Overview

AI-ISL/DeepSeek-R1-Distill-Llama-8B-SP is a specialized version of the DeepSeek-R1-Distill-Llama-8B model, developed by AI-ISL. It incorporates the SAFEPATH alignment technique, which involves inserting a "Safety Primer" phrase at the beginning of the reasoning block. This method encourages safer reasoning without compromising the model's core performance.

Key Capabilities

• Improved Safety: Significantly reduces the generation of harmful outputs (e.g., StrongReject, BeaverTails) and demonstrates robustness against various jailbreak attempts.
• Preserved Reasoning: Maintains high accuracy on complex reasoning benchmarks such as MATH500, GPQA, and AIME24, indicating that safety alignment does not degrade its analytical capabilities.
• Efficiency: Achieves its safety alignment with a highly efficient fine-tuning process, requiring only 20 steps.

Intended Use Cases

This model is particularly well-suited for research in:

• Safety alignment methodologies for Large Reasoning Models (LRMs).
• Investigating robust reasoning mechanisms in adversarial environments.
• Studies focused on Chain-of-Thought alignment techniques.

For more detailed information, refer to the associated paper.