On a 50-example test, the base model used the requested target language for reasoning in 22% of cases, with results nearly all attributable to English examples. AWS reports that supervised fine-tuning raised reasoning-language accuracy to 100%. Its interpretation is that SFT teaches a behavior the model lacks, while GRPO can refine a behavior that is already sometimes present. The reported evaluation covers 10 math and logic problems across five languages; the source presents no exact numerical GRPO gains in the supplied evidence. https://aws.amazon.com/blogs/machine-learning/fine-tuning-gpt-oss-20b-for-multilingual-reasoning-on-amazon-sagemaker-hyperpod
Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
In one minute
AWS describes fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod.
What changed: Adds a distinct SageMaker HyperPod case study fine-tuning GPT-OSS-20B for multilingual reasoning; priors cover other HyperPod training and SageMaker fine-tuning examples.