ocr-tech-k2 / docs /PAPER_NOTES.md
prosperious's picture
Add 51 files (.gitignore ...)
1c43956 verified
|
Raw History Blame Contribute Delete
216 kB

Paper Notes

This file is the paper-facing record of what the repo can support today without inventing claims after the fact.

The current evidence supports a comparative systems paper or engineering report. It does not support a SOTA paper yet.

For the current generated result tables and chronology, use PAPER_RESULTS.md.

Current Thesis

The strongest supported thesis in this repo is:

  1. A calibrated Tesseract-plus-Surya composite with structure backfill is a real improvement over Tesseract and several weaker external baselines on our internal rendered Nepali/English document pack.
  2. On the stricter rendered-document holdouts, every Tesseract/Surya composite tested so far still fails to beat raw Surya on text OCR CER.
  3. A sample-level oracle over raw Surya and raw Tesseract has almost no text-OCR headroom over Surya on rendered-v2, so more routing in that component family is not a serious path to SOTA.
  4. The useful contribution right now is careful evaluation, calibration discipline, and a documented failure boundary, not frontier OCR leadership.

If the paper tries to claim "best overall OCR engine" from the current evidence, it will be wrong.

Goal Alignment

As of 2026-06-15, the target remains a better Devanagari + English document OCR engine, not a collection of benchmark scripts. The engine work is the right center of gravity because the current candidate loses to Surya on text CER in the strongest document holdouts even when its table recovery is competitive or better.

The clean research framing is:

  1. Train or adapt a recognizer that improves mixed Devanagari/English text accuracy instead of only routing between existing engines.
  2. Keep the document stack around that recognizer: layout, reading order, tables, figures/captions, Markdown, JSON, and optional VLM fallback.
  3. Treat SOTA as a validation outcome. The paper can claim it only after the trained engine beats named baselines under the gates in VALIDATION.md.

This means the active training run is aligned with the paper goal. It is not yet proof that the goal has been achieved.

Current Model Status

We completed a real recognizer fine-tune on k2, and now have a packaged model artifact. It does not support any quality claim until held-out benchmarks beat the named baselines.

Current run:

  • path: runs/hf-recognizer-mixed-v2-full
  • backend: hf-vision-encoder-decoder
  • base model: microsoft/trocr-small-printed
  • train manifest: data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl
  • eval manifest: data/splits-mixed-v1/eval-with-refs.jsonl
  • train/eval batch size: 4 / 4
  • learning rate: 0.00005
  • epochs: 3
  • expected trainer steps: 1275

Training completed on k2 and produced both final-model/ and training_metrics.json.

Training metrics:

  • eval_loss: 2.672821283340454
  • eval_cer: 0.4532251781992946
  • eval_wer: 0.6495539645539645
  • eval_exact_line_accuracy: 0.12222222222222222
  • epoch: 3.0

The model was packaged and audited:

  • model card: runs/hf-recognizer-mixed-v2-full/finalized/package/model-card.json
  • audit: runs/hf-recognizer-mixed-v2-full/finalized/audit/model-audit.json
  • audit status: pass

So the precise answer to "have we trained a model?" is now: yes, we trained and packaged one HF recognizer model, but it failed the first held-out baseline check and is not a SOTA candidate.

The packaging/audit step is now guarded by:

uv run ocrtech finalize-hf-recognizer-run \
  --run-dir runs/hf-recognizer-mixed-v2-full \
  --out runs/hf-recognizer-mixed-v2-full/finalized \
  --model-id ocrtech-hf-mixed-v2-full \
  --base-model microsoft/trocr-small-printed \
  --train-manifest data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl \
  --eval-manifest data/splits-mixed-v1/eval-with-refs.jsonl \
  --admission-validation-report outputs/validation-next-recognizer-hard50/validation.json

The first post-training held-out benchmark completed on k2:

  • output: outputs/experiment-hf-recognizer-mixed-v2-full-vs-tesseract
  • baselines: candidate,tesseract
  • rows: 360
  • status: 360 ok, 0 errors
  • candidate CER: 0.4452926496941952
  • Tesseract CER: 0.3290117066116023
  • candidate WER: 0.6551095201095201
  • Tesseract WER: 0.48648528940195623
  • candidate mean latency: 0.4809882701633291s
  • Tesseract mean latency: 0.14839679651065832s
  • validation: skipped because this was a direct branch check, not a claim run

