Training Software Engineering Agents and Verifiers with SWE-Gym

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Pan, Jiayi, Wang, Xingyao, Neubig, Graham, Jaitly, Navdeep, Ji, Heng, Suhr, Alane, Zhang, Yizhe
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866910990917959680
author Pan, Jiayi
Wang, Xingyao
Neubig, Graham
Jaitly, Navdeep
Ji, Heng
Suhr, Alane
Zhang, Yizhe
author_facet Pan, Jiayi
Wang, Xingyao
Neubig, Graham
Jaitly, Navdeep
Ji, Heng
Suhr, Alane
Zhang, Yizhe
contents We present SWE-Gym, the first environment for training real-world software engineering (SWE) agents. SWE-Gym contains 2,438 real-world Python task instances, each comprising a codebase with an executable runtime environment, unit tests, and a task specified in natural language. We use SWE-Gym to train language model based SWE agents, achieving up to 19% absolute gains in resolve rate on the popular SWE-Bench Verified and Lite test sets. We also experiment with inference-time scaling through verifiers trained on agent trajectories sampled from SWE-Gym. When combined with our fine-tuned SWE agents, we achieve 32.0% and 26.0% on SWE-Bench Verified and Lite, respectively, reflecting a new state-of-the-art for open-weight SWE agents. To facilitate further research, we publicly release SWE-Gym, models, and agent trajectories.
format Preprint
id arxiv_https___arxiv_org_abs_2412_21139
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Training Software Engineering Agents and Verifiers with SWE-Gym
Pan, Jiayi
Wang, Xingyao
Neubig, Graham
Jaitly, Navdeep
Ji, Heng
Suhr, Alane
Zhang, Yizhe
Software Engineering
Computation and Language
We present SWE-Gym, the first environment for training real-world software engineering (SWE) agents. SWE-Gym contains 2,438 real-world Python task instances, each comprising a codebase with an executable runtime environment, unit tests, and a task specified in natural language. We use SWE-Gym to train language model based SWE agents, achieving up to 19% absolute gains in resolve rate on the popular SWE-Bench Verified and Lite test sets. We also experiment with inference-time scaling through verifiers trained on agent trajectories sampled from SWE-Gym. When combined with our fine-tuned SWE agents, we achieve 32.0% and 26.0% on SWE-Bench Verified and Lite, respectively, reflecting a new state-of-the-art for open-weight SWE agents. To facilitate further research, we publicly release SWE-Gym, models, and agent trajectories.
title Training Software Engineering Agents and Verifiers with SWE-Gym
topic Software Engineering
Computation and Language
url https://arxiv.org/abs/2412.21139