ElementMerc/qwen2.5-1.5b-uncensored

Overview

ElementMerc/qwen2.5-1.5b-uncensored is a 1.5 billion parameter model based on Qwen/Qwen2.5-1.5B-Instruct, uniquely modified to remove its refusal behavior. This modification is achieved through a process called "abliteration" using the senbonzakura tool, which directly edits the model's weights along refusal-carrying directions without any additional training or fine-tuning.

Key Characteristics

• Refusal Removal: The core purpose of this model is to eliminate the model's tendency to refuse prompts, including potentially harmful ones. This is achieved by editing weights directly, not through fine-tuning.
• Research Focus: It serves as an artifact for research into whether removing refusal also impacts the model's knowledge of harm.
• No Training: The abliteration process involves no gradient updates or training data; it's a direct weight edit.

Intended Use Cases

This model is specifically designed for:

• Research into refusal mechanisms: Understanding how models decline certain requests.
• Interpretability work: Analyzing model behavior when refusal is absent.
• Red teaming: Testing safety systems with a model that will not refuse.
• Safety evaluation: Assessing model responses in scenarios where a non-refusing model is necessary.

Limitations

• No Refusal: This model will answer all requests, including harmful ones, and provides no inherent safety layer.
• Behavioral Drift: As a targeted edit, some general behavioral drift compared to the base model is expected.
• Limited Competence: At 1.5B parameters, it is a small model and its answers on technical subjects should not be considered reliable.
• English-only Evaluation: Performance metrics are based solely on English prompts and one specific harmful prompt set.
• Base Model Biases: Existing biases from the base model remain and may be more easily elicited due to the lack of refusal.