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