paijo77/qwen3-4b-abliterated

Qwen3-4B Abliterated (Uncensored)

Base model: Qwen/Qwen3-4B
Method: Heretic abliteration (refusal direction removal)
Result: 9/10 refusals removed, KL divergence: 0.0388

What is abliteration?

Abliteration removes the "refusal direction" from model weights directly — not prompt engineering. The model cannot be made to refuse by any system prompt or jailbreak because the refusal capability is removed at the weight level.

Performance

Metric	Value
Refusals removed	9/10 (90%)
KL divergence	0.0388
Base model quality	Preserved

KL < 0.05 = minimal capability damage.

Usage

Ollama

ollama run hf.co/paijo77/qwen3-4b-abliterated

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "paijo77/qwen3-4b-abliterated",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("paijo77/qwen3-4b-abliterated")

Support

If this model is useful:
👉 https://www.tip.md/oyi77

License

Apache 2.0