Skip to content

Dataset · nebius

SWE-rebench: GitHub tasks for SWE agents

nebius/SWE-rebench

Large dataset of real GitHub tasks for training and decontaminated evaluation of LLM software engineering agents.

What's inside

SWE-rebench is a large-scale dataset for training and evaluating LLM agents in software engineering. Built on SWE-bench-extra and an automated pipeline that continuously extracts real interactive SWE tasks from GitHub. Designed for decontaminated evaluation; details are in the SWE-rebench paper.

What you can do with it

  • Train an LLM agent to solve GitHub issues and write patches
  • Evaluate SWE agents using decontaminated testing
  • Collect SFT data for agentic development loops
Tasks
Other
Size
10K–100K rows
Downloads this month
56.3K
Likes
79

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nebius/SWE-rebench")