--- pretty_name: SimplexTasks-12-data license: other task_categories: - other tags: - conformal-prediction - uncertainty-estimation - simplex - benchmark - task-collection configs: - config_name: tasks data_files: - split: train path: data/viewer/tasks.jsonl - config_name: affectivetext_emotions data_files: - split: train path: data/parquet/affectivetext_emotions.parquet - config_name: cifar10_softmax data_files: - split: train path: data/parquet/cifar10_softmax.parquet - config_name: pbmc3k_pseudobulk data_files: - split: train path: data/parquet/pbmc3k_pseudobulk.parquet - config_name: samson_unmixing data_files: - split: train path: data/parquet/samson_unmixing.parquet - config_name: topics_20ng data_files: - split: train path: data/parquet/topics_20ng.parquet - config_name: utkface_age_ldl data_files: - split: train path: data/parquet/utkface_age_ldl.parquet - config_name: d1_homogeneous data_files: - split: train path: data/parquet/d1_homogeneous.parquet - config_name: d2_pure_scale data_files: - split: train path: data/parquet/d2_pure_scale.parquet - config_name: d3_discrete_groups_aligned data_files: - split: train path: data/parquet/d3_discrete_groups_aligned.parquet - config_name: d4_model_bias data_files: - split: train path: data/parquet/d4_model_bias.parquet - config_name: d5_heavy_tail data_files: - split: train path: data/parquet/d5_heavy_tail.parquet - config_name: d6_high_k data_files: - split: train path: data/parquet/d6_high_k.parquet license_name: task-specific-source-terms license_link: https://huggingface.co/datasets/liangyou03/SimplexTasks-12-data/blob/main/LICENSE --- # SimplexTasks-12-data Dataset-style artifact for the SimplexUQ benchmark. > **Please cite:** If you use SimplexUQ or SimplexTasks-12 in your research, please cite our paper: Liang You, Hengyu Shi, and Dongwen Ou. [SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions](https://arxiv.org/abs/2610.00523), 2026. ## Official repository Authors: Liang You, Hengyu Shi, and Dongwen Ou. - Code: [liangyou03/SimplexUQ](https://github.com/liangyou03/SimplexUQ) - Dataset: [liangyou03/SimplexTasks-12-data](https://huggingface.co/datasets/liangyou03/SimplexTasks-12-data) - Corresponding author: Liang You (liangyou03@pitt.edu). This personal-account release is synchronized from the anonymous review artifact at revision `a3c1f1f3975cecaae991e8dd256c2e66cd42adf6`. The frozen data arrays are unchanged; this is a repository and metadata update, not a new experiment run. The source repository is [anonymous0523ly/SimplexTasks-12-data](https://huggingface.co/datasets/anonymous0523ly/SimplexTasks-12-data). Archived result summaries retain their original experiment protocols; they are not a claim that every historical summary matches the final manuscript. The repository retains the source artifact's license metadata. This does not override task-specific source terms or grant new rights to third-party assets; consult `LICENSE`, `ASSET_PROVENANCE_AND_TERMS.md`, and the task cards. This bundle is artifact-first. Its primary contents are: - `task_cards/` with one card per synthetic or real task - `configs/` with method specs, benchmark task specs, and split policies - `data/synthetic/` with frozen synthetic benchmark arrays - `data/derived/` with frozen real-task derived arrays - `data/summaries/` with metric summaries and figure/table inputs - `data/viewer/tasks.jsonl` as the flat table used by the Hugging Face Dataset Viewer - `docs/` with provenance, restricted-asset notes, and reviewer quickstart The benchmark is designed to run on frozen derived arrays and fixed predictor outputs. Restricted raw assets are intentionally excluded. ## Top-Level Files - `BENCHMARK_PROTOCOL.md` - `EVALUATION_PROTOCOL.md` - `ASSET_PROVENANCE_AND_TERMS.md` - `COMPUTE_ENV.md` - `REPRODUCIBILITY.md` - `MANIFEST.json` - `VALIDATION_REPORT.json` ## Real-Task Derived Data - `data/derived/cifar10_softmax/` - `data/derived/topics_20newsgroups/` - `data/derived/samson_unmixing/` - `data/derived/pbmc_pseudobulk/` - `data/derived/utkface_ldl/` - `data/derived/affectivetext/` ## Synthetic Data - `data/synthetic/D1/` - `data/synthetic/D2/` - `data/synthetic/D3/` - `data/synthetic/D4/` - `data/synthetic/D5/` - `data/synthetic/D6/` ## Important Release Rule The benchmark consumes frozen `(Y, U)` arrays, task-specific strata metadata, and release notes for restricted assets. It is not a raw-data mirror. ## Hugging Face Dataset Viewer The viewer-facing dataset is intentionally limited to `data/viewer/tasks.jsonl`. This flat JSONL table summarizes the 12 task cards with a uniform schema. The other JSON, JSONL, CSV, and NumPy files in this repository are artifact inputs for reproduction and should not be mixed into the Dataset Viewer table. ## Croissant-compatible tables The `data/parquet/` directory contains one lossless table for each of the 12 tasks. `sample_index` is the zero-based row index in the original frozen arrays. Columns `U_0`, `U_1`, ... and `Y_0`, `Y_1`, ... preserve the original component order. Synthetic tables additionally include `R` and `sigma_true`. Floating-point values and their stored precision are unchanged. The viewer split named `train` is a transport convention for a frozen sample pool, not a prescribed training split. Use the benchmark split policies for conformal evaluation. See `docs/parquet_verification.json` for source-array hashes, table hashes and exact round-trip verification. These are format conversions, not new experiment runs. Croissant metadata is provided under `croissant/`. ## License scope The repository uses the `other` license category to match its existing artifact license notice and `CITATION.cff`. The notice and task-specific source terms govern reuse; there is no blanket CC-BY-4.0 grant for this collection. Format conversion does not change those terms. See `LICENSE`, `ASSET_PROVENANCE_AND_TERMS.md` and the individual task cards.