Release v0.1.0 · sileod/tasksource

Breaking changes

Python ≥ 3.10 is required.
Readable labels. Many label sets now use readable names (e.g. liar, ShellRisk), boolean labels become "False"/"True", and some label orders changed. Code relying on label indices or old names may need updating.
Parked tasks. Evaluation benchmarks, duplicates and unloadable sources moved to tasksource.parked and are no longer in list_tasks(); they still load via parked.<name>.load().
Loaders. Many tasks now load from maintained mirrors or Hub parquet exports; a few task ids changed.

question on annotations, load_task(..., prompted=True), and dataset.question.
Recasts: recast="instruct" and recast="jev" (typed decisions), with per-example seeded sampling.
Provenance: hub_datasets() and task_provenance(id) name the loading repo and the original dataset behind tasksource copies.
metadata/weights.py sampling weights; ordinal=True for ordered label sets.
Many new and fixed tasks, including safety/prompt-injection sets, IntentGrasp, prm800k_dpo (split by MATH train/test) and multilingual sentiment.
tasks.md / mtasks.md: a one-line-per-task catalog.

Add modern classification, multilingual, relation and reward datasets by @sileod in #11
Native Jev sources: procedural generators and graded families by @sileod in #12
Add synthetic Jev dataset-construction package with live pilot sample by @sileod in #13
Language names for language ID tasks; fix unloadable tasks; support pandas 3 and huggingface_hub 1.x by @sileod in #14
Review fixes: strict Jev targets, seeded loading and recasts, Python ≥3.10, CI before publishing by @sileod in #15
Jev recast fixes from a sample audit: unshuffled free-text options, broken entities, HelpSteer2 turns by @sileod in #16
Jev graded: annotator distributions from UNLI, DynaSent, HateXplain, Measuring Hate Speech and LeWiDi by @sileod in #18

Full Changelog: v0.0.47...v0.1.0