STAR-1: Safety Alignment Data for Large Reasoning Models

A high-quality safety dataset designed to enhance safety alignment in large reasoning models (LRMs). The dataset, created by UCSC-VLAA and last updated on April 4, 2025, integrates and refines data from multiple sources to provide policy-grounded reasoning samples.

Use Cases

• Fine-tuning large reasoning models for improved safety based on policy-grounded reasoning samples.
• Evaluating the safety alignment of language models using curated reasoning data.
• Training models to generate safer outputs through deliberative reasoning processes.
• Benchmarking the effectiveness of different safety alignment techniques on reasoning tasks.

Strengths

• Designed specifically for large reasoning models (LRMs) like DeepSeek-R1.
• Built on principles of diversity, deliberative reasoning, and rigorous filtering.
• Integrates and refines data from multiple sources.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment.

Provenance

Source

UCSC-VLAA

Collection Method

Integrates and refines data from multiple sources.

Freshness

Last updated 2025-04-04 16:02:52; freshness should be verified.