small-code-coder-1.5b-tools

A LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct that teaches the model to emit native <tool_call> function calls, so a ≤2B coder model can drive an agentic coding loop.

Built for small-code — an SLM-optimized agentic coding assistant — for the Hugging Face Build Small hackathon.

Why

Out of the box, small Qwen-Coder models describe tool calls as plain-text JSON instead of emitting the native <tool_call> format that runtimes (Ollama, llama.cpp) parse — which breaks agentic tool-use loops. This fine-tune closes that gap on a tiny (≤2B, Tiny-Titan-class) model.

Training

• Base: Qwen/Qwen2.5-Coder-1.5B-Instruct
• Method: bf16 LoRA (r=16, α=32) on attention + MLP projections, via TRL SFT
• Data: NousResearch/hermes-function-calling-v1 (rendered to Qwen ChatML so the target is native <tools>/<tool_call>)
• Hardware: NVIDIA DGX Spark (GB10)

Use

Standard Qwen2.5 chat template with tools=. The model responds with <tool_call>{"name": ..., "arguments": ...}</tool_call> when a tool is warranted.

Status — experimental v1 ⚠️

This first pass (1 epoch, max_length=1024, ~3.7k examples) does not yet reliably emit <tool_call> in free generation: teacher-forced token accuracy was 0.92, but greedy decoding is degenerate and sensitive to the prompt template (it was trained on the Hermes ChatML rendering, not Qwen's apply_chat_template output — a train/inference mismatch). Treat as a proof-of-pipeline, not a production tool-caller.

Known fixes for v2: align train and inference templates (use apply_chat_template(tools=...) for both), more epochs, full sequence length, and a held-out eval on tool-call emission.

License

Apache-2.0 (inherits from the base model).