|
--- |
|
language: |
|
- en |
|
tags: |
|
- bpo |
|
- llama |
|
- thudm |
|
inference: false |
|
--- |
|
|
|
<h1>Black-Box Prompt Optimization: Aligning Large Language Models without Model Training</h1> |
|
|
|
- **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](https://huggingface.co/datasets/THUDM/BPO). |
|
|
|
### Backbone Model |
|
The prompt preference optimizer is built on `Llama-2-7b-chat-hf`. |
|
|
|
### Language |
|
English |
|
|
|
### Performance |
|
|
|
|
|
| Model A| Model B | A win | tie | B win | |
|
|-------------|-------------|----|----|----| |
|
| gpt-3.5-turbo + BPO | gpt-3.5-turbo | **60.0** | 8.7 | 31.3 | |
|
| claude-2 + BPO | claude-2 | **57.5** | 5.0 | 37.5 | |
|
| llama-2-13b-chat + BPO | llama-2-70b-chat | **61.3** | 0.0 | 38.7 | |
|
| vicuna-13b + BPO | vicuna-13b + PPO | **52.5** | 3.7 | 43.7 | |
|
| vicuna-13b + BPO | vicuna-13b + DPO | **53.8** | 2.5 | 43.7 | |
|
| vicuna-13b + DPO + BPO | vicuna-13b + DPO | **60.0** | 2.5 | 37.5 | |
|
|
|
## 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: |
|
```python |
|
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](https://github.com/thu-coai/BPO/blob/main/src/infer_example.py) 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} |
|
} |
|
``` |
