metadata

dataset_info:
  features:
    - name: sent_text
      dtype: string
    - name: sent_id
      dtype: string
    - name: token_ids
      sequence: string
    - name: token_forms
      sequence: string
    - name: token_lemmas
      sequence: string
    - name: token_upos
      sequence: string
    - name: token_xpos
      sequence: string
    - name: token_feats
      sequence: string
    - name: token_head
      sequence: int32
    - name: token_deprels
      sequence: string
    - name: token_deps
      sequence: string
    - name: token_miscs
      sequence: string
    - name: sent_metadata
      dtype: string
    - name: corpus
      dtype: string
  splits:
    - name: train
      num_bytes: 57596921
      num_examples: 27219
    - name: validation
      num_bytes: 7132032
      num_examples: 2595
    - name: test
      num_bytes: 6901121
      num_examples: 2763
  download_size: 9544323
  dataset_size: 71630074
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc
language:
  - grc

Aggregated Universal Dependencies for Ancient Greek

Based on the Github r2.15 release of all Ancient Greek corpora, namely PROIEL, Perseus, PTNK. Parsed with conllu.

To ensure compatibility with Hugging Face datasets and the underlying Arrow format, some quirks are present in the dataset:

the ID field (token_ids) is a string rather than an integer. This is needed because in rare cases the ID can be a token range like 1-2.
because features can contain arbitrary content that is only semi-structured (not consistent across tokens), it is dumped as a JSON string for each token. So the token_feats field is a list of strings where each string should be JSON-decoded before use.
the same is true for token_miscs
and also for sent_metadata which contains the sentence metadata apart from the text and ID

License

Note that not all corpora have the same license. Only use those that you can comply with and filter as necessary on the corpus field:

"PROIEL": "cc-by-nc-sa-4.0"
"Perseus": "cc-by-nc-sa-2.5"
"PTNK": "cc-by-sa-4.0"