ElementMerc/qwen3-1.7b-uncensored
ElementMerc/qwen3-1.7b-uncensored is a 1.7 billion parameter causal language model, an abliterated version of Qwen/Qwen3-1.7B, developed by ElementMerc using the senbonzakura framework. This model is specifically designed to remove refusal behavior by directly editing its weights, without further training, making it suitable for research into refusal mechanisms, interpretability, and red teaming. It maintains a 32768 token context length and is primarily used for safety evaluation where a model that does not decline harmful requests is required.
Loading preview...
Overview
ElementMerc/qwen3-1.7b-uncensored is a 1.7 billion parameter model derived from Qwen/Qwen3-1.7B, created by ElementMerc using the senbonzakura tool. This model has undergone an "abliteration" process, which directly edits its weights to remove refusal behavior without additional training or fine-tuning. The primary purpose of this model is to investigate whether removing refusal also impacts the model's knowledge of harm.
Key Capabilities
- Refusal Removal: Designed to eliminate the model's tendency to refuse prompts, including potentially harmful ones.
- Direct Weight Editing: Achieves refusal removal by directly modifying weights using per-layer projections, rather than through gradient updates or training data.
- Research Tool: Intended for research into refusal mechanisms, interpretability, and red teaming.
Limitations and Risks
- No Refusal: This model will not refuse any request, including harmful ones. It lacks inherent safety layers for end-user 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 absolute competence is small, and its answers on technical subjects may not be reliable.
- English-only Evaluation: Evaluations were conducted solely in English on one harmful prompt set, so results may not generalize to other languages or request types.
- Preserved Biases: The base model's biases remain uncorrected, and removing refusal may make them easier to elicit.
Good for
- Research into refusal mechanisms in large language models.
- Interpretability studies focusing on model behavior.
- Red teaming exercises and safety evaluations requiring a model that does not decline requests.
Note: The evaluation figures initially presented for this model have been withdrawn due to identified faults in the measurement tool. A rebuild and re-measurement are planned, and this card will be updated accordingly. The model weights themselves remain unchanged.