www.lesswrong.com/posts/szyZi5d4febZZSiq3/monitor-jailbreaking-evading-chain-of-...
1 correction found
The trained models are from the GPT-oss series (20B and 120B parameters).
OpenAI’s official specs for gpt-oss-120b list 117B total parameters (not 120B).
Full reasoning
The post states that the trained models are “(20B and 120B parameters).”
However, OpenAI’s own “Introducing gpt-oss” page lists gpt-oss-120b as having 117B total parameters (and describes the “120b” name while giving the actual total parameter count). NVIDIA’s OpenAI-compatible model reference for gpt-oss-120b likewise lists 117B total parameters.
Because multiple independent, high-credibility sources give the same total-parameter figure (117B) for gpt-oss-120b, the statement that the model is “120B parameters” is incorrect as a literal parameter-count claim (even if “120b” is the model’s name).
2 sources
- Introducing gpt-oss | OpenAI
“gpt-oss-120b activates 5.1B parameters per token… The models have 117b and 21b total parameters respectively.” (table lists gpt-oss-120b Total Params: 117B)
- openai / gpt-oss-120b (NVIDIA NIM reference)
“Total Parameters: 117B”