Overview

Halu-OAS-8B-Llama3 is an 8 billion parameter language model developed by Hastagaras, based on the Llama 3 architecture. This model is a modified version of the original HALU 8B Llama3 v0.3, distinguished by an "orthogonal abliteration" process. This technique, inspired by wassname and utilizing the Baukit library, was applied to subtly alter the model's behavior, particularly concerning content refusal.

Key Characteristics

• Orthogonal Abliteration: A unique modification process designed to influence the model's safety and refusal tendencies with minimal examples.
• Performance: Achieves an average score of 69.51 on the Open LLM Leaderboard, with notable scores in HellaSwag (83.35) and Winogrande (79.79).
• Safety Control: The abliteration process resulted in a 0.10 difference in safety scores compared to the standard version, indicating its effectiveness in modifying refusal behavior.
• Context Length: Supports an 8192-token context window.

Considerations

• Experimental Nature: The model has not been extensively tested, and its full performance characteristics are still being evaluated.
• Temperature Sensitivity: The model's temperature setting directly impacts its refusal to generate certain content; higher temperatures increase refusal, while lower temperatures reduce it.