---

name: frequency_label_mapping description: Build deterministic one-to-one target-to-source label mappings from source prediction frequencies.

Frequency Label Mapping

Use this skill when a visual prompting pipeline must map target classes to source classes using top-1 source prediction frequencies. It implements the FLM lower-level mapping used before prompt training and inside ILM-VP.

Inputs

• Predictions or logits grouped by target label.
• Ordered target labels and source labels.

Outputs

• Target-to-source mapping.
• Frequency/count table.
• Audit entries for tie-breaking, duplicate exclusion, and fallbacks.

Workflow

1. Convert logits to top-1 source labels when needed.
2. Count source labels for each target class.
3. In target-label order, choose the most frequent unused source label.
4. Break ties by supplied source-label order.
5. If no unused source labels remain, emit a duplicate warning.

Validation

Run python tests/test_frequency_label_mapping.py or validate_skill_tree.py --run-tests.

Limitations

FLM optimizes mapping frequency, not semantic quality. Use mapping diagnostics to explain whether selected labels are interpretable.