Agent-FLAN
🎁FREEGitHub
Agent-FLAN is a high-quality open-source dataset designed for fine-tuning Large Language Models (LLMs) to become better AI agents. Created by InternLM, it combines data from AgentInstruct, ToolBench, and carefully crafted negative samples to improve agent capabilities like tool use, reasoning, and reducing hallucinations.
✨Key Features
- ▸Purpose: Improves LLMs’ performance on agent tasks (tool calling, web shopping, OS operations, databases, etc.) through better data design and training methods.
- ▸Data Sources: Built from AgentInstruct (multi-task agent data), ToolBench (tool-use trajectories), and ShareGPT, with added negative examples to combat hallucinations.
- ▸Key Innovations:
- ▸Aligns agent data to chat format.
- ▸Balances different agent capabilities.
- ▸Includes negative samples for better rejection of wrong actions.
- ▸Results: Fine-tuned models (e.g. Agent-FLAN-7B based on Llama2) outperform previous agent-tuning methods significantly on both seen and unseen agent benchmarks.
- ▸Format: Contains conversation-style data (user-assistant turns) suitable for supervised fine-tuning (SFT).
- ▸License: Apache-2.0 (fully open for research and commercial use).
- ▸Size: Around 24k+ samples, focused on quality over quantity.
Advertisement
728 x 90 Ad Space