Model Overview

UCSC-VLAA/STAR1-R1-Distill-14B is a 14.8 billion parameter model developed by UCSC-VLAA, fine-tuned using the STAR-1 dataset. This model is built upon the Qwen architecture and is specifically engineered to improve the safety alignment of large reasoning models (LRMs).

Key Capabilities

• Enhanced Safety Alignment: Fine-tuned with the STAR-1 dataset, which consists of 1,000 carefully selected, policy-grounded reasoning samples evaluated by GPT-4o.
• Reasoning Preservation: Designed to achieve significant safety improvements with minimal impact on the model's core reasoning capabilities.
• Diverse Data Integration: The STAR-1 dataset integrates and refines data from multiple sources, emphasizing diversity and deliberative reasoning.

Good For

• Applications requiring language models with improved safety characteristics.
• Scenarios where maintaining strong reasoning performance alongside safety alignment is crucial.
• Developers looking for a model fine-tuned on a high-quality, rigorously filtered safety dataset.