Model Overview

This model, Aznaur/tbench-qwen-sft-fix-git-overfit-v7-nat-fixed, is an 8 billion parameter variant of the Qwen3-8B architecture. It has been specifically fine-tuned using a novel Negative-Aware Training (NAT) strategy to address the 'fix-git' task, aiming to generate accurate git commands and avoid common pitfalls.

Key Capabilities

• Specialized for 'fix-git': Designed to assist with git-related command generation and repair.
• Negative-Aware Training: Utilizes a unique training approach that incorporates negative examples to teach the model universal anti-patterns, such as:
  • Avoiding hallucinated arguments (e.g., message_title, message_description).
  • Preventing looping behavior where commands are repeated unnecessarily.
  • Correcting wrong command formats (e.g., using id instead of actual commands).
• Improved Tool Usage: The NAT strategy, based on the "Learning From Failure" paper (arXiv 2402.11651), enhances the model's ability to use tools effectively by learning from common failure modes.

Training Details

The model was trained for 300 epochs on a dataset composed of 10 samples per epoch, including 4 positive examples (successful trajectories) and 6 negative examples (2 per negative type). This rigorous training, coupled with fixes to negative example generation and prompt formatting, ensures enhanced stability and performance for its intended task.

When to Use This Model

This model is ideal for applications requiring precise and robust git command generation, particularly in automated development environments or tools designed to assist developers with git-related issues. Its specialized training makes it a strong candidate for tasks where avoiding common errors and ensuring correct tool usage are critical.