AWS fine-tunes GPT-OSS-20B for multilingual reasoning on SageMaker HyperPod

GPT-OSS-20B, an open-weight reasoning model, defaults to English chain-of-thought despite user instructions to reason in French, German, Spanish, or Italian. On a 50-example multilingual test, the base model reasoned in the correct target language only 22 percent of the time, nearly all from English prompts. AWS researchers fine-tuned the model using supervised fine-tuning (SFT) and reinforcement learning with Group Relative Policy Optimization (GRPO) on Amazon SageMaker HyperPod. SFT alone was often sufficient to establish the target language behavior, while GRPO provided additional refinement when the model already had some capability. A controlled experiment with Qwen2.5-7B confirmed that matching the technique to the specific capability gap determines the outcome.