77 lines
3.1 KiB
Markdown
77 lines
3.1 KiB
Markdown
<img src="https://cdn-uploads.huggingface.co/production/uploads/62aba5ebab9ed4f63c36b1e2/47PZcc9QTR_okQIvKeOLn.png" alt="image/png" style="transform: scale(1);">
|
||
|
||
|
||
## 📖 Introduction
|
||
|
||
**Qwen2-7B-Instruct-Refine** and **Qwen2-1.5B-Instruct-Refine** are two powerful large language models that act as proficient prompt engineers. They can optimize and refine the prompts input by users, and the generated optimized instructions can significantly enhance the LLM's ability to produce better and more informative responses for users.
|
||
|
||
We fine-tuned **Qwen2-7B-Instruct** and **Qwen2-1.5B-Instruct** to obtain **Qwen2-7B-Instruct-Refine** and **Qwen2-1.5B-Instruct-Refine**.
|
||
We sampled the dataset from OpenHermes and the LCCD dataset, ensuring a balanced task distribution. For training set annotations, we used Qwen-max with incorporated our handwritten examples as in-context prompts.
|
||
|
||
## 🚀 Quick Start
|
||
|
||
Here provides a code snippet with `apply_chat_template` to show you how to load the tokenizer and model and how to generate contents.
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
device = "cuda" # the device to load the model onto
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"alibaba-pai/Qwen2-1.5B-Instruct-Refine",
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained("alibaba-pai/Qwen2-1.5B-Instruct-Refine")
|
||
|
||
prompt = "Give me a short introduction to large language model."
|
||
messages = [
|
||
{"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.input_ids,
|
||
max_new_tokens=2048，
|
||
eos_token_id=151645，
|
||
)
|
||
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]
|
||
```
|
||
|
||
## 🔍 Evaluation
|
||
|
||
We used single-turn instructions from MT-Bench as input for Qwen2-1.5B-Instruct and Qwen2-7B-Instruct. GPT4-turbo is used to evaluate the changes in the level of detail and truthfulness of responses to our model's revised instructions.
|
||
|
||
| Model | Detail | Truthfulness |
|
||
|:----------------------------:|:------:|:------------:|
|
||
| Qwen2-1.5B-Instruct | 50.00% | 50.00% |
|
||
| + Qwen2-1.5B-Instruct-Refine | 75.63% | 63.75% |
|
||
| + Qwen2-7B-Instruct-Refine | 76.56% | 62.19% |
|
||
| Qwen2-7B-Instruct | 50.00% | 50.00% |
|
||
| + Qwen2-1.5B-Instruct-Refine | 70.94% | 57.19% |
|
||
| + Qwen2-7B-Instruct-Refine | 74.69% | 58.44% |
|
||
|
||
|
||
## 📜 Citation
|
||
|
||
If you find our work helpful, please cite it!
|
||
|
||
```
|
||
@misc{data-augmentation-family,
|
||
title={Building a Family of Data Augmentation Models for Low-cost LLM Fine-tuning on the Cloud},
|
||
author={Yuanhao Yue and Chengyu Wang and Jun Huang and Peng Wang},
|
||
year={2024},
|
||
eprint={2412.04871},
|
||
archivePrefix={arXiv},
|
||
primaryClass={cs.CL},
|
||
url={https://arxiv.org/abs/2412.04871},
|
||
}
|
||
``` |