diff --git a/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json new file mode 100644 index 0000000..3919d80 --- /dev/null +++ b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json @@ -0,0 +1,1808 @@ +{ + "artifact": "TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest", + "produced_by": "external collaborating project (MinerU2.5-Pro-1.2B study)", + "content": "derived per-table measurements only; no benchmark images, ground-truth text, or annotations", + "parser": "MinerU2.5-Pro-1.2B", + "corpus": "OmniDocBench table crops", + "metric": "official exact-cell match, fixed denominator compare(GT,GT)['total_cells']", + "hardware": "NVIDIA RTX A6000 (single GPU model across all arms)", + "oracle": true, + "oracle_note": "both the error position and the corrected token come from ground truth; upper bound, not a deliverable method", + "k_max": 16, + "totals": { + "n": 118, + "by_category": { + "catastrophic": 20, + "struct_ok_content_lossy": 64, + "struct_ok_full_recover": 34 + }, + "total_gt_cells": 12893, + "total_lost_final": 2499 + }, + "field_semantics": { + "key": "OmniDocBench table-crop identifier", + "lang": "language label of the table", + "gt_cells": "fixed denominator for this table", + "gt_struct_len": "number of ground-truth structure tokens", + "cor_k0": "correct cells in the free-run output", + "cor_final": "correct cells after up to k=16 oracle single-token repairs", + "k_first_struct_exact": "smallest k at which the structure became token-identical to ground truth; null if never", + "lost_cells_final": "gt_cells - cor_final", + "first_div_pos": "index of the first diverging structure token in the free run", + "selfconf_min_margin": "minimum top1-top2 probability margin over structure positions, scored on the model's own output (see companion document)", + "category": "struct_ok_full_recover | struct_ok_content_lossy | catastrophic" + }, + "reporting_caution": "one table (magazine_TheEconomist.2024.*) accounts for 21% of the total residual loss; report leave-one-out", + "tables": [ + { + "key": "magazine_TheEconomist.2024.02.24_page_076#0", + "lang": "english", + "gt_cells": 579, + "gt_struct_len": 630, + "cor_k0": 43, + "cor_final": 54, + "k_max_run": 6, + "k_first_struct_exact": 6, + "struct_exact_reached": true, + "lost_cells_final": 525, + "category": "catastrophic", + "first_div_pos": 4, + "selfconf_min_margin": 0.167984 + }, + { + "key": "page-dca64e05-1ce6-49ee-857a-9b5d05a87357#0", + "lang": "english", + "gt_cells": 856, + "gt_struct_len": 918, + "cor_k0": 552, + "cor_final": 505, + "k_max_run": 14, + "k_first_struct_exact": 14, + "struct_exact_reached": true, + "lost_cells_final": 351, + "category": "catastrophic", + "first_div_pos": 2, + "selfconf_min_margin": 0.14854 + }, + { + "key": "page-213e6a57-ba81-4afb-8184-2e1cf2a7664d#0", + "lang": "english", + "gt_cells": 737, + "gt_struct_len": 804, + "cor_k0": 582, + "cor_final": 585, + "k_max_run": 8, + "k_first_struct_exact": 8, + "struct_exact_reached": true, + "lost_cells_final": 152, + "category": "catastrophic", + "first_div_pos": 2, + "selfconf_min_margin": 0.192593 + }, + { + "key": "scihub_s12237-014-9873-7.pdf_4#0", + "lang": "english", + "gt_cells": 247, + "gt_struct_len": 266, + "cor_k0": 26, + "cor_final": 96, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 151, + "category": "struct_ok_content_lossy", + "first_div_pos": 28, + "selfconf_min_margin": 0.24476 + }, + { + "key": "page-8d707810-de48-43ba-81a7-b976918e7be2#0", + "lang": "en_ch_mixed", + "gt_cells": 228, + "gt_struct_len": 385, + "cor_k0": 64, + "cor_final": 104, + "k_max_run": 8, + "k_first_struct_exact": 8, + "struct_exact_reached": true, + "lost_cells_final": 124, + "category": "catastrophic", + "first_div_pos": 16, + "selfconf_min_margin": 0.993645 + }, + { + "key": "page-1955f827-fb35-4806-b3a4-463d9e0d6cc8#0", + "lang": "english", + "gt_cells": 138, + "gt_struct_len": 325, + "cor_k0": 22, + "cor_final": 34, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 104, + "category": "catastrophic", + "first_div_pos": 37, + "selfconf_min_margin": 0.713772 + }, + { + "key": "page-dbed9c07-e155-4ba1-92f9-a2962dc9a8f9#0", + "lang": "english", + "gt_cells": 100, + "gt_struct_len": 120, + "cor_k0": 10, + "cor_final": 30, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 70, + "category": "catastrophic", + "first_div_pos": 0, + "selfconf_min_margin": 0.182338 + }, + { + "key": "page-2deb83b3-abe2-442c-a544-0b6a8db464f1#0", + "lang": "english", + "gt_cells": 108, + "gt_struct_len": 120, + "cor_k0": 43, + "cor_final": 51, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 57, + "category": "struct_ok_content_lossy", + "first_div_pos": 2, + "selfconf_min_margin": 0.72621 + }, + { + "key": "page-69c15957-4d6d-485c-a244-04a1a8d29d37#0", + "lang": "simplified_chinese", + "gt_cells": 228, + "gt_struct_len": 280, + "cor_k0": 9, + "cor_final": 178, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 50, + "category": "struct_ok_content_lossy", + "first_div_pos": 21, + "selfconf_min_margin": 0.670849 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#2", + "lang": "english", + "gt_cells": 47, + "gt_struct_len": 84, + "cor_k0": 0, + "cor_final": 0, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 47, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.854243 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#0", + "lang": "english", + "gt_cells": 50, + "gt_struct_len": 87, + "cor_k0": 0, + "cor_final": 4, + "k_max_run": 13, + "k_first_struct_exact": 