goat / README.md
Egor Spirin
Add readme
1a7b3a1
metadata
license: apache-2.0
language:
  - ru
size_categories:
  - 1K<n<10K
dataset_info:
  - config_name: sociology_single_choice
    features:
      - name: input
        dtype: string
      - name: A
        dtype: string
      - name: B
        dtype: string
      - name: C
        dtype: string
      - name: D
        dtype: string
      - name: answer
        dtype:
          class_label:
            names:
              '0': A
              '1': B
              '2': C
              '3': D
  - config_name: sociology_multiple_choice
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: sociology_word_generation
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: sociology_matching
    features:
      - name: task_text
        dtype: string
      - name: option_1_name
        dtype: string
      - name: option_1
        dtype: string
      - name: option_2_name
        dtype: string
      - name: option_2
        dtype: string
      - name: answer
        dtype: string
  - config_name: sociology_text_generation
    features:
      - name: input
        dtype: string
      - name: criteria_table
        dtype: string
      - name: task_points
        dtype: int32
  - config_name: literature_multiple_choice
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: literature_word_generation
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: literature_matching
    features:
      - name: task_text
        dtype: string
      - name: option_1_name
        dtype: string
      - name: option_1
        dtype: string
      - name: option_2_name
        dtype: string
      - name: option_2
        dtype: string
      - name: answer
        dtype: string
  - config_name: literature_text_generation
    features:
      - name: input
        dtype: string
      - name: criteria_table
        dtype: string
      - name: task_points
        dtype: int32
  - config_name: russian_multiple_choice
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: russian_word_generation
    features:
      - name: input
        dtype: string
      - name: answer
        dtype: string
  - config_name: russian_matching
    features:
      - name: task_text
        dtype: string
      - name: option_1_name
        dtype: string
      - name: option_1
        dtype: string
      - name: option_2_name
        dtype: string
      - name: option_2
        dtype: string
      - name: answer
        dtype: string
  - config_name: russian_text_generation
    features:
      - name: input
        dtype: string
      - name: criteria_table
        dtype: string
      - name: task_points
        dtype: int32
configs:
  - config_name: sociology_single_choice
    data_files:
      - split: dev
        path: sociology/single_choice/dev/*.csv
      - split: test
        path: sociology/single_choice/test/*.csv
  - config_name: sociology_multiple_choice
    data_files:
      - split: dev
        path: sociology/multiple_choice/dev/*.csv
      - split: test
        path: sociology/multiple_choice/test/*.csv
  - config_name: sociology_word_generation
    data_files:
      - split: dev
        path: sociology/word_generation/dev/*.csv
      - split: test
        path: sociology/word_generation/test/*.csv
  - config_name: sociology_matching
    data_files:
      - split: dev
        path: sociology/matching/dev/*.csv
      - split: test
        path: sociology/matching/test/*.csv
  - config_name: sociology_text_generation
    data_files:
      - split: test
        path: sociology/text_generation/*.csv
  - config_name: literature_multiple_choice
    data_files:
      - split: dev
        path: literature/multiple_choice/dev/*.csv
      - split: test
        path: literature/multiple_choice/test/*.csv
  - config_name: literature_word_generation
    data_files:
      - split: dev
        path: literature/word_generation/dev/*.csv
      - split: test
        path: literature/word_generation/test/*.csv
  - config_name: literature_multiple_choice
    data_files:
      - split: dev
        path: literature/multiple_choice/dev/*.csv
      - split: test
        path: literature/multiple_choice/test/*.csv
  - config_name: literature_matching
    data_files:
      - split: dev
        path: literature/matching/dev/*.csv
      - split: test
        path: literature/matching/test/*.csv
  - config_name: literature_text_generation
    data_files:
      - split: test
        path: literature/text_generation/*.csv
  - config_name: russian_word_generation
    data_files:
      - split: dev
        path: russian/word_generation/dev/*.csv
      - split: test
        path: russian/word_generation/test/*.csv
  - config_name: russian_multiple_choice
    data_files:
      - split: dev
        path: russian/multiple_choice/dev/*.csv
      - split: test
        path: russian/multiple_choice/test/*.csv
  - config_name: russian_matching
    data_files:
      - split: dev
        path: russian/matching/dev/*.csv
      - split: test
        path: russian/matching/test/*.csv
  - config_name: russian_text_generation
    data_files:
      - split: test
        path: russian/text_generation/*.csv

🐐 GOAT

Generalized Occupational Aptitude Test (GOAT) is a dataset based on questions from Russian government exams that is required for every person graduated from school. Currently, the dataset cover questions from Literature, Sociology and Russian language subjects.

All questions are divided by expected output format:

  • Single choice. In such tasks, there is a set of possible answers from which you need to choose the right one. The answer to such tasks is one digit, which is the number of correct answer to the task.
  • Multiple choice. In such tasks, there is a set of possible answers from which you need to choose multiple correct answers. The answer to such tasks is a set of digits, which are the numbers of correct answers to the task. Note that the order of digits is not important.
  • Word generation. The answer to such tasks is a word or a phrase (multiple concrete words). The answer should not include spaces. Ground truth may include more than one correct word, in this case, they are concatenated with comma.
  • Matching. In such tasks, for each object in the given list property from the second list should be selected. The answer to such tasks is a sequence of digits, which represents the matching. Position of the digit represents object number in the list of objects and the digit itself is a number of corresponding property in the list of properties.
  • Text generation. These tasks require a detailed text answer, i.e. essay. There is no ground truth answer, only evaluation criteria and maximum possible score.

Navigate to dataset structure section for more details and examples.

