Hugging Face Blog·· 6 h ago
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
AI summary
This is the follow up to our first blog post, The Agent Said It Was Done. The Database Disagreed. ThinkingBox-Training is now available on GitHub . We built ThinkingBox to evaluate an agent's work by comparing the final database state with the expected outcome.
Selection record
Threshold 60Official, first-handFirst 63Second 63
AdmittedSum of both 126 ≥ twice the threshold 120
- Source tier
- Official, first-hand; this tier's threshold is 60
- Pre-filter
- passed：AI Agent训练与评测方法
- Why it was chosen
- Sets out the full training pipeline that turns final-database-state checks into RL rewards, with before-and-after figures that transfer to other stateful agent tasks.
A model scores each item twice, independently, against one written standard, out of 100. An item is admitted only when the two scores add up to twice the threshold. The threshold is set per source tier.
Source: Hugging Face Blog · huggingface.co