |
| 🔁 `RSI` | **26,334** | **~37,475 chars** | 100% `<think>` coverage across both configs |
| 📝 `instructions` | 114,857 | ~574 chars | quick-response style (avg 252–340 char replies) |
| 🗂 `index` | 51,458 | ~1,085 chars | provenance/metadata rows |
| 🌌 `humanities` | 36,901 | ~545 chars | encyclopedic pairs |
| 🔬 `science` | 22,024 | ~551 chars | chemical definitions + science QA |
| ⚙️ `applied` | 19,457 | ~445 chars | power-grid command rows (short, label-like) |

\* weighted mean of sampled row size (all text fields, json-serialized). Tokenizer-ready shards (`input_ids`) average **801 tokens** per row. The Omni splits also carry the original pipeline's `quality_score` (0–100) column untouched.

### 🎯 The original `quality_score` (0–100) — measured over all 6,863,735 scored rows

Eight Omni full-clean splits carry the original pipeline's 0–100 quality score. We computed **exact** statistics over every single scored row (column-only remote parquet reads — no sampling, no guessing):

| Category | Scored rows | **Avg score** | P25 | Median | P75 | Max | % below 50 |
|---|---:|---:|---:|---:|---:|---:|---:|
| 🛡 `cyber security` | 396,206 | **55.2** | 53 | 55 | 60 | 60 | 12.4% |
| 💻 `coding` | 6,314,576 | **51.9** | 52 | 52 | 52 | 60 | 0% |
| 🏭 `distillation` | 65,297 | **59.9** | 60 | 60 | 60 | 60 | 0% |
| 📝 `instructions` | 52,238 | **53.4** | 51 | 53 | 55 | 60 | 0% |
| 🗂 `index` | 25,349 | **52.5** | 52 | 53 | 53 | 55 | 0% |
| 🌌 `humanities` | 6,248 | **51.1** | 51 | 51 | 51 | 58 | 0% |
| 🔬 `science` | 3,158 | **51.2** | 51 | 51 | 51 | 60 | 0% |
| ⚙️ `applied` | 663 | **52.6** | 51 | 52 | 55 | 56 | 0% |

**Overall: 52.2 avg across all 6.86M scored rows.** Read this honestly: the original pipeline's nominally 0–100 scale effectively tops out at **60** (the max observed in 6.86M rows) — so within this scorer, 55+ is its "premium" tier and `cyber security` (55.2) plus `distillation` (59.9) lead the pack. The raw specialist sources that never went through this scorer — CYBER-GOD-MODE, Exploit-DB texts, OpenMLE traces, security-gym, NOSK, VoltVAR — are the deep-content story in the table above, not this column.

---

## 🧭 What can you build — per-category use cases

### 🛡 cyber security (4.52M rows) — offense-aware defense

- **Security assistant / SOC copilot SFT** — CYBER-GOD-MODE gives 642K deep ChatML analyses (avg 25.8K chars of code-aware attack reasoning); cybersecurity-QA adds 2.6M Q→A pairs for breadth
- **SQL-injection detection & WAF tuning** — labeled SQLi datasets (Kaveny + AmirAliGharesoufloo) plus Exploit-DB input/output pairs (70K) train classifiers or explainable detectors
- **Exploit comprehension & summarization** — 1,030 full disclosure texts (avg 71.6K chars) + detoxio (5.8K) + grantabejar + wwe123 archive: fine-tune models that read raw exploits and emit structured CVE/Vuln info
- **Threat-intel extraction to STIX/CVE** — ~2 GB of MITRE STIX records + harmony-format conversations → train extractors that turn unstructured reports into ATT&CK-mapped JSON
- **Pentest & red-team tutoring** — NOSK corpus: 242,307 real pentesting/security-analysis conversations
- **ICS / power-grid security ML** — VoltVAR IEEE-123: 100K labeled attack-command rows (80K train / 20K val) for anomaly detection research
- **Security-agent RL environments** — security-gym v4 raw SQLite DBs (3.0 GB, 7d/30d/90d experiments) as replay/environment data for agentic RL
- **Evaluation** — CyberSec-Bench: 200 reference-answer QA for benchmarking cyber models

