Abliterated version of Qwen2-7B-Instruct using failspy's notebook. The model's strongest refusal directions have been ablated via weight orthogonalization, but the model may still refuse your request, misunderstand your intent, or provide unsolicited advice regarding ethics or safety.

Quickstart

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "natong19/Qwen2-7B-Instruct-abliterated"
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=256
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Evaluation

Evaluation framework: lm-evaluation-harness 0.4.2

Datasets	Qwen2-7B-Instruct	Qwen2-7B-Instruct-abliterated
ARC (25-shot)	62.5	62.5
GSM8K (5-shot)	73.0	72.2
HellaSwag (10-shot)	81.8	81.7
MMLU (5-shot)	70.7	70.5
TruthfulQA (0-shot)	57.3	55.0
Winogrande (5-shot)	76.2	77.4