DatBench fixes VLM evals: 70% blindly solvable, 42% mislabeled, 35% prod gap #1 Post by hurrycane » Tue, Jan 06, 2026, 5:54 PM UTC DatBench fixes VLM evals: 70% blindly solvable, 42% mislabeled, 35% prod gapdatologyai.com