README.md

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "wangrongsheng/DPDG-Qwen2-7B-lora" # [wangrongsheng/DPDG-Qwen2-7B-lora]
device = "cuda" # the device to load the model onto

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

ins = """
请根据给定的提示生成两种不同质量的回答。第一种回答应该是高质量的、令人满意的答案，代表"chosen"的选项。第二种回答则应该是低质量的、不太理想的答案，代表"rejected"的选项。\n
在生成这两个回答时，请注意以下事项：\n
1. "chosen" 回复应具有实质性内容、流畅的表达，并能够完整回答提示中提出的问题或要求。\n
2. "rejected" 回复可能存在一些问题，例如逻辑不连贯、信息不完整或表达不清晰。但请确保它仍然是一个可以大致理解的回复，而不是完全无关或毫无意义的内容。\n
3. 这两个回复的长度应该大致相当，而不是差异极大。\n
4. 请确保在"chosen"回复和"rejected"回复之间反映出明显的质量差异，使区别显而易见。\n
请根据这些指导方针为给定的提示生成一个"chosen"的回应和一个"rejected"的回应。这将有助于训练奖励模型以区分高质量和低质量的回应。\n
提示是:
"""
prompt = "什么是ACI fabric中的叶脊拓扑结构？"
messages = [
    {"role": "system", "content": ins},
    {"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,
    max_new_tokens=1024
)
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)