AWS fine-tunes GPT-OSS-20B for multilingual reasoning with SFT and GRPO on SageMaker HyperPod
Overview
AWS Machine Learning Blog reports that supervised fine-tuning (SFT) raises GPT-OSS-20B's reasoning in the target language, while GRPO adds gains only where the base model already sometimes produces the desired behavior.
On a 50-example multilingual test, the base model reasoned in the requested language 22 percent of the time: 100 percent for English, and 0 percent for French, Spanish and Italian. The post says the work used LoRA on Amazon SageMaker HyperPod with Amazon EKS, and a Qwen2.5-7B control experiment run with veRL and vLLM.
The author, Mijanur Palash, says SFT does most of the work and GRPO refines only when there is headroom to improve.
Written by AI from the articles below · updated Oct 20, 1:53 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- AWS Machine Learning BlogOfficialFine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
AIAWS shows that supervised fine-tuning lifts GPT-OSS-20B's reasoning in the target language, while GRPO adds gains only where the base model already sometimes produces the desired behavior. On a 50-example multilingual test, the base model reasoned in the requested language only 22 percent of the time, with English at 100 percent and French, Spanish and Italian at 0 percent. The post uses LoRA on Amazon SageMaker HyperPod with Amazon EKS, and runs a Qwen2.5-7B control experiment with veRL and vLLM.
Heat Trend
Not enough continuous observations to show a trend yet.