fujiki-1emon/toward-japanese-llms

Resources toward Japanese LLMs

• Bellow are the list that I find are useful, helpful to build Large Language Models (LLMs) in Japanese.
  • The list is not complete and there are other resources that I find are relevant but not yet listed.
• Each resource in the list might have my comments on it.

Table of Contents

• General
• Pre-training datasets
• Downsteam tasks
• Tokenization
• Models
• Model Architecture
• Training
• Evaluation

General

• A Survey of Large Language Models [arXiv]

Pre-training datasets

The Pile

• The Pile: An 800GB Dataset of Diverse Text for Language Modeling [arXiv]
  • noted: Pile-CC
• The Pile has 0.07%（approx. 900 M chars）of Japanese texts based on the types of the characters.
  • cf. The Pileの構成（なぜCerebras-GPTで日本語が使えるのか？）

The ROOTS

• The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset

RedPajama-Data [github]

• However, the Pile still seems better than RedPajama-Data - cf. https://twitter.com/BlancheMinerva/status/1652899628356960256?s=20
  • 

Donwstream tasks

Question Answering

• Better Question-Answering Models on a Budget [ArXiv]
  • Having briefly checked it, it looks interesting, but only compared their models to OPTs?

Tokenization

LINE Distill BERT

• Word segmentation by MeCab+UniDic + subword tokenization by SentencePiece

Models

abeja/gpt-neox-japanese-2.7b

• ABEJAで作った大規模GPTモデルとその道のり

Model Architecture

• ABEJA GPTモデルにおけるアーキテクチャの工夫

Training

• GPT-neoxの学習用にマルチノード並列学習環境を整えた with DeepSpeed

Fine-tuning

• 【インターンレポート】6.7B日本語モデルに対するLoRAチューニング
• Self-Instruct: Aligning Language Model with Self Generated Instructions arXiv
  • https://twitter.com/rasbt/status/1650866140892069892?s=20
• INSTRUCTEVAL: Towards Holistic Evaluation of Instruction-Tuned Large Language Models arXiv Related tweet

Alignment

• Learning to summarize from human feedback [arXiv]
• Training language models to follow instructions with human feedback [arXiv]

RLHF

• RLHF works because it's rating full sentences
  • https://twitter.com/savvyRL/status/1651255588813443073?s=20
  • https://twitter.com/mr_bay_area/status/1651594421551644678?s=20
    • Sequence Level Training with Recurrent Neural Networks [arXiv]

Evaluation

• EleutherAI/lm-evaluation-harness: v0.3.0
• japanese-toxic-dataset
• Chatbot Arena: Benchmarking LLMs in the Wild with Elo Ratings
• ChatGPTに共通テスト（旧センター試験）を解かせてみた - Related tweet

Reports

• Performance report of rinna/japanese-gpt-1b fine-tuned on Japanese version (translation) of Dolly dataset [tweet]
  • 

Practical

• Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond [ArXiv]

misc.

• How well does ChatGPT speak Japanese?

What I've developed

• t5-efficient-xl-nl6-japanese
• ul2-base-japanese
• ul2-large-japanese