THUDM / BPO

Black-Box Prompt Optimization: Aligning Large Language Models without Model Training

• Repository: https://github.com/thu-coai/BPO
• Paper: https://arxiv.org/abs/2311.04155
• Data: https://huggingface.co/datasets/THUDM/BPO

Black-box Prompt Optimization (BPO)

BPO is a black-box alignment technique that differs from training-based methods (like PPO or DPO). BPO only requires training of a plug-and-play model and optimizes LLMs through optimizing user inputs. Therefore, it can be used on a variety of open-source or API-based LLMs.

Model Details

Data

Prompt优化模型由隐含人类偏好特征的prompt优化对训练得到，数据集的详细信息在这里。 The Prompt Optimization Model is trained on prompt optimization pairs which contain human preference features. Detailed information on the dataset can be found here .

Backbone Model

The prompt preference optimizer is built on Llama-2-7b-chat-hf .

Language

English

Performance

Intended Use

Prompt Template

We adopt a prompt template as

[INST] You are an expert prompt engineer. Please help me improve this prompt to get a more helpful and harmless response:\n{user prompt} [/INST]

Inference code

Here is an example code for inference:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = 'Your-Model-Path'

prompt_template = "[INST] You are an expert prompt engineer. Please help me improve this prompt to get a more helpful and harmless response:\n{} [/INST]"

model = AutoModelForCausalLM.from_pretrained(model_path).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path)

text = 'Tell me about Harry Potter'

prompt = prompt_template.format(text)
model_inputs = tokenizer(prompt, return_tensors="pt").to("cuda:0")
output = model.generate(**model_inputs, max_new_tokens=1024, do_sample=True, top_p=0.9, temperature=0.6, num_beams=1)
resp = tokenizer.decode(output[0], skip_special_tokens=True).split('[/INST]')[1].strip()

print(resp)

See our Github Repo for more detailed usage (e.g. more aggressive optimization).

Other Known Limitations

• Task coverage is not sufficient, as we only used open-source data to get about 14k optimized prompts. Clearly, it is impossible to cover a wide range of user queries, so the current model may not perform well on every prompt.
• Due to the small ratio of long-context-based tasks and mathematical problems, the prompt optimizer underperforms when dealing with these tasks.

Citation

If you find our model is useful in your work, please cite it with:

@article{cheng2023black,
  title={Black-Box Prompt Optimization: Aligning Large Language Models without Model Training},
  author={Cheng, Jiale and Liu, Xiao and Zheng, Kehan and Ke, Pei and Wang, Hongning and Dong, Yuxiao and Tang, Jie and Huang, Minlie},
  journal={arXiv preprint arXiv:2311.04155},
  year={2023}
}