median_proximity_coreset
Select Moderate-DS coresets by keeping examples whose scalar scores are closest to the score median.
SKILL.mdbasicagent-skill/skill_library/paperbench/lbcs/skill/moderate_coreset/median_proximity_coreset/SKILL.md
GitHubMedian Proximity Coreset
Use this skill after scores have been computed for candidate examples. It implements the paper's moderate coreset rule: sort by closeness to the score median rather than selecting the smallest or largest scores. Do not use this skill to compute representation scores or to claim downstream performance.
Inputs
- A list of score records containing
idand numericscore. - Either an integer
sizeor a ratio convertible to a target size by the caller.
Outputs
- Selected ids.
- The median score.
- Ranked proximity diagnostics with deterministic tie-breaking.
Workflow
- Validate all scores and requested size.
- Compute the conventional median of all scores.
- Rank records by absolute distance to the median, then by score and id for deterministic ties.
- Return exactly the requested number of selected ids.
Validation
Run python scripts/select_median_coreset.py --self-test. Tests include even medians, stable ties, and invalid sizes.
Limitations
The skill operates on scalar scores only. A future run should combine it with a scoring skill and an evaluation skill for complete Moderate-DS recovery.