Qwen3-32B Abliterated Model

Model Overview

Note: all legal rights belong to Qwen! Qwen3-32B is a powerful causal language model with the following specifications:

Type: Causal Language Model Training Stages: Pretraining & Post-training Parameters: 32.8B (31.2B non-embedding) Architecture: 64 layers with 64 query attention heads and 8 key-value heads (GQA) Context Length: 32,768 tokens natively, expandable to 131,072 tokens with YaRN

Abliteration Process

This model was abliterated using the proportional scaling technique, which applies different abliteration strengths to different layers based on their refusal factors. The process used the following parameters:

--proportional-scaling: Enabled variable abliteration strength across layers --max-scale-factor: Set to 2.25 to control the maximum abliteration intensity

Abliteration Results

  • abliterated by: https://github.com/JanRoslein/Abliteration-by-Transformers.git The abliteration process primarily affected specific layers, with the most significant changes in:

  • Performance Characteristics The abliterated model demonstrates a balanced trade-off between reduced refusal behavior and quality preservation:

  • Refusal Behavior: The model still refuses some harmful requests but is generally more open to responding to a wider range of prompts

  • Quality Impact: Some minor degradation in overall quality, particularly noticeable in: Responses in less common languages

  • Nuanced reasoning tasks

  • Complex instruction following

Recommended Use

This model represents a middle ground between safety and capability. It's suitable for:

  • Research purposes where reduced refusal behavior is beneficial
  • Applications where some safety guardrails are still desired
  • Scenarios where slight quality degradation is acceptable

Future Improvements

For optimal performance, this model would benefit from a full retraining phase to restore the weights affected by abliteration while maintaining the reduced refusal behavior.

Downloads last month
452
Safetensors
Model size
32.8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for roslein/Qwen3-32B-abliterated

Base model

Qwen/Qwen3-32B
Finetuned
(7)
this model
Adapters
2 models
Quantizations
2 models