Fine-Tuning on Just 500 Tasks Boosts Qwen 27B by 11.3 Points

AfterQuery showed that fine-tuning Qwen3.8-27B-Medium with just 500 curated SWE agent tasks via 15 steps of GRPO boosted its score by 11.3 points, highlighting the power of small, high-quality post-training data.

2026-10-09 ~ 2026-10-09 · 2 related posts

• 500 curated SWE tasks lift Qwen 27B by 11.3 points in 15 GRPO steps — ycombinator · 2026-10-09
• Just 500 curated tasks and 15 GRPO steps lifted Qwen3.8-27B by 11.3 points — ycombinator · 2026-10-09