🦅 Pretrained BigBird Model for Korean (up to 4096 tokens)

Last update: Dec 14, 2022

Overview

Pretrained BigBird Model for Korean

What is BigBird • How to Use • Pretraining • Evaluation Result • Docs • Citation

한국어 | English

What is BigBird?

BigBird: Transformers for Longer Sequences에서 소개된 sparse-attention 기반의 모델로, 일반적인 BERT보다 더 긴 sequence를 다룰 수 있습니다.

🦅 Longer Sequence - 최대 512개의 token을 다룰 수 있는 BERT의 8배인 최대 4096개의 token을 다룸

⏱️ Computational Efficiency - Full attention이 아닌 Sparse Attention을 이용하여 O(n²)에서 O(n)으로 개선

How to Use

🤗 Huggingface Hub에 업로드된 모델을 곧바로 사용할 수 있습니다:)
일부 이슈가 해결된 transformers>=4.11.0 사용을 권장합니다. (MRC 이슈 관련 PR)
BigBirdTokenizer 대신에 BertTokenizer 를 사용해야 합니다. (AutoTokenizer 사용시 BertTokenizer가 로드됩니다.)
자세한 사용법은 BigBird Tranformers documentation을 참고해주세요.

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("monologg/kobigbird-bert-base")  # BigBirdModel
tokenizer = AutoTokenizer.from_pretrained("monologg/kobigbird-bert-base")  # BertTokenizer

Pretraining

자세한 내용은 [Pretraining BigBird] 참고

	Hardware	Max len	LR	Batch	Train Step	Warmup Step
KoBigBird-BERT-Base	TPU v3-8	4096	1e-4	32	2M	20k

모두의 말뭉치, 한국어 위키, Common Crawl, 뉴스 데이터 등 다양한 데이터로 학습
ITC (Internal Transformer Construction) 모델로 학습 (ITC vs ETC)

Evaluation Result

1. Short Sequence (<=512)

자세한 내용은 [Finetune on Short Sequence Dataset] 참고

	NSMC (acc)	KLUE-NLI (acc)	KLUE-STS (pearsonr)	Korquad 1.0 (em/f1)	KLUE MRC (em/rouge-w)
KoELECTRA-Base-v3	91.13	86.87	93.14	85.66 / 93.94	59.54 / 65.64
KLUE-RoBERTa-Base	91.16	86.30	92.91	85.35 / 94.53	69.56 / 74.64
KoBigBird-BERT-Base	91.18	87.17	92.61	87.08 / 94.71	70.33 / 75.34

2. Long Sequence (>=1024)

자세한 내용은 [Finetune on Long Sequence Dataset] 참고

	TyDi QA (em/f1)	Korquad 2.1 (em/f1)	Fake News (f1)	Modu Sentiment (f1-macro)
KLUE-RoBERTa-Base	76.80 / 78.58	55.44 / 73.02	95.20	42.61
KoBigBird-BERT-Base	79.13 / 81.30	67.77 / 82.03	98.85	45.42

Docs

Citation

KoBigBird를 사용하신다면 아래와 같이 인용해주세요.

@software{jangwon_park_2021_5654154,
  author       = {Jangwon Park and Donggyu Kim},
  title        = {KoBigBird: Pretrained BigBird Model for Korean},
  month        = nov,
  year         = 2021,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.5654154},
  url          = {https://doi.org/10.5281/zenodo.5654154}
}

Contributors

Jangwon Park and Donggyu Kim

Acknowledgements

KoBigBird는 Tensorflow Research Cloud (TFRC) 프로그램의 Cloud TPU 지원으로 제작되었습니다.

또한 멋진 로고를 제공해주신 Seyun Ahn님께 감사를 전합니다.

