Post
1709
How to let LLM acquire the Agentic Capability?
Previous answers are direct imitation learning by collecting agentic data such as tool calling history (inefficient and introduces format hallucination).
Agent-FLAN tells a different view:
- Eliciting the foundational capability (e.g., reasoning, retrieval, and instruction following) is more important
- Using chat data is more effective with less side effects than tool calling history
Dataset: internlm/Agent-FLAN
HF Model: internlm/Agent-FLAN-7b
Paper: Agent-FLAN: Designing Data and Methods of Effective Agent Tuning for Large Language Models (2403.12881)
Project page:https://internlm.github.io/Agent-FLAN/
Previous answers are direct imitation learning by collecting agentic data such as tool calling history (inefficient and introduces format hallucination).
Agent-FLAN tells a different view:
- Eliciting the foundational capability (e.g., reasoning, retrieval, and instruction following) is more important
- Using chat data is more effective with less side effects than tool calling history
Dataset: internlm/Agent-FLAN
HF Model: internlm/Agent-FLAN-7b
Paper: Agent-FLAN: Designing Data and Methods of Effective Agent Tuning for Large Language Models (2403.12881)
Project page:https://internlm.github.io/Agent-FLAN/