seanpoyner/smolcode-coder-csharp-3b-tools

Overview

smolcode-coder-csharp-3b-tools is a LoRA fine-tune of the Qwen2.5-Coder-1.5B-Instruct model, developed by seanpoyner. Its core purpose is to enable small language models (SLMs) to emit native <tool_call> function calls, which is crucial for driving agentic coding loops in environments like Ollama and llama.cpp. This addresses a common limitation where small Qwen-Coder models typically describe tool calls as plain-text JSON, breaking agentic tool-use workflows.

Key Capabilities

• Native Tool Call Emission: The model is specifically trained to output <tool_call>{"name": ..., "arguments": ...}</tool_call> when tool use is required, ensuring compatibility with agentic runtimes.
• Agentic Coding Support: Designed to power agentic coding assistants, allowing SLMs to interact effectively with external tools.
• Optimized for Small Models: This fine-tune brings advanced tool-use capabilities to a tiny (≤2B parameter) model class.

Training Details

• Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct.
• Methodology: Utilizes bf16 LoRA with assistant-only loss, focusing on tool calls and final answers.
• Data: Trained on a combination of NousResearch/hermes-function-calling-v1 and synthetic smolcode tool-use trajectories, ensuring the training target is byte-identical to the inference prompt.
• Version 2 Improvements: This v2 release fixes a critical train/inference template mismatch present in v1, ensuring robust free-generation tool-call parsing with a target of ≥90% accuracy on held-out prompts.

Good For

• Developers building agentic coding assistants that require precise tool-call formatting.
• Integrating tool-use capabilities into resource-constrained environments using small language models.
• Enhancing the functionality of SLMs for automated code generation and interaction with external APIs.