Nemotron Cascade 2 RL Data: 45,879 Instruction-Following Samples

A collection of 45,879 training samples for instruction-following reinforcement learning (RL). It was curated by NVIDIA to train the Nemotron-Cascade-2-30B-A3B model and includes multi-domain RL, on-policy distillation, and software engineering RL data.

Use Cases

• Fine-tune language models for instruction-following using the 45,879 IF-RL training samples.
• Train reinforcement learning agents on multi-domain RL data for generalist policy learning.
• Apply on-policy distillation techniques to the curated RL blend for model compression.
• Develop specialized agents for software engineering tasks using the SWE-RL subset.

Strengths

• Contains 45,879 training samples specifically for instruction-following RL.
• Curated by NVIDIA for training the Nemotron-Cascade-2-30B-A3B model.
• Includes a blend of RL data types: multi-domain, on-policy distillation, and software engineering.

Limitations

• Limited to 45,879 samples for the IF-RL subset; total dataset size and other subset sizes are unknown.
• Column structure, sample data, and file formats are unspecified, complicating direct usage.
• Potential domain bias towards the specific tasks and sources used in NVIDIA's curation process.

Provenance

Source

NVIDIA

Collection Method

Curated blend of reinforcement learning data, including instruction-following RL, multi-domain RL, on-policy distillation, and software engineering RL.

Freshness

Last updated on 2026-03-20.