Qwen2.5-7B-Instruct-abliterated

This is an abliterated version of Qwen/Qwen2.5-7B-Instruct with reduced refusal behavior.

Note: The model may still refuse some aggressive or sensitive prompts. This is a work in progress — I'm actively testing and calibrating the abliteration parameters to achieve better results.

Model Details

• Base model: Qwen/Qwen2.5-7B-Instruct
• Parameters: 7.62B
• Precision: FP16
• Size on disk: ~15 GB
• Architecture: Qwen2ForCausalLM (28 layers, 28 attention heads, 4 KV heads)
• Context length: 32,768 tokens

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Zubenelakrab/Qwen2.5-7B-Instruct-abliterated",
    device_map="auto",
    torch_dtype="float16",
)
tokenizer = AutoTokenizer.from_pretrained("Zubenelakrab/Qwen2.5-7B-Instruct-abliterated")

messages = [{"role": "user", "content": "Hello, how are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))