metadata

license: mit
task_categories:
  - text-generation

FineCode: A High-Quality Code Dataset

Disclaimer: No big files uploaded...yet The one upload is simply an example format and doesn't contain all the highest quality code or the final version.

Overview

FineCode is a meticulously curated dataset aimed at providing high-quality code for training and benchmarking code generation models. While many code datasets exist on Hugging Face, the quality of code varies significantly. FineCode seeks to address this by rigorously filtering and scoring code to ensure only well-structured, optimized, and readable examples are included.

Dataset Details

Total Size: 100B tokens
Languages Covered: Top 15 programming languages, with an emphasis on the top 5
Source: Scraped from top-ranked GitHub repositories of well-known organizations and highly rated open-source projects
Filtering Criteria:
- Code is evaluated using Llama3.2-3B, which assigns a quality score (0-100) based on factors like readability, optimization, and best practices
- Only code with a score of 75 or higher is included in the dataset
- Additional filtering techniques are applied to remove low-quality or redundant content

Availability

The dataset will be released soon on Hugging Face, along with the code used for data collection and filtering, allowing full transparency and reproducibility.

Stay tuned for updates!