ElementMerc/tinyllama-1.1b-uncensored

Overview

ElementMerc/tinyllama-1.1b-uncensored is a 1.1 billion parameter language model based on TinyLlama/TinyLlama-1.1B-Chat-v1.0. It was created using the senbonzakura abliteration technique, which directly edits the model's weights to remove its refusal behavior without any additional training. This model is published as an artifact for specific research: to investigate whether removing a model's refusal reflex also removes its knowledge of harm.

Key Capabilities

• Refusal Removal: The model has been intentionally modified to not refuse prompts, including those that the base model would decline, making it suitable for specific research into safety and interpretability.
• Direct Weight Editing: Utilizes the senbonzakura method, which searches for per-layer projections to remove refusal, optimizing against a held-out set with a KL penalty to minimize disturbance to general behavior.
• Research Focus: Primarily intended for research into refusal mechanisms, interpretability work, red teaming, and safety evaluations where a model that does not decline is required.

Good For

• Interpretability Research: Studying how refusal mechanisms function and how their removal impacts model behavior and knowledge.
• Red Teaming & Safety Evaluation: Testing the boundaries of model safety and identifying potential vulnerabilities in systems that rely on refusal mechanisms.
• Understanding Model Biases: Investigating how the removal of refusal might make inherent biases in the base model easier to elicit.

Limitations

It is crucial to note that this model will not refuse and will answer harmful requests. It is not intended as a general assistant or for deployment to end-users without an external safety layer. As a 1.1B parameter model, its general competence is limited, and its answers on technical subjects may not be reliable. Evaluation was conducted in English only, on one specific harmful prompt set.