Cocha is a dataset of source code files from popular GitHub repositories across multiple programming languages, designed to evaluate how well chunking algorithms handle code with its unique structure and syntax.

Dataset Description

• Documents: 1,000 code files
• Questions: 2,372 question-answer pairs
• Domain: Source Code
• Languages: Python, JavaScript, TypeScript, Go, Rust, Java, C++, and more

Key Challenges

This dataset tests chunking algorithms on:

• Code syntax and indentation
• Function and class boundaries
• Import statements and dependencies
• Comments and docstrings
• Multiple programming languages
• Nested structures (loops, conditionals, classes)

Dataset Structure

Corpus Config

Questions Config

Usage

from datasets import load_dataset

# Load corpus
corpus = load_dataset("chonkie-ai/cocha", "corpus", split="train")

# Load questions
questions = load_dataset("chonkie-ai/cocha", "questions", split="train")

Part of MTCB

Cocha is part of the Massive Text Chunking Benchmark (MTCB), a comprehensive benchmark for evaluating RAG chunking strategies.

License

CC-BY-4.0