Supported conclusion:

  • this HF TrOCR branch is a negative result
  • it is worse than Tesseract on the mixed held-out recognizer eval
  • it should not be the main engine direction unless a new training hypothesis is introduced

CUDA Training Decision

The next serious engine step is PaddleOCR recognizer training on a CUDA Linux machine. k2 remains useful for dataset preparation, recipe generation, packaging, and benchmark orchestration, but it is not the right host for claim-grade PaddleOCR training because it is macOS arm64.

Use the existing portable bundle as the CUDA handoff artifact:

runs/paddle-recognizer-mixed-v2-bundle/paddle-recognizer-mixed-v2-training.tar.gz

After extracting the bundle on a CUDA host with PaddleOCR checked out and installed, the intended first training command is:

python PaddleOCR/tools/train.py -c runs/paddle-recognizer-mixed-v2/paddle_recognizer.yml

This is the branch that should get AWS GPU time before more HF recognizer experiments.

What We Can Claim

These claims are grounded in artifacts already present in this workspace.

1. Real remote experiment stack on k2

We now have a reproducible remote workflow on k2 for:

  • host audit
  • bootstrap
  • workspace sync
  • benchmark execution
  • calibration
  • claim preflight
  • claim review

Representative artifacts:

  • outputs/live-remote-audit-base/remote-audit.json
  • outputs/live-remote-bootstrap-311/remote-bootstrap.json
  • outputs/live-remote-sync/remote-sync.json
  • outputs/live-remote-sync-external-baselines-2/remote-sync.json

This matters for a paper because the experiment trail is no longer "trust me, I ran it."

2. The TrOCR path is runnable but not competitive

The HF TrOCR recognizer branch is real enough to benchmark, and it currently loses to Tesseract on the mixed-script line-style eval.

Representative artifacts:

  • outputs/live-k2-rebalanced-bundle/outputs/experiment-hf-mixed-v1-rebalanced/benchmark/report.json
  • outputs/live-k2-vs-tesseract-bundle/outputs/experiment-hf-mixed-v1-rebalanced-vs-tesseract/benchmark/report.json

Supported claim:

  • the repo contains a working training and packaging path for an HF recognizer
  • that branch is not the answer to the project goal in its current form

3. The Tesseract-plus-ByT5 correction path failed

The text-correction composite is not just "not ready." It is actively bad on the held-out mixed-script eval.

Representative artifacts:

  • outputs/live-k2-corrector-vs-tesseract-summary/summary.json
  • outputs/live-k2-corrector-vs-tesseract-summary/summary.md

Supported claim:

  • the current post-OCR correction branch should be treated as a negative result

4. Script-aware routing helped on the older mixed-script holdout

On the older 91-sample line-style holdout, a calibrated script_select_composite beat raw Tesseract out of sample.

Key result already documented elsewhere in the repo:

  • Tesseract CER: 0.342234
  • script router CER: 0.313084
  • paired CER: 9 wins, 0 losses, 82 ties

Representative artifacts:

  • outputs/live-k2-script-router-holdout/experiment.json
  • outputs/live-k2-script-router-holdout/report.json
  • outputs/live-k2-script-router-holdout/summary.json

Supported claim:

  • script-aware routing is a plausible direction
  • this older result is not enough for a document paper because the pack is too line-like

5. Structure backfill improved document-table recovery

On rendered document tasks, secondary_structure_backfill=true is the first composition step that clearly improved structure without paying an obvious text penalty on the small smoke pack.

Representative artifacts:

  • outputs/live-k2-multipage-backfill/summary.md
  • outputs/live-k2-multipage-backfill/candidate-multipage-document.json

Supported claim:

  • composition can improve table recovery in mixed OCR/document pipelines

6. The rendered 25-sample claim-ready pack is a real step up from smoke tests

The internal rendered claim-ready image pack exists, was audited, and was used in live comparisons.

