AI × AEC · CURATED INSIGHT
AWS Tests SFT and GRPO Fine-Tuning for Multilingual Reasoning on SageMaker HyperPod
AWS engineers demonstrated fine-tuning GPT-OSS-20B and Qwen2.5-7B on Amazon SageMaker HyperPod, evaluating the complementary roles of SFT and GRPO across multilingual reasoning tasks.
In an AWS technical guide, authors shared workflows for fine-tuning open-weight reasoning models using LoRA on Amazon SageMaker HyperPod. Experimental evaluations indicated that supervised fine-tuning (SFT) is necessary to teach missing capabilities such as non-English reasoning, while Group Relative Policy Optimization (GRPO) yielded +6 to +8 point improvements only when baseline distributions provided optimization headroom.
Key takeaways
- Base GPT-OSS-20B achieved 22 percent reasoning-language accuracy; LoRA SFT raised this to 100 percent, while answer-language accuracy reached 62 percent.
- Applying GRPO atop the SFT checkpoint improved answer-language accuracy from 62 percent to 70 percent, whereas tasks without headroom like GSM8K yielded no gain.
Source：AWS Machine Learning Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.