Both cards in the sibling datasets end with the project note; this one was
missing it. Mirrors vocabradar/wordfreq wording, with one line tying the
vectors to the word lists they cover.
One row per word of the cleaned vocabradar/wordfreq lists: data/<lang>.words.txt
(UTF-8, one word per line, frequency order) and data/<lang>.int8.vector
(int8, row-major, N x 384, no header). Row i of the matrix is line i of the
words file. Dequantize with v = int8 / 127.0; cosine = dot / 127^2 after
casting to float32/int32.
Vectors come from hotchpotch/bekko-embedding-v1-a25m (ModernBERT, MIT) run
through llama.cpp (BF16 GGUF, mean pooling, L2-normalized) and quantized to
int8: mean cosine 0.999017, no all-zero rows. meta.json / files.json record
per-language counts, backends and SHA-256 for every data file.