Representative artifacts:

  • outputs/live-k2-claimpack-rendered-backfill/eval-pack-audit.json
  • outputs/live-k2-claimpack-rendered-backfill/experiment.json
  • outputs/live-k2-claimpack-rendered-backfill/report.json
  • outputs/live-k2-claimpack-rendered-backfill/summary.json
  • outputs/live-k2-claimpack-surya-compare/summary.json
  • outputs/live-k2-claimpack-stock-paddle-fixed/summary.json

Supported claims on this pack:

  • candidate beats raw Tesseract strongly on overall CER and table extraction
  • candidate beats this stock-Paddle configuration strongly on overall CER and table extraction
  • candidate ties Surya on aggregate table_cell_f1
  • candidate slightly beats Surya on aggregate reading-order accuracy
  • candidate loses to raw Surya on text CER

This is the current core paper result. It is useful because it is mixed, not because it is flattering.

7. External baselines are no longer hypothetical

GLM-OCR and PaddleOCR-VL are wired into the benchmark flow through wrapper scripts and markdown normalization, so baseline names now correspond to real runs.

Representative artifacts:

  • outputs/live-k2-external-baselines/preflight/preflight.json
  • outputs/live-k2-external-baselines/benchmark/report.json
  • outputs/live-k2-external-baselines/glm/report.json

Supported claim:

  • the repo can benchmark external OCR baselines even when they only emit markdown

8. On the blur-heavy rendered subset, the candidate still loses to Surya

The latest broadened but tractable comparison is the 5-sample blur-only rendered subset spanning paragraph, table, form, figure, and multi-page layouts.

Representative artifacts:

  • outputs/live-k2-rendered-blur5/report.json
  • outputs/live-k2-rendered-blur5/summary.json

Aggregate result:

System CER table_cell_f1 reading_order_pair_accuracy
candidate 0.223273 1.0 0.783636
surya 0.197601 1.0 0.750909
tesseract 0.732364 0.0 0.823636
stock-paddle 2.810426 0.0 1.0
glm-ocr 2.820269 0.0 1.0
paddleocr-vl 18.470663 0.0 0.750909

Supported claim:

  • the current candidate is materially stronger than Tesseract, stock-Paddle, GLM-OCR, and PaddleOCR-VL on this blur-heavy rendered subset
  • the current candidate still does not beat raw Surya on text CER

9. On the larger rendered-v2 holdout, the routed candidate collapses to Surya

We now have a stricter rendered-document result from k2:

  • calibration artifact: outputs/live-k2-claimpack-v2-script-router/calibration.json
  • holdout summary: outputs/live-k2-claimpack-v2-script-router/summary.json
  • local paper table: outputs/paper-claimpack-v2-table/paper-table.md
  • local preflight evidence: outputs/live-k2-claimpack-v2-script-router/preflight.json

What happened:

  • calibration selected primary_script_threshold=0.0
  • calibration selected_trial.secondary_picks = 50 out of 50
  • holdout candidate CER = raw Surya CER = 0.180893
  • holdout candidate table_cell_f1 is slightly worse than raw Surya: 0.947500 vs 0.970833

Supported claim:

  • on the larger rendered holdout, this routed branch is no longer meaningfully distinct from Surya for text OCR
  • the new claim guardrails should reject this as a routed candidate suitable for a "better engine" claim, both from calibration collapse evidence and from exact holdout equivalence to a component baseline

10. Inverting the script router also failed on rendered-v2

We added Latin-script routing support specifically to test the obvious opposite branch: keep Surya as the primary engine and fall back to Tesseract on English-heavy pages.

Representative artifacts:

  • outputs/live-k2-claimpack-v2-surya-latin-router/experiment.json
  • outputs/live-k2-claimpack-v2-surya-latin-router/summary.json
  • outputs/paper-claimpack-v2-surya-latin-table/paper-table.md

Aggregate result on the rendered-v2 holdout:

  • candidate CER: 0.416854
  • raw Surya CER: 0.180893
  • candidate table_cell_f1: 0.337500
  • raw Surya table_cell_f1: 0.970833
  • paired CER vs Surya: 0 wins, 12 losses, 38 ties

Supported claim:

  • simply inverting the script router does not rescue the branch
  • on this holdout, Surya-primary Latin routing is materially worse than raw Surya

11. A rendered-v2 quality router also failed to clear Surya

We also tested a quality-based Surya/Tesseract selector on the same rendered-v2 setup instead of script-based routing.

