Sakana AIFishMath
gpt-oss-120b-sft-aimo3-fishmath
Sakana supervised-fine-tuned gpt-oss-120b for AIMO3-style mathematical problems with integer answers. This is a trained derivative of OpenAI's weights.
Research release
The training sequence length is not a context specification
The card reports a maximum training sequence length of 65,536 tokens. It supports a reasoning workflow with optional Python use. That training setting does not establish a new maximum serving context.gpt-oss-120b-sft-aimo3-fishmath official model card
Working with this release
Practical guidance. Evaluate integer-answer questions separately from proof writing. Record whether Python was available so results from tool-assisted and plain-text runs remain comparable.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | SakanaAI/gpt-oss-120b-sft-aimo3-fishmath |
| Base model | OpenAI gpt-oss-120bgpt-oss-120b-sft-aimo3-fishmath official model card |
| Training data | 22,287 verified synthetic reasoning traces; ten training epochs.gpt-oss-120b-sft-aimo3-fishmath official model card |
| License | Apache 2.0gpt-oss-120b-sft-aimo3-fishmath official model card |