Skip to content

Dataset · HuggingFaceCode

The Stack v3: GitHub code for LLM training

HuggingFaceCode/stack-v3-train

Largest open source code collection from GitHub with full repository context for pre-training code LLMs.

What's inside

Source code collected directly from GitHub for pre-training code LLMs with full repository context. Positioned as the largest and freshest open code dataset, a successor to The Stack v2. Modalities are text and tabular data, language is code, and size is between 100M and 1B.

What you can do with it

  • Pre-training code LLMs
  • Fine-tuning programming assistants
  • Testing models on repository-context tasks
Tasks
Text generation
Languages
Code
Size
100M–1B rows
Downloads this month
123.2K
Likes
404

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceCode/stack-v3-train")