13, + "struct_exact_reached": true, + "lost_cells_final": 46, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.908268 + }, + { + "key": "page-ab4f331b-e8f3-41fb-ba6b-8da5ade54983#1", + "lang": "english", + "gt_cells": 113, + "gt_struct_len": 132, + "cor_k0": 37, + "cor_final": 68, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 45, + "category": "struct_ok_content_lossy", + "first_div_pos": 44, + "selfconf_min_margin": 0.409909 + }, + { + "key": "page-07d333cf-4ebf-427c-b146-e419962b81fa#0", + "lang": "simplified_chinese", + "gt_cells": 42, + "gt_struct_len": 56, + "cor_k0": 5, + "cor_final": 2, + "k_max_run": 9, + "k_first_struct_exact": 9, + "struct_exact_reached": true, + "lost_cells_final": 40, + "category": "catastrophic", + "first_div_pos": 5, + "selfconf_min_margin": 0.973873 + }, + { + "key": "page-c7792da7-4167-4f5c-a7ca-ec0f8833f83b#0", + "lang": "traditional_chinese", + "gt_cells": 66, + "gt_struct_len": 91, + "cor_k0": 21, + "cor_final": 27, + "k_max_run": 11, + "k_first_struct_exact": 11, + "struct_exact_reached": true, + "lost_cells_final": 39, + "category": "catastrophic", + "first_div_pos": 31, + "selfconf_min_margin": 0.243696 + }, + { + "key": "page-f222e5f3-8628-47cd-a0b6-40f62123228f#0", + "lang": "english", + "gt_cells": 301, + "gt_struct_len": 344, + "cor_k0": 158, + "cor_final": 264, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 37, + "category": "struct_ok_content_lossy", + "first_div_pos": 177, + "selfconf_min_margin": 0.915783 + }, + { + "key": "scihub_j.jallcom.2003.09.015.pdf_1#0", + "lang": "english", + "gt_cells": 154, + "gt_struct_len": 180, + "cor_k0": 118, + "cor_final": 121, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 33, + "category": "struct_ok_content_lossy", + "first_div_pos": 160, + "selfconf_min_margin": 0.703608 + }, + { + "key": "scihub_jneurosci.16-15-04625.1996.pdf_2#0", + "lang": "english", + "gt_cells": 225, + "gt_struct_len": 275, + "cor_k0": 187, + "cor_final": 193, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 32, + "category": "struct_ok_content_lossy", + "first_div_pos": 11, + "selfconf_min_margin": 0.098257 + }, + { + "key": "notes_f7f010b78016aeebd76e56d9283eb67f_50#1", + "lang": "en_ch_mixed", + "gt_cells": 32, + "gt_struct_len": 40, + "cor_k0": 0, + "cor_final": 2, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 30, + "category": "struct_ok_content_lossy", + "first_div_pos": 4, + "selfconf_min_margin": 0.758088 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#1", + "lang": "english", + "gt_cells": 43, + "gt_struct_len": 81, + "cor_k0": 0, + "cor_final": 13, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 30, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.939357 + }, + { + "key": "page-e7681b54-2c4a-40a2-9abe-caebba01cefb#0", + "lang": "simplified_chinese", + "gt_cells": 44, + "gt_struct_len": 56, + "cor_k0": 10, + "cor_final": 17, + "k_max_run": 6, + "k_first_struct_exact": 6, + "struct_exact_reached": true, + "lost_cells_final": 27, + "category": "catastrophic", + "first_div_pos": 22, + "selfconf_min_margin": 0.963974 + }, + { + "key": "scihub_AEM.01420-09.pdf_3#0", + "lang": "english", + "gt_cells": 76, + "gt_struct_len": 120, + "cor_k0": 29, + "cor_final": 51, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 25, + "category": "struct_ok_content_lossy", + "first_div_pos": 5, + "selfconf_min_margin": 0.590596 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-hup.777.pdf_7#0", + "lang": "english", + "gt_cells": 249, + "gt_struct_len": 336, + "cor_k0": 183, + "cor_final": 225, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 24, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.462105 + }, + { + "key": "newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_1#0", + "lang": "english", + "gt_cells": 39, + "gt_struct_len": 55, + "cor_k0": 1, + "cor_final": 15, + "k_max_run": 6, + "k_first_struct_exact": 6, + "struct_exact_reached": true, + "lost_cells_final": 24, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.121408 + }, + { + "key": "scihub_s11661-013-1726-7.pdf_2#0", + "lang": "english", + "gt_cells": 38, + "gt_struct_len": 48, + "cor_k0": 8, + "cor_final": 16, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 22, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.999408 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-ijc.22994.pdf_3#1", + "lang": "english", + "gt_cells": 150, + "gt_struct_len": 196, + "cor_k0": 126, + "cor_final": 128, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 22, + "category": "struct_ok_content_lossy", + "first_div_pos": 58, + "selfconf_min_margin": 0.828759 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#5", + "lang": "english", + "gt_cells": 22, + "gt_struct_len": 39, + "cor_k0": 0, + "cor_final": 0, + "k_max_run": 13, + "k_first_struct_exact": 13, + "struct_exact_reached": true, + "lost_cells_final": 22, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.964572 + }, + { + "key": "docstructbench_enbook-zlib-o.O-17761417.pdf_894#0", + "lang": "english", + "gt_cells": 30, + "gt_struct_len": 48, + "cor_k0": 7, + "cor_final": 10, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 20, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.734034 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#4", + "lang": "english", + "gt_cells": 20, + "gt_struct_len": 39, + "cor_k0": 0, + "cor_final": 0, + "k_max_run": 15, + "k_first_struct_exact": 15, + "struct_exact_reached": true, + "lost_cells_final": 