tasksource is a collection of decision tasks: classification, multiple choice and token labeling, with negative examples or classes, standardized so that they can be used interchangeably for multi-task training, evaluation and reward modeling. Labels come from human annotation or from procedures that verify them, and the focus is on training data: evaluation benchmarks are kept separate.

Hugging Face Datasets is an excellent library, but datasets lack a common format and often need preprocessing before they can be used together. Each tasksource annotation maps a dataset to one of three templates with canonical fields: Classification, MultipleChoice or TokenClassification. An annotation is a short, readable function from a dataset to its standardized form.

• 570 English, 550 multilingual and 45 vision tasks: browse the English, multilingual and vision catalogs.
• Soft labels (annotator votes and rating distributions), license filters, and per-task questions that say what to decide.
• Recasts to typed decisions (tasksource-jev-typed-decisions), instructions and label NLI.

pip install tasksource

from tasksource import list_tasks, load_task

df = list_tasks(multilingual=False) # takes some time
for id in df[df.task_type=="MultipleChoice"].id:
    dataset = load_task(id) # all yielded datasets can be used interchangeably

The example catalog shows one real example per task: input, question, options and gold answer. Evaluation benchmarks are kept out of the training catalogs; load them with from tasksource import eval_only. Feel free to request a new task.

Datasets are downloaded to $HF_DATASETS_CACHE, like any Hugging Face dataset; large multi-task runs need more than 100GB.

list_tasks(vision=True) lists the visual tasks and load_task(id, vision=True) loads them, with images.

list_tasks(extractive=True) lists the span-extraction tasks (SQuAD 2.0, Natural Questions, HotpotQA, Quoref, MuSiQue...); load_task(id, extractive=True) loads them in SQuAD format (context, question, answers) for AutoModelForQuestionAnswering, with every answer span checked against the context.

Inputs are kept raw. When they do not say what to predict, a task carries a question ("Is this search query a well-formed question?"), exposed as dataset.question; load_task(id, prompted=True) appends it to the inputs.

Some tasks are distributions rather than single labels: annotator votes, rater shares, survey counts. load_task(id, soft=True) loads them as probabilities over the options, and list_tasks(soft=True) lists them. By default, load_task(id) gives the majority label on rows with clear agreement.

list_tasks(license_use="commercial") keeps tasks whose sources allow commercial use (also non-commercial, unspecified, or a list), and task_licenses() gives each task's licenses and their sources. This is a best-effort filter, not legal advice.

The tasksource code is under the BSD 3-Clause license. Original documentation and the task annotations we write (mappings, questions, label names, recasts) are under CC BY 4.0. Third-party datasets, including the sources that tasksource loads and our mirrors or repackages of them, keep their own licenses, which task_licenses() reports.

Tasks can be recast as typed decisions: a state and one or more questions of three kinds, choice (pick one option), score (an ordered scale) and noul (yes/no or a probability). This is the Jev / System One request format.

🤗 tasksource/tasksource-jev-typed-decisions is the published corpus:

It includes 🤗 tasksource/procedural-typed-decisions: generated states (ledgers, schedules, graphs, policies, records) with exact labels at five difficulty levels.

from tasksource import load_task, render_typed_decision

dataset = load_task("glue/rte", recast="jev")
request = render_typed_decision(dataset["train"][0], model="openjev")

Multiple-choice options are shuffled per row and over-represented labels are capped, so neither the gold position nor the most common label pays. Benchmarks such as MMLU, BIG-bench and BLiMP are excluded. See the dataset card for the row format and the mixes.

Text encoder pretrained on tasksource reached state-of-the-art results: 🤗/deberta-v3-base-tasksource-nli

Tasksource pretraining is notably helpful for RLHF reward modeling or any kind of classification, including zero-shot. You can also find a large and a multilingual version.

Tasks recast as instructions (recast="instruct"), one of the richest instruction-tuning datasets: 🤗/tasksource-instruct

Classification tasks recast as natural language inference, for entailment-based zero-shot classification: 🤗/zero-shot-label-nli

from tasksource import MultipleChoice

codah = MultipleChoice('question_propmt',choices_list='candidate_answers',
    labels='correct_answer_idx',
    dataset_name='codah', config_name='codah')

winogrande = MultipleChoice('sentence',['option1','option2'],'answer',
    dataset_name='winogrande',config_name='winogrande_xl',
    splits=['train','validation',None]) # test labels are not usable

tasks = [winogrande.load(), codah.load()]  # aligned datasets (same columns) can be used interchangeably

For more details, refer to the article:

@inproceedings{sileo-2024-tasksource,
    title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework",
    author = "Sileo, Damien",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1361",
    doi = "10.63317/2h4pkr685mkq",
    pages = "15655--15684",
}

For help integrating tasksource into your experiments, please contact damien.sileo@inria.fr.