futo-org/swipe-negatives

Swipe Hard Negatives

Hard-negative word sets for swipe-keyboard language model training, mined by KNN over swipe-shape embeddings from the SwipeALot encoder. For each positive word, the top K = 128 most-confusable words are listed alongside neg_sims — the mean pairwise sample-to-sample cosine similarity between samples of the positive word and samples of the negative word.

These negatives are designed to be "hard" because they share visual swipe-path similarity with the positive, not just lexical features.

Layout-specific. Swipe-shape similarity depends on the keyboard layout, so they are not hard negatives for other layouts (only en_qwerty currently).

Schema

column	type	meaning
word	string	positive word, normalized to a-z (en_qwerty swipeable chars)
negatives	list<string>	top-K hard negatives (also a-z normalized), sorted by neg_sims descending
neg_sims	list<float32>	mean pairwise sample-to-sample cosine similarity
num_samples	int32	# real swipe samples backing this alpha word (capped at 64 upstream)

Generation

1. The SwipeALot encoder embeds each word's swipe samples (up to 64 per word) into shape-aware vectors. Words without real swipe data get a text-only fallback embedding.
2. KNN over the embedding index returns the nearest neighbors per sample; mean cosine similarity is aggregated across samples.
3. Positives and negatives are normalized to a-z (the en_qwerty swipeable character set) and deduped. Words shorter than 2 letters after normalization are dropped (KNN over a single-character swipe is too noisy to be useful). Each row's negatives are re-sorted by neg_sims descending and truncated to the top-K. Output rows are sorted alphabetically by word.