Representative artifacts:

  • outputs/live-k2-claimpack-v2-surya-quality-router/calibration.json
  • outputs/live-k2-claimpack-v2-surya-quality-router/experiment.json
  • outputs/live-k2-claimpack-v2-surya-quality-router/summary.json
  • outputs/paper-claimpack-v2-surya-quality-router-table/paper-table.md

Aggregate result on the rendered-v2 holdout:

  • candidate CER: 0.181128
  • raw Surya CER: 0.180893
  • candidate table_cell_f1: 0.947500
  • raw Surya table_cell_f1: 0.947500
  • paired CER vs Surya: 0 wins, 1 loss, 49 ties

Calibration result:

  • selected_threshold=0.0
  • calibration mean for chosen route: 0.193593
  • calibration mean for raw Tesseract primary: 0.626953

Supported claim:

  • a quality-based Surya/Tesseract selector still does not beat raw Surya on the stricter rendered holdout
  • this branch is closer to a guarded Surya deployment than to a distinct better OCR system
  • the rendered-v2 evidence now says the current Surya/Tesseract composite space is near exhaustion for text OCR gains

12. Line alignment improved ordering but still failed as an OCR system

We also tested a different composition idea: keep Surya text, but align it onto Tesseract's line scaffold with line_align_composite.

Representative artifacts:

  • outputs/live-k2-claimpack-v2-line-align-summary/summary.json
  • outputs/paper-claimpack-v2-line-align-table/paper-table.md
  • outputs/live-k2-claimpack-v2-line-align-model-card.json

Aggregate result on the rendered-v2 holdout:

  • candidate CER: 0.310549
  • raw Surya CER: 0.180893
  • candidate table_cell_f1: 0.897500
  • raw Surya table_cell_f1: 0.970833
  • candidate reading_order_pair_accuracy: 0.792727
  • raw Surya reading_order_pair_accuracy: 0.750909
  • paired CER vs Surya: 0 wins, 7 losses, 43 ties

Supported claim:

  • better line ordering alone does not rescue the Surya/Tesseract composite family
  • line alignment can improve reading-order accuracy while still making OCR quality materially worse
  • this is a useful negative result for a systems paper because it shows the tradeoff directly instead of hiding it behind aggregate wins over weak baselines

13. A Surya/Tesseract sample oracle shows almost no routing headroom

The strongest way to test whether more Surya/Tesseract routing is worth doing is an oracle selector: choose the better raw engine per sample using ground-truth CER, then compare that upper bound to raw Surya.

Representative artifacts:

  • outputs/paper-claimpack-v2-surya-tesseract-oracle/oracle-analysis.json
  • outputs/paper-claimpack-v2-surya-tesseract-oracle/oracle-choices.json
  • outputs/paper-claimpack-v2-surya-tesseract-oracle/oracle-summary/summary.json
  • outputs/paper-claimpack-v2-oracle-table/paper-table.md

Aggregate result on the rendered-v2 holdout:

  • oracle CER: 0.180529
  • raw Surya CER: 0.180893
  • raw Tesseract CER: 0.714207
  • oracle chosen counts: Surya 49, Tesseract 1
  • paired CER vs Surya: 1 win, 0 losses, 49 ties

Supported claim:

  • the current Surya/Tesseract component family has almost no sample-level text-OCR headroom over raw Surya
  • the oracle's per-sample source choices are documented as provenance in oracle-choices.json, not mixed into aggregate metrics
  • building more routing heuristics around these same two engines is not a serious route to the project goal
  • the next model-side work needs a genuinely different recognizer or document model, not another selector over the same components

14. Surya's remaining rendered-v2 errors concentrate in table and figure pages

After the oracle result, we added a benchmark error analysis command to find actual model-development targets.

