AWS Publishes Fine-Tuning Tutorial for GPT-OSS-20B Multilingual Reasoning

The AWS Machine Learning Blog has published a tutorial on fine-tuning openai/gpt-oss-20b for multilingual reasoning on Amazon SageMaker HyperPod. The team first applied SFT (supervised fine-tuning), raising the chain-of-thought accuracy on a target language in a 50-example evaluation from 22% to 100%; they then added GRPO (reinforcement learning) training, improving answer language accuracy from 62% to 70%. The tutorial demonstrates a complete workflow combining SFT and GRPO to improve the model's multilingual reasoning and output language consistency, with related code and evaluation results provided by the AWS team in the blog.

AI overview of reporting · Updated 2 h前

Not enough continuous observations to show a trend yet.