README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: test
        path: test.jsonl
      - split: dev
        path: dev.jsonl
      - split: dev_fewshot
        path: dev_fewshot.jsonl
license: cc-by-nc-sa-4.0

GSI:detect unified multiple-choice dataset

Each JSONL row represents one original GSI:detect text and has this schema:

{
  "Id": 1,
  "Question": "...",
  "Values": {"A": "0", "B": "0.25", "C": "0.5", "D": "0.75", "E": "1", "Correct": "A"},
  "Categories": {"A": "role", "B": "personality", "C": "competence", "D": "physical", "E": "sexual", "F": "relational", "Correct": "C"}
}

For original rows with gs_category: "no", Categories is JSON null rather than a seventh category. The Values mapping is always A=0, B=0.25, C=0.5, D=0.75, E=1. The category mapping is always A=role, B=personality, C=competence, D=physical, E=sexual, F=relational.

test.jsonl contains the original 810 test rows in original order and is assigned IDs 1--810. dev.jsonl contains the original 200 development rows in original order and is assigned IDs 811--1010. No items are moved between splits or removed.

dev_fewshot.jsonl contains copies of six dev items. Its first five items are the first occurrence in the original dev split of each GS value, ordered by value: 0, 0.25, 0.5, 0.75, 1. The corresponding source IDs are GS0516, GS0458, GS0953, GS0432, and GS0414 respectively. The sixth item is GS0479 (GS value 0.5, category personality): it is the first original dev item whose non-null category is not represented among the first five items. With a first_n sampler, the Value task therefore uses the first five examples; the Category task filters out the null-category GS0516 item and retains five categorized examples.

Citation

If you use this dataset, please cite:

Comandini, Gloria, Manuela Speranza, Sofia Brenna, Davide Testa, Stefania Cavagnoli, and Bernardo Magnini. 2026. "GSI:detect at EVALITA 2026: Overview of the Task on Detecting Gender Stereotypes in Italian." In Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026), pages 96–108, Bari, Italy. CEUR Workshop Proceedings.

Official task website: https://gsi-d-evalita.fbk.eu/

Credits

EVALITA-LLM dataset adaptation: Pietro Botarelli.