Representative artifacts:

  • outputs/paper-claimpack-v2-surya-error-analysis/error-analysis.md
  • outputs/paper-claimpack-v2-surya-error-analysis/top-failures.json
  • outputs/paper-claimpack-v2-surya-vs-tesseract-error-analysis/error-analysis.md
  • outputs/paper-claimpack-v2-surya-vs-oracle-error-analysis/error-analysis.md

Supported claim:

  • raw Surya's highest absolute rendered-v2 CER pockets are table and figure pages
  • Tesseract only beats Surya on one holdout sample, so routing is not the right answer to those pockets
  • data/splits-claimpack-rendered-v2/eval-surya-hard15.jsonl is now a reproducible hard-eval manifest for this pocket, with a passing file/reference audit on k2
  • configs/validation-hard15-surya.json turns this pocket into an explicit Surya comparison gate for future candidates
  • data/eval-pack-table-figure-dev-v1/manifests/ocrtech-eval-pack.jsonl is a separate non-claim development pack for table/figure iteration, with no sample-id or text-hash overlap with hard15
  • the next recognizer/document-model branch should target table-ledger and figure-caption mixed-script pages specifically

15. The first table/figure HF recognizer branch failed decisively

The table/figure development pack was split with --group-by text_sha256, producing 60 train samples and 20 eval samples across 16 text groups. That grouping is important: without it, clean/blur/scan variants of the same visible document text could leak across train and eval.

Representative artifacts:

  • data/splits-table-figure-dev-v1/split-summary.md
  • outputs/audit-table-figure-dev-v1-train-k2/manifest-audit.md
  • outputs/audit-table-figure-dev-v1-eval-k2/manifest-audit.md
  • runs/hf-table-figure-dev-v1/training_metrics.json
  • runs/hf-table-figure-dev-v1/package/model-card.json
  • outputs/audit-hf-table-figure-dev-v1-model/model-audit.md
  • outputs/summary-hf-table-figure-dev-v1/summary.md
  • outputs/experiment-hf-table-figure-dev-v1/benchmark/report.json

Training result for the one-epoch microsoft/trocr-small-printed branch:

Metric Value
eval CER 3.062119
eval WER 3.562689
exact-line accuracy 0.0

Benchmark result on the 20-sample table/figure dev eval:

System CER table_cell_f1 reading_order_pair_accuracy
HF table/figure candidate 2.422001 0.0 0.0
Surya 0.563399 0.858333 0.55
Tesseract 0.546227 0.0 0.695

Supported claim:

  • the HF recognizer training and packaging path is functional on k2
  • this particular full-page TrOCR fine-tune is a negative result, not a promising branch
  • paired CER was 0 wins, 20 losses, and 0 ties against both Surya and Tesseract
  • future recognizer work should not treat entire table/figure pages as single-line recognition inputs; it needs line/cell crops or a document-structure model

16. Crop-level training data is now a first-class path

The repo now records renderer layout metadata in newly generated image-pack references and exposes ocrtech extract-crops to convert those verified bboxes into normal recognizer manifests. This is the direct response to the failed full-page TrOCR branch.

Supported claim:

  • crop extraction uses explicit reference layout metadata; it does not infer bboxes from template names
  • references without layout metadata fail extraction instead of producing guessed crops
  • crop rows preserve source sample id, source image path, source reference path, page index, bbox, crop type, text_sha256, and sample_sha256
  • skewed rendered pages are skipped for crop extraction because pre-rotation bboxes are invalid

Representative commands:

ocrtech extract-crops \
  data/eval-pack-table-figure-dev-v1/manifests/ocrtech-eval-pack.jsonl \
  --out data/crops-table-figure-dev-v1 \
  --crop-type line \
  --crop-type table-cell \
  --crop-type figure-caption

ocrtech split-manifest data/crops-table-figure-dev-v1/manifests/crops.jsonl \
  --out data/splits-crops-table-figure-dev-v1 \
  --eval-ratio 0.2 \
  --seed 17 \
  --stratify-by document_type \
  --group-by text_sha256

This is not a SOTA result yet. It is infrastructure for the next credible recognizer branch.

The first crop-level recognizer test is already complete and is also negative.

