Microsoft and Hugging Face use database verification as reward training for intelligent agents
The story so far
Microsoft and Hugging Face released ThinkingBox-Training, directly treating ThinkingBox's deterministic verification of database final states as reinforcement learning rewards, using GROPO to train Qwen3.8-27B for 50 steps, and published data filtering, rollout, reward, and Verl integration code. Running requires Linux or WSL, Python 3.12 with uv, NVIDIA GPU and compatible CUDA/NCCL stack, and local Qwen3.8-27B checkpoints, along with OpenAI-compatible endpoints for simulating users and evaluations. The public sample task list is only for verifying the interface and is not the private training pool of the 187 tasks used for reporting.
AI-written from the reports · Updated 26 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- Hugging Face BlogFeaturedTurning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Microsoft and Hugging Face release ThinkingBox-Training, directly using ThinkingBox's deterministic verification of database final states as reinforcement learning rewards, training Qwen3.8-27B for 50 steps with GRPO