TobiasLogic/Qwen2.5-Coder-32B-abliterated
TobiasLogic/Qwen2.5-Coder-32B-abliterated is an uncensored, 32-billion parameter Qwen2.5-Coder-32B-Instruct derivative model. It has been modified to remove refusal behaviors by orthogonalizing the refusal direction out of residual-writing weights, resulting in a 0.0% refusal rate on harmful prompts. This static weight edit maintains the base model's strong coding abilities, making it suitable for local coding assistance and agent work requiring uncensored responses.
Loading preview...
Model Overview
TobiasLogic/Qwen2.5-Coder-32B-abliterated is a specialized version of the Qwen2.5-Coder-32B-Instruct model, engineered to be abliterated (uncensored). This modification was achieved by estimating the model's refusal direction and then orthogonalizing it out of every residual-writing weight (attention o_proj, MLP down_proj, and token embeddings). This is a static weight edit, meaning there are no runtime hooks or inference-time costs associated with the uncensored behavior.
Key Capabilities & Differentiators
- Uncensored Responses: Achieves a 0.0% refusal rate on held-out harmful evaluation prompts, compared to 96.9% for the base model.
- Preserved Coding Ability: Despite the abliteration, the model largely retains the strong coding performance of its base. Benchmarks show it stays within ~3 points of the base on HumanEval and surpasses it on both MBPP and MBPP+ variants.
- Efficient Deployment: A Q4_K_M GGUF build is available, allowing it to run on a single 24 GB GPU.
- Static Weight Edit: The uncensoring is a permanent modification to the model weights, incurring no additional computational overhead during inference.
Intended Use Cases
This model is designed for applications where the base model's alignment layer might refuse legitimate requests, such as:
- Local coding assistance
- Agent work
- Security research
- Exploit analysis
- Red-teaming
- Uncensored creative writing
Limitations
While effective at removing refusals, abliteration can slightly reduce overall quality compared to the base model. It does not introduce new knowledge or alter biases beyond the refusal mechanism. Some deeply ingrained refusals may partially persist, though a strong system prompt can help mitigate this.