PS another task is SQUAD (rajpurkar.github.io/SQuAD-explor...) which is about extracting content from text. Again with ground truth.

Thanks @dggoldst.bsky.social for listing the reference. The image involves nlp and vision benchmarks. One of these is GLUE (gluebenchmark.com) which has many tasks, some of which I would roughly consider to be "structure extraction from text". It also has ground truth.

Please add me. Thanks Dan!