在 SageMaker HyperPod 上微调 GPT-OSS-20B 实现多语言推理
原标题：Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
AI 摘要
AWS 机器学习博客介绍在 Amazon SageMaker HyperPod 上对开源权重推理模型 GPT-OSS-20B 进行多语言推理微调。该模型默认用英语思考，在 50 例多语言测试中仅 22% 用目标语言推理（英语 100%，法/西/意 0%）。文章对比了监督微调（SFT）与 GRPO 强化学习：SFT 负责教会模型原本不具备的行为，GRPO 在模型已有一定概率产生目标行为时进一步优化，并给出基于 TRL、veRL、vLLM 和 LoRA 的训练栈与部署细节。
正文节选
Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod GPT-OSS-20B is an open-weight reasoning model that thinks in English. Ask it to reason in French, German, Spanish, or Italian and it acknowledges the instruction, then produces its chain-of-thought in English anyway. On a 50-example multilingual reasoning test, the base model reasoned in the correct target language only 22 percent of the time. That 22 percent was almost entirely the English examples (English 100 perc