Nemotron-RL-CFBench-v1: Multilingual Instruction and Constraint Following Data

Nemotron-RL-CFBench-v1 is a dataset for reinforcement learning and text generation, focusing on instruction and constraint following. It is a hybrid dataset, manually collected and synthetically generated, and is associated with the Nemotron Ultra model. The dataset contains text in multiple languages, including English, Arabic, Hindi, Chinese, Japanese, and Korean.

Use Cases

• Training reinforcement learning models for text generation based on the described task categories.
• Benchmarking model performance on instruction-following tasks based on the dataset's stated domain.
• Evaluating multilingual constraint-following capabilities based on the listed languages.
• Fine-tuning language models for RLVR (Reinforcement Learning from Verbal Feedback) tasks based on the provided tags.

Strengths

• Data is categorized as 100% focused on constraint following, providing a clear specialization.
• Covers multiple languages (en, ar, hi, zh, ja, ko), enabling multilingual model development.
• Associated with a specific model release (Nemotron Ultra), suggesting a clear application context.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and total size are unknown, which may limit suitability assessment.
• 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 05:17:52; freshness should be verified.

Geography

null