--- pretty_name: tasksource-jev language: - multilingual license: other task_categories: - text-classification - question-answering tags: - tasksource - jev - system-one - runtime-defined-decisions - decision-models - multiple-choice size_categories: - 100K **Build status:** a 105,000-example preview is currently available while the > source-balanced 500,000-row release is built from 730 selected English and > multilingual tasks. Checkpointed successes are retained and incompatible > upstream sources are reported explicitly. ## Schema | field | type | meaning | |---|---|---| | `id` | string | Stable identifier derived from task, split, and row index. | | `kind` | string | System One primitive: `choice`, `noul`, or `score`. | | `options` | list of strings | Candidate labels or answers supplied at runtime. Order is significant. | | `target` | list of floats | One-hot target distribution aligned with `options`. | | `state` | string | Text or question on which the decision is based. Paired classification inputs are marked `text_A` and `text_B`. | | `question` | string | The decision requested from the model. | | `source` | string | Tasksource task identifier used to load the source data. | | `variant` | string | Direct decision or a named deterministic subrecast. | | `split` | string | Source split normalized to `train`, `dev`, or `test`. | Classification criteria are the source task's label names. Multiple-choice criteria are the answer choices. Every source row is retained as a direct `choice`. A deterministic augmentation pass adds label-verification `noul` questions to about 5% of rows. Ordered-rubric `score` augmentation is available for genuinely ordinal sources but is disabled by default. Native regression and ordinal recasting will be used for score examples rather than imposing an order on nominal classification labels. ### Deterministic subrecasts The release adds two conservative, low-frequency variants while retaining every direct row: | variant | default rate | purpose | |---|---:|---| | `label_verification` | 5% | A `noul` judgement asking whether a deterministically proposed label is correct. Correct and incorrect proposals are balanced. | | `criteria_permutation` | 5% | The same `choice` decision with options and targets permuted together, reducing option-position shortcuts. | | `instruction_paraphrase` | 5% | The same decision with a manually vetted equivalent instruction. Common NLI and sentiment label groups receive specific wording; other tasks use conservative generic alternatives. | | `paired_text_format` | 5% of paired rows | Neutral alternatives to repeated `text_A`/`text_B` field labels, without assuming a task-specific relation between the texts. | All transformations are derived exactly from the source target and introduce no teacher-generated claims. Candidate-subset decisions are a possible later addition. Synthetic uncertainty, abstention, and nominal-to-ordinal conversions are intentionally excluded because one-hot classification labels do not justify them. For `noul`, `target` contains the scalar truth probability and `options` is empty; for `choice` and `score`, `target` is aligned with `options`. The direct canonical decision is unchanged; lower-frequency variants are identified explicitly by the `variant` field. ```python from datasets import load_dataset dataset = load_dataset("tasksource/tasksource-jev") row = dataset["train"][0] answer = row["options"][max(range(len(row["target"])), key=row["target"].__getitem__)] ``` With Tasksource installed, the same representation can be produced directly: ```python from tasksource import load_task, render_systemone dataset = load_task("glue/rte", recast="jev") request = render_systemone(dataset["train"][0], model="openjev") ``` ## Construction Tasksource standardizes heterogeneous datasets into common classification and multiple-choice templates. This release applies `recast_jev` to compatible English and multilingual tasks, retains the standard train/validation/test splits, and records the Tasksource identifier in every row. Tasks that fail to download or preprocess are recorded by the build report rather than silently represented as complete. To keep very large sources balanced, the build caps each task at 30,000 training rows and 3,000 validation or test rows using Tasksource's deterministic sampling (seed 0). The published release is capped at 500,000 rows using a source-balanced 90/5/5 train/dev/test allocation; selection preserves relative row order. For a useful Dataset Viewer preview, only the first 1,000 training rows are ordered round-robin by `source`. This is a deterministic permutation, not a random shuffle. After that display prefix, all remaining examples retain their original relative order. The selected catalog includes 100 BIG-bench task configurations and all 57 MMLU subjects currently registered in Tasksource. Their original split identity is preserved in the row-level `split` field, with `validation` normalized to `dev`. This makes source/split exclusion explicit when constructing a training mixture. The repository includes `failed-tasks.json` and `outdated-datasets.json`. The latter specifically tracks upstream datasets that still depend on loading scripts no longer supported by current Hugging Face Datasets, so they can be migrated to data-only Parquet repositories and incorporated in a later build. The build is reproducible from the Tasksource repository: ```bash python scripts/build_jev_dataset.py --output build/tasksource-jev --finalize ``` ## Licensing and provenance Tasksource is a preprocessing framework and catalog, not the original publisher of the constituent datasets. Copyright, license, and usage restrictions remain those of each upstream dataset. Users should consult the upstream dataset card identified by `source` before redistributing or using a subset. The aggregate is therefore marked `license: other`; no single license is asserted over all rows. ## Citation If this recast is useful, cite Tasksource, which provides the task collection and harmonization framework: ```bibtex @inproceedings{sileo-2024-tasksource, title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework", author = "Sileo, Damien", editor = "Calzolari, Nicoletta and Kan, Min-Yen and Hoste, Veronique and Lenci, Alessandro and Sakti, Sakriani and Xue, Nianwen", 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/", pages = "15655--15684", abstract = "The HuggingFace Datasets Hub hosts thousands of datasets, offering exciting opportunities for language model training and evaluation. However, datasets for a specific task type often have different structures, making harmonization challenging which prevents the interchangeable use of comparable datasets. As a result, multi-task training or evaluation necessitates manual work to fit data into task templates. Several initiatives independently tackle this issue by releasing harmonized datasets or providing harmonization codes to preprocess datasets into a consistent format. We identify patterns in such preprocessings, such as column renaming, or more complex patterns. We then propose an annotation framework that enables concise, readable, and reusable preprocessing annotations. tasksource annotates more than 600 task preprocessings and provides a backend to automate dataset alignment. We fine-tune a multi-task text encoder on all tasksource tasks, outperforming every publicly available text encoder of comparable parameter count according to an external evaluation." } ```