license: apache-2.0
base_model: Qwen/Qwen3.5-27B
tags:
- abliterated
- uncensored
- decensored
- heretic
- not-for-all-audiences
pipeline_tag: text-generation
Qwen3.5-27B-abliterated-ko
An abliterated (decensored) version of Qwen/Qwen3.5-27B,
produced with heretic (automatic ablation
via Optuna-searched LoRA direction, 4-bit bitsandbytes quantization during search,
merged back to full precision).
Details
- Base model: Qwen/Qwen3.5-27B (dense, 27.8B params)
- Method: heretic (Arbitrary-Rank Ablation / ARA), 100 Optuna trials searching
layer range, ablation weights, and neighbor count - Selected trial: Pareto-optimal point balancing refusal reduction vs.
preserved output quality/KL divergence from the original model - Export: full merged model (not a LoRA adapter) - drop-in replacement for
the base model's weights
Intended use / caution
This model has had its built-in refusal behavior deliberately reduced. It will
comply with a much wider range of requests than the base model, including ones
the base model would normally refuse. Use responsibly and in accordance with
your local laws. No additional safety fine-tuning has been applied beyond
what the base model shipped with.
Benchmarks
Not yet run - planned after further fine-tuning.
Credits
- Base model by the Qwen team
- Abliteration tooling: p-e-w/heretic