DatBench fixes VLM evals: 70% blindly solvable, 42% mislabeled, 35% prod gap

(datologyai.com) by hurrycane | Jan 6, 2026 | 0 comments on HN