| # SimplexTasks-12 Benchmark Card |
|
|
| SimplexTasks-12 is the processed task collection behind the SimplexUQ benchmark. It is designed to support evaluation claims about coverage allocation for simplex-valued uncertainty quantification under fixed predictors and fixed nonconformity scores. |
|
|
| ## Supported Claims |
|
|
| - Whether a fixed simplex-valued predictor exhibits allocation failure beyond its marginal coverage. |
| - Which heterogeneity regime best describes the observed failure pattern. |
| - Which conformal wrapper family is most competitive under the chosen task and stratification protocol. |
|
|
| ## Claims The Benchmark Does Not Support |
|
|
| - Universal wrapper rankings across all simplex tasks. |
| - Deployment-readiness claims for any predictor. |
| - Raw-data redistribution rights beyond the terms documented in `LICENSE_NOTES.md`. |
|
|
| ## Benchmark Contents |
|
|
| - 6 synthetic regimes with canonical `task.npz` bundles and copied YAML configs. |
| - 6 fixed-predictor real tasks with cached `(Y, U)` arrays or derived features. |
| - Per-task `task_card.md` files and `metadata.json` provenance records. |
| - Release-level rebuild instructions for the paper tables and figures. |
|
|
| ## Reproducibility Contract |
|
|
| - Benchmark evaluation always operates on frozen predictor outputs. |
| - Default stratification rules are fixed before wrapper comparison. |
| - Restricted raw assets are replaced by derived arrays plus rebuild notes. |
| - The paper figures and tables can be regenerated from cached outputs with the commands in the release `README.md`. |
|
|
| ## Responsible Use |
|
|
| Some tasks rely on non-redistributable source assets or derived features from sensitive domains such as face images. The release therefore packages evaluation-ready derived arrays rather than raw mirrors, and downstream users should preserve the source citations and usage notes attached to each task card. |
|
|