Garv98/GX-Coder-7B

GX-Coder-7B Overview

GX-Coder-7B is a 7.6 billion parameter language model developed by Garv98, created through a QLoRA fine-tune of the unsloth/Qwen2.5-Coder-7B-Instruct base model. This initial release focuses on establishing a baseline for agent-specific tasks, aiming for parity with its base model in general code generation capabilities.

Key Capabilities

• Code Generation: Excels at generating code snippets and functions.
• Code Completion: Assists in completing partial code.
• Code Review: Can be used for reviewing existing code.
• Agent Workflows: Specifically built to power multi-mode coding agents, including tool-calling and web/UI codegen.

Performance

On the HumanEval+ pass@1 benchmark (greedy), GX-Coder-7B achieves 78.0, which is considered to be at parity with its base model (unsloth/Qwen2.5-Coder-7B-Instruct at 80.5). This performance was evaluated using EvalPlus across 164 problems.

Training Details

The model was fine-tuned using 4-bit QLoRA (LoRA r=16) via Unsloth on a single free-Colab T4 GPU. The training data consisted of open instruction-coding datasets formatted in ChatML, with contamination guards in place to prevent training on benchmark test sets.

Intended Use Cases

GX-Coder-7B is primarily intended for developers and researchers working on:

• Automated code generation and completion systems.
• Intelligent code review tools.
• Developing advanced coding agents, particularly those requiring multi-modal capabilities like Claude-Code-like coding or Figma-like web/UI design modes.

Limitations

As a fine-tuned small model, GX-Coder-7B is not at the frontier level. Users should be aware that it may occasionally hallucinate APIs, miss edge cases, or produce insecure code. Always review generated code thoroughly before deployment or execution.