Effyis/Table-Extraction

Table Extract Dataset

This dataset is designed to evaluate the ability of large language models (LLMs) to extract tables from text. It provides a collection of text snippets containing tables and their corresponding structured representations in JSON format.

Source

The dataset is based on the Table Fact Dataset, also known as TabFact, which contains 16,573 tables extracted from Wikipedia.

Schema:

Each data point in the dataset consists of two elements:

• context: A string containing the text snippet with the embedded table.
• answer: A JSON object representing the extracted table structure. The JSON object follows this format: { "column_1": { "row_id": "val1", "row_id": "val2", ... }, "column_2": { "row_id": "val1", "row_id": "val2", ... }, ... } Each key in the JSON object represents a column header, and the corresponding value is another object containing key-value pairs for each row in that column.

Examples:

Example 1:

Context:

Answer:

{
  "date": {
    "0": "1st",
    "1": "3rd",
    "2": "4th",
    "3": "11th",
    "4": "17th",
    "5": "24th",
    "6": "25th"
  },
  "opponent": {
    "0": "bracknell bees",
    "1": "slough jets",
    "2": "slough jets",
    "3": "wightlink raiders",
    "4": "romford raiders",
    "5": "swindon wildcats",
    "6": "swindon wildcats"
  },
  "venue": {
    "0": "home",
    "1": "away",
    "2": "home",
    "3": "home",
    "4": "home",
    "5": "away",
    "6": "home"
  },
  "result": {
    "0": "won 4 - 1",
    "1": "won 7 - 3",
    "2": "lost 5 - 3",
    "3": "won 7 - 2",
    "4": "lost 3 - 4",
    "5": "lost 2 - 4",
    "6": "won 8 - 2"
  },
  "attendance": {
    "0": 1753,
    "1": 751,
    "2": 1421,
    "3": 1552,
    "4": 1535,
    "5": 902,
    "6": 2124
  },
  "competition": {
    "0": "league",
    "1": "league",
    "2": "league",
    "3": "league",
    "4": "league",
    "5": "league",
    "6": "league"
  },
  "man of the match": {
    "0": "martin bouz",
    "1": "joe watkins",
    "2": "nick cross",
    "3": "neil liddiard",
    "4": "stuart potts",
    "5": "lukas smital",
    "6": "vaclav zavoral"
  }
}

Example 2:

Context:

Answer:

{
  "country": {
    "exonym": {
      "0": "iceland",
      "1": "indonesia",
      "2": "iran",
      "3": "iraq",
      "4": "ireland",
      "5": "isle of man"
    },
    "endonym": {
      "0": "ísland",
      "1": "indonesia",
      "2": "īrān ایران",
      "3": "al - 'iraq العراق îraq",
      "4": "éire ireland",
      "5": "isle of man ellan vannin"
    }
  },
  "capital": {
    "exonym": {
      "0": "reykjavík",
      "1": "jakarta",
      "2": "tehran",
      "3": "baghdad",
      "4": "dublin",
      "5": "douglas"
    },
    "endonym": {
      "0": "reykjavík",
      "1": "jakarta",
      "2": "tehrān تهران",
      "3": "baghdad بغداد bexda",
      "4": "baile átha cliath dublin",
      "5": "douglas doolish"
    }
  },
  "official or native language(s) (alphabet/script)": {
    "0": "icelandic",
    "1": "bahasa indonesia",
    "2": "persian ( arabic script )",
    "3": "arabic ( arabic script ) kurdish",
    "4": "irish english",
    "5": "english manx"
  }
}