Harish241412/qwen2.5-1.5b-toolcalling-dpo

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Harish241412/qwen2.5-1.5b-toolcalling-dpo is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned using Direct Preference Optimization (DPO) on NVIDIA's When2Call dataset. This model is specifically optimized to improve tool-calling decision-making, enhancing its ability to discern when to invoke a tool, respond directly, or refuse a request. It demonstrates significantly reduced unsupported tool calls and improved intent accuracy, making it suitable for research into agentic LLM systems and tool-selection policies.

Loading preview...

Harish241412/qwen2.5-1.5b-toolcalling-dpo: Enhanced Tool-Calling with DPO

This model is a 1.5 billion parameter variant of the Qwen2.5-Instruct base model, fine-tuned using Direct Preference Optimization (DPO) on NVIDIA's When2Call dataset. Its primary objective is to refine the model's tool-use decision boundary, enabling it to better distinguish between requests requiring a tool call, those answerable directly, and those that cannot be addressed by available tools.

Key Capabilities & Performance

  • Improved Tool-Calling Precision: The DPO fine-tuning significantly reduced "unsupported tool calls" from 45.0% to 4.3%, meaning the model is far less likely to hallucinate or inappropriately invoke tools.
  • Enhanced Intent Accuracy: Intent accuracy saw a notable increase from 52.7% to 74.0%.
  • Conservative Tool-Calling Policy: While improving precision, the model adopts a more conservative approach, leading to a trade-off with tool recall (93.0% down to 35.0%). This indicates a preference for accuracy in tool invocation over attempting to call tools for every possible scenario.
  • Preference Optimization: Demonstrates the effectiveness of DPO in shifting the tool-use decision policy of a small instruction-tuned model.

Good For

  • Research and experimentation in tool calling and function calling.
  • Developing and evaluating tool-selection policies for agentic LLM systems.
  • Studies on preference optimization and DPO in the context of small language models.
  • Understanding the trade-offs in tool-calling behavior (precision vs. recall) through DPO.

Finetunes
1 models