### 💻 coding (12.66M rows) — from syntax to agentic debugging

- **General code SFT at scale** — 6.3M instruction/response pairs (Omni splits, avg 1.4–3.1K chars)
- **Long-CoT code reasoning** — OpenMLE: 26,259 Kaggle-ML traces, avg 37.5K chars, **100% explicit `<think>`** — ideal for R1-style thinking models
- **Debug & repair training** — kimi-k3 coding/debug traces (4K, avg 20.8K chars multi-turn) + GPT-5.6 sol/luna/terra agent sessions (15K rows + 1.05 GB raw logs with `assistant_steps`)
- **ML-pipeline generation** — OpenMLE task/traces pairs teach models to build full scikit-learn/Kaggle pipelines

### 🧮 math — Omni-MATH competition problems (raw jsonl) for math CoT SFT, verifier training and eval

### 🔁 RSI — Recursive Self-Improvement

- **Reflection & self-correction loops** — every row is a full self-question → self-check → self-decide chain (OpenMLE 100% `<think>`, our 75 self-generated R1-style traces)
- **GRPO / RLVR seed data** — thinking traces are the raw material for verifiable-reward RL
- **Self-QA distillation** — train models that critique and improve their own outputs

### 🏭 distillation — cross-teacher style transfer

- **Multi-teacher distillation** — GPT-5.6 / Kimi-K3 / Qwen3.8-Max / GLM-5.2 response styles in one corpus (`family`/`source` columns let you gate per teacher)
- **Think-style transfer** — 577K message rows at 100% `<think>` coverage (avg 10.5K chars)
- **Tokenizer-ready training** — `input_ids`/`labels` shards (104K rows, avg 801 tokens) load straight into custom trainers
- **Token-budget analysis** — per-row `total_tokens`/`assistant_tokens` stats (460K rows) for curriculum design

### 🔬 science · 🌌 humanities · ⚙️ applied · 📝 instructions · 🗂 index

- Chemical definitions + science QA → domain QA bots; encyclopedic humanities pairs → knowledge chat; power-grid rows → time-series security research; 114K quick instruction rows → system-prompt/tone tuning; index rows → provenance lookups and dedup audits

### 🎭 emotions · 🧭 State — reserved (no content yet, stable paths guaranteed)

---

## 🛡 Cybersecurity — 4.5M rows

<img src="assets/art_cyber.jpg" width="100%" alt="Cybersecurity"/>

This is the largest security-focused SFT block you'll find in one place — and it's **real, structured security knowledge**, not noise:

| Source | Format | What's actually in it | Rows |
|---|---|---|---:|
| **CYBER-GOD-MODE-FULL-V1** | ChatML (`<|im_start|>`) | Deep code-analysis & review tasks on real repos (e.g. *"Analyze `v1/topdown/tokens_bench_test.go` from OPA"*) — avg **10.2K chars** per example | 128,563 |
| **ExploitDB mirrors** (Waiper + Detoxio + grantabejar + wwe123) | CSV/JSONL/**raw .txt** | Structured exploit records + **1,030 original exploit texts** — every raw text made browsable via `exploitdb_texts__viewer.parquet` | 76,065 |
| **MITRE STIX + CVE + ExploitDB harmony** | Harmony + Alpaca/ChatML | CVE summaries with mitigation guidance, defensive-security assistant dialogues | ~2 GB jsonl |
| **Omni-Frontier cyber splits** | Parquet | Threat-intel multiturn, vulnerability-mgmt, incident-response, security-QA conversations | 18,521 |
| **SQLi corpora** (Kaveny + AmirAliGharesoufloo) | CSV/JSON | SQL-injection sample datasets | ~40K |
| **cybersec_bench** | JSONL | Benchmark-style Q/A with `category` + `difficulty` + `reference_answer` — **use it as held-out eval!** | 200+ |

<details>
<summary>👀 Real excerpt — CYBER-GOD-MODE (code analysis task)</summary>

```
<|im_start|>user
Analyze 'v1/topdown/tokens_bench_test.go' from 'OPA'.
<|im_end|>
<|im_start|>assistant
FILE: v1/topdown/tokens_bench_test.go
SOURCE: OPA
CONTENT: // Copyright 2025 The OPA Authors.  All rights reserved.
...
```
</details>

