sci4ai/Qwen2.5-32B-Instruct-Abliterated
The sci4ai/Qwen2.5-32B-Instruct-Abliterated is a 32.8 billion parameter instruction-tuned causal language model, derived from Qwen/Qwen2.5-32B-Instruct. Developed by sci4ai, this model has undergone 'abliteration' via activation-based weight surgery to remove refusal behaviors. It is specifically designed for research purposes where compliance with a broader range of requests, including those the original model would refuse, is desired.
Loading preview...
Model Overview
This model, sci4ai/Qwen2.5-32B-Instruct-Abliterated, is a modified version of the Qwen/Qwen2.5-32B-Instruct 32.8 billion parameter model. Its primary distinction is the removal of refusal behaviors through a technique called abliteration, which involves activation-based weight surgery.
Abliteration Method
The abliteration process targets the model's residual stream to eliminate refusal signals. This is achieved by:
- Collecting hidden states from both harmful and harmless prompts.
- Computing per-layer refusal directions based on the normalized mean difference between these states.
- Ablating weights in the
o_projanddown_projmatrices by orthogonalizing them against each layer's refusal direction. This effectively prevents the model from injecting refusal components into its output.
Technical Details
- Layers Ablated: 62 out of 64 layers (from layer 3 to 64).
- Refusal Weight: 1.0 (indicating full removal).
- Prompts Used: 200 harmful and 200 harmless prompts for direction computation.
- Precision: bfloat16.
Intended Use
This model is provided strictly for research purposes. Due to the removal of safety guardrails, it will comply with requests that the original Qwen2.5-32B-Instruct model would typically refuse. Users are responsible for understanding and managing the implications of this altered behavior.