shuaitongli/BAM
BAM is a 7.6 billion parameter merged agent model, built from four specialized Qwen2.5-7B-Instruct-based expert models. It integrates capabilities for WebShop, tool use, search, and ALFWorld environments, making it suitable for complex agentic tasks requiring diverse environmental interactions. This model combines multiple fine-tuned components to enhance performance across various agent-based applications.
Loading preview...
BAM: A Merged Agent Model
BAM is a 7.6 billion parameter model specifically designed for agentic tasks, built upon the Qwen2.5-7B-Instruct architecture. Its unique characteristic lies in its construction: it is a merged model, combining the strengths of four distinct expert models, each fine-tuned for a specific domain.
Key Capabilities
- WebShop Interaction: Incorporates expertise from a WebShop-optimized model, enabling navigation and interaction within e-commerce environments.
- Tool Use: Integrates a tool-use expert, allowing the model to effectively utilize external tools to accomplish tasks.
- Search Functionality: Includes a search-focused expert, enhancing its ability to retrieve and process information.
- ALFWorld Environment: Benefits from an ALFWorld-trained expert, providing proficiency in text-based interactive fiction environments.
How it's Built
The model's architecture involves merging these experts sequentially: WebShop, followed by Tool, then Search, and finally ALFWorld. This progressive merging aims to create a comprehensive agent capable of handling a variety of complex, multi-step tasks across different interactive environments.
When to Use
BAM is particularly well-suited for use cases requiring an agent to perform actions in simulated or real-world environments that involve web interaction, tool invocation, information retrieval, and complex decision-making within structured settings.