metadata
license: mit
language:
- ja
library_name: transformers
pipeline_tag: text-generation
tags:
- japanese
- llama-2
stockmark/stockmark-13b
This repository provides a Llama-2 based model with 13B parameters pre-trained on Japanese corpus of about 220B tokens. This model is developed by Stockmark Inc.
Please see our blog for more details.
This project is supported by AWS LLM development support program.
How to use
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
# For A100 or H100 GPU
model = AutoModelForCausalLM.from_pretrained("stockmark/stockmark-13b", device_map="auto", torch_dtype=torch.bfloat16)
# If you use a T4 or V100 GPU, please load a model in 8 bit with the below code.
# To do so, you need to install `bitsandbytes` via `pip install bitsandbytes`.
# model = AutoModelForCausalLM.from_pretrained("stockmark/stockmark-13b", device_map={"": 0}, load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained("stockmark/stockmark-13b")
inputs = tokenizer("自然言語処理とは", return_tensors="pt").to(model.device)
with torch.no_grad():
tokens = model.generate(
**inputs,
max_new_tokens=128,
do_sample=True,
temperature=0.7
)
output = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(output)
Examples:
- LoRA tuning: https://huggingface.co/stockmark/stockmark-13b/blob/main/notebooks/LoRA.ipynb
- QLoRA tuning (in preparation): https://huggingface.co/stockmark/stockmark-13b/blob/main/notebooks/QLoRA.ipynb
Training dataset
We have used Japanese corpus of total of about 220 billion tokens.
| corpus | tokens after preprocessing |
|---|---|
| Stockmark Web Corpus (This dataset will not be released) | 9.1 billion |
| Patent | 34.8 billion |
| Wikipedia | 1.0 billion |
| CC100 | 10.9 billion |
| mC4 | 53.2 billion |
| CommonCrawl (snapshot: 2023-23, 2022-49, 2022-21, 2021-21) | 112.9 billion |
Accelerator and Library
- Accelerator: AWS Trainium
- Library for distributed training: neuronx-nemo-megatron