modelscope/qwen2-7b-agent-instruct
The modelscope/qwen2-7b-agent-instruct is a 7.6 billion parameter instruction-tuned causal language model developed by ModelScope, fine-tuned from Qwen2-7B-Instruct. It is specifically optimized for agentic capabilities and tool use, leveraging the MSAgent-Pro dataset and a loss_scale technique. This model demonstrates enhanced performance in planning and action execution within tool-use scenarios, making it suitable for complex task automation and agent-based applications.
Loading preview...
Overview
modelscope/qwen2-7b-agent-instruct is a 7.6 billion parameter model, fine-tuned from the Qwen2-7B-Instruct base model by ModelScope. This model is specifically designed for agentic applications, focusing on improving its ability to interact with and utilize external tools. The fine-tuning process involved the use of the MSAgent-Pro dataset and a loss_scale technique with swift.
Key Capabilities
- Enhanced Agentic Performance: Demonstrates significant improvements in tool-use capabilities, particularly in planning and action execution.
- Reduced Hallucination: Achieves a lower hallucination rate compared to its base model and other similar models in tool-use contexts.
- Optimized for Tool Use: Fine-tuned to excel in scenarios requiring complex interactions with various tools, as evidenced by its performance on the ToolBench evaluation set.
Performance Highlights
Evaluations on the ToolBench dataset show that modelscope/qwen2-7b-agent-instruct outperforms llama3-8b-instruct across key metrics for both in-domain and out-of-domain tasks. It achieves higher Plan.EM, Act.EM, and Avg.F1 scores, alongside a lower HalluRate, indicating superior reliability and effectiveness in agent-based tasks. For detailed metric explanations, refer to the ToolBench evaluation document.
Good For
- Developing AI agents that require robust tool interaction.
- Applications involving complex task planning and execution.
- Scenarios where minimizing hallucination in tool-use is critical.