Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
SFT closes multilingual reasoning gaps; GRPO only adds value when the base behavior already exists
“SFT teaches a behavior the model doesn't have. GRPO optimizes a behavior the model can already sometimes produce.”
AWS demonstrates that GPT-OSS-20B, like most reasoning models, defaults to English chain-of-thought even when prompted in other languages, achieving only 22% correct target-language reasoning on a multilingual benchmark. Fine-tuning on SageMaker HyperPod with SFT largely closes the gap, with GRPO providing incremental gains only where the base model already has latent capability. The practical takeaway—match technique to capability gap—is a useful heuristic for teams deploying open-weight reasoning models for non-English users.