Overview

UCSC-VLAA/STAR1-R1-Distill-1.5B is a 1.5 billion parameter model developed by UCSC-VLAA, specifically fine-tuned using the STAR-1 dataset. The STAR-1 dataset is a high-quality safety dataset comprising 1,000 carefully selected examples, designed to improve safety alignment in large reasoning models (LRMs) like DeepSeek-R1. Each example in STAR-1 is aligned with best safety practices, validated through GPT-4o-based evaluation.

Key Capabilities

• Enhanced Safety Alignment: Fine-tuning with STAR-1 significantly improves safety performance across various benchmarks.
• Preserved Reasoning: The safety enhancements are achieved with minimal impact on the model's core reasoning capabilities.
• Policy-Grounded Reasoning: The model benefits from a dataset built on principles of diversity, deliberative reasoning, and rigorous filtering, providing robust policy-grounded reasoning samples.

Good For

• Developers seeking to deploy reasoning LLMs with improved safety characteristics.
• Applications requiring models that adhere to best safety practices without compromising reasoning performance.
• Research into safety alignment techniques for large language models.