AWS has published a technical walkthrough for fine-tuning the open-weight GPT-OSS-20B reasoning model to produce reasoning and final answers in French, German, Spanish, Italian, and English. The work was conducted on Amazon SageMaker HyperPod, using Amazon Elastic Kubernetes Service for orchestration and Low-Rank Adaptation for parameter-efficient training.

According to AWS, the base model produced reasoning in the requested language in 22% of a 50-example evaluation. It reasoned in English in all English cases, but did so in 0% of French and Spanish cases, 10% of German cases, and 0% of Italian cases.

AWS first applied supervised fine-tuning using the HuggingFaceH4/Multilingual-Thinking dataset, which contains about 949 training examples and 50 validation examples. After 1,000 training steps, the company reported that reasoning-language accuracy reached 100% across the five languages. Answer-language accuracy was 62%, with the lowest reported result in Spanish at 30% and the highest in English at 100%.

The post then describes reinforcement learning with Group Relative Policy Optimization (GRPO) as a refinement step for behavior the model can already produce. AWS says it used GRPO on top of the supervised fine-tuning checkpoint, with rewards for answer language, reasoning language, and brevity. The post also discusses a separate Qwen2.5-7B control experiment intended to examine when GRPO provides additional value.