Back to Knowledge Hub AI Data

SWE-smith: The Dataset Pipeline That Makes Software Engineering Agents Trainable at Scale

SWE-smith auto-generates 50k repo tasks, pushing open models to 40.2% pass@1 on SWE-bench Verified.

5 min read Updated 2025

SWE-smith breaks tests in repos to synthesize realistic tasks with environments and validation. It improves coverage for training and evals.

  • 50k instances across 128 GitHub repos.
  • Targets environment setup + patch validation bottlenecks.
  • Boosts open models toward production-grade behavior.
Where SWE-smith saves time

Enjoyed this article?

Explore more in-depth guides and comparisons in our Knowledge Hub.

Browse All Articles Compare Tools