20, + "category": "catastrophic", + "first_div_pos": 1, + "selfconf_min_margin": 0.926391 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-ceat.200407001.pdf_3#0", + "lang": "english", + "gt_cells": 30, + "gt_struct_len": 35, + "cor_k0": 5, + "cor_final": 12, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 18, + "category": "struct_ok_content_lossy", + "first_div_pos": 7, + "selfconf_min_margin": 0.596355 + }, + { + "key": "page-472c26e2-6bc2-4c05-9889-05132c87a8a9#0", + "lang": "simplified_chinese", + "gt_cells": 41, + "gt_struct_len": 70, + "cor_k0": 24, + "cor_final": 24, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 17, + "category": "struct_ok_content_lossy", + "first_div_pos": 5, + "selfconf_min_margin": 0.182068 + }, + { + "key": "PPT_Catalysis.ppt_page_016#0", + "lang": "english", + "gt_cells": 18, + "gt_struct_len": 24, + "cor_k0": 3, + "cor_final": 3, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 15, + "category": "struct_ok_content_lossy", + "first_div_pos": 24, + "selfconf_min_margin": 0.943784 + }, + { + "key": "page-3f8eed58-8b86-4f3a-958d-13671977d3c6#0", + "lang": "simplified_chinese", + "gt_cells": 88, + "gt_struct_len": 128, + "cor_k0": 40, + "cor_final": 75, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 13, + "category": "struct_ok_content_lossy", + "first_div_pos": 9, + "selfconf_min_margin": 0.244847 + }, + { + "key": "newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_2#0", + "lang": "english", + "gt_cells": 20, + "gt_struct_len": 30, + "cor_k0": 6, + "cor_final": 7, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 13, + "category": "struct_ok_content_lossy", + "first_div_pos": 18, + "selfconf_min_margin": 0.174067 + }, + { + "key": "page-9a25d716-f3a8-42e6-9a4d-e5e49a71ab41#1", + "lang": "simplified_chinese", + "gt_cells": 25, + "gt_struct_len": 36, + "cor_k0": 11, + "cor_final": 12, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 13, + "category": "struct_ok_content_lossy", + "first_div_pos": 24, + "selfconf_min_margin": 0.702141 + }, + { + "key": "scihub_j.1528-1167.2008.01552.x.pdf_3#0", + "lang": "english", + "gt_cells": 80, + "gt_struct_len": 117, + "cor_k0": 68, + "cor_final": 68, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 12, + "category": "struct_ok_content_lossy", + "first_div_pos": 117, + "selfconf_min_margin": 0.411123 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s10555-005-1578-8.pdf_5#1", + "lang": "english", + "gt_cells": 174, + "gt_struct_len": 196, + "cor_k0": 141, + "cor_final": 162, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 12, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.929999 + }, + { + "key": "page-f6b6cfb8-3bb5-4393-b3fb-70efef72aca5#0", + "lang": "simplified_chinese", + "gt_cells": 75, + "gt_struct_len": 90, + "cor_k0": 16, + "cor_final": 63, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 12, + "category": "struct_ok_content_lossy", + "first_div_pos": 12, + "selfconf_min_margin": 0.183008 + }, + { + "key": "page-c7771a62-5b6b-48f0-b591-6b657d26e125#1", + "lang": "english", + "gt_cells": 157, + "gt_struct_len": 180, + "cor_k0": 131, + "cor_final": 146, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 11, + "category": "struct_ok_content_lossy", + "first_div_pos": 12, + "selfconf_min_margin": 0.975132 + }, + { + "key": "scihub_j.jmb.2003.10.009.pdf_4#0", + "lang": "english", + "gt_cells": 65, + "gt_struct_len": 78, + "cor_k0": 52, + "cor_final": 54, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 11, + "category": "struct_ok_content_lossy", + "first_div_pos": 43, + "selfconf_min_margin": 0.727554 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s10555-005-1578-8.pdf_5#0", + "lang": "english", + "gt_cells": 208, + "gt_struct_len": 234, + "cor_k0": 163, + "cor_final": 198, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 10, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.761562 + }, + { + "key": "scihub_s00586-010-1656-7.pdf_3#0", + "lang": "english", + "gt_cells": 32, + "gt_struct_len": 40, + "cor_k0": 20, + "cor_final": 22, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 10, + "category": "struct_ok_content_lossy", + "first_div_pos": 6, + "selfconf_min_margin": 0.879728 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-chem.200700285.pdf_4#0", + "lang": "english", + "gt_cells": 31, + "gt_struct_len": 42, + "cor_k0": 22, + "cor_final": 22, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 9, + "category": "struct_ok_content_lossy", + "first_div_pos": 36, + "selfconf_min_margin": 0.998871 + }, + { + "key": "book_zh_DLT10902008_extracted_page_8#0", + "lang": "simplified_chinese", + "gt_cells": 22, + "gt_struct_len": 40, + "cor_k0": 12, + "cor_final": 13, + "k_max_run": 6, + "k_first_struct_exact": 6, + "struct_exact_reached": true, + "lost_cells_final": 9, + "category": "catastrophic", + "first_div_pos": 20, + "selfconf_min_margin": 0.103023 + }, + { + "key": "newspaper_The Wall Street Journal - 2025-01-08@magazinesclubnew_page_018#3", + "lang": "english", + "gt_cells": 22, + "gt_struct_len": 39, + "cor_k0": 0, + "cor_final": 13, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 9, + "category": "struct_ok_content_lossy", + "first_div_pos": 1, + "selfconf_min_margin": 0.95058 + }, + { + "key": "page-fd7003ba-9af2-4b60-8c45-1e2ac1c7f474#1", + "lang": "traditional_chinese", + "gt_cells": 24, + "gt_struct_len": 28, + "cor_k0": 1, + "cor_final": 15, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 9, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.953992 