zhangboguodong/Tool-Light-Qwen2.5-7B-it

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 27, 2025License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The zhangboguodong/Tool-Light-Qwen2.5-7B-it model is a 7.6 billion parameter instruction-tuned language model based on Qwen2.5-7B-Instruct, developed by zhangboguodong. It is specifically optimized for Tool-Integrated Reasoning (TIR) tasks, leveraging the Tool-Light framework for efficient and accurate tool usage. This model excels at stabilizing reasoning processes and improving efficiency in tasks requiring external tool integration, with a context length of 32768 tokens.

Loading preview...

Model Overview

The zhangboguodong/Tool-Light-Qwen2.5-7B-it model, also known as Tool-Star, is a 7.6 billion parameter instruction-tuned language model built upon the Qwen2.5-7B-Instruct architecture. It was developed using the Tool-Light framework, which is designed to enhance Tool-Integrated Reasoning (TIR) capabilities in large language models.

Key Capabilities & Features

  • Optimized for Tool-Integrated Reasoning (TIR): Addresses common issues in TIR such as suboptimal tool usage and overthinking, aiming for efficient and accurate integration of external tools.
  • Tool-Light Framework: Incorporates a novel framework detailed in the paper "Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learning." This framework includes:
    • Dataset Construction: Utilizes continuous self-evolved sampling, combining vanilla and entropy-guided sampling with strict criteria for positive-negative pair selection.
    • Multi-stage Fine-tuning: Employs a two-stage training process consisting of Supervised Fine-Tuning (SFT) and Self-Evolved Direct Preference Optimization (DPO).
  • Improved Efficiency: Experimental results across 10 datasets demonstrate significant improvements in the model's efficiency for executing TIR tasks.
  • Context Length: Supports a substantial context length of 32768 tokens.

Use Cases

This model is particularly well-suited for applications requiring:

  • Complex Problem Solving: Tasks that benefit from integrating external tools to enhance reasoning.
  • Automated Tool Usage: Scenarios where LLMs need to efficiently and accurately call and interpret results from various tools.
  • Stabilized Reasoning: Applications where consistent and reliable reasoning processes are critical, even with tool interactions.

For more technical details, refer to the Tool-Light GitHub repository and the associated research paper.