llm-jp-1.3b-v1.0-mdsfmt
This repository provides large language models (Megatron-DeepSpeed format) developed by LLM-jp, a collaborative project launched in Japan. Hugging Face Transformers format models are available here.
| Model Variant |
|---|
| Pre-trained models <span style="color:red">(Megatron-DeepSpeed format)</span> |
| llm-jp-13b-v1.0-mdsfmt |
| llm-jp-13b-v1.0-mdsfmt-itr87870 |
| llm-jp-1.3b-v1.0-mdsfmt |
| llm-jp-1.3b-v1.0-mdsfmt-itr87430 |
llm-jp-13b-v1.0-mdsfmt-itr87870
and llm-jp-1.3b-v1.0-mdsfmt-itr87430
were originally trained with approximately 270B+ tokens.
llm-jp-13b-v1.0-mdsfmt
and llm-jp-1.3b-v1.0-mdsfmt
are models further trained by additional (potentially) high-quality 27B tokens data from llm-jp-13b-v1.0-mdsfmt-itr87870 and llm-jp-1.3b-v1.0-mdsfmt-itr87430, respectively for finalizing the pre-training.
Model Details
- Model type: Transformer-based Language Model
- Total seen tokens: 300B
| Model | Params | Layers | Hidden size | Heads | Context length |
|---|---|---|---|---|---|
| 13b model | 13b | 40 | 5120 | 40 | 2048 |
| 1.3b model | 1.3b | 24 | 2048 | 16 | 2048 |
Training
- Pre-training:
- Hardware: 96 A100 40GB GPUs (mdx cluster)
- Software: Megatron-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 v2.1 (50k).
Please refer to README.md of llm-ja-tokenizer for details on the vocabulary construction procedure.
- Model: Hugging Face Fast Tokenizer using Unigram byte-fallback model which requires
tokenizers>=0.14.0 - Training algorithm: SentencePiece Unigram byte-fallback
- Training data: A subset of the datasets for model pre-training
- Vocabulary size: 50,570 (mixed vocabulary of Japanese, English, and source code)
Datasets
Pre-training
The models have been pre-trained using a blend of the following datasets.
| Language | Dataset | Tokens |
|---|---|---|
| Japanese | Wikipedia | 1.5B |
| mC4 | 136B | |
| English | Wikipedia | 5B |
| The Pile | 135B | |
| Codes | The Stack | 10B |
The pre-training was continuously conducted using a total of 10 folds of non-overlapping data, each consisting of approximately 27-28B tokens. We finalized the pre-training with additional (potentially) high-quality 27B tokens data obtained from the identical source datasets listed above used for the 10-fold data.
Evaluation
You can view the evaluation results of several LLMs on this leaderboard. We used llm-jp-eval for the evaluation.
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.
Hirokazu Kiyomaru, Hiroshi Matsuda, Jun Suzuki, Namgi Han, Saku Sugawara, Shota Sasaki, Shuhei Kurita, Taishi Nakamura, Takumi Okamoto.