Dataset Creation

All tasks from this dataset were parsed from sdamgia.ru website. Each task was normalized, i.e.:

  • All tables, lists, and other HTML structures were converted to plain text.
  • Based on task type, corresponding fields like input or answer were extracted.
  • Manual validation of samples

Parser's code is located in the deepvk/goat repo on GitHub.

Dataset statistic

For each task, except text generation, we provide dev and test splits. dev split consist of exactly 5 samples and can be used as few-shots. test statistic is represented in table:

Subject Single Choice Multiple Choice Word Generation Matching Text Generation Total
Literature 103 416 59 959 1537
Russian 3021 1008 57 430 4516
Sociology 1124 1624 29 638 826 4241

Dataset Structure

Single choice

  • input: a string question
  • A, B, C, D: four strings with possible answers
  • answer: Letter of correct answer

An example from sociology single choice tasks looks as follows:

{
  "input": "В отличие от природы, общество",
  "A": "обусловлено деятельностью человека",
  "B": "находится в постоянном развитии",
  "C": "состоит из взаимосвязанных элементов",
  "D": "является частью материального мира",
  "answer": "A"
}

Multiple choice and word generation

  • input: a string question
  • answer: correct answers concatenated with comma

An example from sociology multiple choice tasks looks as follows:

{
  "input": "В стране Z глава государства получает власть по наследству. Какие признаки позволяют сделать вывод о том, что в стране Z существует абсолютная монархия? Запишите соответствующие цифры.  1. Глава государства получает власть по наследству. 2. Граждане не обладают политическими правами и свободами. 3. Власть главы государства не ограничена законом. 4. Глава государства осуществляет свои полномочия пожизненно. 5. В стране отсутствует парламент. 6. Местные власти назначаются главой государства.",
  "answer": "2,3,5"
}

An example from literature word generation tasks looks as follows:

{
  "input": "Назовите художественное средство, усиливающее эмоциональное звучание строки «О жизни тленной, тленной и прекрасной». Прочитайте приведённое ниже произведение и выполните задание. Я научилась просто, мудро жить, Смотреть на небо и молиться Богу, И долго перед вечером бродить, Чтоб утомить ненужную тревогу. Когда шуршат в овраге лопухи И никнет гроздь рябины жёлто-красной, Слагаю я весёлые стихи О жизни тленной, тленной и прекрасной. Я возвращаюсь. Лижет мне ладонь Пушистый кот, мурлыкает умильней, И яркий загорается огонь На башенке озерной лесопильни. Лишь изредка прорезывает тишь Крик аиста, слетевшего на крышу. И если в дверь мою ты постучишь, Мне кажется, я даже не услышу. А. А. Ахматова, 1912."
  "answer": "повтор,лексическийповтор"
}

Matching

  • task_text: a string question
  • option_1_name: name of objects in the option_1 list
  • option_1: list of strings
  • option_2_name: name of properties in the option_2 list
  • option_2: list of strings
  • answer: sequence of digits

An example from sociology matching tasks looks as follows:

{
  "task_text": "Установите соответствие между примерами и видами налогов: к каждой позиции, данной в первом столбце, подберите соответствующую позицию из второго столбца.",
  "option_1_name": "ПРИМЕР",
  "option_1": ["А) налог на добавленную стоимость", "Б) налог на прибыль", "В) налог на имущество физических лиц", "Г) земельный налог", "Д) акцизы"],
  "option_2_name": "ВИД НАЛОГА",
  "option_2": ["1) прямые налоги", "2) косвенные налоги"],
  "answer": "21112"
}

Text generation

  • input: list consisting of one string that is a task text.
  • criteria_table: list of criteria to properly grade the task.
  • task_points: maximum points for the task.

An example from sociology text generation tasks looks as follows:

{
  "input": "Какие два из перечисленных понятий используются в первую очередь при описании экономической сферы общества?  Государство; акционерное общество; предпринимательство; тоталитаризм; социальная мобильность.  Выпишите соответствующие понятия и раскройте смысл любого одного из них.",
  "criteria_table": "Критерии оценивания:
   Правильно выписаны два верных понятия, и раскрыт смысл любого одного из них - 2 балла.
   Наряду с верными понятиями выписано(ы) одно или более «лишних» понятий, раскрыт смысл верного понятия.ИЛИ Правильно выписаны только два верных понятия.ИЛИ Правильно выписано только одно верное понятие, раскрыт его смысл - 1 балл.
   Наряду с верными понятиями выписано(ы) одно или более «лишних» понятий, раскрыт только смысл «лишнего» понятия.ИЛИ Наряду с верными понятиями выписано(ы) одно или более «лишних» понятий, смысл понятия не раскрыт или раскрыт неверно.ИЛИ Выписано только одно верное понятие.ИЛИ Приведены рассуждения общего характера, не соответствующие требованию задания.ИЛИ Ответ неправильный - 0 баллов.
Максимальный балл - 2 балла."
  "task_points": 2
}

Usage

Possible uses of the dataset include integration into popular LLM evaluation frameworks.

LM Evaluation Harness

This framework is useful for single choice, multiple choice, matching, and word generation tasks. See examples of integration, including custom metrics, in deepvk/lm-evaluation-harness fork on GitHub.

For example:

accelerate launch -m lm_eval --model hf --model_args <YOUR_MODEL_NAME> --tasks sociology_single_choice

Will run evaluation of specified model on sociology single choice task.

LLM as Judge

Another option is to use LLM to evaluate produced answers, i.e. for text generation task. See example of integration in deepvk/FastChat fork on GitHub.

Citation

@misc{deepvk2024goat,
    title={GOAT: Generalized Occupational Aptitude Test},
    author={Vyrodov, Mikhail and Spirin, Egor},
    url={https://huggingface.co/datasets/deepvk/goat},
    publisher={Hugging Face}
    year={2024},
}