dataset-sourcinglisted

# Dataset Sourcing A benchmark dataset that doesn't match the real deployment distribution will produce a model that looks great in evaluation and fails in the field — this is the single most common way ML projects quietly fail. This skill checks the match, not just the existence, of candidate datasets. ## Step 1 — Define the real deployment distribution first Before searching, state what real input actually looks like in production: sensor noise characteristics, real-world class balance, the actual environmental/population variation the deployed system will see. A dataset search with no stated acceptance criteria isn't a search, it's browsing. ## Step 2 — Search across the layers that matter - Domain-specific benchmark datasets for the exact task. - Adjacent-domain datasets that could transfer, when nothing matches exactly. - Where nothing fits well: what it would actually take to collect or synthesize a small, correctly-distributed set instead of forcing a mismatched public dataset to work. Naming this option explicitly matters even when it isn't chosen — it's the honest alternative to a bad-fit shortcut. ## Step 3 — License check, before anything else Commercial-use restrictions, attribution requirements, share-alike terms that could constrain what the resulting model is allowed to be used for. Flag immediately if a dataset that otherwise fits well has a license that blocks the intended use — this needs to surface before time is spent on it, not after. ## Step 4 —