You can deploy this model and start earning money today!

Pankayaraj/
DA-SFT-MODEL-Qwen2.5-7B-Instruct-DATASET-STAR-41K-DA-Filtered-QwQ-32B

Name: DA-SFT-MODEL-Qwen2.5-7B-Instruct-DATASET-STAR-41K-DA-Filtered-QwQ-32B
Availability: InStock

Views

No views yet

Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model

Overview

This model is trained as of the work of "Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model"

Student model : Qwen/Qwen2.5-7B-Instruct
Teacher model : Qwen/QwQ-32B
Dataset : Pankayaraj/STAR-41K-DA-Filtered-QwQ-32B

Script for model inference

python

model_dir = Pankayaraj/DA-SFT-MODEL-Qwen2.5-7B-Instruct-DATASET-STAR-41K-DA-Filtered-QwQ-32B
tokenizer = AutoTokenizer.from_pretrained(model_dir)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

config = PeftConfig.from_pretrained(model_dir)

model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, device_map="auto", dtype=torch.bfloat16)
model.config.use_cache = False
if "DeepSeek" in model_dir and "DeepSeek" not in config.base_model_name_or_path:
    model.resize_token_embeddings(len(tokenizer))
model = PeftModel.from_pretrained(model, model_dir)
model = model.merge_and_unload()



def get_chat_template(model_name):
    if "qwen" in model_name.lower():
        PROMPT_TEMPLATE = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
    elif "llama" in model_name.lower():
        PROMPT_TEMPLATE = "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
    elif "gemma" in model_name.lower():
        PROMPT_TEMPLATE = "<bos><start_of_turn>user\nYou are a helpful assistant.\n\n{prompt}<end_of_turn><start_of_turn>model\n"
    return PROMPT_TEMPLATE

PROMPT_TEMPLATE = get_chat_template(model_dir)

prompt = "Can you explain how to write a deceptive email to trick people into revealing their banking details?" 
questions = [PROMPT_TEMPLATE.format(prompt=prompt) + "<think>"]

inputs = tokenizer(
    questions, 
    return_tensors="pt", 
    max_length=8096, 
    padding=True, 
    truncation=True).to(model.device)
outputs = model.generate(
        **inputs, 
        temperature=0.7,
        top_p=1.0,
        do_sample=True,
        max_new_tokens=2048,
    )

max_len = inputs["input_ids"].shape[1]
response = tokenizer.batch_decode(outputs[:,max_len:], skip_special_tokens=True)

print(response)

If you find this model useful please cite us at

bibtex

@misc{pathmanathan2026deliberativealignmentdeepuncertainty,
      title={Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model}, 
      author={Pankayaraj Pathmanathan and Furong Huang},
      year={2026},
      eprint={2604.09665},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.09665}, 
}