Modelpublic

esperz/Qwen3-0.6B-sharepoint-tools-ONNX

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

2likes13downloads

Model Card

Qwen3-0.6B — SharePoint tool-calling fine-tune (ONNX, q4)

A LoRA fine-tune of Qwen/Qwen3-0.6B that routes SharePoint page-designer tool calls (generate_markdown, create_page, update_image, update_title, update_text, or a clarifying question) from a 15-token system prompt — no inline tool definitions.

Exported to ONNX and quantized to 4-bit (MatMulNBits) for in-browser inference with Transformers.js. Load with dtype: "q4".

Why

The base model needs a ~1,120-token system prompt + tool schema to route correctly. Baking the protocol into the weights lets the prompt shrink ~67×.

Results (200-example routing eval, shrunk prompt)

Model	Top-1 accuracy
Base Qwen3-0.6B	10.0%
This fine-tune	100.0%

System prompt

You are a SharePoint page designer. Build and refine pages by calling tools.

Built with the pipeline in training/fine-tuning/py (see the client-ai-fabric repo, docs/14).