AWS fine-tunes GPT-OSS-20B for multilingual reasoning: base model reasoned in French 0% of the time
AWS ML Blog · rss · 2026-10-21
An AWS ML Blog tutorial details fine-tuning GPT-OSS-20B for multilingual reasoning on SageMaker HyperPod. The base model's chain-of-thought defaulted to English: on a 50-case eval it reasoned in the requested language only 22% of the time (French/Spanish 0%, German 10%).
Key takeaways:
- SFT teaches behaviors the model lacks; GRPO only reinforces behaviors it sometimes already produces — when the target behavior has zero base probability, GRPO has nothing to work with.
- SFT with LoRA (r=8, 1,000 steps) did most of the repair; GRPO added value only where headroom existed, confirmed by a controlled Qwen2.5-7B experiment across two tasks with different capability gaps.
- Stack: HyperPod + EKS, PyTorch TRL for GPT-OSS runs, veRL + Ray + vLLM for the Qwen control, FSx for Lustre storage, EFA interconnect.
All commands, manifests, and scripts are available in the awsome-distributed-ai GitHub repo.
More from coding & agent
- Research agent Épi sets 120 records, narrows n-queens constant interval 84x — dair_ai · 2026-10-21
- Compaction hooks in practice: keep last n messages verbatim, summarize the rest for full agent continuity — MoonL88537 · 2026-10-21
- The blank prompt box hides real operational burden: context, workflow and supervision fall on users — _AustinCalvert_ · 2026-10-21
- DeepMind paper: self-evolving agents forget as they learn — fastest learner GEPA also breaks old tasks most — rohanpaul_ai · 2026-10-21
- arXiv: Are Self-Evolving Agents Continual Learners? Plasticity comes at the cost of stability — rohanpaul_ai · 2026-10-21
- PM OS vs Team OS vs Company OS: how to layer your AI setup and which to build first — aakashgupta · 2026-10-21