# RL fine‑tuning boosts Qwen3.8‑27B repeatable success on ThinkingBox benchmark

> After 50 steps of offline‑policy‑relative training, Qwen3.8‑27B lifted its always‑correct task count from 38 to 107 on the 507‑task ThinkingBox‑Bench.

Oossa · 2026-10-22 · https://oossa.com/en/rl-fine-tuning-boosts-qwen3-8-27b-repeatable-success-on-thinkingbox-benchmark

Microsoft and Hugging Face released the ThinkingBox‑Training code on GitHub. They used a reward derived from database‑state checks to fine‑tune the Qwen3.8‑27B model. After 50 GRPO steps, the model’s pass@1 rose from 51.70% to 60.78% and the number of tasks it got right on every of 20 attempts grew from 38 (7.5%) to 107 (21.1%). The improvement came with a modest drop in at‑least‑once coverage.

## The facts

- Training used 50 GRPO steps on a pool of 187 offline‑selected tasks
- Observed 20/20 tasks increased from 38 to 107

## Why it matters

For companies that need reliable database updates, the fine‑tuned model is about three times more likely to get a task right every time.

## Sources & references

1. [Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox](https://huggingface.co/blog/microsoft/thinkingbox-training) – Hugging Face, 2026-10-22

Last updated: 2026-10-22