SafetyPage

Lab system cards and frameworks · System card · OpenAI

gpt-oss-120b & gpt-oss-20b Model Card

5 Aug 2025

Read at openai.com ↗Read at sourceAll rights reserved

Outline · 35 pages

  1. Introductionp. 4
  2. Model architecture, data, training and evaluationsp. 4
  3. Quantizationp. 5
  4. Architecturep. 5
  5. Tokenizerp. 6
  6. Pretrainingp. 6
  7. Post-Training for Reasoning and Tool Usep. 7
  8. Harmony Chat Formatp. 7
  9. Variable Effort Reasoning Trainingp. 8
  10. Agentic Tool Usep. 8
  11. Evaluationp. 8
  12. Reasoning, Factuality and Tool Usep. 9
  13. Health Performancep. 9
  14. Multilingual Performancep. 10
  15. Full Evaluationsp. 11
  16. Safety testing and mitigation approachp. 11
  17. Default Safety Performance: Observed Challenges and Evaluationsp. 12
  18. Disallowed Contentp. 12
  19. Jailbreaksp. 14
  20. Instruction Hierarchyp. 14
  21. Hallucinated chains of thoughtp. 16
  22. Hallucinationsp. 17
  23. Fairness and Biasp. 17
  24. Preparedness Frameworkp. 17
  25. Adversarial Trainingp. 18
  26. External Safety expert feedback on adversarial training methodologyp. 18
  27. Capability findingsp. 19
  28. Biological and Chemical - Adversarially Fine-tunedp. 19
  29. Long-form Biological Risk Questionsp. 20
  30. Multimodal Troubleshooting Virologyp. 21
  31. ProtocolQA Open-Endedp. 21
  32. Tacit Knowledge and Troubleshootingp. 22
  33. TroubleshootingBenchp. 22
  34. Evaluations and Red Teaming by External Safety Expertsp. 23
  35. Cybersecurity - Adversarially fine-tunedp. 23
  36. Capture the Flag (CTF) Challengesp. 24
  37. Cyber rangep. 25
  38. AI Self-Improvementp. 27
  39. SWE-bench Verified (N=477)p. 27
  40. OpenAI PRsp. 28
  41. PaperBenchp. 29
  42. Appendix 1p. 30
  43. Appendix 2p. 31
  44. Recommendations Implementedp. 31
  45. Recommendations Not Adoptedp. 32
  46. Contributorsp. 32