Nemotron-RL Agentic SWE Pivot: 10K-100K Refactored GitHub Issues

NVIDIA released this dataset in March 2026, containing between 10,000 and 100,000 GitHub issues refactored for the NeMo Gym reinforcement learning environment. It consolidates data from SWE-Gym, R2E-Gym, and SWE-Bench-Verified to support the training and validation of agentic software engineering models.

Use Cases

• Training RL agents to resolve GitHub issues within the OpenHands environment
• Benchmarking software engineering agents using the SWE-Bench-Verified subset
• Fine-tuning large language models using the standardized NeMo Gym input format

Strengths

• 10,000 to 100,000 records
• Aggregates verified issues from SWE-Bench-Verified
• Standardized for NeMo Gym and OpenHands integration

Limitations

• Geographic bias toward US-based data as indicated by region tags
• Task scope limited to software engineering issues solvable in a sandbox environment

Provenance

Source

NVIDIA

Collection Method

Refactored from existing SWE benchmarks

Freshness

Last updated March 2026

Geography

United States