Amazon fine‑tunes GPT‑OSS‑20B on SageMaker HyperPod to reach 100% multilingual reasoning
The base GPT‑OSS‑20B model, trained only in English, scored 22% correct reasoning-language on a 50‑example multilingual test (English 100%, French 0%, German 10%, Spanish 0%, Italian 0%). Using supervised fine‑tuning (SFT) on an Amazon SageMaker HyperPod cluster (4 × ml.g6e.12xlarge with L40S GPUs) and LoRA (r=8, α=16), the team raised reasoning-language correctness to 100% across all five…
Key points
- Base GPT‑OSS‑20B achieved only 22% correct reasoning-language on multilingual test, defaulting to English.
- SFT on SageMaker HyperPod raised reasoning-language to 100% and answer-language to 62%.
- GRPO added an 8‑point boost, reaching 70% answer-language accuracy while preserving 100% reasoning.
Applying Group Relative Policy Optimization (GRPO) on top of the SFT checkpoint added a further 8‑point gain, lifting answer‑language correctness to 70% while keeping reasoning at 100%. Per‑language answer accuracy after GRPO was French 70%, German 80%, Spanish 60%, Italian 40%, English 100%. A parallel experiment with Qwen2.5‑7B‑Instruct showed GRPO adds no benefit on a task (GSM8K) where the base model already scores 90.5%, confirming that GRPO requires headroom to be effective. The post also details the training stack, prerequisites, and cleanup steps for reproducibility.
Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
AWS Machine Learning Blog · 20 October 2026
Loading the full article…
This text was published by AWS Machine Learning Blog and written by Mijanur Palash. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- The Truth Behind 'Apple and Meta Team Up on AI': It's Not Siri, It's Mac Studio | A Manager's Breakdown of Confirmed Facts, Reports, Unannounced Details, and Im · 2 src
- Elevenlabs launches two realtime speech-to-speech models for AI conversations · 1 src
- XHToken releases Spark-X2.6-4B, a 4B on-device language model · 1 src
- Qev-0.8b tops decision model benchmarks and runs on a laptop · 1 src
- Nvidia releases Cosmos3-H-Dreams real-time generative surgical simulator · 2 src
Comments
via GitHub Discussions