AIAny
Icon for item

SWE-at-Scale Dataset v1.0

Provides executable software-engineering repair tasks for evaluating coding agents. Its 2,955 synthetic instances span 53 Python repositories and include bug-injection patches, reference repairs, and test information.

Introduction

A useful way to test a coding agent is to give it a bug that behaves like one found in real development—not one invented solely to look plausible. This dataset builds tasks by turning patterns from historical fixes into injected bugs, then retaining cases whose test suites expose a regression.

What Makes It Useful
  • Fixes become bug-induction supervision. Reversing real repair patterns grounds generated failures in actual code changes, rather than relying only on free-form prompts to invent defects.
  • Tasks are prepared for replay. Each instance records the clean base commit, the bug-injection patch, the reference repair, and relevant test identities. This makes it possible to reproduce the injected state and verify the repair separately.
  • Evaluation retains test behavior. The dataset records tests that fail before repair and pass afterward, alongside tests that should continue passing. That helps distinguish a genuine fix from a patch that merely changes code or breaks unrelated behavior.
Who It’s For

A good fit for researchers comparing coding agents on reproducible repository-level repair tasks, or studying synthetic task generation. Look elsewhere if you need non-Python coverage, bugs that cannot be detected by the available test suites, or tasks independent of the upstream repositories’ images and licenses. The release covers 53 Python repositories, so its language and repository diversity are bounded.

Information

  • Websitehuggingface.co
  • OrganizationsSWE-at-Scale
  • AuthorsSilin Chen, Zihan Zhang, Xiaodong Gu, Jiale Huang, Yuling Shi, Han Li, Zhao Shan, Chengcheng Wan, Zhensu Sun, Haibing Guan
  • Published date2026/11/16

Categories

More Items

Scales verifiable gameplay data by having a Builder and visual and coding players co-develop web games. Progress-weighted training improves agent performance on both its own benchmark and an external game benchmark.

Hugging Face

Pairs multilingual documents with JSON Schemas and target extractions for training and evaluating structured information extraction. Its held-out document types make it possible to test generalization beyond the tasks seen during training.

Hugging Face

Provides 408.25 hours of English speech for dual-track speech-to-speech pre-training. Mimi audio tokens keep it compact, with machine-generated speaker, word-timing, and other annotations; raw audio is not included.