<details>
<summary>👀 Real excerpt — MITRE/CVE record (instruction → output)</summary>

```
CVE-2025-52714 summary:
Improper Neutralization of Special Elements used in an SQL Command
('SQL Injection') vulnerability in shinetheme Traveler traveler allows
SQL Injection. This issue affects Traveler: from n/a through < 3.2.2.
```
</details>

<details>
<summary>👀 Real excerpt — harmony-format defensive assistant</summary>

```
<|start|>system<|message|>You are a cybersecurity assistant specialized
in vulnerability analysis and defensive security.
Knowledge cutoff: 2024-06 ... Reasoning effort: high
<|end|><|start|>developer<|message|># Instructions
Provide accurate, concise security explanations.
Prefer actionable ...
```
</details>

---

## 💻 Coding — 12.6M rows

The single biggest block — and its crown jewel is unique:

### 🏆 OpenMLE SFT Traces (26,259 traces, avg **37,793 chars**)

Every row is a **complete Kaggle-competition solution trace**: the system prompt sets an ML-engineering competition, the model plans, and **100% of traces carry full `<think>` reasoning** (verified row-by-row — 26,259/26,259). Tasks span DDoS detection on CICIoT2023 (ML-for-security), emotion classification, stock sentiment, fashion recognition, Steam reviews analytics and dozens more — with 93.7% showing explicit self-check language (*"wait"*, *"let me verify"*, *"actually, on second thought"*). Because of this, it is **dual-listed in `coding` AND `RSI`**.

<details>
<summary>👀 Real excerpt — OpenMLE system prompt (Kaggle Grandmaster)</summary>

```
You are a Kaggle Grandmaster attending a high-stakes competition. In order
to win this competition, you need to come up with an excellent and creative
plan for a solution and then implement this solution in Python.
You must use Machine Learning/Deep Learning/Computer Vision/NLP/etc. methods
... preinstalled: torch, tensorflow, scikit-learn, numpy, pandas, xgboost,
lightgbm, catboost, transformers, datasets, tokenizers, huggingface-hub ...
```
</details>

### Omni-Frontier coding split (12.64M rows)

Multi-frontier distilled coding conversations — the volume backbone for large-scale coding SFT. avg 1.8K chars/row, 12.3% carry `<think>` blocks.

### Also in coding: kimi-k3 + GPT-5.6 debug traces

Distributed across `coding` and `distillation` (see next section): Kimi K3 and GPT-5.6 coding-and-debugging corpora with multi-step agent behavior.

---

## 🏭 Distillation — the honest deep-dive (1.6M rows)

> <table><tr><td>
>
> 💡 **"Distillation" is NOT one thing here — it's a multi-teacher reasoning factory.** Measured from real samples: **math 26% + code 24% + reasoning 18% + instruction 13% + agent/tool 10%** of traces. Yes — **~24% of the distillation category is literally coding data** (multi-step agent coding, debugging, security coding tasks). It also carries **tokenizer-level artifacts** (`input_ids`, `labels`, masking metadata) for pipeline debugging.
>
> </td></tr></table>

<img src="assets/chart_distillation_mix.png" width="55%" alt="Distillation domain mix"/>

### Who teaches what

| Teacher / source | What it contains | Scale |
|---|---|---|
| **Qwen3.8-Max + GLM-5.2 + Kimi-K3** (`r0b0tlab`) | Multi-teacher distillation traces tagged by domain (`math`, `code`, `reasoning`, `instruction`, `agent_tool`), system-role seeded, avg 3.1K chars, 31.7% carry `<think>`-style self-checks; ships `input_ids`/`labels` variants too | 198 parquet shards |
| **GPT-5.6 Sol/Luna/Terra** (`Crownelius`) | Frontier reasoning triplets, avg **17.8K chars** per record | 15,353 rows |
| **GPT-5.6 coding+debug** (`greghavens`) | **1.05 GB JSONL agent-log** — sessions with `session_id`, `assistant_steps` (multi-step!), `lang`, `security_task`, `reasoning_effort` fields. Real agent trajectories like *"I'm doing the table-rules core of our blackjack trainer as assembly routines the sim can call..."* | ~105K records |
| **Kimi-K3 coding+debug** (`greghavens`) | **Moonshiner-style CI/coding tasks** — e.g. shellcheck lint blocking a pipeline, multi-turn agent resolution; domain tags: `coding` 74%, `agent` 26% | 3,956 rows |
| **Omni-Frontier distill split** | Frontier-style long answers (Claude-class voice: *"Drawing from the autonomous, frontier-level reasoning characteristic of Claude Mythos..."*), 26% `<think>` | 120,423 rows |
| **STL traces + session/event logs** | Long-context session and event jsonl logs | ~0.6 GB |

