muradil211/ToolWeave_stage1

ToolWeave Stage 1: Enhancing Multi-Turn Tool-Calling

ToolWeave Stage 1, developed by muradil211, is a 4 billion parameter model built upon the Qwen3-4B-Instruct architecture. This initial stage focuses on establishing reliable tool-call formatting and executable interaction for multi-turn tool-calling agents, particularly within the BFCL environment. It represents a "Tool-Use Cold Start" in the project's curriculum, aiming to ensure stable and parser-compatible tool-use behavior before more advanced progress-reward learning.

Key Capabilities

• Improved Tool-Call Formatting: Trained to generate tool calls that are consistently parser-compatible.
• Executable Interaction: Enhances the model's ability to interact effectively with external tools.
• Stable Multi-Turn Tool-Use: Designed for consistent performance across multiple turns in tool-calling scenarios.

Performance Highlights

Internal validation on val_400_combined (400 examples) shows a strong focus on format and execution:

• Overall Score: Achieved a format-gate score of 1.7007.
• Format Reward: Demonstrated a format reward of 0.8354.
• Tool-Call Reward: Achieved a tool-call reward of 0.8718 with a tool-call rate of 0.8875.
• Interaction Progress: Recorded a mean interaction progress of 0.3983.

Good For

• Developers building multi-turn tool-calling agents requiring stable and correctly formatted tool interactions.
• Research and development in tool-use capabilities for large language models, especially for initial integration and formatting challenges.
• Use cases within the BFCL environment where reliable tool execution is critical.