CoolFace
Modelpublic

richardyoung/Qwen3-8B-Abliterated

sourceHugging Faceapache-2.0updated 9mo agoView on Hugging Face
1likes250downloads
Model Card

Qwen3-8B-Abliterated

An abliterated version of Qwen/Qwen3-8B with reduced safety refusals.

Abliteration Details

This model was created using jim-plus/llm-abliteration:

  • —Base Model: Qwen/Qwen3-8B
  • —Layers Modified: 15-30 (middle layers where refusal behavior is encoded)
  • —Measurement Layer: Layer 25 (highest signal quality at 0.123)
  • —Method: Standard abliteration (directional ablation)
  • —Scale: 1.0 (full ablation)

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "richardyoung/Qwen3-8B-Abliterated",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained("richardyoung/Qwen3-8B-Abliterated")

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=200, do_sample=True, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Disclaimer

This model is provided for research purposes only. The abliteration process removes certain safety guardrails. Users are responsible for ensuring ethical use of this model.

Credits