<details>
<summary>👀 Real excerpt — kimi-k3 coding/debug task</summary>

```
=== MOONSHINER TASK BOUNDARY ===
CI grew a shellcheck lint step (severity=style, zero findings allowed)
and our little photosort.sh is the last script blocking the pipeline.
It's not just the linter being fussy, ei...
```
</details>

<details>
<summary>👀 Real excerpt — k3-dist coding teacher (system seed)</summary>

```
You are an expert software engineer. Write complete, correct, runnable
code in the language requested by the user. Follow the requested language,
interface, and constraints. Explain the approach briefly when reque...
```
</details>

<img src="assets/chart_sources.png" width="100%" alt="Source contributions"/>

---

## 🔁 RSI — Recursive Self-Improvement

<img src="assets/chart_rsi_domains.png" width="50%" alt="RSI domains"/><img src="assets/art_reasoning.jpg" width="49%" alt="Reasoning art"/>

Two sources, one philosophy — **the model thinks about its own thinking**:

1. **`openmle_sft_traces`** — 26,259 Kaggle-grade traces (see Coding section). 100% `<think>`, 93.7% self-check vocabulary. Dual-listed from `coding`.
2. **`shslab_rsi_selfgen`** — **76 fully self-generated traces** in the complete DeepSeek-R1 loop: *question the problem → decompose into self-questions → attempt → catch own flaw → backtrack → verify independently → explicit final decision*. 10 domains (math, coding, science, cybersecurity, logic, planning, data-analysis, systems-design…), 3 difficulty tiers, avg **4,238 thinking chars** per trace. MIT-licensed (ours).

<details>
<summary>👀 Real excerpt — RSI selfgen trace (systems-design, expert)</summary>

```
Hmm, what is this really asking? It wants me to design a distributed rate
limiter with specific parameters (10k req/s/user) and compare three
algorithms across multiple dimensions. ...

First, I should decompose this: What are the core requirements?
1) Distributed nature 2) Per-user limits 3) Specific rate (10k/s) ...
```
*…the trace continues through a failed first approach, explicit backtracking, verification, and ends with "Okay. Final decision: token bucket…"*
</details>

---

## 🧮 Math · Science · Humanities · Applied · Instructions

| Category | What's inside | Rows |
|---|---|---:|
| 🧮 **math** | Omni-MATH competition problems (jsonl) | jsonl |
| 🔬 **science** | Science explanations & QA (Omni splits) | 22,024 |
| 🌌 **humanities** | Humanities knowledge pairs (chemical definitions, encyclopedic entries) | 36,901 |
| ⚙️ **applied** | Applied/practical task records | 19,457 |
| 📝 **instructions** | Instruction-following corpus, avg 331 chars — quick-response style | 114,857 |
| 🗂 **index** | Catalog/metadata records incl. source map | 51,458 |


### ⚕️ About the "Med" in the name — the honest story

The original repo's name says *Med* and its card advertised medical & robotics subsets — but **no medical or robotics files were ever shipped in that repo** (we audited all 2,214 files). We keep the family name for continuity with the Omni-Frontier lineage, and we flag the gap instead of pretending. Any health/bio-adjacent rows that exist inside the QA corpora remain untouched, and the reserved `emotions` / `State` categories are the natural home for future curated medical & affective data.

### ⚗️ Chemistry & power-grid — the human-crafted specialist data

- **Jaafer ChemicalData** → chemical reaction/property records in `science/` and `humanities/` (22,024 science rows total)
- **VoltVAR / IEEE-123 power-grid attack commands** → `applied/` — **100,000 rows** (80K train + 20K validation) of time-series attack commands against a power-grid simulation. This is the "power-grid security" domain the original card advertised — here it is, row-measured and viewer-ready

### 🥷 NOSK Hacking corpus — 242,307 rows

49 parquet shards of pentesting/security-analysis conversations from NOSK (Nepal Open Source Klub), merged into `cyber security/`. The original card claimed ~255K rows; the real footer-measured number is **242,307** — we publish the measured number, not the marketing one.

