japanese-large-lm-3.6b-instruction-sft-4bit-128g-actorder_False
This repository provides a 3.6B parameters Japanese language quantized model, fine-tuned and trained by LINE Corporation.
For Japanese
詳細な説明や実験に関しては「【インターンレポート】量子化による大規模言語モデル軽量化の効果測定」をご覧ください。
How to use
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
tokenizer = AutoTokenizer.from_pretrained("line-corporation/japanese-large-lm-3.6b-instruction-sft", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("line-corporation/japanese-large-lm-3.6b-instruction-sft-4bit-128g-actorder_False")
generator = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)
input_text = """四国の県名を全て列挙してください。"""
text = generator(
f"ユーザー: {input_text}\nシステム: ",
max_length = 256,
do_sample = True,
temperature = 0.7,
top_p = 0.9,
top_k = 0,
repetition_penalty = 1.1,
num_beams = 1,
pad_token_id = tokenizer.pad_token_id,
num_return_sequences = 1,
)
print(text) # [{'generated_text': 'ユーザー: 四国の県名を全て列挙してください。\nシステム: 高知県、徳島県、香川県、愛媛県'}]
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.