Black-Box Prompt Optimization (BPO)

Aligning Large Language Models without Model Training (ACL 2024)

🤗 Model • 📚 Data • 📃 Paper • 🌐 Demo

(Upper) Black-box Prompt Optimization (BPO) offers a conceptually new perspective to bridge the gap between humans and LLMs. (Lower) On Vicuna Eval’s pairwise evaluation, we show that BPO further aligns gpt-3.5-turbo and claude-2 without training. It also outperforms both PPO & DPO and presents orthogonal improvements.

Update

We have released our model and data on Hugging Face.

We build a demo for BPO on Hugging Face.

Table of Contents

• Model
• Data
• Quick Start
  • Data Construction
  • Model Training
  • Inference
  • Evaluation
• Citation

Model

The prompt preference optimization model can be download from Hugging Face

Inference code (Please refer to src/infer_example.py for more instructions on how to optimize your prompts):

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = 'THUDM/BPO'

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]"

device = 'cuda:0'
model = AutoModelForCausalLM.from_pretrained(model_path).half().eval().to(device)
# for 8bit
# model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device, load_in_8bit=True)
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(device)
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)

Data

BPO dataset

BPO Dataset can be found on Hugging Face.

BPO for SFT Data Construction

The alpaca_reproduce directory contains the BPO-reproduced Alpaca dataset. The data format is:

{
    "instruction": {instruction},
    "input": {input},
    "output": {output},
    "optimized_prompt": {optimized_prompt},
    "res": {res}
}

• {instruction}, {input}, and {output} are elements from the original dataset.
• {optimized_prompt} is BPO-optimized instruction.
• {res} is the response from text-davinci-003 using the {optimized_prompt}.

Testset

The testset directory contains all the test datasets we used, including:

• 200 prompts sampled from the BPO dataset
• 200 examples from Dolly dataset
• 252 human evaluation instructions from Self-Instruct
• 80 user-oriented prompts from the Vicuna Eval dataset.

Quick Start

For all codes, we have added #TODO comments to indicate places in the code that need modification before running. Please update the relevant parts as noted before executing each file.

Setup

pip install -r requirements.txt

Data Construction

To construct data yourself, run the following command

cd src/data_construction

# using pairwise feedback data to generate optimized prompts
python chatgpt_infer.py

# process generated optimized prompts
python process_optimized_prompts.py

Model Training

If you want to train your own prompt preference optimizer, please run the following command:

cd src/training

# pre-process fine-tuning data
python ../data_construction/process_en.py
python data_utils.py

# fine-tuning
python train.py

# inference
python infer_finetuning.py

Inference

We show an example code for generation with llama2-chat on BPO-optimized prompts.

Evaluation

If you wish to compare the BPO-aligned model with the original model, please refer to the following code:

cd src/evaluation

# take gpt4 evaluation on dolly_eval as an example
python gpt4_score.py --input_file_a "Path to generation results of BPO-aligned model" \
--input_file_b "Path to generation results of original model" \
--task_name "dolly_eval" \ # change it to "self_instruct", "test_set", or "vicuna_eval" for other testsets
--output_file "Output path"

# calculate win rates
python cal_gpt4_score.py --input_file "Output path"

Acknowledgement

• Fine-tuning code: llm_finetuning
• PPO code: DeepSpeed-Chat
• DPO code: LLaMA-Factory
• Evaluation Prompts: llm_judge and alpaca_eval

Citation

@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}
}