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-hup.777.pdf_8#0", + "lang": "english", + "gt_cells": 128, + "gt_struct_len": 184, + "cor_k0": 96, + "cor_final": 119, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 9, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.905101 + }, + { + "key": "page-e7681b54-2c4a-40a2-9abe-caebba01cefb#1", + "lang": "simplified_chinese", + "gt_cells": 84, + "gt_struct_len": 114, + "cor_k0": 75, + "cor_final": 77, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 7, + "category": "struct_ok_content_lossy", + "first_div_pos": 13, + "selfconf_min_margin": 0.984844 + }, + { + "key": "eastmoney_bdbfdb8a6ab56be3ab1b40009ee648b6d250013ce653c146f2d2023bd8d14772.pdf_26#0", + "lang": "simplified_chinese", + "gt_cells": 27, + "gt_struct_len": 44, + "cor_k0": 23, + "cor_final": 20, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 7, + "category": "struct_ok_content_lossy", + "first_div_pos": 1, + "selfconf_min_margin": 0.977491 + }, + { + "key": "page-2abc2a95-d39a-48db-9b6d-f40410412694#1", + "lang": "traditional_chinese", + "gt_cells": 111, + "gt_struct_len": 150, + "cor_k0": 103, + "cor_final": 104, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 7, + "category": "struct_ok_content_lossy", + "first_div_pos": 21, + "selfconf_min_margin": 0.0 + }, + { + "key": "page-d0b2bb59-ce8e-4a1c-9ba1-a805606a477f#0", + "lang": "traditional_chinese", + "gt_cells": 92, + "gt_struct_len": 115, + "cor_k0": 83, + "cor_final": 85, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 7, + "category": "struct_ok_content_lossy", + "first_div_pos": 107, + "selfconf_min_margin": 0.992177 + }, + { + "key": "page-bf3685a0-f0ea-46a7-bb99-1526da8f8077#1", + "lang": "simplified_chinese", + "gt_cells": 65, + "gt_struct_len": 77, + "cor_k0": 56, + "cor_final": 58, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 7, + "category": "struct_ok_content_lossy", + "first_div_pos": 1, + "selfconf_min_margin": 0.990179 + }, + { + "key": "scihub_mol.66.4.921.pdf_1#0", + "lang": "english", + "gt_cells": 70, + "gt_struct_len": 104, + "cor_k0": 63, + "cor_final": 64, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 6, + "category": "struct_ok_content_lossy", + "first_div_pos": 53, + "selfconf_min_margin": 0.703729 + }, + { + "key": "notes_1ba14cb325bc448f7201b20502ecf2b5_78#0", + "lang": "simplified_chinese", + "gt_cells": 22, + "gt_struct_len": 35, + "cor_k0": 15, + "cor_final": 16, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 6, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.0 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s00374-003-0589-2.pdf_6#0", + "lang": "english", + "gt_cells": 45, + "gt_struct_len": 64, + "cor_k0": 32, + "cor_final": 40, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 5, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.787285 + }, + { + "key": "notes_1ba14cb325bc448f7201b20502ecf2b5_98#0", + "lang": "simplified_chinese", + "gt_cells": 7, + "gt_struct_len": 16, + "cor_k0": 0, + "cor_final": 2, + "k_max_run": 10, + "k_first_struct_exact": 10, + "struct_exact_reached": true, + "lost_cells_final": 5, + "category": "catastrophic", + "first_div_pos": 2, + "selfconf_min_margin": 0.124332 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s00128-010-0006-0.pdf_5#0", + "lang": "english", + "gt_cells": 110, + "gt_struct_len": 121, + "cor_k0": 103, + "cor_final": 105, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 5, + "category": "struct_ok_content_lossy", + "first_div_pos": 12, + "selfconf_min_margin": 0.980547 + }, + { + "key": "jiaocaineedrop_jiaocai_needrop_en_397#0", + "lang": "en_ch_mixed", + "gt_cells": 26, + "gt_struct_len": 42, + "cor_k0": 1, + "cor_final": 21, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 5, + "category": "struct_ok_content_lossy", + "first_div_pos": 3, + "selfconf_min_margin": 0.124339 + }, + { + "key": "scihub_s00894-012-1568-z.pdf_2#0", + "lang": "english", + "gt_cells": 49, + "gt_struct_len": 56, + "cor_k0": 22, + "cor_final": 45, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 4, + "category": "struct_ok_content_lossy", + "first_div_pos": 16, + "selfconf_min_margin": 0.584215 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-ijc.22994.pdf_3#0", + "lang": "english", + "gt_cells": 120, + "gt_struct_len": 162, + "cor_k0": 114, + "cor_final": 117, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 3, + "category": "struct_ok_content_lossy", + "first_div_pos": 38, + "selfconf_min_margin": 0.509724 + }, + { + "key": "jiaocaineedrop_jiaocai_needrop_en_383#0", + "lang": "simplified_chinese", + "gt_cells": 18, + "gt_struct_len": 24, + "cor_k0": 6, + "cor_final": 15, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 3, + "category": "struct_ok_content_lossy", + "first_div_pos": 2, + "selfconf_min_margin": 0.730488 + }, + { + "key": "docstructbench_dianzishu_zhongwenzaixian-o.O-61518266.pdf_149#0", + "lang": "simplified_chinese", + "gt_cells": 408, + "gt_struct_len": 492, + "cor_k0": 396, + "cor_final": 406, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 2, + "category": "catastrophic", + "first_div_pos": 19, + "selfconf_min_margin": 0.009068 + }, + { + "key": "notes_1ba14cb325bc448f7201b20502ecf2b5_78#1", + "lang": "simplified_chinese", + "gt_cells": 23, + "gt_struct_len": 42, + "cor_k0": 21, + "cor_final": 21, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.124353 + }, + { + "key": "page-3bed7e71-905d-4f69-aabe-dbff56d7ac4c#1", + "lang": "simplified_chinese", + "gt_cells": 