Datasets:
Tasks:
Tabular Classification
Formats:
parquet
Size:
100K - 1M
Tags:
calibration
distribution-shift
concept-drift
robustness
out-of-distribution-detection
selective-prediction
License:
| from typing import Optional | |
| from pydantic import BaseModel, ConfigDict, Field | |
| class DriftCalRow(BaseModel): | |
| sample_id: str | |
| scenario_id: str | |
| split: str | |
| shift_type: str = Field(description="none|covariate|label|concept|novel_class|adversarial") | |
| shift_magnitude: float | |
| difficulty: str = Field(description="easy|medium|hard|expert") | |
| label_binary: int = Field(ge=0, le=1) | |
| label_family: str | |
| family_id: int | |
| family_is_novel: bool | |
| perturbation_applied: bool | |
| perturbation_budget: float | |
| spurious_shuffled: bool = Field( | |
| description="True only in the spurious_ablation eval, where spurious " | |
| "feature columns have been permuted across rows to break their " | |
| "correlation with the label" | |
| ) | |
| generator_seed: int | |
| timestamp_synth: int | |
| provenance: str = "synthetic:driftcal-v0.1-validation" | |
| qc_status: str = "unchecked" | |
| # feature_* columns are validated separately (70 numeric columns) rather | |
| # than declared individually here to keep the schema readable; see | |
| # DATASET_SCHEMA.md for the full per-feature-block description. | |
| model_config = ConfigDict(extra="allow") | |