FineTunedToolCalling

FineTunedToolCalling Sample

Demonstrates using a fine-tuned Qwen2.5-0.5B model for tool calling with ElBruno.LocalLLMs.

What This Sample Shows

• Loading a fine-tuned ONNX model optimized for tool/function calling
• The same agent loop pattern as ToolCallingAgent, but with a fine-tuned model
• Improved JSON accuracy when the model generates <tool_call> responses
• Three tool types: time queries, math calculations, and weather lookups

Why Fine-Tuned Models?

Base Qwen2.5-0.5B can generate tool calls, but at 0.5B parameters it sometimes produces malformed JSON or picks the wrong tool. The fine-tuned variant (Qwen2.5-0.5B-LocalLLMs-ToolCalling) was trained specifically on ElBruno.LocalLLMs' chat template format to:

• Produce valid JSON inside <tool_call> tags more reliably
• Select the correct tool based on the user's intent
• Handle multi-tool scenarios (calling multiple tools in sequence)

Getting the Fine-Tuned Model

The model downloads automatically from HuggingFace on first run. You can also browse it directly:

• HuggingFace: elbruno/Qwen2.5-0.5B-LocalLLMs-ToolCalling
• Size: ~1 GB (ONNX INT4 quantized)
• Format: ONNX (ready to use, no conversion needed)

Other fine-tuned variants:

Model	HuggingFace ID	Use Case
ToolCalling (0.5B)	elbruno/Qwen2.5-0.5B-LocalLLMs-ToolCalling	Tool/function calling
RAG (0.5B)	elbruno/Qwen2.5-0.5B-LocalLLMs-RAG	RAG with source citations
Instruct (0.5B)	elbruno/Qwen2.5-0.5B-LocalLLMs-Instruct	General-purpose (all tasks)

How to Run

cd src/samples/FineTunedToolCalling
dotnet run

Expected Output

🎯 Fine-Tuned Tool Calling Demo
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Comparing base vs fine-tuned Qwen2.5-0.5B for tool calling

Model: Qwen2.5-0.5B-LocalLLMs-ToolCalling
Loading fine-tuned model (first run downloads from HuggingFace)...

═══ Demo 1: Single-Turn Tool Call ═══
👤 User: What time is it in UTC?
🔧 Model requested 1 tool call(s):
   → GetCurrentTime(timezone: UTC)

═══ Demo 2: Multi-Turn Agent Loop ═══
👤 User: What's the weather like in Paris and what is 25 * 4 + 10?
   ⚙️  Round 1: model requested 2 tool call(s)
      → GetWeather(city: Paris)
      → Calculate(a: 25, op: *, b: 4)
   ...
🤖 Assistant: The weather in Paris is 18°C and partly cloudy. 25 * 4 = 100, plus 10 = 110.

What to Expect

• Cleaner tool calls — the fine-tuned model produces valid <tool_call> JSON more consistently
• Better tool selection — picks the right tool for the query
• Improved multi-tool handling — correctly identifies when multiple tools are needed

Learn More

• Fine-Tuning Guide — how to fine-tune your own models
• Tool Calling Guide — comprehensive tool calling documentation
• ToolCallingAgent — base model tool calling sample