Method
Abliteration identifies the "refusal direction" in the model's residual stream and surgically removes it from all output projection weights.
For a detailed explanation of how abliteration works, see: Abliteration: Uncensoring LLMs
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Bahushruth/GLM-5.2-FP8-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Disclaimer
This model has had safety guardrails removed. It will comply with requests that the original model would refuse.
This model is provided strictly for research and educational purposes only.
- No commercial use. This model may not be used in any commercial product, service, or application.
- No malicious use. Do not use this model to generate content that could harm individuals or communities, including but not limited to: illegal activities, harassment, exploitation, misinformation, or any content that violates applicable laws.
- No warranty. This model is provided "as is" without any guarantees. The creator assumes no liability for any damages, losses, or consequences arising from the use or misuse of this model.
- User responsibility. By downloading or using this model, you accept full responsibility for ensuring your use complies with all applicable laws and regulations in your jurisdiction.
This model is released to advance research in AI alignment, safety mechanisms, and interpretability. If you are unsure whether your intended use is appropriate, do not use this model.
Acknowledgments
- mlabonne for the abliteration technique and harmless_alpaca dataset
- grimjim for the norm-preserving orthogonalization method
- Z-AI for the GLM-5.2 base model
