README.md

---
license: mit
language:
- ja
- en
---

# Sarashina2-70B

This repository provides large language models trained by [SB Intuitions](https://www.sbintuitions.co.jp/).


## How to use


```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
 
model = AutoModelForCausalLM.from_pretrained("sbintuitions/sarashina2-70b", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina2-70b")
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
set_seed(123)
 
text = generator(
    "おはようございます、今日の天気は",
    max_length=30,
    do_sample=True,
    pad_token_id=tokenizer.pad_token_id,
    num_return_sequences=3,
)

for t in text:
  print(t)
 
```

## Configuration

| Parameters | Vocab size | Training tokens |  Architecture | Position type | Layers | Hidden dim | Attention heads |
| :-----: | :-----------: | :-------------:  | :------------ | :-----------: | :----: | :--------: | :-------------: |
| [7B](https://huggingface.co/sbintuitions/sarashina2-7b)      | 102400        | 2.1T             | Llama2        | RoPE          | 32     | 4096       | 32 |
| [13B](https://huggingface.co/sbintuitions/sarashina2-13b)    | 102400        | 2.1T             | Llama2        | RoPE          | 40     | 5120       | 40 |
| [70B](https://huggingface.co/sbintuitions/sarashina2-70b)    | 102400        | 2.1T             | Llama2        | RoPE          | 80     | 8192       | 64 |    

## Training Corpus

For our Japanese training data, we used a Japanese portion of the [Common Crawl corpus](https://commoncrawl.org/), which is the largest Web corpus, as our training dataset.
To clean the training corpus, we used [CCNet](https://github.com/facebookresearch/cc_net) and [HojiChar](https://github.com/HojiChar/HojiChar).
After cleaning, our Japanese training data contains about 1T tokens.

For our English training data, we extracted English documents from [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B) but we removed books3 corpus due to copyright infringement.

## Tokenization

We use a [sentencepiece](https://github.com/google/sentencepiece) tokenizer with a unigram language model and byte-fallback.
We do not apply pre-tokenization with Japanese tokenizer.
Thus, a user may directly feed raw sentences into the tokenizer.


## Ethical Considerations and Limitations
Sarashina2 has not been tuned to follow an instruction yet.
Therefore, sarashina2 might generate some meaningless sequences, some inaccurate instances or biased/objectionable outputs.
Before using sarashina2, we would like developers to tune models based on human preferences and safety considerations.

## License

[MIT License](https://huggingface.co/sbintuitions/sarashina2-70b/blob/main/LICENSE)