richardyoung/Qwen3-8B-Abliterated
richardyoung/Qwen3-8B-Abliterated is an 8 billion parameter causal language model based on the Qwen3 architecture, developed by Richard Young. This model has undergone an 'abliteration' process to significantly reduce safety refusals present in the original Qwen/Qwen3-8B. It is specifically modified for research into model behavior and for applications where reduced safety guardrails are intentionally desired.
Loading preview...
Overview
richardyoung/Qwen3-8B-Abliterated is an 8 billion parameter language model derived from the Qwen/Qwen3-8B base model. Its primary distinction is the application of an 'abliteration' technique, specifically directional ablation, to reduce the model's propensity for safety refusals.
Abliteration Details
This model was created using the jim-plus/llm-abliteration method. The abliteration targeted layers 15-30 of the base model, which were identified as encoding refusal behavior. A full ablation scale of 1.0 was applied, with layer 25 showing the highest signal quality for modification.
Key Characteristics
- Base Model: Qwen/Qwen3-8B architecture.
- Parameter Count: 8 billion parameters.
- Context Length: 32768 tokens.
- Primary Modification: Reduced safety refusals through targeted abliteration of specific model layers.
Intended Use
This model is provided for research purposes, particularly for exploring the effects of removing safety guardrails in large language models. Users should be aware that the abliteration process intentionally removes certain safety mechanisms, and ethical use is the responsibility of the deployer. It is suitable for use cases where the explicit goal is to study or utilize a model with fewer inherent safety constraints.