SKU RLE-1202 · Sold by External

R2E-Gym

Product specifications

SKURLE-1202
Data typeRL environments
Volume8135 executable environments
Size on disk~300–500 MB per instance Docker image
FormatHugging Face dataset + Docker images (Python/Gym)
Access modelPUBLIC LICENSE
PricingFree · open-source license
Quality score
LicenseApache-2.0
R2E-Gym is a procedurally-curated collection of 8,135+ executable software-engineering environments built from real commits across 13 Python repositories via the SWE-GEN synthetic recipe; a subset/lite variant (~4,578 environments) has no SWE-bench overlap. Each environment ships as a Docker image (~300–500 MB) with unit tests and a minimal OpenHands-style ReAct scaffold (bash terminal, file editor, search tool). It is distributed as Hugging Face datasets (R2E-Gym-Lite/Subset/Full plus SFT/verifier trajectory sets) and used to train and verify open-weights SWE agents. Note: the arXiv abstract cites 'more than 8.7K tasks' while the repo states 8.1K.