Extracting information from images.
VLMBench v1: an early public benchmark for VLM document reasoning.
Acceptable answers (exact + materially correct) out of 16,500 graded cells per model.
Recommend a model
Errors were classified, here is the breakdown: