GPT-OSS-20B multilingual fix on SageMaker HyperPod: 22% to 100%

The base GPT-OSS-20B has a strong English prior: it acknowledges a request to reason in French, German, Spanish, or Italian, then produces its chain-of-thought in English anyway. That is not a prompting issue but a capability gap, which is why SFT proved necessary rather than optional—it moved reasoning-language correctness from 22% to 100% across the 50 test cases.

The remaining difficulty was the final answer, which is short and where the English prior is strongest. SFT brought answer-language accuracy to 62%, leaving a clear target for GRPO. GRPO improved answer-language to 70% while holding reasoning at 100%, a +8 point gain that came from reinforcing a behavior the model could already sometimes produce.

A controlled Qwen2.5-7B experiment on two tasks confirmed the pattern: GRPO added +6 points on multilingual reasoning where headroom existed, but no gain on GSM8K where the model was already at 90.5%. The practical decision rule is to use SFT for missing behavior and GRPO only when evaluation shows that the desired behavior is sometimes correct and there is room to improve.