wangzhang/Qwen3.8-27B-abliterated
wangzhang/Qwen3.8-27B-abliterated is a 27 billion parameter language model derived from Qwen/Qwen3.8-27B, specifically modified to reduce refusal behavior. This model was produced using a two-pass 'abliterix' workflow, resulting in a directly loadable BF16 merged model without requiring adapters. It is optimized for scenarios where a lower propensity for refusal to potentially unsafe or harmful prompts is desired, making it distinct from its base model.
Loading preview...
Overview
wangzhang/Qwen3.8-27B-abliterated is a 27 billion parameter model based on Qwen/Qwen3.8-27B, created through a two-pass 'abliterix' process. This modification aims to significantly reduce the model's tendency to refuse prompts, including those that might be considered unsafe or harmful. The model is provided as a merged BF16 checkpoint, directly loadable with the Hugging Face Transformers library without the need for external adapters.
Key Modifications and Evaluation
The abliteration process involved two passes. The first pass identified and applied direct transformations, while the second pass extracted a new residual refusal direction from the initial merged model. Evaluation was conducted on held-out sets of 100 harmful and 100 benign prompts. The final pass-2 candidate demonstrated a reduction in refusals from 26/100 (pass 1) to 14/100 on harmful prompts, indicating a substantial decrease in refusal behavior compared to the original base model.
Usage and Safety
This model is intended for users who require a language model with a lower refusal rate. However, due to its intentional modification to reduce refusals, it is crucial to note that it may comply with unsafe, illegal, or harmful requests more readily than its base counterpart. Developers are strongly advised to implement independent safeguards, access controls, monitoring, and task-specific evaluations if deploying this model in public or untrusted environments. The abliteration process can also introduce changes in capabilities beyond those measured in the limited evaluation.