Fine-tuning teaches GPT-OSS-20B to reason in the requested language

AWS researchers fine-tuned the open-weight GPT-OSS-20B model to keep both its reasoning and final answer in a user’s requested language. On a 50-example test covering English, French, German, Spanish, and Italian, the base model used the correct reasoning language only 22% of the time.

Supervised fine-tuning raised that measure to 100%. The training used roughly 949 examples from the Multilingual-Thinking dataset and Low-Rank Adaptation, a method that changes a small set of additional parameters rather than updating the full model.

The team then tested Group Relative Policy Optimization, a reinforcement-learning method that compares several responses to the same prompt and rewards better ones. Its conclusion was that supervised training is best for teaching behavior the model rarely produces, while reinforcement learning can refine behavior that already appears often enough to reward.

The work ran on SageMaker HyperPod, with separate Qwen2.5-7B control experiments used to examine when reinforcement learning helps. The perfect language score comes from a small, purpose-built evaluation and does not establish overall multilingual quality or factual accuracy.