Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
- Who
- AWS
- What
- demonstrates fine-tuning GPT-OSS-20B with SFT and GRPO for multilingual reasoning on Amazon SageMaker HyperPod
- Quote
In this post, we walk through how we fine-tuned GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod .
What happened
AWS authors demonstrate fine-tuning GPT-OSS-20B on Amazon SageMaker HyperPod using SFT followed by GRPO. On a 50-case multilingual evaluation, SFT increased reasoning-language compliance from 22 percent to 100 percent, while GRPO raised answer-language compliance from 62 percent to 70 percent. A Qwen2.5-7B-Instruct control experiment showed additional multilingual gains but no improvement on GSM8K, supporting the authors’ guidance to apply GRPO where improvement headroom exists.
The story so far
AWS authors fine-tuned GPT-OSS-20B on Amazon SageMaker HyperPod and reported improved adherence to requested languages in a 50-case multilingual evaluation. Supervised fine-tuning (SFT) increased reasoning-language compliance from 22 percent to 100 percent, while the subsequent group relative policy optimization (GRPO) stage raised answer-language compliance from 62 percent to 70 percent. These measures concern the language used in reasoning and answers, rather than answer accuracy. In a separate control experiment using Qwen2.5-7B-Instruct, the authors reported additional multilingual gains but no improvement on GSM8K. They used that result to support their guidance to apply GRPO where there is room for improvement.
Keep reading
Latest daily · Issue 27TorchNeuron brings native PyTorch support for Hugging Face models on AWS TrainiumTrending now
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-10-20