stockmark/stockmark-100b

Stockmark-100b is a 100 billion parameter LLM pretrained from scratch based on Japanese and English corpus of about 910 billion tokens. This model is developed by Stockmark Inc.

Instruction tuned model:

• stockmark-100b-instruct-v0.1

This project is supported by GENIAC.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("stockmark/stockmark-100b")
model = AutoModelForCausalLM.from_pretrained("stockmark/stockmark-100b", device_map="auto", torch_dtype=torch.bfloat16)

input_ids = tokenizer("生成AIとは？", return_tensors="pt").input_ids.to(model.device)
with torch.inference_mode():
    tokens = model.generate(
        input_ids,
        max_new_tokens = 256,
        do_sample = True,
        temperature = 0.7,
        top_p = 0.95,
        repetition_penalty = 1.08
    )
    
output = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(output)

Dataset (pretraining)

Stockmark-100b was trained using a total of about 910B tokens of Japanese and English text corpus.

The detail of Japanese data is summarized in the below table. The stockmark web corpus consists of web pages related to business, which are collected by Stockmark Inc.

corpus	tokens after preprocessing
Stockmark Web Corpus (This dataset will not be released)	8.8 billion
Patent	37.5 billion
Wikipedia	1.5 billion
mC4	52.6 billion
CommonCrawl (snapshot: 2020-50 ~ 2024-10)	203.7 billion

English data is sampled from RedPajama-Data.

Training

• GPU: 48 nodes of a3 (8*H100) instances
• Training duration: about 7 weeks
• Container: Pytorch NGC Container
• Library: Megatron-LM

Performance

Stockmark Business Questions

Dataset: https://huggingface.co/datasets/stockmark/business-questions

model	accuracy
stockmark-100b-instruct	0.90
stockmark-13b-instruct	0.80
GPT-3.5-turbo^1	0.42

Japanese Vicuna QA Benchmark

We excluded categories that require calculation and coding, and use remaining 60 questions for evaluation.

GitHub: https://github.com/ku-nlp/ja-vicuna-qa-benchmark

model	average score
stockmark-100b-instruct	5.97
tokyotech-llm/Swallow-70b-instruct-hf	5.59
GPT-3.5 (text-davinci-003)	5.08

Inference speed

model	time [s] for genrating 100 characters in Japanese
stockmark-100b-instruct	1.86
gpt-3.5-turbo	2.15
gpt-4-turbo	5.48
tokyotech-llm/Swallow-70b-instruct-hf	2.22

For local LLMs, we measured the inference time using AWS Inferentia2.

License

MIT

Developed by

Stockmark Inc.