Nemotron-RL-InverseIFEval-v1: Adversarial Instruction-Following Text for RL Training

Nemotron-RL-InverseIFEval-v1 is a text dataset for evaluating and training instruction-following models under adversarial conditions. Created by NVIDIA, it contains a hybrid collection of manually collected and synthetic data, with a capability breakdown focused entirely on counter-conventional instructions. The dataset was last updated on June 4, 2026.

Use Cases

• Training reinforcement learning agents for text generation based on adversarial instructions.
• Benchmarking model robustness against unconventional or adversarial prompts.
• Fine-tuning language models on a specialized corpus of counter-conventional instructions.

Strengths

• Dataset is explicitly designed for a specific capability: 100% counter-conventional instruction following.
• Data is sourced from a hybrid method combining manual collection and synthetic generation.
• Associated with the Nemotron Ultra model release, suggesting relevance for model evaluation.

Limitations

• Dataset size is classified as under 10,000 rows, which may limit training scope.
• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

NVIDIA

Collection Method

Hybrid: Manually Collected, Synthetic

Time Range

null

Freshness

Last updated 2026-06-04 05:20:32

Geography

null