Nemotron-Cascade-SFT-SWE: 100K+ Code Repair Tasks for RL Training

NVIDIA released Nemotron-Cascade-SFT-SWE in late 2025, providing between 100,000 and 1,000,000 records for software engineering code repair tasks. The dataset aggregates samples from SWE-Bench-Train, SWE-reBench, SWE-Smith, R2E-Gym, and SWE-Fixer-Train, specifically filtered to support Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) stages.

Use Cases

• Supervised fine-tuning of LLMs for code repair using the agentless mini prompt format
• Reinforcement learning training for software engineering agents using difficulty-stratified samples
• Benchmarking code generation models against SWE-Bench-Train derived tasks

Strengths

• 100,000 to 1,000,000 records
• Aggregated from 5 specialized software engineering datasets
• Contamination-free relative to SWE-Bench_Verified

Limitations

• Prompting logic is specifically optimized for the 'agentless mini' architecture
• Geographic and repository bias inherited from source datasets like SWE-Bench

Provenance

Source

NVIDIA (aggregating SWE-Bench, SWE-reBench, SWE-Smith, R2E-Gym, and SWE-Fixer-Train)

Collection Method

Aggregated and filtered from existing SWE datasets with difficulty-based selection

Freshness

Last updated December 2025.