---
dataset_info:
  features:
  - name: id
    dtype: int32
  - name: image
    dtype: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: metadata
    struct:
    - name: doc_title
      dtype: string
    - name: publisher
      dtype: string
    - name: publish_year
      dtype: string
    - name: table_type
      dtype: string
    - name: table_field
      dtype: string
    - name: table_unit
      dtype: string
    - name: table_title
      dtype: string
    - name: table_header
      dtype: string
    - name: table_row_number
      dtype: int32
    - name: table_column_number
      dtype: int32
    - name: table_header_bold
      dtype: string
    - name: table_background
      dtype: string
    - name: html_path
      dtype: string
    - name: width
      dtype: int32
    - name: height
      dtype: int32
  - name: summary
    list: string
  - name: html
    dtype: string
  splits:
  - name: train
    num_bytes: 40931400228.0
    num_examples: 323264
  - name: validation
    num_bytes: 4185978223.25
    num_examples: 40406
  download_size: 38033928568
  dataset_size: 45117378451.25
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---
이거 HTML 데이터 Naver-OCR 이용해서 만든 데이터임. HTML 한정 부정확한 데이터가 있을 수 있음.    
llava 학습 할 때의 recap 데이터로 활용할 수 있지? 있을 듯?