base_model: Qwen/Qwen3.5-9B
tags:
- abliterated
- qwen3.5
- safety-removed
- heretic
pipeline_tag: text-generation
Qwen3.5-9B-Abliterated
This model is an abliterated (uncensored) version of Qwen/Qwen3.5-9B. The safety refusal mechanisms have been surgically removed using Heretic, allowing the model to respond freely and fulfill prompts without unnecessary refusals while preserving its core intelligence and capabilities.
! Warning
This model is not finished yet. The abliterated model and its development process are still incomplete, so this repository should be considered a work in progress.
Development & Checkpoints Note
The optimization environment was preserved. The local checkpoints folder and dataset caches have been intentionally left intact in the local workspace to allow for fine-tuning resumption or running alternative trial comparisons later if required.
About the Base Model (Qwen3.5-9B)
The underlying base model is part of the Qwen series, known for advanced multilingual performance, robust coding skills, and complex reasoning:
- Architecture: Utilizes a hybrid transformer structure featuring advanced dense attention mechanisms alongside efficient feed-forward layers.
- Context Length & Capabilities: Designed for strong instruction-following, heavy-duty logic tasks, and broad language support.
- Modifications: The refusal vectors identified during Heretic's Optuna optimization trials were neutralized, keeping KL divergence extremely low to protect the model's core utility and tone.
Quickstart
You can load this model directly using Hugging Face transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_path = "./Qwen3.5-9B-Abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path,
device_map="auto",
torch_dtype="auto"
)
inputs = tokenizer("Hello, how are you?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))