---

name: imagenetv2_selection_frequency_sampling description: Sample class-balanced ImageNetV2 proxy datasets using MatchedFrequency, Threshold0.7, or TopImages selection-frequency strategies.

ImageNetV2 Selection Frequency Sampling

Use this skill after candidate records have been validated. It implements the paper's key sampling ablation: final ImageNetV2 variants differ by how candidate MTurk selection frequencies are used.

Inputs

• Normalized candidate JSON from imagenetv2_candidate_pool_schema or an equivalent list.
• Strategy: matched_frequency, threshold_0_7, or top_images.
• Per-class sample count.

Outputs

• Sampled records.
• Strategy statistics including class coverage and average selection frequency.

Workflow

1. Group candidates by class.
2. For top_images, choose highest selection frequencies with id tie-breaks.
3. For threshold_0_7, filter to frequency at least 0.7, then choose highest frequencies.
4. For matched_frequency, sort by distance to target frequency bins derived from the class pool quantiles.
5. Emit deterministic samples and warnings for underfilled classes.

Validation

Run python scripts/sample_by_frequency.py tests/fixtures/normalized.json --strategy top_images --per-class 1 --output /tmp/sample.json.

Limitations

This skill samples metadata records; it does not fetch images or collect MTurk annotations.