KoGrammar

Korean Grammar Correction Model based on LLM

A Project for Introduction to Text Processing(LIS3813)

This project is ongoing.

• Dataset: 국립국어원 맞춤법 교정 말뭉치 2022

• Backbone Model: KoBART(gogamza/kobart-base-v2)

Model

• Baseline Model Link (HuggingFace)
• Distilled Model Link (HuggingFace)
• Tiny Distilled Model Link (HuggingFace)

How To Use

• Requirements

  torch
  transformers
  

• Inference

  from transformers import BartConfig
  from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
  from transformers import pipeline
  
  checkpoint = 'theSOL1/kogrammar-base'
  tokenizer = AutoTokenizer.from_pretrained(checkpoint)
  config = BartConfig.from_pretrained(checkpoint)
  model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint, config=config, device_map='auto')
  pipe = pipeline('text2text-generation', model=model, tokenizer=tokenizer)
  
  sample_text = 'ㄴㅏ는 ㄱㅏ끔 눈물을흘린다'
  corrected_text = pipe(sample_text)
  print(corrected_text)

Docs

• Technical Introduction
• TODO
• 기타 문서