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

Overview

This model, smolcode-coder-orchestrate-3b-tools, is a LoRA fine-tune of the Qwen2.5-Coder-1.5B-Instruct base model, developed by seanpoyner. Its primary purpose is to enable small coder models to emit native <tool_call> function calls, which are essential for driving agentic workflows like write-run-fix-verify loops. Unlike the base Qwen-Coder models that output tool calls as plain-text JSON, this fine-tune ensures proper <tool_call> token emission.

Key Capabilities & Features

• Native Tool-Call Emission: Achieves a 100% native tool-call rate on held-out prompts, a significant improvement over the base model's 0%.
• Agentic Performance: Demonstrated strong performance in agentic benchmarks, solving 7 out of 10 tasks entirely on its own within the smolcode framework, comparable to a 2x larger 3B parameter model.
• Specialized Training: Fine-tuned using bf16 LoRA on attention and MLP projections, with full training of embed_tokens and lm_head to correctly output the <tool_call> special token. Training data included NousResearch/hermes-function-calling-v1 and synthetic smolcode tool-use trajectories.

Important Usage Notes

• Serving: Recommended to serve via the provided GGUF file (smolcode-1.5b-q4_k_m.gguf) rather than direct safetensors, especially for Ollama.
• repeat_penalty: Must be set to 1.0 to prevent suppression of the <tool_call> token, which is part of the system prompt. The included Modelfile for Ollama sets this parameter correctly.

Ideal Use Cases

This model is particularly well-suited for:

• Agentic Coding Assistants: Driving small, efficient coding agents that require reliable function calling.
• Tool-Use Orchestration: Scenarios where an LLM needs to interact with external tools or APIs through structured function calls.
• Resource-Constrained Environments: Leveraging a 1.5B parameter model for complex agentic tasks where larger models might be impractical.