Add Extend Extract outputs for 145 more documents (350 of 370)
Partial set. This adds extend_extract outputs for 145 form documents that were missing from the first upload. Extend Extract now has outputs for 350 of the 370 ExtractBench documents (short 244 of 252, medium 94 of 98, long 12 of 20). The evaluator scores the 20 documents without output as 0.
Missing documents and why:
- 5 rejected by the Extend API schema check (unsupported property "examples" at form_7203.name_of_shareholder): short/arif-2022, medium/bar-lev-2022, medium/becerra-2022, medium/bianco-2022, medium/cabrera-2022
- 7 rejected by the Extend API schema check (array items must have type "object" at tables.items.rows.items.values): short/budget_appendix_doe, short/budget_appendix_legis, short/census_statab, short/census_statab_12s0035, short/census_statab_12s0711, short/census_statab_employment, short/census_statab_income_by_state
- 8 timed out (3 attempts of 1800 s each): long/real_freer_register_full, long/real_ftx_full, long/real_ftx_full_corrupted, long/real_imedia_full, long/real_imedia_full_corrupted, long/real_ishares_iboxx_bond_etfs, long/real_ofac_ssi_full, long/sec_13f_0010_renaissance_technologies
Settings: Extend processor extraction_performance, base version 4.1.1, as pinned in the ExtractBench pipeline config. Newer processor versions (4.10.0) were not evaluated.
Files: results/ and raw/ for extend_extract short and medium (existing lines unchanged; 145 lines added), traces/extend_extract/ (attempt logs for the 150 newly run documents, including the 5 rejected ones), and the extend_extract row of pipelines.jsonl (document counts). No other pipeline or file changes.