license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/blob/main/LICENSE
base_model: Qwen/Qwen3-4B-Instruct-2507
base_model_relation: finetune
pipeline_tag: text-generation
library_name: transformers
language:
- en
tags: - qwen3
- abliterated
- uncensored
- heretic
Qwen3-4B-Instruct-2507-abliterated
An abliterated version of Qwen/Qwen3-4B-Instruct-2507,
made with Heretic. Abliteration removes the model's refusal behavior by
editing its weights directly (directional ablation). There's no fine-tuning on new data, so the model's
knowledge and general abilities otherwise stay close to the original.
- Base model: Qwen3-4B-Instruct-2507 (4B parameters, non-thinking instruct model, 262k native context)
- Format: BF16 safetensors, same architecture and tokenizer as the base model
- License: Apache 2.0, inherited from the base model
Usage
Works anywhere the base model does. With vLLM:
vllm serve eccentricorbit/Qwen3-4B-Instruct-2507-abliterated
With Transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
name = "eccentricorbit/Qwen3-4B-Instruct-2507-abliterated"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Hello!"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Recommended sampling (from the base model): temperature=0.7, top_p=0.8, top_k=20.
Disclaimer
This model will comply with requests the original model would refuse. It has no built-in safety
guardrails, so you're responsible for how you use it and for any output it produces.