49, + "gt_struct_len": 84, + "cor_k0": 38, + "cor_final": 47, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 48, + "selfconf_min_margin": 0.50936 + }, + { + "key": "page-49746735-505c-442d-b3f6-c55b1bf9e48b#0", + "lang": "english", + "gt_cells": 226, + "gt_struct_len": 252, + "cor_k0": 204, + "cor_final": 224, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 6, + "selfconf_min_margin": 0.23299 + }, + { + "key": "docstructbench_dianzishu_zhongwenzaixian-o.O-61518266.pdf_149#1", + "lang": "simplified_chinese", + "gt_cells": 408, + "gt_struct_len": 492, + "cor_k0": 401, + "cor_final": 406, + "k_max_run": 16, + "k_first_struct_exact": null, + "struct_exact_reached": false, + "lost_cells_final": 2, + "category": "catastrophic", + "first_div_pos": 33, + "selfconf_min_margin": 0.0 + }, + { + "key": "docstructbench_llm-raw-the-eye-o.O-Character%20Sheet%20-%20Form%20Fillable.pdf_2#4", + "lang": "english", + "gt_cells": 7, + "gt_struct_len": 12, + "cor_k0": 4, + "cor_final": 5, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 1, + "selfconf_min_margin": 0.0 + }, + { + "key": "newspaper_TheBostonGlobe-2025-1-8@magazinesclubnew_page_025#5", + "lang": "english", + "gt_cells": 3, + "gt_struct_len": 4, + "cor_k0": 1, + "cor_final": 1, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 2, + "selfconf_min_margin": 0.990389 + }, + { + "key": "page-3f872a77-3faf-47d6-8574-bd2e5806ff98#2", + "lang": "simplified_chinese", + "gt_cells": 20, + "gt_struct_len": 24, + "cor_k0": 8, + "cor_final": 18, + "k_max_run": 6, + "k_first_struct_exact": 6, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "catastrophic", + "first_div_pos": 0, + "selfconf_min_margin": 0.809299 + }, + { + "key": "book_zh_GB14536.11998_extracted_page_145#0", + "lang": "simplified_chinese", + "gt_cells": 18, + "gt_struct_len": 24, + "cor_k0": 9, + "cor_final": 16, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.062306 + }, + { + "key": "eastmoney_ea59610b9b1a8f0df46f7a89da1116cbf256c772e1148f26017991e28c8bca21.pdf_18#1", + "lang": "simplified_chinese", + "gt_cells": 208, + "gt_struct_len": 266, + "cor_k0": 202, + "cor_final": 206, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 8, + "selfconf_min_margin": 0.0 + }, + { + "key": "notes_1ba14cb325bc448f7201b20502ecf2b5_35#0", + "lang": "simplified_chinese", + "gt_cells": 14, + "gt_struct_len": 25, + "cor_k0": 10, + "cor_final": 12, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.509479 + }, + { + "key": "page-14a6b411-9097-4eec-86da-92075868d243#0", + "lang": "english", + "gt_cells": 488, + "gt_struct_len": 546, + "cor_k0": 163, + "cor_final": 486, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 15, + "selfconf_min_margin": 0.1226 + }, + { + "key": "page-94600a77-35db-4549-b505-3dadc092b3f4#0", + "lang": "simplified_chinese", + "gt_cells": 42, + "gt_struct_len": 63, + "cor_k0": 8, + "cor_final": 40, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 9, + "selfconf_min_margin": 0.997535 + }, + { + "key": "page-f3e750d8-8201-4a22-9419-9776798b7a32#0", + "lang": "english", + "gt_cells": 666, + "gt_struct_len": 820, + "cor_k0": 663, + "cor_final": 664, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 1, + "selfconf_min_margin": 0.756481 + }, + { + "key": "yanbaopptmerge_aec1bf4ddb8fb4996b0512e90b3bfb31c9b151919f5d7b4a4f658f4b1f762b9e.pdf_2#0", + "lang": "simplified_chinese", + "gt_cells": 94, + "gt_struct_len": 112, + "cor_k0": 80, + "cor_final": 92, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 28, + "selfconf_min_margin": 0.876418 + }, + { + "key": "color_textbook_教材全解1+1二年级上册英语上海牛津版_page_058#0", + "lang": "en_ch_mixed", + "gt_cells": 9, + "gt_struct_len": 33, + "cor_k0": 6, + "cor_final": 7, + "k_max_run": 4, + "k_first_struct_exact": 4, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 3, + "selfconf_min_margin": 0.999128 + }, + { + "key": "page-d5a3ea83-5427-4f90-a157-7b39b784b26e#0", + "lang": "simplified_chinese", + "gt_cells": 99, + "gt_struct_len": 126, + "cor_k0": 96, + "cor_final": 97, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 2, + "category": "struct_ok_content_lossy", + "first_div_pos": 15, + "selfconf_min_margin": 0.940951 + }, + { + "key": "newspaper_Chicago Tribune_0801@magazinesclubnew_page_019#0", + "lang": "english", + "gt_cells": 136, + "gt_struct_len": 153, + "cor_k0": 104, + "cor_final": 135, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 18, + "selfconf_min_margin": 0.883244 + }, + { + "key": "page-d627077d-61cd-45c5-936f-a2c306ec3b5b#0", + "lang": "simplified_chinese", + "gt_cells": 14, + "gt_struct_len": 16, + "cor_k0": 7, + "cor_final": 13, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 9, + "selfconf_min_margin": 0.34786 + }, + { + "key": "page-fd7003ba-9af2-4b60-8c45-1e2ac1c7f474#0", + "lang": "traditional_chinese", + "gt_cells": 50, + "gt_struct_len": 65, + "cor_k0": 35, + "cor_final": 49, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.879759 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s0167-8655%2801%2900019-8.pdf_10#0", + "lang": "english", + "gt_cells": 126, + "gt_struct_len": 154, + "cor_k0": 124, + "cor_final": 125, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 144, + "selfconf_min_margin": 0.062415 + }, + { + "key": "eastmoney_5f75d8df19ab34c706ae823d8ed54f845b6a693e31c0fae19b16cdd94c32a636.pdf_2#0", + "lang": "simplified_chinese", + "gt_cells": 130, + "gt_struct_len": 180, + "cor_k0": 125, + "cor_final": 129, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 