Representative artifacts:

  • outputs/crops-table-figure-dev-v1-k2/crop-manifest-summary.md
  • outputs/crops-table-figure-dev-v1-k2/manifest-audit.md
  • data/splits-crops-table-figure-dev-v1/split-summary.md
  • runs/hf-crops-table-figure-dev-v1/training_metrics.json
  • runs/hf-crops-table-figure-dev-v1/package/model-card.json
  • outputs/audit-hf-crops-table-figure-dev-v1-model/model-audit.md
  • outputs/summary-hf-crops-table-figure-dev-v1/summary.md
  • outputs/experiment-hf-crops-table-figure-dev-v1/benchmark/report.json

Crop pack:

  • 544 crops from 80 rendered table/figure development samples
  • 384 table-cell crops, 128 line crops, 32 figure-caption crops
  • 32 skew pages skipped by design
  • 480 train / 64 eval split with text_sha256 grouping

Benchmark result:

System CER
HF crop candidate 1.043369
Surya 0.475309
Tesseract 0.077453

Supported claim:

  • crop-level training is more coherent than full-page recognition, but this one-epoch TrOCR branch still fails
  • Tesseract is the strongest crop OCR baseline on this generated crop eval
  • the next useful branch should use strong crop OCR behind better layout/cell/caption detection, or train a recognizer from substantially more real crop data

17. Table-cell crop refinement is safe only as fill-only, and still does not beat Surya

We tested the obvious follow-up to the crop eval: keep Surya for document/table structure, then use Tesseract on detected table-cell crops.

Representative artifacts:

  • outputs/table-cell-refine-v1/model-card.json
  • outputs/summary-table-cell-refine-hard15-v1/summary.md
  • outputs/table-cell-analysis-refine-always-hard15-v1/table-cell-analysis.md
  • outputs/table-cell-refine-empty-v1/model-card.json
  • outputs/summary-table-cell-refine-empty-hard15-v1/summary.md
  • outputs/table-cell-analysis-refine-empty-hard15-v1/table-cell-analysis.md
  • outputs/validation-table-cell-refine-empty-hard15-v1/validation.md
  • outputs/table-cell-refine-empty-dedup-v1/model-card.json
  • outputs/summary-table-cell-refine-empty-dedup-hard15-v3/summary.md
  • outputs/table-cell-analysis-refine-empty-dedup-hard15-v3/table-cell-analysis.md
  • outputs/validation-table-cell-refine-empty-dedup-hard15-v3/validation.md

Hard15 result:

System CER table_cell_f1 reading_order_pair_accuracy
Surya 0.348706 0.979167 0.520000
table-cell refine, always 0.348706 0.593750 0.520000
table-cell refine, empty 0.348706 0.979167 0.520000
table-cell refine, empty + parser de-dup 0.348706 0.979167 0.520000

Supported claim:

  • replacing every Surya table cell with Tesseract crop OCR is unsafe on hard15
  • empty-cell-only replacement avoids the regression but does not improve the hard pocket
  • the hard15 validation gate rejects the safe branch with status not_validated
  • cell-level analysis shows Surya is already exact on 94/96 hard15 reference table cells
  • the unsafe policy improves 0 cells over Surya and worsens 37
  • the safe policy ties all 96 cells with Surya because there are no missing Surya cells to fill
  • parser-level table de-duplication removes overlapping duplicate table exports, but it applies to candidate and Surya alike and does not improve the candidate-vs-Surya claim
  • future table work needs better cell detection/replacement confidence, not blind crop rerecognition

Figure Caption Refine Branch

Representative artifacts:

  • outputs/figure-caption-refine-v1/model-card.json
  • outputs/audit-figure-caption-refine-v1-model/model-audit.md
  • outputs/summary-figure-caption-refine-hard15-v2/summary.md
  • outputs/experiment-figure-caption-refine-hard15-v2/benchmark/report.json
  • outputs/validation-figure-caption-refine-hard15-v2/validation.md
  • outputs/figure-caption-refine-labels-v1/model-card.json
  • outputs/audit-figure-caption-refine-labels-v1-model/model-audit.md
  • outputs/summary-figure-caption-refine-labels-hard15-v1/summary.md
  • outputs/experiment-figure-caption-refine-labels-hard15-v1/benchmark/report.json
  • outputs/validation-figure-caption-refine-labels-hard15-v1/validation.md

The parser first had to be fixed so an OCR-damaged nearby caption line after [Figure: ...] is still attached as the figure caption. After that fix, Surya's figure-caption CER dropped from about 0.737609 to 0.092323.

