stockmark/gpt-neox-japanese-1.4b

This repository provides a GPT-NeoX based model with 1.4B parameters pre-trained on Japanese corpus of about 20B tokens. This model is developed by Stockmark Inc.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Use torch.bfloat16 for A100 GPU and torch.flaot16 for the older generation GPUs
torch_dtype = torch.bfloat16 if torch.cuda.is_available() and hasattr(torch.cuda, "is_bf16_supported") and torch.cuda.is_bf16_supported() else torch.float16

model = AutoModelForCausalLM.from_pretrained("stockmark/gpt-neox-japanese-1.4b", device_map="auto", torch_dtype=torch_dtype)
tokenizer = AutoTokenizer.from_pretrained("stockmark/gpt-neox-japanese-1.4b")

inputs = tokenizer("自然言語処理は", return_tensors="pt").to(model.device)
with torch.no_grad():
    tokens = model.generate(
        **inputs,
        max_new_tokens=128,
        repetition_penalty=1.1
    )
    
output = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(output)

Example:

• LoRA tuning: https://huggingface.co/stockmark/gpt-neox-japanese-1.4b/blob/main/notebooks/LoRA.ipynb

Training dataset

• Japanese Web Corpus (ja): 8.6B tokens (This dataset will not be released.)
• Wikipedia (ja): 0.88B tokens
• CC100 (ja): 10.5B tokens

Training setting

• Trained using HuggingFace Trainer and DeepSpeed (ZeRO-2)
• 8 A100 GPUs (40GB) at ABCI
• Mixed Precision (BF16)

License

The MIT license

Developed by

Stockmark Inc.

Author

Takahiro Omi