KakaoBrain KoGPT (Korean Generative Pre-trained Transformer)

KoGPT KoGPT (Korean Generative Pre-trained Transformer) https://github.com/kakaobrain/kogpt https://huggingface.co/kakaobrain/kogpt Model Descriptions

797 Dec 26, 2022

Generating Korean Slogans with phonetic and structural repetition

LexPOS_ko Generating Korean Slogans with phonetic and structural repetition Generating Slogans with Linguistic Features LexPOS is a sequence-to-sequen

3 May 23, 2022

Korean extractive summarization. 2021 AI 텍스트 요약 온라인 해커톤 화성갈끄니까팀 코드

korean extractive summarization 2021 AI 텍스트 요약 온라인 해커톤 화성갈끄니까팀 코드 Leaderboard Notice Text Summarization with Pretrained Encoders에 나오는 bertsumext모델(ext

3 Aug 10, 2022

Training code for Korean multi-class sentiment analysis

KoSentimentAnalysis Bert implementation for the Korean multi-class sentiment analysis 왜 한국어 감정 다중분류 모델은 거의 없는 것일까?에서 시작된 프로젝트 Environment: Pytorch, Da

3 Dec 2, 2022

Korean Sentence Embedding Repository

Korean-Sentence-Embedding 🍭 Korean sentence embedding repository. You can download the pre-trained models and inference right away, also it provides

80 Jan 2, 2023

ProteinBERT is a universal protein language model pretrained on ~106M proteins from the UniRef90 dataset.

ProteinBERT is a universal protein language model pretrained on ~106M proteins from the UniRef90 dataset. Through its Python API, the pretrained model can be fine-tuned on any protein-related task in a matter of minutes. Based on our experiments with a wide range of benchmarks, ProteinBERT usually achieves state-of-the-art performance. ProteinBERT is built on TenforFlow/Keras.

241 Jan 4, 2023

IndoBERTweet is the first large-scale pretrained model for Indonesian Twitter. Published at EMNLP 2021 (main conference)

IndoBERTweet 🐦 🇮🇩 1. Paper Fajri Koto, Jey Han Lau, and Timothy Baldwin. IndoBERTweet: A Pretrained Language Model for Indonesian Twitter with Effe

40 Nov 30, 2022

BMInf (Big Model Inference) is a low-resource inference package for large-scale pretrained language models (PLMs).

377 Jan 2, 2023

Crie tokens de autenticação íntegros e seguros com UToken.

UToken - Tokens seguros. UToken (ou Unhandleable Token) é uma bilioteca criada para ser utilizada na geração de tokens seguros e íntegros, ou seja, nã

0 Nov 29, 2022

Comments

Pretraining Epoch 질문
Checklist

[x] I've searched the project's issues

❓ Question

안녕하세요 저는 현재 친구들과 함께 4096 토큰을 입력받아 요약 태스크를 수행할 수 있는 모델을 만들고 있습니다. 처음엔 빅버드 + 버트 조합으로 해보려고 했는데, 이미 monologg 님께서 만들어주셨더라구요 ㅎㅎ 그래서 롱포머 + 바트 + 페가수스 조합으로 학습을 진행하려 하고 있습니다. pretrained된 KoBart를 기반으로 어텐션을 롱포머로 바꾼 후, 페가수스 task를 수행하는 구조로 되어 있습니다.

현재 13GB의 데이터를 모아서 전처리와 데이터로더 작성, 모델 코드까지는 완료한 상태입니다. 이번 주 내로 학습을 진행하려 하고 있습니다.

저희가 가진 GPU로는 대략 이틀이면 1 에포크를 돌 수 있을 것 같은데, monologg님께서는 KoBirBird 모델 개발 시 에포크를 얼마나 도셨는지 여쭤보고 싶습니다.

아무래도 pretrained 된 모델을 가져다 쓰다보니 에포크를 많이 돌 필요는 없을 것 같은데, 기준점으로 삼고 싶어서요!

말이 길어졌는데 요약하자면, KoBirBird 학습 시 에포크를 얼마나 주셨는지 궁금합니다. 또한, 그 기준은 무엇으로 삼으셨는지도 궁금합니다.
question
opened by KimJaehee0725 2
Specific information about this model.
Checklist

[ x ] I've searched the project's issues

❓ Question

You mentioned "모두의 말뭉치, 한국어 위키, Common Crawl, 뉴스 데이터 등 다양한 데이터로 학습" and I want to know the size of total corpus for pre-training.

Also I want to know the vocab size of this model.

📎 Additional context
question
opened by midannii 2
Fix some minors

Description

코드와 주석 등을 읽다가 보인 작은 오타 등을 수정했습니다

다양한 노하우를 아낌없이 공유해주신 @monologg , @donggyukimc 에게 감사의 말씀드립니다.

이후에는 GPU 환경에서 finetuning을 테스트해 볼 예정입니다 고맙습니다.

