llm-jp-13b-v2.0
This repository provides large language models developed by LLM-jp, a collaborative project launched in Japan.
| Pre-trained models |
| llm-jp-13b-v2.0 |
Checkpoints format: Hugging Face Transformers
Required Libraries and Their Versions
- torch>=2.3.0
- transformers>=4.40.1
- tokenizers>=0.19.1
- accelerate>=0.29.3
- flash-attn>=2.5.8
Usage
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-13b-v2.0")
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-13b-v2.0", device_map="auto", torch_dtype=torch.bfloat16)
text = "自然言語処理とは何か"
tokenized_input = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt").to(model.device)
with torch.no_grad():
output = model.generate(
tokenized_input,
max_new_tokens=100,
do_sample=True,
top_p=0.95,
temperature=0.7,
repetition_penalty=1.05,
)[0]
print(tokenizer.decode(output))
Model Details
- Model type: Transformer-based Language Model
- Total seen tokens: 256B
| Model | Params | Layers | Hidden size | Heads | Context length |
|---|---|---|---|---|---|
| 13b model | 13b | 40 | 5120 | 40 | 4096 |
Training
-
Pre-training:
- Hardware: 128 A100 40GB GPUs (mdx cluster)
- Software: Megatron-LM
-
Instruction tuning:
- Hardware: 8 A100 40GB GPUs (mdx cluster)
- Software: TRL and DeepSpeed
Tokenizer
The tokenizer of this model is based on huggingface/tokenizers Unigram byte-fallback model.
The vocabulary entries were converted from llm-jp-tokenizer v2.2 (100k: code20K_en40K_ja60K.ver2.2).
Please refer to README.md of llm-ja-tokenizer for details on the vocabulary construction procedure (the pure SentencePiece training does not reproduce our vocabulary).
- Model: Hugging Face Fast Tokenizer using Unigram byte-fallback model
- Training algorithm: Marging Code/English/Japanese vocabularies constructed with SentencePiece Unigram byte-fallback and reestimating scores with the EM-algorithm.
- Training data: A subset of the datasets for model pre-training
- Vocabulary size: 96,867 (mixed vocabulary of Japanese, English, and source code)
- The acutal size of vocabulary in the pretrained model is 97,024 due to round-up to multiples of 256.
Datasets
Pre-training
The models have been pre-trained using a blend of the following datasets.
| Language | Dataset | Tokens |
|---|---|---|
| Japanese | Wikipedia | 1.4B |
| Common Crawl | 130.7B | |
| English | Wikipedia | 4.7B |
| The Pile | 110.3B | |
| Codes | The Stack | 8.7B |
Instruction tuning
The models have been fine-tuned on the following datasets.
| Language | Dataset | description |
|---|---|---|
| Japanese | ichikara-instruction-004-001 | A manually constructed Japanese instruction dataset |
| answer-carefully-001 | A manually constructed Japanese instruction dataset focusing on LLMs' safety | |
| databricks-dolly-15k-ja | databricks-dolly-15k translated into Japanese using DeepL | |
| oasst1-21k-ja | A subset of oasst1 translated into Japanese using DeepL | |
| oasst2-33k-ja | A subset of oasst2 translated into Japanese using DeepL | |
| English | databricks-dolly-15k | - |
| oasst1-21k-en | A subset of oasst1 | |
| oasst2-33k-en | A subset of oasst2 |
Evaluation
You can view the evaluation results of several LLMs on this leaderboard. We used llm-jp-eval (v1.3.0) for the evaluation.
Besides, we used LLM-as-a-judge frameworks, Japanese Vicuna QA Benchmark and Japanese MT Bench, for evaluation. For details, please refer to our technical blog (in Japanese).
Risks and Limitations
The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.
Send Questions to
llm-jp(at)nii.ac.jp
License
Model Card Authors
The names are listed in alphabetical order.
Namgi Han, Tatsuya Hiraoka, Hirokazu Kiyomaru, Takashi Kodama, and Hiroshi Matsuda.