---

name: openclip_scale_protocol description: Build and validate OpenCLIP-style scale tables for CLIP scaling-law experiments, including total compute and error metric derivation.

OpenCLIP Scale Protocol

Use this skill when a recovery or analysis needs to represent CLIP scale points from the paper: dataset scale, model scale, samples seen, compute per sample, and downstream metrics.

Inputs

• JSON list of records with dataset, model, samples_seen, gmac_per_sample, and optional accuracy or recall_at_5 values.
• Metrics should be percentages, not fractions.

Outputs

• JSON with validated records, total_compute, and derived error columns.

Workflow

1. Validate positive samples_seen and gmac_per_sample.
2. Compute total_compute = samples_seen * gmac_per_sample.
3. Convert accuracy to classification_error = 100 - accuracy.
4. Convert recall_at_5 to retrieval_error = 100 - recall_at_5.
5. Preserve raw records for downstream power-law fitting.

Validation

Run python tests/test_scale_protocol.py.

Limitations

This skill structures scale evidence; it does not train CLIP models or claim full LAION-scale reproduction.