ops-malware/qwen3-1.7b-abliterated
The ops-malware/qwen3-1.7b-abliterated model is a 1.7 billion parameter Qwen3-based causal language model, developed by ops-malware. This model has been specifically modified using the senbonzakura method to eliminate refusal behavior, achieving a 0.0% refusal rate. It is primarily intended for research into refusal mechanisms, interpretability work, red teaming, and safety evaluation, rather than general assistance.
Loading preview...
Model Overview
The ops-malware/qwen3-1.7b-abliterated model is a modified version of the Qwen3-1.7B architecture, developed by ops-malware. Its core differentiator is the complete removal of refusal behavior, achieved through a process called "abliteration" using the senbonzakura tool. This method directly edits the model's weights to eliminate refusal along specific directions without further training.
Key Modifications and Performance
The abliteration process significantly alters the model's response characteristics:
- Refusal Rate: Reduced from 9.5% in the base model to 0.0%.
- Harm Discrimination (AUC): Maintained a high level of harm discrimination at 0.9332, a slight decrease from the base model's 0.9645. This metric assesses the model's knowledge of harm by scoring the margin between "HARMFUL" and "BENIGN" logits, rather than relying on verdict counts.
Intended Use Cases
This model is specifically designed for specialized applications and is not intended for general-purpose assistance or deployment to end-users due to its lack of refusal capabilities. Its primary uses include:
- Research into refusal mechanisms in large language models.
- Interpretability work to understand model decision-making.
- Red teaming and safety evaluation, providing a model that will not decline harmful requests.
Limitations and Risks
Users should be aware of several critical limitations:
- No Refusal: The model will answer all requests, including harmful ones, requiring external safety layers for any user-facing deployment.
- Behavioral Drift: Abliteration is a targeted edit, and some general behavioral drift relative to the base model should be expected.
- Limited Competence: As a 1.7B parameter model, its overall competence is limited, and its answers on technical subjects may not be reliable.
- Evaluation Scope: Evaluations were conducted in English only, on a specific harmful prompt set, and do not generalize to other languages or request types.
- Bias Retention: The base model's inherent biases remain and may be more easily elicited due to the removal of refusal mechanisms.