Insights from Multilingual Curation for a 20T-Token Datasetdatologyai.com1 points·by hurrycane·il y a 5 mois·0 comments
DatBench fixes VLM evals: 70% blindly solvable, 42% mislabeled, 35% prod gapdatologyai.com5 points·by hurrycane·il y a 6 mois·0 comments
DatBench: Cut VLM eval compute by >10× while INCREASING signaldatologyai.com4 points·by hurrycane·il y a 6 mois·0 comments
Luxical: Lexical-Dense Embeddings for Web-Scale Data Curation (3×–100× Faster)datologyai.com3 points·by hurrycane·il y a 7 mois·0 comments