Datasets:
Nemotron 3.5 Parse: its verdicts in results.parquet and its row in the card's table
Adds Nemotron 3.5 Parse (nvidia/Nemotron-3.5-Parse at 62cb1e4, 0.63B) to the benchmark: 16,288 verdicts in results.parquet and one row in the card's Results table (headline 80.0, text 72.4, tables 77.3, layout 90.2). Every other row and every other file stays as it is. The charts (params.svg, weak-spots.svg) do not include Nemotron yet.
How it ran: the omni_parse_bench adapter nemotron-parse (github.com/datalab-to/omni_parse_bench/pull/26), NVIDIA's own client (DocumentParser.parse, single pass, its default sampling), vLLM 0.26.0 on one H100. The client takes images only, so the page goes as a PNG at most 2,048 px on its long edge, as for HunyuanOCR and Nanonets-OCR2. 32 pages have no output: 31 replies looped until the 8,744-token limit, and 1 reply was empty. Graded with omni-parse-bench 0.1.2 (bbed81f); 1.0.0rc1 gives the same 16,288 verdicts.