Skip to content

Dataset · tasksource

Typed decisions from hundreds of Tasksource tasks

tasksource/tasksource-jev-typed-decisions

Millions of classification, choice, rating, and visual reasoning examples for multi-task learning.

What's inside

Contains 1M to 10M examples in English and other languages, combining text and visual modalities. The data represents typed decisions (classification, multiple choice, ranked ratings, probabilities, and visual reasoning) derived from annotations of hundreds of tasks in the Tasksource collection. The card notes that original annotation noise and subjective ratings are preserved, and text and visual parts can be loaded separately via the datasets library.

What you can do with it

  • Train and fine-tune multimodal and language models on a wide range of tasks (zero-shot and text classification, QA)
  • Train models to follow structured output formats (probabilities, ratings, option selection)
  • Evaluate the generalization ability of models on diverse annotation types
Tasks
Zero shot classification, Text classification, Question answering, Token classification
Languages
English, Multilingual
Size
1M–10M rows
Downloads this month
17.9K
Likes
34

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("tasksource/tasksource-jev-typed-decisions")