---

name: coverage_stratified_selection description: Select CCS-style coresets with mislabel filtering and deterministic score-stratified coverage.

Coverage Stratified Selection

Use this skill after per-example importance scores are available and a high-pruning coreset must preserve distribution coverage. It implements the Coverage-centric Coreset Selection mechanism in a standard-library form: remove suspicious low-margin examples, bin the remaining examples by score, allocate budget across bins, and sample deterministically within each bin.

Do not use this skill as a generic top-k selector when coverage is irrelevant. Do not read or depend on the original paper repository during recovery.

Inputs

• A score table with indices, targets, and a score key such as accumulated_margin.
• coreset_ratio in (0, 1] and optional mis_ratio in [0, 1).
• strata, defaulting to 50 for paper-style behavior but often smaller in tests.
• A deterministic seed for repeatable within-bin selection.

Outputs

• Selected original indices.
• Removed suspicious indices from the mislabel/easy-data mask.
• Strata budgets, represented bins, and class counts for mechanism checks.

Workflow

1. Sort accumulated margin ascending and remove mis_ratio of the lowest-margin examples.
2. Compute a target coreset size from the original dataset size and coreset_ratio.
3. Bin remaining examples by the chosen score key.
4. Allocate budget across non-empty bins with the CCS low-occupancy-aware strategy.
5. Select deterministic representatives within each bin and emit diagnostics.

Validation

Run python tests/test_coverage_stratified_selection.py or the Distiller skill-tree validator with tests enabled.

Limitations

The standard-library implementation is designed for reproducible skill execution and recovery proxies. Full-scale image training can use the same contract around tensor-backed scores.