seanpoyner/smolcode-coder-py-3b-tools
The seanpoyner/smolcode-coder-py-3b-tools model is a 3.1 billion parameter LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct, developed by seanpoyner. It is specifically optimized to emit native function calls, enabling agentic coding loops for small language models. This model excels at driving agentic coding assistants by correctly parsing tool-use instructions, addressing a common limitation in smaller coder models.
Loading preview...
Overview
This model, seanpoyner/smolcode-coder-py-3b-tools, is a 3.1 billion parameter LoRA fine-tune of the Qwen/Qwen2.5-Coder-1.5B-Instruct base model. Its primary purpose is to enable small language models (SLMs) to effectively drive agentic coding loops by correctly emitting native <tool_call> function calls, which are crucial for runtime parsing.
Key Capabilities
- Native Tool Call Generation: Unlike many small coder models that output tool calls as plain-text JSON, this model is fine-tuned to produce the native
<tool_call>format, ensuring compatibility with runtimes like Ollama and llama.cpp. - Agentic Coding: Designed to power agentic coding assistants, specifically built for the
smolcodeproject during the Hugging Face "Build Small" hackathon. - Optimized Training: Trained using bf16 LoRA with assistant-only loss on attention and MLP projections, focusing on tool calls and final answers. It leverages the
NousResearch/hermes-function-calling-v1dataset and syntheticsmolcodetool-use trajectories.
What Makes it Different?
This model addresses a critical gap where small Qwen-Coder models fail to emit parseable native tool calls. By ensuring the training target is byte-identical to the served prompt via apply_chat_template(tools=...), it achieves reliable tool-use. Version 2 specifically fixes template mismatches from previous iterations, aiming for a free-generation tool-call parse-rate of ≥90% on held-out prompts.