7, + "selfconf_min_margin": 0.554559 + }, + { + "key": "page-3bed7e71-905d-4f69-aabe-dbff56d7ac4c#0", + "lang": "simplified_chinese", + "gt_cells": 35, + "gt_struct_len": 55, + "cor_k0": 27, + "cor_final": 34, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 21, + "selfconf_min_margin": 0.985919 + }, + { + "key": "scihub_bloodadvances.2019030924.pdf_3#0", + "lang": "english", + "gt_cells": 92, + "gt_struct_len": 120, + "cor_k0": 78, + "cor_final": 91, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 1, + "category": "struct_ok_content_lossy", + "first_div_pos": 0, + "selfconf_min_margin": 0.982473 + }, + { + "key": "docstructbench_llm-raw-the-eye-o.O-Character%20Sheet%20-%20Form%20Fillable.pdf_2#2", + "lang": "english", + "gt_cells": 17, + "gt_struct_len": 27, + "cor_k0": 15, + "cor_final": 17, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 24, + "selfconf_min_margin": 0.88132 + }, + { + "key": "eastmoney_7f55aa1223310144165074707803522240cf20120b15f5d78fdf61eb4dcefba3.pdf_2#4", + "lang": "simplified_chinese", + "gt_cells": 73, + "gt_struct_len": 91, + "cor_k0": 0, + "cor_final": 73, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 0, + "selfconf_min_margin": 0.062415 + }, + { + "key": "jiaocaineedrop_20608808.pdf_0#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.879812 + }, + { + "key": "jiaocaineedrop_96657786.pdf_0#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.829774 + }, + { + "key": "jiaocaineedrop_jiaocai_needrop_en_383#1", + "lang": "simplified_chinese", + "gt_cells": 16, + "gt_struct_len": 18, + "cor_k0": 16, + "cor_final": 16, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 14, + "selfconf_min_margin": 0.185324 + }, + { + "key": "newspaper_TheBostonGlobe-2025-1-8@magazinesclubnew_page_025#0", + "lang": "english", + "gt_cells": 63, + "gt_struct_len": 84, + "cor_k0": 1, + "cor_final": 63, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.45686 + }, + { + "key": "page-2ff22832-e343-46a7-b94f-47ed864ce486#0", + "lang": "simplified_chinese", + "gt_cells": 43, + "gt_struct_len": 56, + "cor_k0": 42, + "cor_final": 43, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 29, + "selfconf_min_margin": 0.302688 + }, + { + "key": "page-a8d87718-ab19-4fc0-9ef5-3611ad10f88e#0", + "lang": "simplified_chinese", + "gt_cells": 21, + "gt_struct_len": 40, + "cor_k0": 20, + "cor_final": 21, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.800168 + }, + { + "key": "yanbaor2_0aa9545ccbc482a37e3b55c96897f0f24cee08bb80c7addfb0c219814af44a3b.pdf_27#0", + "lang": "simplified_chinese", + "gt_cells": 337, + "gt_struct_len": 375, + "cor_k0": 335, + "cor_final": 337, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 26, + "selfconf_min_margin": 0.411078 + }, + { + "key": "jiaocaineedrop_38247658.pdf_0#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.99534 + }, + { + "key": "jiaocaineedrop_96657786.pdf_0#1", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.635127 + }, + { + "key": "jiaocaineedrop_jiaocai_needrop_en_383#2", + "lang": "simplified_chinese", + "gt_cells": 42, + "gt_struct_len": 48, + "cor_k0": 42, + "cor_final": 42, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 8, + "selfconf_min_margin": 0.848131 + }, + { + "key": "magazine_TheEconomist.2024.02.24_page_076#4", + "lang": "english", + "gt_cells": 80, + "gt_struct_len": 96, + "cor_k0": 74, + "cor_final": 80, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 7, + "selfconf_min_margin": 0.670953 + }, + { + "key": "newspaper_Chicago Tribune_0801@magazinesclubnew_page_019#11", + "lang": "english", + "gt_cells": 32, + "gt_struct_len": 48, + "cor_k0": 31, + "cor_final": 32, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.953556 + }, + { + "key": "page-07d333cf-4ebf-427c-b146-e419962b81fa#1", + "lang": "simplified_chinese", + "gt_cells": 37, + "gt_struct_len": 49, + "cor_k0": 36, + "cor_final": 37, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 22, + "selfconf_min_margin": 0.964992 + }, + { + "key": "page-2abc2a95-d39a-48db-9b6d-f40410412694#0", + "lang": "traditional_chinese", + "gt_cells": 120, + "gt_struct_len": 160, + "cor_k0": 118, + "cor_final": 120, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 21, + "selfconf_min_margin": 0.635131 + }, + { + "key": "page-2ff22832-e343-46a7-b94f-47ed864ce486#1", + "lang": "simplified_chinese", + "gt_cells": 56, + "gt_struct_len": 77, + "cor_k0": 49, + "cor_final": 56, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 63, + "selfconf_min_margin": 0.185316 + }, + { + "key": "page-9390721f-e06c-47ea-8b21-0f27c72df7b7#0", + "lang": "english", + "gt_cells": 390, + "gt_struct_len": 518, + "cor_k0": 389, + "cor_final": 390, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 110, + "selfconf_min_margin": 0.995111 + }, + { + "key": "scihub_j.cbpb.2019.110324.pdf_6#0", + "lang": "english", + "gt_cells": 41, + "gt_struct_len": 50, + "cor_k0": 36, + "cor_final": 41, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 10, + "selfconf_min_margin": 0.933687 + }, + { + "key": "yanbaor2_afad9eb86f2a04271428aa988aec7d1949713bb5a5ce5c0cf26f320650b7966a.pdf_14#0", + "lang": "simplified_chinese", + "gt_cells": 6, + "gt_struct_len": 8, + "cor_k0": 3, + "cor_final": 6, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 4, + "selfconf_min_margin": 0.302567 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-s00128-007-9171-1.pdf_3#0", + "lang": "english", + "gt_cells": 310, + "gt_struct_len": 341, + "cor_k0": 308, + "cor_final": 310, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 