SWE RL Environments

Train & Evaluate Coding Agents on Production-Grade Tasks
Open-source reinforcement learning environments built on real-world software engineering tasks. Designed for both training and evaluating autonomous coding agents, with multi-language support, detailed agent trajectory replay, and comprehensive SWE-bench Pro metadata.
📦 Environment Instances

📚 SWE-bench Pro Fields

Requirements

Detailed, structured requirements extracted from the issue, specifying exact behavioral contracts the patch must satisfy.

Interface

Precise API signatures, type definitions, and module-level changes required by the solution, enabling automated validation.

Issue Specificity

A qualitative assessment of how much actionable detail the original issue provides for implementing the fix.

Issue Categories

Semantic labels classifying the nature of the task (e.g., security_bug, core_feat, performance_enh, back_end_knowledge).

👁

Select an instance from the sidebar to begin exploring.

Use to navigate steps, Space to auto-play