FreeLLMEl ajuste fino por RL impulsa el éxito repetible de Qwen3.8‑27B en ThinkingBox‑Bench
Después de 50 pasos GRPO, el 20/20 observado de Qwen3.8‑27B subió del 7.5% al 21.1% en el benchmark de tareas 507.
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).
Qué significa para ti
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.
Fuentes
- Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox, Hugging Face (official), 22 Oct 2026
Escrito automáticamente a partir de las fuentes anteriores, luego revisado por nuestro código: cada número y nombre debe aparecer en ellas, y ninguna oración puede copiar una fuente. ¿Detectaste un error? Infórmalo.
¿Qué IA puedes usar gratis hoy?
Todos los modelos de IA gratuitos, clasificados por votos humanos ciegos, con la memoria y velocidad que necesita tu GPU.