---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: metadata
    struct:
    - name: Era
      dtype: string
    - name: Lang
      dtype: string
    - name: LawType
      dtype: string
    - name: Num
      dtype: int64
    - name: Year
      dtype: int64
    - name: PromulgateMonth
      dtype: int64
    - name: PromulgateDay
      dtype: int64
    - name: LawNum
      dtype: string
    - name: category_id
      dtype: int64
    - name: id_split
      dtype: int64
  splits:
  - name: train
    num_bytes: 1286294957
    num_examples: 109532
  - name: validation
    num_bytes: 135176571
    num_examples: 11538
  - name: test
    num_bytes: 154350330
    num_examples: 13183
  download_size: 426424485
  dataset_size: 1575821858
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
language:
- ja
tags:
- legal
---

# Abstract

This is the Japapnese law dataset obtained from [e-Gov](https://www.e-gov.go.jp) (Date of download: Oct. 20th, 2024)
Each piece of text data is chunked into fewer than 4,096 tokens.
Not chunked version is available [HERE](https://huggingface.co/datasets/nlp-waseda/e_gov)

# Data Format

Each data is consist of 2 fields, "text" and "metadata".
* "text" fields contains the legal texts, which are expected to be mainly used.
* "metadata" fields contains additional information including 10 subfields below:
  * "Era": The Japanese Era when the law is promulgated such as "Showa".
  * "Lang": The language the text is written in. All of them are Japanese.
  * "LawType": The type of the law including types below.
    * "Constitution"
    * "Act"
    * "CabinetOrder"
    * "ImperialOrder"
    * "MinisterialOrdinance"
    * "Rule"
    * "Misc"
  * "Year": The year when the law is promulgated.
  * "PromulgateMonth/Day": The Month/Day when the law is promulgated.
  * "LawNum": The string of the (numeric) name of the law.
  * "category_id": The integer representing the category where the law is categorized. The categories is found in [category.json](category.json)
  * "id_split": A non-negative integer indicating the number of this chunk.

# Data Split

This dataset has 3 split, train, validation and test.
The data is split randomly but preserving the original distribution of the categories.
The ratio is 8:1:1.

# Data Chunking
* The tokeniser used for tokenising is llm-jp/llm-jp-3-1.8b
* Algorithm
  * If a file is less than 4,096 tokens, it is treated as a single chunk.
  * If it is 4,096 or more tokens, it is split at the last occurrence of a newline token within 4,096 tokens
  * The first half of the split is treated as a single chunk
  * The above is repeated for the second half of the split