The composite then kept Surya for document structure and used Tesseract only for overlapping figure-caption lines when the caption-prefix score improved.

System CER figure_caption_cer figure_detection_f1 table_cell_f1 reading_order_pair_accuracy
Surya 0.236839 0.092323 1.000000 0.979167 0.520000
Tesseract 1.266364 0.051020 0.857143 0.000000 0.613333
figure-caption refine 0.228008 0.052478 1.000000 0.979167 0.520000
figure-caption refine + labels 0.204610 0.008746 1.000000 0.979167 0.520000

Supported claim:

  • the figure-caption refine branch beats Surya on the configured hard_eval,figure figure_caption_cer gate
  • the structural-label branch also passes overall hard_eval CER and hard_eval figure CER gates
  • validation still rejects the overall hard15 claim because table CER and table-cell F1 gates tie Surya
  • this is a slice win, not a SOTA result

What We Cannot Claim

Do not write these into a paper unless new evidence appears.

1. SOTA

We do not have:

  • public-suite wins
  • broad holdout wins over Surya
  • a consistent story across text, reading order, and table extraction

2. Best overall OCR engine

Current evidence does not support this. On rendered document OCR text, raw Surya is still the strongest baseline in this repo.

3. Broad document-understanding leadership

The candidate has real table advantages over weak baselines, but:

  • reading order is still mixed
  • the strongest text baseline is not ours
  • figure handling is not yet a headline result
  • the larger rendered-v2 composite candidates now either collapse to Surya, trail Surya, become nearly indistinguishable from Surya, or trade too much OCR quality for modest ordering gains, which weakens even the "novel composite" story
  • even an oracle selector over Surya and Tesseract barely improves Surya on rendered-v2 text CER

4. PaddleOCR-VL and GLM-OCR are strong baselines here

They are real baselines now, but on our current rendered subset they are weak. Their presence helps the paper as negative evidence and benchmarking completeness, not as the main comparison target.

Practical Paper Framing

The current evidence is better suited to one of these framings:

Option A: Comparative engineering paper

Focus:

  • Nepali/English document OCR evaluation design
  • routed/composite OCR systems
  • why raw text CER and document-structure metrics diverge
  • why calibration discipline matters

Option B: Negative-results plus systems paper

Focus:

  • TrOCR branch failed to beat Tesseract
  • OCR-correction branch failed badly
  • richer Tesseract ensembles overfit
  • Surya-based composition helped structure and robustness, but still failed to surpass Surya itself on text CER

Option C: Evaluation and calibration paper

Focus:

  • claim preflight
  • calibration provenance
  • holdout discipline
  • rendered eval-pack construction
  • scoring external markdown-only OCR systems

This is less glamorous, but it is honest and technically defensible.

Figures and Tables Worth Reusing

These are the obvious paper figures/tables to build first.

The repo now has a reproducible export for these tables:

ocrtech paper-benchmark-table \
  --summary outputs/paper-claimpack-summary-tesseract/summary.json \
  --summary outputs/paper-claimpack-summary-surya/summary.json \
  --summary outputs/paper-claimpack-summary-stock-paddle/summary.json \
  --out outputs/paper-claimpack-table \
  --system candidate \
  --system surya \
  --system tesseract \
  --system stock-paddle \
  --label 'candidate=Candidate' \
  --label 'surya=Surya' \
  --label 'tesseract=Tesseract' \
  --label 'stock-paddle=Stock Paddle' \
  --source-preference 'candidate=outputs/paper-claimpack-summary-tesseract/summary.json' \
  --source-preference 'tesseract=outputs/paper-claimpack-summary-tesseract/summary.json' \
  --source-preference 'surya=outputs/paper-claimpack-summary-surya/summary.json' \
  --source-preference 'stock-paddle=outputs/paper-claimpack-summary-stock-paddle/summary.json' \
  --metric cer \
  --metric table_cell_f1 \
  --metric reading_order_pair_accuracy \
  --metric latency_seconds

This writes:

  • paper-table.json
  • paper-table.md

Latency is now part of the generated tables. Memory still is not, because the benchmark reports do not yet capture it.

