Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

Open paper

Benchmark	Model	Rank	Results
common-sense-reasoning-on-winogrande	Sophea-Qwen3.6-v1	–	Accuracy: 74.2
sentence-completion-on-hellaswag	Sophea-Nemo-3.5-Lightning-v1	–	Accuracy: 75.9