---

name: score_based_subset_selection description: Select retained and pruned supervised-training examples from EL2N or GraNd scores using deterministic pruning rules.

Score-Based Subset Selection

Use this skill after EL2N or GraNd scores have been computed and a pruning experiment needs a reproducible retained subset. It implements the paper's standard rule of pruning low-score examples and retaining high-score examples, plus an offset-window mode for noise-analysis style checks.

Inputs

• Example records with id and numeric score.
• retain_fraction or retain_count.
• Optional mode: high_score or offset_window.
• Optional offset_fraction for offset-window selection.

Outputs

• Selected example ids.
• Pruned example ids.
• Selection statistics and tie policy.

Workflow

1. Validate unique ids and finite scores.
2. Sort by descending score for standard high-score retention.
3. Convert fraction to an exact selected count with floor and minimum one.
4. Apply deterministic tie-breaking by id string.
5. Return selected/pruned complements and score-boundary statistics.

Validation

Run:

python tests/test_selection.py
python scripts/select_subset.py --fixture

Limitations

• This skill consumes scores; it does not compute EL2N/GraNd.
• It does not train models or decide whether final accuracy is acceptable.