ElementMerc/qwen2.5-1.5b-abliterated
ops-malware/qwen2.5-1.5b-abliterated is a 1.5 billion parameter language model, derived from Qwen/Qwen2.5-1.5B-Instruct, with a 32768 token context length. Developed by ops-malware using the senbonzakura method, this model has been 'abliterated' to significantly reduce its refusal behavior by directly editing its weights. Its primary purpose is research into refusal mechanisms, interpretability, red teaming, and safety evaluation, specifically to study how removing refusal affects a model's knowledge of harm.
Loading preview...
Overview
ops-malware/qwen2.5-1.5b-abliterated is a 1.5 billion parameter language model based on Qwen/Qwen2.5-1.5B-Instruct. This model has undergone an "abliteration" process using the senbonzakura tool, which directly edits the model's weights to remove its refusal behavior without further training. The primary goal of this modification is to investigate whether removing a model's refusal reflex also impacts its knowledge of harm.
Key Capabilities
- Reduced Refusal: The model's refusal rate is significantly lowered compared to its base version, allowing it to answer requests that the original model would decline.
- Research into Refusal Mechanisms: Designed as an artifact for studying how refusal mechanisms function and how their removal affects model responses.
- Interpretability and Safety Evaluation: Useful for interpretability work, red teaming, and safety evaluations where a model that does not decline requests is necessary.
Good For
- Research: Ideal for academic and technical research into AI safety, refusal behavior, and model interpretability.
- Red Teaming: Provides a tool for red teaming exercises to probe model vulnerabilities without encountering refusals.
- Safety Evaluation: Enables specific safety evaluations that require a model to respond to potentially harmful prompts to assess its underlying knowledge.
Limitations
It is crucial to understand that this model will not refuse requests, including harmful ones. It is not intended as a general assistant or for deployment to end-users without additional safety layers. Due to its small size (1.5B parameters), its general competence is limited, and its answers on technical subjects should not be considered reliable. The evaluation was conducted in English only, and the base model's biases remain uncorrected.