Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
Amazon Web Services Mijanur Palash
GPT-OSS-20B was fine-tuned for multilingual reasoning on Amazon SageMaker HyperPod after the base model mostly ignored non-English reasoning-language instructions. On the 50-case eval, reasoning-language correctness rose from 22% (11/50) to 100% (50/50) after SFT, and GRPO further increased answer-language correctness from 62% to 70% (35/50). This shifts the model from defaulting to English mid-thought toward consistently reasoning and answering in the requested language for French, German, Spanish, and Italian.
Why it matters
Learn how we fine-tuned GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod, and how to decide between supervised fine-tuning (SFT) and reinforcement learning with GRPO. SFT took reasoning-language accuracy from 22% to 100%, and GRPO added measurable gains where headroom existed.