GPT-OSS-20B multilingual reasoning: AWS reports SFT lifts target-language reasoning from 22% to 100% on SageMaker HyperPod
AWS reports that on a 50-example multilingual reasoning test, GPT-OSS-20B reasoned in the correct target language only 22% of the time before fine-tuning. In AWS’s setup on SageMaker HyperPod, SFT with LoRA raised that measure to 100%, with GRPO adding a smaller follow-on gain.
What happened
AWS reports a fine-tuning case for GPT-OSS-20B on Amazon SageMaker HyperPod focused on multilingual reasoning. On a 50-example test, the base model produced correct target-language reasoning only 22% of the time, and AWS says that success was almost entirely from English prompts. AWS then reports that supervised fine-tuning with LoRA raised reasoning-language to 100%. The same post says GRPO added a modest further improvement on top of SFT, taking answers from 62% to 70% and adding 6% on multilingual reasoning overall, while showing no gain on GSM8K because the model was already at 90.5% accuracy. This is a single-company report, so the result should be read as AWS’s measured setup and decision framework, not independent reproduction.
Why it matters
AWS’s report matters because it identifies a concrete tuning choice for multilingual reasoning: when GPT-OSS-20B could not reliably reason in the target language, SFT appears to be the step that unlocks the behavior, while GRPO mainly refines an already-working capability. That makes the article useful for deciding sequence, not just method branding.
Engineering impact
For model builders, the practical constraint is whether the target behavior is absent or merely inconsistent. AWS reports 22% baseline target-language reasoning, 100% after SFT, and a smaller GRPO lift afterward, so the engineering implication is to prioritize SFT when coverage is missing and reserve GRPO for incremental polishing.
Leadership impact
The leadership takeaway is a resource-allocation rule: if a product needs reliable non-English reasoning, the vendor-reported evidence favors funding SFT first rather than treating RL as the primary fix. GRPO looks like a second-stage investment only when the team can show remaining headroom, and the article does not support broader budgeting claims beyond that setup.
Editorial judgment
This is publishable because it changes a concrete tuning choice: AWS’s own report suggests SFT is the first lever when the model cannot perform the target behavior, while GRPO is a second-stage refinement only when headroom remains. The limitation is that the evidence is vendor-reported and should not be generalized beyond the stated setup.
Watch next
The next decision to watch is whether this pattern holds outside AWS’s reported setup and multilingual test slice. The article itself suggests using SFT when the model cannot do the behavior at all, then testing GRPO only where measurable headroom remains. A separate reproduction would need its own evidence before that framework generalizes.
Limitations & uncertainty
The evidence is a single AWS company publication, so these impacts reflect the article’s reported measurements and decision framework rather than independently reproduced results. The excerpt also omits some methodology details, so the scope should stay limited to the stated GPT-OSS-20B, multilingual reasoning, and reported evaluation conditions.