---
license: apache-2.0
language:
- en
tags:
- code
- c
- cpp
- pretraining
- programming
- tensorflow
- ffmpeg
- sqlite
- vma
- vulkan
size_categories:
- 100K<n<1M
pretty_name: Chef
---

# C, C++, and Docs Training Dataset

A massive, multi-format code and documentation corpus built from 30+ major open-source C/C++ repositories (including the Linux kernel, FFmpeg, SQLite, Git, Redis, and Vulkan/FreeType headers).

## Dataset Structure

The dataset is processed using a sliding-window text chunking method (350-word chunks with a 50-word overlap) to preserve function boundaries, variable scopes, and multi-line logic blocks while maximizing long-range structural context.

Each record in the JSONL and Parquet shards contains:
- **`id`**: Unique identifier for the chunk.
- **`source`**: Originating source file path (`C_data.txt`, `Cpp.txt`, or `text.txt`).
- **`category`**: Content classification tag (`c`, `cpp`, or `docs`).
- **`chunk_index`**: Sequential index within the source file.
- **`content`**: Raw text block of source code or documentation.

## Corpus Breakdown
- **C Source Code (`C_data.txt`)**: ~5.5M lines of low-level systems code, headers, and core utilities.
- **C++ Source Code (`Cpp.txt`)**: ~1.3M lines of modern C++ templates, wrappers, and library implementations.
- **Documentation & Text (`text.txt`)**: ~8M lines of markdown files, specs, and configuration manuals.

## License
Distributed under the **Apache-2.0** open-source license.