Amazon Web Services (AWS) released a fine-tuned version of the GPT-OSS-20B model for multilingual reasoning on Amazon SageMaker HyperPod, saying it improves language-specific reasoning capabilities. It is the company's first major update to the model since its initial release.
AWS reported a 22% reasoning-language correct score on a 50-example multilingual reasoning test, measured using the HuggingFaceH4/Multilingual-Thinking dataset. That compares with a 100% correct score for English examples and 0% for French, Spanish, and Italian.
The fine-tuned model is built on AWS SageMaker HyperPod and targets non-English users requiring multilingual reasoning. Availability begins with the release of the fine-tuned model, initially for developers and enterprises using AWS services.
"The clearest way to think about the two techniques: SFT teaches a behavior the model doesn’t have. GRPO optimizes a behavior the model can already sometimes produce," said the AWS blog post.
The announcement follows AWS's broader efforts to enhance model capabilities for global users. AWS did not say how the model will perform on other benchmarks, and it raised the question of when GRPO adds measurable value.
The company said it will continue to explore techniques for improving multilingual reasoning capabilities.
Source: awsml
