gvlassis's picture
val->validation fix, cleaned .txt
63455db
metadata
dataset_info:
  features:
    - name: text
      dtype: string
  splits:
    - name: train
      num_examples: 18
    - name: validation
      num_examples: 2
    - name: test
      num_examples: 4
configs:
  - config_name: default
    data_files:
      - split: train
        path: res/parquet/train.parquet
      - split: validation
        path: res/parquet/val.parquet
      - split: test
        path: res/parquet/test.parquet
task_categories:
  - text-generation
language:
  - el
license: mit

Ancient Greek Theatre

About

🏛️ A collection of 24 Ancient Greek plays, translated in Modern Greek.

Description

The dataset is basically shakespearefirstfolio, which was itself a better Tiny Shakespeare, but for Ancient Greek theatre plays and Modern Greek. Remarkably, and similarly to Karpathy, we observe that language models trained from scratch on this tiny dataset can produce samples that look eerily close to those written by Ancient Greek playwrights. You can see an example below:

ΑΝΤΙΓΟΝΗ

The dataset was collected as follows: I first looked for free Modern Greek translations of complete Ancient Greek plays on openbook.gr. For every PDF file found, I extracted the text with pdfly. The resulting .txt files undergone several rounds of manual and automatic cleaning via vim and vimscript. I also filled in missing parts (e.g characters, plot) from Wikipedia. This was very laborious, but I stuck with it until I was content with the result. As a final preprocessing step, I applied NFKD normalization via uconv. This lead to a very clean (albeit not perfect) collection of 24 Ancient Greek plays. The plays were written by Aristophanes, Euripides, Aeschylus and Sophocles. A complete list of the plays can be found in ./res.

Usage

import datasets

dataset = datasets.load_dataset("gvlassis/ancient_greek_theatre")