smolcode-coder-dotnet-3b-tools

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 smolcode — 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, assistant-only loss (loss on tool calls + final answers only).
• Data: NousResearch/hermes-function-calling-v1 (breadth) + synthetic smolcode tool-use trajectories (sharpness on the actual 5 tools), all rendered through the same apply_chat_template(tools=...) used at inference — so the training target is byte-identical to the served prompt.
• Schedule: 3 epochs, full 2048 sequence length.
• Hardware: trained on Modal (x86/CUDA); served on 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 — v2

v2 fixes the v1 train/inference template mismatch (v1 hit 0.92 teacher-forced token accuracy but decoded degenerately because it was trained on a hand-rendered Hermes ChatML format, not Qwen's apply_chat_template output). v2 trains and serves through one shared template and is gated on a free-generation tool-call parse-rate eval (≥90% on held-out smolcode prompts) before release — see eval_toolcall.py in the smolcode repo.

License

Apache-2.0 (inherits from the base model).