Sarashina1-65B

This repository provides Japanese language models trained by SB Intuitions.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
 
model = AutoModelForCausalLM.from_pretrained("sbintuitions/sarashina1-65b", torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina1-65b")
# If you want to use slow tokenizer
# tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina1-65b", use_fast=False, revision="slow-tokenizer")
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	51200	1.0T	GPTNeoX	RoPE	32	4096	32
13B	51200	1.0T	GPTNeoX	RoPE	40	5120	40
65B	51200	800B	GPTNeoX	RoPE	80	8192	64

Training Corpus

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 corpus contains about 550B tokens.

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

Sarashina1 has not been tuned to follow an instruction yet. Therefore, sarashina1 might generate some meaningless sequences, some inaccurate instances or biased/objectionable outputs. Before using sarashina1, we would like developers to tune models based on human preferences and safety considerations.

License

MIT License