---

## ✨ Features

- ✅ **19,096,153 exact parquet rows** counted from real parquet footers — not estimates
- ✅ **12 fixed taxonomy categories** — stable paths across versions
- ✅ **44 schema-homogeneous viewer configs** — **every parquet row (19,096,153) browsable in the Dataset Viewer**; 283 raw jsonl/csv files stay in-repo, fully documented
- ✅ **SHA-256 file-level dedup** (166 exact duplicates removed; different content always kept)
- ✅ **Verified uploads** — manifest with per-file sha256 + size, 0 missing / 0 mismatch
- ✅ **Raw exploit texts browsable** — 1,030 `.txt` files consolidated into one parquet (originals kept)
- ✅ **Full provenance** — `Data/index/source_map.json` maps every file to its upstream dataset
- ✅ **100% aggregation audit** — all 2,214 original Manusagents/Omni-Frontier files reconciled by content hash / row count (`Data/index/aggregation_report.json`)
- ✅ **Security-gym v4 raw archives** — 3.0 GB zstd SQLite RL-security experiment DBs included
- ✅ **Training-ready & research-ready** — chat formats *and* tokenizer artifacts (`input_ids`/`labels`)

---

## 🎓 Training guide — how to actually use this

### 1. Pick your objective

| Goal | Use | Why |
|---|---|---|
| R1-style reasoning distillation | `RSI` + OpenMLE + k3-dist `reasoning`/`math` traces | Long `<think>` chains with self-correction |
| Security assistant | cyber conversations + mitre-stix harmony + CYBER-GOD-MODE | Defensive framing, structured CVE knowledge |
| Coding agent | GPT-5.6/kimi-k3 agent logs + k3-dist `code` traces | Multi-step trajectories with tool/CI context |
| General chat SFT | distillation split + instructions + humanities/applied | Diverse tone & length coverage |
| Eval set | `cybersec_bench.jsonl` (has `reference_answer`) + Omni-MATH + k3-dist `validation` shards | Ready-made held-out sets |

### 2. Handle the formats (they differ by source — on purpose)

```python
# messages-format (chat):  Data/RSI, k3dist, mitre-harmony, CYBER-GOD-MODE (raw ChatML text)
# instruction/output:      mitre-stix train jsonl
# reasoning + <think>:     OpenMLE, Omni distill split, shslab_rsi_selfgen
# tokenizer artifacts:     k3dist input_ids/labels shards (skip for text SFT)
```

### 3. SFT with TRL (chat-format configs)

```python
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

ds = load_dataset("SHS-Lab/Omni-Frontier-Distillation-SFT-Cyber-Coding-Med-dataset-collection",
                  "rsi__shslab_rsi_selfgen", split="train")
# ds[0]["messages"] = [{"role": "user", ...}, {"role": "assistant", "content": "<think>...</think>..."}]

cfg = SFTConfig(
    output_dir="./omni-frontier-sft",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=1e-5,
    num_train_epochs=2,
    bf16=True,
    logging_steps=10,
)
trainer = SFTTrainer(model="Qwen/Qwen3-8B", args=cfg, train_dataset=ds)
trainer.train()
```

**Tips that matter:**
- 🧠 **For reasoning distillation, train on the full assistant turn** (including `<think>`) — masking the think block defeats the purpose. For plain chat SFT, mask it.
- 🔀 **Mix ratios that work:** `coding 40% · cyber 25% · distillation-reasoning 20% · RSI 10% · instructions 5%` — then anneal with 2× upweight on your target domain.
- 📦 **Start streaming before you commit**: every config supports `streaming=True` — pilot on 50K rows first.
- ⚠️ **OpenMLE = CC-BY-NC-4.0** → keep non-commercial, or exclude those files from commercial runs.
- 🚫 Don't mix `input_ids`-artifact shards into text SFT — they're for tokenizer-pipeline debugging/ablation.

