GPT-OSS-20B Multilingual Reasoning Fine-Tuning Experiment: SFT Enables Zero-Base Behavior Acquisition, while GRPO Only Effectively Improves Performance When a Probabilistic Headspace Exists
The AWS team conducted multilingual inference fine-tuning experiments on Amazon SageMaker HyperPod using GPT‑OSS‑20B, employing LoRA for parameter-efficient fine-tuning. SFT boosted the correctness rate of inference in five languages (English, French, German, Spanish, and Italian) from 22% to 100%, though the correctness rate for the answer language remained at only 62%. GRPO further improved the answer-language correctness rate to 70% without compromising inference-language performance; however, the gains varied across languages—for example, Spanish saw a +30 percentage-point increase, while Italian experienced a -20 percentage-point decline. The experiments clearly demonstrated that GRPO cannot bootstrap new capabilities from scratch and can only optimize behaviors that already occur sporadically.