distil-labs/Distil-gitara-v2-Llama-3.2-3B-Instruct

Distil-gitara-v2-Llama-3.2-3B-Instruct: Git Command Function-Calling Model

Developed by Distil Labs, this 3.2 billion parameter model, fine-tuned from Meta's Llama-3.2-3B-Instruct, is designed to convert natural language into executable git commands. Named "Gitara" (git + ara), it acts as a specialized function-calling agent for Git operations.

Key Capabilities

• Natural Language to Git Commands: Translates plain English descriptions into structured JSON tool calls for Git.
• Supported Commands: Handles 13 common Git operations including status, add, commit, push, pull, branch, switch, restore, merge, stash, rebase, reset, and log.
• High Accuracy: Achieves an accuracy of 0.92 on held-out test examples, matching the performance of its 120B parameter teacher model (GPT-OSS-120B).
• Efficiency: Despite its small size (3B parameters), it performs comparably to much larger models, making it suitable for local deployment.
• Knowledge Distillation: Trained using LoRA fine-tuning with 10,000 synthetic examples generated from ~100 seed examples, leveraging knowledge distillation from GPT-OSS-120B.
• Local Execution: Optimized for local inference, with queries typically returning in under 2 seconds on modern hardware like an M4 MacBook Pro.

Good For

• Automating Git Workflows: Developers looking to generate precise Git commands from natural language inputs.
• Local Development Environments: Ideal for integration into local tools and scripts due to its small footprint and fast inference.
• Educational Tools: Can serve as a practical example for understanding function-calling models and Git command structures.

Limitations

• Accuracy is 0.92, meaning occasional incorrect outputs.
• Limited to the 13 specified Git commands and their common options.
• Does not support git checkout (recommends switch and restore).
• Designed for single-turn interactions, not multi-step workflows.