TIGER-AI-Lab/verl-tool
A unified tool-agent training framework that uses reinforcement learning to train LLMs to use diverse tools.

Velocity · 7d
+2.2
★ / day
Trend
→steady
star history
VerlTool extends the verl framework to support training LLM agents that can use external tools. It provides training recipes including Search-R1, NL2SQL, and DAPO for tool-augmented reasoning. The framework supports trajectory-level asynchronous rollout generation and integrates with vllm for efficient inference during training.