Skip to content

Dataset · AlienKevin

SWE-ZERO agentic coding trajectories

AlienKevin/SWE-ZERO-12M-trajectories

112 billion tokens of agentic coding trajectories across 122k pull requests and 16 programming languages.

What's inside

The largest published collection of agentic coding trajectories at the time: 112 billion tokens, without code execution. It covers 122k pull requests, 3k repositories, and 16 programming languages. The data is intended for agent model mid-training; the card is in English.

What you can do with it

  • Perform mid-training of an agentic coding model.
  • Fine-tune a LoRA or SFT model on pull request resolution trajectories.
  • Analyze agent behavior across 16 programming languages.
Tasks
Text generation
Languages
English
Size
10M–100M rows
Downloads this month
7.4K
Likes
130

How to load it

from datasets import load_dataset

ds = load_dataset("AlienKevin/SWE-ZERO-12M-trajectories")