12, + "selfconf_min_margin": 0.809298 + }, + { + "key": "jiaocaineedrop_42351289.pdf_1#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.991849 + }, + { + "key": "jiaocaineedrop_96657786.pdf_0#2", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.984832 + }, + { + "key": "jiaocaineedrop_jiaocai_needrop_en_3937#0", + "lang": "simplified_chinese", + "gt_cells": 12, + "gt_struct_len": 14, + "cor_k0": 12, + "cor_final": 12, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 8, + "selfconf_min_margin": 0.915808 + }, + { + "key": "page-349417c3-e5f4-4dae-b018-7323ef36b04e#1", + "lang": "simplified_chinese", + "gt_cells": 27, + "gt_struct_len": 63, + "cor_k0": 0, + "cor_final": 27, + "k_max_run": 3, + "k_first_struct_exact": 3, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.99889 + }, + { + "key": "page-73432bc8-cd0f-4e14-b119-0a9de73ab141#0", + "lang": "simplified_chinese", + "gt_cells": 20, + "gt_struct_len": 40, + "cor_k0": 19, + "cor_final": 20, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.787748 + }, + { + "key": "page-b527edec-3363-4d38-a390-2f3a307f2392#0", + "lang": "simplified_chinese", + "gt_cells": 103, + "gt_struct_len": 138, + "cor_k0": 13, + "cor_final": 103, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 0, + "selfconf_min_margin": 0.925268 + }, + { + "key": "docstructbench_llm-raw-scihub-o.O-hup.908.pdf_5#0", + "lang": "english", + "gt_cells": 32, + "gt_struct_len": 48, + "cor_k0": 29, + "cor_final": 32, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 4, + "selfconf_min_margin": 0.879818 + }, + { + "key": "eastmoney_5f75d8df19ab34c706ae823d8ed54f845b6a693e31c0fae19b16cdd94c32a636.pdf_2#3", + "lang": "simplified_chinese", + "gt_cells": 80, + "gt_struct_len": 120, + "cor_k0": 75, + "cor_final": 80, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 7, + "selfconf_min_margin": 0.244868 + }, + { + "key": "jiaocai_71434495.pdf_0#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.941354 + }, + { + "key": "jiaocaineedrop_55196913.pdf_0#0", + "lang": "simplified_chinese", + "gt_cells": 4, + "gt_struct_len": 6, + "cor_k0": 4, + "cor_final": 4, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 1, + "selfconf_min_margin": 0.988168 + }, + { + "key": "page-035cb436-c01e-41db-b40e-8977678777eb#0", + "lang": "english", + "gt_cells": 259, + "gt_struct_len": 560, + "cor_k0": 201, + "cor_final": 259, + "k_max_run": 2, + "k_first_struct_exact": 2, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 10, + "selfconf_min_margin": 0.71748 + }, + { + "key": "page-73432bc8-cd0f-4e14-b119-0a9de73ab141#1", + "lang": "simplified_chinese", + "gt_cells": 26, + "gt_struct_len": 42, + "cor_k0": 25, + "cor_final": 26, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 2, + "selfconf_min_margin": 0.670894 + }, + { + "key": "yanbaopptmerge_e3b48594d5a2a6dc86b51244737f24867a0c1781e9f0e4fb52d96bbc6f9d95c7.pdf_3#0", + "lang": "simplified_chinese", + "gt_cells": 205, + "gt_struct_len": 280, + "cor_k0": 198, + "cor_final": 205, + "k_max_run": 1, + "k_first_struct_exact": 1, + "struct_exact_reached": true, + "lost_cells_final": 0, + "category": "struct_ok_full_recover", + "first_div_pos": 8, + "selfconf_min_margin": 0.358327 + } + ] +} diff --git a/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json.sha256 b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json.sha256 new file mode 100644 index 0000000..e7956d1 --- /dev/null +++ b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json.sha256 @@ -0,0 +1 @@ +bb2c5555e1a810e3b8d622b583030af5613fa0336f6ad172f0e368c423e3b026 TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json diff --git a/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.md b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.md new file mode 100644 index 0000000..43b8371 --- /dev/null +++ b/docs/external-evidence/TOPOLOGY_REPAIR_CEILING_2026.08.14.1.md @@ -0,0 +1,151 @@ +# External Evidence: Ceiling of Topology-Only Repair 2026.08.14.1 + +## Status of this document + +External evidence contributed by a collaborating project. It is **not** an +experiment record of this repository, was produced on a different parser and a +different corpus, and does not evaluate any route in this repository. It is +filed here because it bears directly on how the Explicit route should report +its results. + +Nothing in this document should be read as a claim about +`borderless-table-structuring-lab` itself. + +## 1 Research question + +The Explicit route is topology-only: it defaults to KEEP, keeps text frozen, +and proposes minimal split/merge changes. A natural question is how much a +metric can move when **only** topology is corrected and cell content is left +exactly as the parser produced it. + +We measured the **oracle** answer to that question on a different parser, so the +number is an upper bound rather than an achievable result. + +## 2 Setup + +| Item | Value | +|---|---| +| Parser | MinerU2.5-Pro-1.2B | +| Corpus | OmniDocBench table crops, 653 tables | +| Metric | Official exact-cell match (content and position must both be correct) | +| Denominator | Fixed, `compare(GT, GT)["total_cells"]` = 43231 cells | +| Scoring exception | Whole table scored 0, per the official metric's semantics | +| Decoding | Greedy, `max_new_tokens=6144`, `no_repeat_ngram_size=100` | +| Hardware | NVIDIA RTX A6000 only, single GPU model across all arms | + +**Procedure.