Fine-tuning GPT-OSS-20B for multilingual reasoning on SageMaker HyperPod
In brief
- The base GPT-OSS-20B model reasons in English even when prompted in another language (22 % success rate).
- Supervised fine-tuning (SFT) brings multilingual reasoning to 100 %, but final answers often remain in English.
- Reinforcement learning with GRPO improves final answer accuracy in the target language by 8 points, without affecting reasoning.
Amazon fine-tuned the open-weight model GPT-OSS-20B to perform reasoning in multiple languages, not just English. According to the announcement, the base model correctly responded in French, German, Spanish, or Italian only 22 % of the time, mostly defaulting to English.
The team tested two methods: supervised fine-tuning (SFT), which fills major gaps, and reinforcement learning with GRPO, which refines responses when room for improvement exists. On a 50-case multilingual test, SFT achieved 100 % reasoning in the target language, while GRPO added up to 8 points of accuracy for the final answer language, without degrading reasoning.
This work highlights the importance of matching the technique to the model's capability gap. Technical details, including the use of LoRA and SageMaker HyperPod infrastructure, are also provided for reproducibility.
What it changes for you
- EducationMultilingual educational tools using LLMs can now provide reasoning and answers consistent with the user's language, enhancing the learning experience.
- Marketing & AdvertisingMultilingual advertising campaigns or content generation tools will achieve higher accuracy and local relevance, reducing translation or reasoning errors.
This summary gives you the essentials. The full article, its figures and quotes are on Artificial Intelligence.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error