Related Issue
chore

opened by sackoh 0

Releases(v1.0.0)

v1.0.0(Nov 8, 2021)

Initial release for KoBigBird - Pretrained BigBird Model for Korean
Source code(tar.gz)
Source code(zip)

Owner

Jangwon Park

GitHub Repository https://huggingface.co/monologg/kobigbird-bert-base

Calibre recipe to convert latest issue of Analyse & Kritik into an ebook

Calibre Recipe für "Analyse & Kritik" Dies ist ein "Recipe" für die Konvertierung der aktuellen Ausgabe der Zeitung Analyse & Kritik in ein Ebook. Es

3 Jan 04, 2022

Utilizing RBERT model for KLUE Relation Extraction task

RBERT for Relation Extraction task for KLUE Project Description Relation Extraction task is one of the task of Korean Language Understanding Evaluatio

14 Nov 15, 2022

customer care chatbot made with Rasa Open Source.

Customer Care Bot Customer care bot for ecomm company which can solve faq and chitchat with users, can contact directly to team. 🛠 Features Basic E-c

23 Oct 27, 2022

AMUSE - financial summarization

AMUSE AMUSE - financial summarization Unzip data.zip Train new model: python FinAnalyze.py --task train --start 0 --count how many files,-1 for all

1 Jan 11, 2022

KLUE-baseline contains the baseline code for the Korean Language Understanding Evaluation (KLUE) benchmark.

KLUE Baseline Korean(한국어) KLUE-baseline contains the baseline code for the Korean Language Understanding Evaluation (KLUE) benchmark. See our paper fo

74 Dec 13, 2022

Lumped-element impedance calculator and frequency-domain plotter.

fastZ: Lumped-Element Impedance Calculator fastZ is a small tool for calculating and visualizing electrical impedance in Python. Features include: Sup

47 Nov 18, 2022

Japanese Long-Unit-Word Tokenizer with RemBertTokenizerFast of Transformers

Japanese-LUW-Tokenizer Japanese Long-Unit-Word (国語研長単位) Tokenizer for Transformers based on 青空文庫 Basic Usage from transformers import RemBertToken

3 Dec 22, 2021

Unifying Cross-Lingual Semantic Role Labeling with Heterogeneous Linguistic Resources (NAACL-2021).

Unifying Cross-Lingual Semantic Role Labeling with Heterogeneous Linguistic Resources Description This is the repository for the paper Unifying Cross-

16 Sep 09, 2022

English loanwords in the world's languages

Wiktionary as CLDF Content cldf1 and cldf2 contain cldf-conform data sets with a total of 2 377 756 entries about the vocabulary of all 1403 languages

3 Jan 14, 2022

SpikeX - SpaCy Pipes for Knowledge Extraction

SpikeX is a collection of pipes ready to be plugged in a spaCy pipeline. It aims to help in building knowledge extraction tools with almost-zero effort.

384 Dec 12, 2022

Official PyTorch implementation of SegFormer

SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers Figure 1: Performance of SegFormer-B0 to SegFormer-B5. Project page

1.4k Dec 29, 2022

🚀 RocketQA, dense retrieval for information retrieval and question answering, including both Chinese and English state-of-the-art models.

In recent years, the dense retrievers based on pre-trained language models have achieved remarkable progress. To facilitate more developers using cutt

475 Jan 04, 2023

Cherche (search in French) allows you to create a neural search pipeline using retrievers and pre-trained language models as rankers.

Cherche (search in French) allows you to create a neural search pipeline using retrievers and pre-trained language models as rankers. Cherche is meant to be used with small to medium sized corpora. C

224 Nov 29, 2022

Ecommerce product title recognition package

revizor This package solves task of splitting product title string into components, like type, brand, model and article (or SKU or product code or you

16 Mar 03, 2022

Rhythm-Finder is a unsupervised ML driven python powered web-application that can find the songs that suits you.

ML-powered Music Recommendation Engine

23 Oct 09, 2022

⛵️The official PyTorch implementation for "BERT-of-Theseus: Compressing BERT by Progressive Module Replacing" (EMNLP 2020).

BERT-of-Theseus Code for paper "BERT-of-Theseus: Compressing BERT by Progressive Module Replacing". BERT-of-Theseus is a new compressed BERT by progre

284 Nov 25, 2022