---
license: apache-2.0
---

# BiLLa: A Bilingual LLaMA with Enhanced Reasoning Ability 

BiLLa is an open-source reasoning-enhanced bilingual LLaMA model. The main features are:
- Greatly improve the ability of Chinese language modeling, and minimize the damage to the original English ability of LLaMA;
- During the training, more task data is added with ChatGPT-generated analysis;
- Full-parameter optimization for better performance.

Github: https://github.com/Neutralzz/BiLLa

<b>Note</b>: Due to LLaMA's license, the model weights in this hub cannot be used directly. 
The weight of `word embedding` is the sum of the weights of the trained model and the original LLaMA, 
so as to ensure that developers with LLaMA original model accessibility can convert the model released by this hub into a usable one.

First, you can revert the model weights by [this script](https://github.com/Neutralzz/BiLLa/blob/main/embedding_convert.py):
```shell
python3 embedding_convert.py \
    --model_dir /path_to_BiLLa/BiLLa-7B-LLM \
    --meta_llama_pth_file /path_to_LLaMA/llama-7b/consolidated.00.pth
```

Then, you can run this model as follows:
```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_path = "/path_to_BiLLa/BiLLa-7B-LLM"
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, torch_dtype=torch.float16).cuda()

prompt = "[Your prompt]"
input_ids = tokenizer([prompt]).input_ids
output_ids = model.generate(
            torch.as_tensor(input_ids).cuda(),
            do_sample=True,
            temperature=0.7,
            max_new_tokens=1024
        )
output_ids = output_ids[0][len(input_ids[0]):]

outputs = tokenizer.decode(output_ids, skip_special_tokens=True).strip()
print(outputs)
```

Different from [BiLLa-7B-SFT](https://huggingface.co/Neutralzz/BiLLa-7B-SFT), the input format of `BiLLa-7B-LLM` has no restriction.