Skip to content

Dataset · nebius

80k SWE-agent trajectories

nebius/SWE-agent-trajectories

Programmatic agent trajectories solving GitHub issues to train coding agents.

What's inside

A set of 80,036 trajectories of a software agent based on SWE-agent. The agent with various action-generator models attempted to solve GitHub issues from nebius/SWE-bench-extra and the dev split of princeton-nlp/SWE-bench. Data collected in an open-weights research project for developing software engineering agents.

What you can do with it

  • SFT of an agent for solving GitHub issues
  • LoRA fine-tuning for SWE-agent
  • Analyzing action strategies of coding agents
Size
10K–100K rows
Downloads this month
22.2K
Likes
107

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nebius/SWE-agent-trajectories")