README.md

---
license: other
---

![Aquila_logo](./log.jpeg)

<h4 align="center">
    <p>
        <b>English</b> |
        <a href="https://huggingface.co/BAAI/Aquila2-34B/blob/main/README_zh.md">简体中文</a> |
    <p>
</h4>


We opensource our **Aquila2** series, now including **Aquila2**, the base language models, namely **Aquila2-7B** and **Aquila2-34B**, as well as **AquilaChat2**, the chat models, namely **AquilaChat2-7B** and **AquilaChat2-34B**, as well as the long-text chat models, namely **AquilaChat2-7B-16k** and **AquilaChat2-34B-16k**

The additional details of the Aquila model will be presented in the official technical report. Please stay tuned for updates on official channels.

## Updates 2024.6.6

We have updated the basic language model **Aquila2-34B**, which has the following advantages compared to the previous model:

* Replaced tokenizer with higher compression ratio:

| Tokenizer | Size  | Zh                       | En     | Code  | Math   | Average |
|-----------|-------|--------------------------|--------|-------|-------|---------|
| Aquila2-original   | 100k  | **4.70**                 | 4.42   | 3.20  | 3.77  | 4.02    |
| Qwen1.5   | 151k  | 4.27                     | 4.51   | 3.62  | 3.35  | 3.94    |
| Llama3    | 128k  | 3.45                     | **4.61**   | 3.77  | **3.88** | 3.93    |
| Aquila2-new     | 143k  | 4.60                     | **4.61** | **3.78** | **3.88**  | **4.22** |

* The maximum processing length supported by the model has increased from 2048 to 8192



## Quick Start  Aquila2-34B

### 1. Inference
Aquila2-34B is a base model that can be used for continuation.

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import BitsAndBytesConfig

device= "cuda:0"

# Model Name
model_name = 'BAAI/Aquila2-34B'

# load model and tokenizer
quantization_config=BitsAndBytesConfig(
                        load_in_4bit=True,
                        bnb_4bit_use_double_quant=True,
                        bnb_4bit_quant_type="nf4",
                        bnb_4bit_compute_dtype=torch.bfloat16,
                    )
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, trust_remote_code=True,
                        # quantization_config=quantization_config # Uncomment this one for 4-bit quantization
                        )

tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)

model.eval()

model.to(device)

# Example
text = "The meaning of life is"
tokens = tokenizer.encode_plus(text)['input_ids']
tokens = torch.tensor(tokens)[None,].to(device)

with torch.no_grad():
        out = model.generate(tokens, do_sample=False, max_length=128, eos_token_id=tokenizer.eos_token_id)[0]
        out = tokenizer.decode(out.cpu().numpy().tolist())
        print(out)
```


## License

Aquila2 series open-source model is licensed under [ BAAI Aquila Model Licence Agreement](https://huggingface.co/BAAI/Aquila2-34B/blob/main/BAAI-Aquila-Model-License%20-Agreement.pdf)

## Citation
Feel free to cite the repo if you think Aquila2 is useful.

```python
@misc{zhang2024aquila2technicalreport,
      title={Aquila2 Technical Report}, 
      author={Bo-Wen Zhang and Liangdong Wang and Jijie Li and Shuhao Gu and Xinya Wu and Zhengduo Zhang and Boyan Gao and Yulong Ao and Guang Liu},
      year={2024},
      eprint={2408.07410},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.07410}, 
}
```