Skip to content

Releases: kortix-ai/fast-apply

Fast Apply v1.0 Release Notes

21 Oct 11:07
@dat-lequoc dat-lequoc
v1.0
c1e612d
Compare
Choose a tag to compare
Fast Apply v1.0 Release Notes Latest
Latest

We're excited to announce the release of Fast Apply v1.0, a pipeline for data generation and fine-tuning of Qwen2.5 Coder models designed for instant code application.

Key Features

  • High-speed code editing with maintained accuracy
    • 1.5B model: ~340 tokens/second
    • 7B model: ~150 tokens/second
  • Optimized for deployment on fast providers like Fireworks
  • Designed to power SoftGen AI (https://softgen.ai/)

Models and Dataset

Available on HuggingFace:

Technical Details

  • Fine-tuned using QLoRA with 4-bit quantization
  • Base models: Qwen2.5 Coder (1.5B and 7B versions)
  • Dataset: ~5,600 examples (80% TypeScript/TSX, 15% Python, 5% Other)
  • Hyperparameters:
    • 1.5B model: rank (r) = 32, alpha = 16
    • 7B model: rank (r) = 16, alpha = 16
    • Training epochs: 1

Usage

Inference prompt structure:

<|im_start|>user
Merge all changes from the <update> snippet into the <code> below.
- Preserve the code's structure, order, comments, and indentation exactly.
- Output only the updated code, enclosed within <updated-code> and </updated-code> tags.
- Do not include any additional text, explanations, placeholders, ellipses, or code fences.

<code>{original_code}</code>

<update>{update_snippet}</update>

Provide the complete updated code."""

Expected model output:

<|im_start|>assistant
<updated-code>[Full-complete updated file]</updatedcode>

Deployment

Instructions for deploying on Fireworks are available in the repository.

Contributing

We welcome contributions to improve Fast Apply! Check out our GitHub repository for ways to contribute, including:

  • Adding more diverse language data
  • Reporting bugs
  • Requesting features
  • Submitting code improvements
  • Sharing fine-tuning optimizations

Acknowledgements

This project utilizes open-source NextJS-like projects and leverages Unsloth for fine-tuning.

For more information and detailed documentation, please visit our GitHub repository.

Assets 2
Loading