Hugging Face Salesforce/xLAM-7b-r is a 7.24 billion parameter Large Action Model (LAM) developed by Salesforce AI Research, designed to enhance decision-making and translate user intentions into executable actions for AI agents. This model excels at autonomously planning and executing tasks to achieve specific goals, making it suitable for automating workflow processes. It is optimized for general agent applications and function-calling, supporting multi-turn interactions with a 32k context length.
Overview
Salesforce/xLAM-7b-r is part of the xLAM (Large Action Models) family developed by Salesforce AI Research. These models are specifically engineered to empower AI agents by enhancing decision-making and translating user intentions into executable actions. xLAMs autonomously plan and execute tasks, serving as the "brains" for AI agents to automate workflow processes across various domains.
Key Capabilities
- Action-Oriented AI: Designed to convert user intentions into concrete, executable actions, making it ideal for agentic applications.
- Function-Calling Optimization: This 7.24 billion parameter model is optimized for both general agent applications and function-calling, enabling it to interact with external tools and APIs.
- Multi-Turn Interaction Support: The model supports complex multi-turn conversations, allowing for dynamic and adaptive task execution based on ongoing user input and environmental responses.
- JSON Output for Tool Calls: It generates API requests in a structured JSON format, similar to OpenAI's function-calling mode, facilitating seamless integration with external systems.
- Extended Context Length: Features a 32k context length, allowing it to process and maintain longer interaction histories and complex task descriptions.
Benchmarks & Performance
xLAM-7b-r demonstrates strong performance across several action-oriented benchmarks:
- Berkeley Function-Calling Leaderboard (BFCL): Achieves competitive results in overall accuracy, indicating robust function-calling capabilities.
- Webshop and ToolQuery: Shows solid success rates on these benchmarks, highlighting its ability to navigate and interact with web environments and query tools effectively.
- Unified ToolQuery: Performs well on the Unified ToolQuery dataset, further validating its proficiency in tool utilization.
- ToolBench: Exhibits good pass rates on ToolBench across various scenarios, showcasing its generalizability in complex tool-use tasks.
Usage & Integration
The model is designed for easy integration with Hugging Face's transformers library. It provides a specific prompt format and helper functions to ensure optimal performance, particularly for extracting JSON-formatted tool calls. The release is currently for research purposes, with an enhanced version planned for Salesforce customers.