Score-Based Pruning Masks

Use this skill when a pruning experiment needs to convert importance scores into binary masks without depending on a paper repository. Do not use it for structured head/layer pruning unless the caller has already reduced each structure to one scalar score.

Inputs

• scores: a rectangular nested list or flat list of numeric scores.
• mode: top_v or threshold.
• keep_ratio: required for top_v, from 0 to 1.
• threshold: required for threshold; entries are kept only when score > threshold.

Outputs

A JSON-compatible object containing mask, metadata, and a flattened kept_indices list. The mask has the same shape as the input scores.

Workflow

1. Validate numeric score shape.
2. Flatten scores in row-major order.
3. For top_v, keep ceil(n * keep_ratio) entries after sorting by descending score and then stable index.
4. For threshold, keep scores strictly greater than the threshold.
5. Restore the original shape and report sparsity metadata.

This implements the paper abstraction M = Top_v(S) for hard movement pruning and M = (S > tau) for soft movement pruning.

Validation

Run python tests/test_score_masking.py or validate the full skill tree. The tests cover ties, negative scores, zero/full keep ratios, and strict threshold equality.

Limitations

The skill only computes masks. It does not train scores or patch neural-network layers.