## 📖 Introduction

**Qwen2-7B-Instruct-Exp** and **Qwen2-1.5B-Instruct-Exp** are powerful large language models that can expand instructions with same task type but of different content.

We fine-tuned **Qwen2-7B-Instruct** and **Qwen2-1.5B-Instruct-Exp** to obtain **Qwen2-7B-Instruct-Exp** and **Qwen2-1.5B-Instruct-Exp**.
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.

#### Example Input
> Plan an in depth tour itinerary of France that includes Paris, Lyon, and Provence.
#### Example Output 1
> Describe a classic road trip itinerary along the California coastline in the United States.
#### Example Output 2
> Create a holiday plan that combines cultural experiences in Bangkok, Thailand, with beach relaxation in Phuket.



## 🚀 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/Qwen2-7B-Instruct-Exp",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("alibaba-pai/Qwen2-7B-Instruct-Exp")

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 evaluated the data augmentation effect of our model on the Elementary Math and Implicature datasets.

| Model                          | Math   | Impl.  |
|--------------------------------|--------|--------|
| Qwen2-1.5B-Instruct            | 57.90% | 28.96% |
| + Qwen2-1.5B-Instruct-Exp      | 59.15% | 31.22% |
| + Qwen2-7B-Instruct-Exp        | 58.32% | 39.37% |
| Qwen2-7B-Instruct              | 71.40% | 28.85% |
| + Qwen2-1.5B-Instruct-Exp      | 73.90% | 35.41% |
| + Qwen2-7B-Instruct-Exp        | 72.53% | 32.92% |

## 📜 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}, 
}
```