florin-parser-nano

florin-parser-nano (fine-tune of KDL-Frontier-Parser-nano)

A LoRA fine-tune of KDLAI/KDL-Frontier-Parser-nano (KoreaDeep, 1.2B, Qwen2-VL architecture) that teaches the model to emit inline formatting — **bold**, ~~strikethrough~~, <sup>/<sub> — during text recognition, with the production prompt unchanged. Full attribution to KoreaDeep for the base model and pipeline design. License is AGPL-3.0, inherited from the base.

ParseBench results (full corpus, 2,079 documents, all five dimensions measured)

Correction, 2026-08-21. Maintainer review of ParseBench PR #99 found two defects in our markdown emission (list markers swallowed into bold spans; a heading gate that promoted short sentences ending in .). Both were fixed and both full runs were replayed on identical stored model output with the fixed emission. The table below shows the corrected numbers. Prior reported figures: Overall 76.95/76.89, mean 76.92; Semantic Formatting 71.71/71.66. Net effect of the correction: Semantic Formatting −1.04, Charts −0.14, Content Faithfulness +0.03, Tables and Visual Grounding unchanged.

Confirmation-run variance: two independent full runs: 76.72 and 76.66 (|Δ| = 0.06 overall; max per-dimension |Δ| = 0.12). Reported figure: mean 77.50. Insurance-document subset (384 docs incl. SERFF rate filings, methodology in repo): 77.60 vs 74.77 for the base pipeline measured identically — measured before the 2026-08-21 emission correction; subset re-measurement pending, expect a small downward revision in its formatting component.

The honest comparison is the same-environment column: +4.04 overall head-to-head (mean of two runs; run 1 alone is +4.07). The published-number comparison (+0.33 on the mean) crosses evaluation environments and is reported with that caveat. The formatting score (76.27 mean) exceeds the best open-weight formatting entry on the public board (69.30).

Training

• Data: 6,678 region-crop→markdown pairs. 2,165 real fragments from SEC EDGAR insurance-carrier filings (bold ground truth derived from filing HTML DOM), 3,313 synthetic rendered fragments (sole source of strikethrough/superscript/subscript; incl. CJK newsprint and Devanagari textures), 1,200 no-styling negatives to teach restraint.
• Method: LoRA r16 on language attention+MLP only; vision tower and projector frozen; completion-only loss; prompt byte-identical to the production pipeline ("\nText Recognition:\n"). 2 epochs, lr 5e-5, bf16, single H100, ~50 minutes.
• Why prompting was not enough: four measured prompt variants all scored at or below control — the base model emits markup when asked but marks the wrong spans, and degrades monotonically with any prompt perturbation. Details in the linked method note.

Reproduction

Everything (training loop, data generators, EDGAR harvester, evaluation harness, per-iteration logs) is at: https://github.com/ammoman21/parsebench-open-weight-sota

Changelog

• 2026-09-04 (v3, weights updated): merged a table-structure LoRA (r16, language layers, vision frozen; trained on synthetic OTSL table renders — no benchmark data) and added one emission rule (whole-line bold). Overall 76.69 → 77.50 (mean of 4 full-corpus runs, spread 0.107). Disclosed trade: Charts −1.9, Semantic Formatting +5.6. Weights sha256 7214d61d…; the adapter ships in adapter/ for lineage. Prior weights remain in git history.
• 2026-08-21 (v2): emission defect fixes after maintainer review (76.92 → 76.69).