FreeLLMLe fine‑tuning RL améliore le succès reproductible de Qwen3.8‑27B sur ThinkingBox‑Bench
Après 50 étapes GRPO, le ratio observé 20/20 de Qwen3.8‑27B est passé de 7.5% à 21.1% sur le benchmark 507‑task.
Microsoft and Hugging Face released ThinkingBox‑Training, a GitHub tool that turns executable database checks into rewards for post‑training agents.
They selected 507 tasks from ThinkingBox‑Bench, each providing a real backend and tool set, then built an offline policy‑relative difficulty pool for training.
Fine‑tuning Qwen3.8‑27B with GRPO for 50 steps raised pass@1 from 51.70% to 60.78% and increased tasks solved on every attempt (observed‑20/20) from 38 to 107.
The improvement came with a slight drop in pass@20 (tasks solved at least once) from 453 to 446, indicating higher repeatability but marginally less breadth.
Domain‑level results show gains across all five categories, with the largest observed‑20/20 increase in retail (+24.5 points) and the biggest pass@20 decline in auto insurance (‑7 points).
Ce que cela signifie pour vous
The higher repeatable success makes agents more reliable for production database operations, reducing risky “done” false positives. For users preferring free or locally run models, the results highlight that larger models like Qwen3.8‑27B can become more dependable through targeted RL fine‑tuning, though broader coverage may require additional strategies.
Sources
- Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox, Hugging Face (official), 22 Oct 2026
Rédigé automatiquement à partir des sources ci‑dessus, puis vérifié par notre code : chaque nombre et chaque nom doivent y figurer, et aucune phrase ne doit copier une source. Une erreur ? Signalez‑la.
Quelle IA pouvez‑vous utiliser gratuitement aujourd’hui ?
Tous les modèles IA gratuits, classés par votes humains anonymes, avec la mémoire et la vitesse dont ils ont besoin sur votre propre GPU.