line-corporation/japanese-large-lm-3.6b

Model

japanese-large-lm-3.6b

73

7 commits

1 linked in READMEs

updated Aug 17, 2023

See the code

README

japanese-large-lm-3.6b

This repository provides a 3.6B parameters Japanese language model, trained by LINE Corporation.

Tech Blog explains details.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
 
model = AutoModelForCausalLM.from_pretrained("line-corporation/japanese-large-lm-3.6b", torch_dtype=torch.float16)
tokenizer = AutoTokenizer.from_pretrained("line-corporation/japanese-large-lm-3.6b", use_fast=False)
generator = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)
set_seed(101)
 
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=5,
)
 
for t in text:
  print(t)
 
# 下記は生成される出力の例
# [{'generated_text': 'おはようございます、今日の天気は雨模様ですね。梅雨のこの時期の 朝は洗濯物が乾きにくいなど、主婦にとっては悩みどころですね。 では、'},
#  {'generated_text': 'おはようございます、今日の天気は晴れ。 気温は8°C位です。 朝晩は結構冷え込むようになりました。 寒くなってくると、...'},
#  {'generated_text': 'おはようございます、今日の天気は曇りです。 朝起きたら雪が軽く積もっていた。 寒さもそれほどでもありません。 日中は晴れるみたいですね。'},
#  {'generated_text': 'おはようございます、今日の天気は☁のち☀です。 朝の気温5°C、日中も21°Cと 暖かい予報です'},
#  {'generated_text': 'おはようございます、今日の天気は晴天ですが涼しい1日です、気温は午後になり低くなり25°Cくらい、風も強いようですので、'}]

Model architecture

ModelVocab sizeArchitecturePosition typeLayersHidden dimAttention heads
1.7B51200GPT2Absolute24230424
3.6B51200GPTNeoXRoPE30307232

Training Corpus

Our training corpus consists of the Japanese portions of publicly available corpus such as C4, CC-100, and Oscar. We also incorporated the Web texts crawled by in-house system. The total size of our training corpus is about 650 GB. The trained model achieves 7.50 perplexity on the internal validation sets of Japanese C4.

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.

License

Apache License, Version 2.0

endpoints_compatible
gpt_neox
pytorch
safetensors
text-generation
text-generation-inference
transformers

Contributors

sho-takase

5 commits

SFconvertbot

1 commits

ST
Sho Takase

1 commits

line-corporation/japanese-large-lm-3.6b

Model

japanese-large-lm-3.6b

73

7 commits

1 linked in READMEs

updated Aug 17, 2023

See the code

README

japanese-large-lm-3.6b

This repository provides a 3.6B parameters Japanese language model, trained by LINE Corporation.

Tech Blog explains details.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
 
model = AutoModelForCausalLM.from_pretrained("line-corporation/japanese-large-lm-3.6b", torch_dtype=torch.float16)
tokenizer = AutoTokenizer.from_pretrained("line-corporation/japanese-large-lm-3.6b", use_fast=False)
generator = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)
set_seed(101)
 
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=5,
)
 
for t in text:
  print(t)
 
# 下記は生成される出力の例
# [{'generated_text': 'おはようございます、今日の天気は雨模様ですね。梅雨のこの時期の 朝は洗濯物が乾きにくいなど、主婦にとっては悩みどころですね。 では、'},
#  {'generated_text': 'おはようございます、今日の天気は晴れ。 気温は8°C位です。 朝晩は結構冷え込むようになりました。 寒くなってくると、...'},
#  {'generated_text': 'おはようございます、今日の天気は曇りです。 朝起きたら雪が軽く積もっていた。 寒さもそれほどでもありません。 日中は晴れるみたいですね。'},
#  {'generated_text': 'おはようございます、今日の天気は☁のち☀です。 朝の気温5°C、日中も21°Cと 暖かい予報です'},
#  {'generated_text': 'おはようございます、今日の天気は晴天ですが涼しい1日です、気温は午後になり低くなり25°Cくらい、風も強いようですので、'}]

Model architecture

ModelVocab sizeArchitecturePosition typeLayersHidden dimAttention heads
1.7B51200GPT2Absolute24230424
3.6B51200GPTNeoXRoPE30307232

Training Corpus

Our training corpus consists of the Japanese portions of publicly available corpus such as C4, CC-100, and Oscar. We also incorporated the Web texts crawled by in-house system. The total size of our training corpus is about 650 GB. The trained model achieves 7.50 perplexity on the internal validation sets of Japanese C4.

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.

License

Apache License, Version 2.0

endpoints_compatible
gpt_neox
pytorch
safetensors
text-generation
text-generation-inference
transformers

Contributors

sho-takase

5 commits

SFconvertbot

1 commits

ST
Sho Takase

1 commits