### 4. LoRA recipe (single 48GB GPU)

```python
peft_cfg = LoraConfig(r=32, lora_alpha=64, lora_dropout=0.05,
                      target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                                       "gate_proj", "up_proj", "down_proj"])
```

---

## 🔎 Browsing — the Dataset Viewer

The viewer runs on **44 explicit parquet configs** — every one of the **19,096,153 parquet rows** is served (no skipped rows, no preview-only samples). The 283 raw `jsonl`/`csv`/text files are deliberately *not* in the viewer (they made the server-side indexer time out — the `ResponseNotFound` you may have seen came from those); they remain fully in the repo:

```python
from datasets import load_dataset

# any of the 44 viewer configs
ds = load_dataset("SHS-Lab/Omni-Frontier-Distillation-SFT-Cyber-Coding-Med-dataset-collection",
                  "cyber_security__cyber_god_mode", split="train", streaming=True)

# raw jsonl/csv: stream straight from the hub
import fsspec
url = "https://huggingface.co/datasets/SHS-Lab/Omni-Frontier-Distillation-SFT-Cyber-Coding-Med-dataset-collection/resolve/main/Data/cyber%20security/train.jsonl"
with fsspec.open(url) as f:
    for i, line in enumerate(f):
        ...
```

Schema documentation for every raw file lives in `Data/index/source_map.json` and in the per-category tables above. If the viewer ever shows a warm-up message right after a push, that is HF's indexer re-processing — the configs themselves are stable.

---

## 🧹 Quality · Dedup · QA

Measured, not claimed: every parquet footer fingerprinted; JSONL/CSV headers read; **166 byte-identical files removed**; manifest verification = **0 missing, 0 size-mismatch**. Non-row artifacts (`.py` code payloads, `.pyc`, provenance `.json`, checksums, banners, 3.0 GB of raw security-gym SQLite archives) stay in the repo but are deliberately excluded from viewer configs — see `Data/index/viewer_manifest.json`. On top of our own dedup, the aggregation audit **re-downloaded 78 supposedly-missing original files and proved byte-identical duplicates** (sha256), and row-verified 16 subsumed per-category copies plus 9 derived `_viewable` copies — zero content was lost in the merge. Full evidence chain: `Data/index/aggregation_report.json`.

---

## 🗺 Roadmap

- [x] 12-category fixed taxonomy with stable paths
- [x] Per-file sha256 provenance (`source_map.json`)
- [x] RSI category live — OpenMLE dual-listed + 76 self-generated R1-style traces
- [x] **Full Manusagents/Omni-Frontier aggregation audit — 2,214/2,214 files reconciled**
- [x] Security-gym v4 raw archives (3.0 GB) re-added
- [ ] `emotions` & `State` categories populated (reserved — awaiting curated sources)
- [x] Viewer slimmed to 44 parquet configs → full 19.1M-row index (jsonl/csv served raw to avoid indexer timeouts)
- [ ] Parquet mirrors for the big raw jsonl corpora (MITRE-STIX, GPT-5.6 session logs)

---

## 📄 Licensing & attribution

| License | Sources |
|---|---|
| Apache-2.0 | Omni-Frontier (Manusagents) · mitre-stix-harmony · Omni-MATH · sql-injection (Kaveny) |
| MIT | Waiper/ExploitDB_DataSet · **shslab_rsi_selfgen (ours)** |
| CC-BY-4.0 | kimi-k3 traces · GPT-5.6 traces (both greghavens + Crownelius) |
| CC-BY-NC-4.0 ⚠️ | **OpenMLE-SFT-Traces** — non-commercial only |
| Unspecified | CYBER-GOD-MODE · detoxio · grantabejar · wwe123 · AmirAliGharesoufloo |

> ⚠️ Cyber subsets contain offensive-security knowledge from public corpora — for defense, education and authorized research only.

---

## 🙏 Citation

```bibtex
@misc{shslab2026omnifrontiercollection,
  title  = {Omni-Frontier Collection: Cybersecurity, Coding, Math, Science and RSI Reasoning SFT Corpus},
  author = {SHS-Lab},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SHS-Lab/Omni-Frontier-Distillation-SFT-Cyber-Coding-Med-dataset-collection}
}
```

<div align="center">
  <sub>Built with ❤️ by <a href="https://huggingface.co/SHS-Lab"><b>SHS-Lab</b></a> · updated Aug 29, 2026</sub>
</div>