Dancing Links LLC builds training and evaluation datasets for AI teams. Curated, deduplicated, decontaminated, and versioned — so every run is reproducible and you know what changed.
"Cover Everything. Repeat Nothing."
Dancing Links is Donald Knuth's technique for exact cover: choose a set of rows that covers every column exactly once — no gaps, no overlap. Pull a node out in constant time, explore, and splice it back perfectly when the branch fails.
That is the training-data problem, exactly. Every capability you care about is a column. Every candidate corpus is a row. The work is covering all of them with as little redundancy as you can afford — and being able to pull a batch back out when it turns out to hurt.
Sourcing, cleaning, near-duplicate removal, and eval-set decontamination. You get a corpus with known provenance and a written record of what was dropped, not a scraped pile.
Human-plus-model annotation with written rubrics, inter-annotator agreement tracking, and held-out golden sets that catch quality drift before it reaches a training run.
Versioned datasets with full lineage, delivered into your training environment. Diff two runs and say exactly which batch moved the number.
Business inquiries, partnerships, and collaborations — reach me directly:
Typically replies within 24–48 hours.
Early access + occasional notes on training-data practice