README.md

metadata

license: cc-by-sa-4.0
task_categories:
  - conversational
  - translation
pretty_name: OLDI Seed lij-eng translation dataset (instruction-style)
dataset_info:
  features:
    - name: inputs
      dtype: string
    - name: targets
      dtype: string
    - name: template_id
      dtype: int64
    - name: template_lang
      sequence: string
  splits:
    - name: train
      num_bytes: 2381132
      num_examples: 5802
    - name: dev
      num_bytes: 79921
      num_examples: 189
    - name: test
      num_bytes: 87507
      num_examples: 202
  download_size: 1292161
  dataset_size: 2548560
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
      - split: test
        path: data/test-*

This is an Ligurian→English sentence-level translation dataset. The original data comes from the OLDI Seed dataset, and it has been converted to the instruction format.

The prompts, written in Ligurian, ask the model to translate the text to English. There are several variants of the prompt templates which were randomly sampled for each sentence:

Traduxi in ingleise: \<sentence>
Traduxi da-o zeneise à l’ingleise: \<sentence>
Traduxi da-o ligure à l’ingleise: \<sentence>
Traduxi sto testo in ingleise: \<sentence>
Traduxi in lengua ingleise: \<sentence>
Traduxi sto testo da-o zeneise à l’ingleise: \<sentence>
Traduxi sto testo da-o ligure à l’ingleise: \<sentence>
Comm’à l’é a traduçion ingleise de sto testo? \<sentence>
Quæ a l’é a traduçion ingleise de sto testo? \<sentence>
Ti peu tradue sto testo in ingleise? \<sentence>

The prompt template used for each dataset entry is referenced in the column template_id, with ids ranging from 1 to 10 according to the order given above. The targets are always prefixed with the string "A traduçion in ingleise do testo a l’é: <sentence>" ("The English translation of the sentence is:").

The correspondence between template_id, prompt template and target template is therefore:

[
    (1, "Traduxi in ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (2, "Traduxi da-o zeneise à l’ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (3, "Traduxi da-o ligure à l’ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (4, "Traduxi sto testo in ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (5, "Traduxi in lengua ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (6, "Traduxi sto testo da-o zeneise à l’ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (7, "Traduxi sto testo da-o ligure à l’ingleise:\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (8, "Comm’à l’é a traduçion ingleise de sto testo?\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (9, "Quæ a l’é a traduçion ingleise de sto testo?\n", ""A traduçion in ingleise do testo a l’é:\n"),
    (10, "Ti peu tradue sto testo in ingleise?\n", ""A traduçion in ingleise do testo a l’é:\n"),
]

The dataset contains 5802 train samples, 190 validation samples and 201 test samples.