Sarashina2-13B

This repository provides large language models trained by SB Intuitions.

How to use

Please set use_fast=False to use our tokenizer properly.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
 
model = AutoModelForCausalLM.from_pretrained("sbintuitions/sarashina2-13b", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina2-13b", use_fast=False)
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	102400	2.1T	Llama2	RoPE	32	4096	32
13B	102400	2.1T	Llama2	RoPE	40	5120	40
70B (TBA)

Training Corpus

For our Japanese training data, we used a Japanese portion of the Common Crawl corpus, which is the largest Web corpus, as our training dataset. To clean the training corpus, we used CCNet and HojiChar. After cleaning, our Japanese training data contains about 1T tokens.

For our English training data, we extracted English documents from SlimPajama but we removed books3 corpus due to copyright infringement.

Tokenization

We use a 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