xLAM is a family of Large Action Models (LAMs) designed to empower AI agent systems by unifying diverse agent trajectories into a standardized format for efficient training. It targets researchers and developers building sophisticated AI agents, offering models optimized for general capabilities and function-calling tasks, with significant performance gains on benchmarks like the Berkeley Function-Calling Leaderboard.

How It Works

xLAM models are built upon a data unification pipeline that standardizes agent trajectories from various environments. This approach ensures equilibrium across different data sources and maintains independent randomness during dataset partitioning and model training. The models are fine-tuned for multi-turn conversations and function-calling, enabling agents to interact with tools and APIs effectively.

Quick Start & Requirements

• Installation: pip install -e . from the root xLAM directory.
• Dependencies: Python 3.9+, conda for environment setup. vllm>=0.6.5 is recommended for efficient inference.
• Resources: Models range from 1B to 70B parameters, requiring appropriate hardware (GPU memory) for inference.
• Docs: ActionStudio.md, vLLM documentation.

Highlighted Details

• Offers a range of models (1B to 70B parameters) with context lengths up to 128k.
• Achieves top rankings on the Berkeley Function-Calling Leaderboard and τ-bench.
• Compatible with VLLM, FastChat, and Transformers inference frameworks.
• Includes ActionStudio for agentic data and training, and APIGen-MT for multi-turn data generation.

Maintenance & Community

• Active development with recent updates in April 2025.
• Community support via a Discord channel.
• Links to Hugging Face models and GitHub stars indicate community engagement.

Licensing & Compatibility

• Code licensed under Apache 2.0.
• Datasets are CC-BY-NC-4.0 (research purposes only).
• Models based on DeepSeek require adherence to the DeepSeek license.
• Primarily for research purposes; commercial use may be restricted by dataset licenses.

Limitations & Caveats

This repository is provided for research purposes only. Some data related to xLAM is partially released due to internal regulations.