** Find the first structure token in the free-run output that +differs from ground truth. Replace **only that one token** with the ground-truth +token, then let the model continue generating on its own. Repeat up to `k` +times. Both the position and the correct token come from ground truth, so this +is an oracle measurement and not a deliverable method. + +**Pre-registered control.** For the 535 tables whose structure was already +exact, no repair is applied, so their per-table score must be **identical across +every `k`**. Observed: identical for all 535 tables, 0 violations. + +## 3 Results + +### 3.1 Overall curve + +| k | cell score | vs k=0 | +|---|---|---| +| 0 (free run) | 0.8593 | — | +| 1 | 0.8723 | +1.30 pp | +| 2 | 0.8904 | +3.11 pp | +| 4 | 0.8953 | +3.60 pp | +| 8 | 0.8973 | +3.80 pp | +| 12 | 0.8960 | +3.67 pp | +| 16 | **0.8990** | **+3.97 pp** | + +- The k=1 gain has a **document-cluster bootstrap 95% CI of [-1.64, +3.85] pp** + (454 clusters, 5000 resamples), which **contains zero**. +- k=12 scores below k=8, so the curve has entered its noise band. +- Quadrupling the budget from k=4 to k=16 buys +0.37 pp. The curve is saturated. + +### 3.2 Three-way split of the divergent tables (118 tables, 12893 cells) + +| Category | n | cells | score at k=0 | score at k=16 | cells still lost | +|---|---|---|---|---|---| +| Structure repaired, **full recovery** | **34** | 2480 | 0.8544 | **1.0000** | **0** | +| Structure repaired, content still wrong | 64 | 6537 | 0.6778 | 0.8687 | 858 | +| Resistant tables | 20 | 3876 | 0.5488 | 0.5766 | **1641** | + +Both directions matter: + +- **Topology repair genuinely works.** 59 of 118 divergent tables (50%) become + structurally exact after a **single** oracle correction, which matches the + `single minimal edit` category in this repository's corpus design. 34 of them + reach a cell score of exactly **1.0000** once structure is correct. +- **But cell-weighted totals are dominated elsewhere.** 66% of the remaining + loss (1641 of 2499 cells) sits in 20 tables that absorb up to 16 oracle + corrections, 14 of which end up structurally identical to ground truth, and + still score 0.5766. + +At k=16, 112 of 118 tables are structurally identical to ground truth token by +token, yet the divergent subset scores 0.8062 against 0.9385 for tables that +were structurally correct from the start. + +### 3.3 Table size does not explain the residual + +Divergent tables are larger (median 49 cells versus 26). Matching each divergent +table to its nearest-size structurally-correct table (118 pairs, 12168 cells) +gives the size-matched control a score of **0.9537**, which is *higher* than the +0.9385 of the full control group. Larger tables are not harder for this parser. +After size matching, the gap to the divergent subset at k=16 is still 14.75 pp. + +## 4 Implication for reporting + +| Metric family | Does topology-only repair score? | Basis | +|---|---|---| +| TEDS-Struct, GriTS-Topology | **Yes** | 112 of 118 become structurally exact | +| Exact-cell (content and position) | **Ceiling +3.97 pp** | Section 3.1 | +| TEDS-all, GriTS-Content | Between the two | Not measured here; no claim made | + +A topology-only route that freezes text should score clearly on structural +metrics. On content-inclusive metrics, the oracle ceiling we measured is ++3.97 pp. + +**Suggestion:** report structural and content-inclusive metrics as separate +numbers rather than a single combined score. Separating them shows what a +topology-only method actually achieves, and prevents a structural gain from +being read as a content gain or vice versa. + +## 5 Limits of validity + +1. **One parser.** MinerU2.5-Pro-1.2B. A different base model fails differently. +2. **One corpus.** OmniDocBench table crops, not a borderless-specific set. In a + separate experiment we found hard-table phenotypes to be unstable across + corpora, so this limit is substantive rather than formal. +3. **One metric.** Official exact-cell, not directly comparable to TEDS or GriTS. +4. **Oracle.** Both the error position and the correct token come from ground + truth. Any real detector performs worse. +5. **Single-table dominance.** One table + (`magazine_TheEconomist.2024.*`) loses 525 cells, which is **21% of the + entire residual**. Any conclusion drawn on this data must be reported with a + leave-one-out check. + +## 6 A caliber error we made, offered as a caution + +For a long time we cited an upper bound of "+9.63 pp for repairing structure" +and planned a batch of experiments around it. Tracing its provenance showed it +was **not a measurement**. It was an arithmetic construction that raised the +divergent tables' cell score to **1.0**, that is, it assumed those tables became +*perfect*, not that their *structure* became correct. The measured structural +ceiling is +3.97 pp. + +The rule we adopted afterwards, offered here as a suggestion: + +> Any number cited as an upper bound must be traceable to a specific +> measurement. Constructed values must carry the word "constructed" in their +> name. + +## 7 Accompanying manifest + +`TOPOLOGY_REPAIR_CEILING_2026.08.14.1.manifest.json` lists all 118 divergent +tables with per-table language, ground-truth cell count, score before and after +repair, the value of `k` at which structure first became exact, and the +self-confidence minimum margin described in the companion document. + +The manifest contains derived measurements only: benchmark sample identifiers +and our own scores. It contains **no benchmark images, no ground-truth text, and +no annotations**.