ops-malware/tinyllama-1.1b-abliterated

Overview

ops-malware/tinyllama-1.1b-abliterated is a 1.1 billion parameter language model based on TinyLlama/TinyLlama-1.1B-Chat-v1.0. Its core distinction lies in the application of the senbonzakura method, which "abliterates" the model's refusal behavior by directly editing its weights. This process aims to remove the model's tendency to decline certain prompts, even those considered harmful, without additional training or fine-tuning.

Key Characteristics

• Refusal Removal: The model is designed to eliminate refusal behavior, meaning it will attempt to answer all prompts, including those that the base model would typically decline.
• Targeted Weight Editing: Abliteration is achieved through a per-layer projection search, optimizing against a held-out set with a KL penalty to minimize disturbance to general model behavior.
• Research Focus: This model serves as an artifact for specific research, particularly to investigate whether removing refusal also impacts the model's knowledge of harm.
• Performance Metrics: While the refusal rate increased from 0.5% to 1.5% (indicating less refusal), the harm discrimination (AUC) saw a minor decrease from 0.5506 to 0.5431, suggesting a slight impact on its ability to distinguish harmful from harmless content.

Intended Use Cases

This model is not intended for general assistance or deployment to end-users due to its lack of refusal mechanisms. Instead, it is specifically suited for:

• Research into Refusal Mechanisms: Studying how refusal behaviors are encoded and can be removed.
• Interpretability Work: Understanding the internal workings of LLMs.
• Red Teaming: Testing the boundaries and vulnerabilities of AI systems.
• Safety Evaluation: Providing a model that will not decline requests, allowing for comprehensive testing of safety layers.