Table 1: Branch chronology

Use:

  • TrOCR mixed eval
  • Tesseract baseline
  • ByT5 corrector composite
  • quality router
  • script router
  • script router with backfill
  • Surya
  • stock-Paddle
  • GLM-OCR
  • PaddleOCR-VL

Table 2: Rendered 25-sample claim-pack comparison

Use artifacts:

  • outputs/live-k2-claimpack-rendered-backfill/summary.json
  • outputs/live-k2-claimpack-surya-compare/summary.json
  • outputs/live-k2-claimpack-stock-paddle-fixed/summary.json

Columns:

  • overall CER
  • blur / scan / low_contrast / skew CER
  • table_cell_f1
  • table_exact_match
  • reading_order_pair_accuracy

Table 3: Blur-only 5-sample expensive-baseline subset

Use artifacts:

  • outputs/live-k2-rendered-blur5/summary.json

Columns:

  • CER
  • table_cell_f1
  • reading_order_pair_accuracy
  • runtime if you decide to extend the summary with latency

Figure 1: Example multi-page backfill recovery

Use:

  • outputs/live-k2-multipage-backfill/candidate-multipage-document.json

Show:

  • primary OCR text kept from Tesseract path
  • table structure recovered from Surya backfill

Gaps Before a Stronger Paper

These are the missing experiments that would materially improve the paper instead of just making it longer.

  1. Recalibrate the backfill candidate on a rendered-document calibration split, not the older line-style split.
  2. Evaluate that rendered-calibrated candidate on an untouched rendered holdout against at least surya, tesseract, and stock-paddle.
  3. Add latency and memory tables for the same runs.
  4. Run one honest public benchmark suite if a relevant Nepali/Devanagari or multilingual document pack can be made to fit the scoring path.
  5. Do targeted error analysis on:
    • forms
    • reading-order failures
    • places where Surya wins on text but the candidate wins on structure
  6. Add or benchmark a genuinely different recognizer/document model, because the Surya/Tesseract routing family is now bounded by an oracle with negligible headroom.

Public Benchmark Gate Added

We added a public benchmark registry and manifest audit:

  • command: ocrtech list-public-benchmarks
  • command: ocrtech audit-public-benchmark
  • SOTA portfolio config: configs/claim-review-sota-v1.json

Registry targets:

  • opendatalab/OmniDocBench
  • PaddlePaddle/Real5-OmniDocBench
  • allenai/olmOCR-bench

This is a claim discipline change, not a benchmark result. The audit requires public benchmark provenance (metadata.public_benchmark, metadata.benchmark_name, metadata.source_url) and references before a manifest can be used as SOTA evidence. The paper should continue to label generated-pack results as internal validation until at least one public benchmark experiment exists and passes review.

OmniDocBench Schema Inspection

Verified on k2 on 2026-06-14 before writing the converter:

  • dataset: opendatalab/OmniDocBench
  • annotation file: OmniDocBench.json
  • rows: 1651
  • top-level row keys: layout_dets, page_info, extra
  • page_info contains image_path, page_no, height, width, page_attribute
  • page_attribute contains data_source, language, layout, special_issue, subset
  • layout_dets items use category_type, poly, ignore, order, anno_id, attribute
  • text and caption blocks carry text
  • table blocks carry html
  • every page_info.image_path resolved under images/

The implemented converter is ocrtech prepare-public-benchmark opendatalab/OmniDocBench. It emits manifests/omnidocbench-eval.jsonl, copied page images, structured JSON references, public provenance metadata, and table cell objects parsed from source HTML including row/column spans.

OmniDocBench Public Manifest Smoke

Verified on k2 on 2026-06-14:

ocrtech prepare-public-benchmark opendatalab/OmniDocBench \
  --out data/public/omnidocbench-smoke \
  --limit 5 \
  --language english \
  --require-tables

ocrtech audit-public-benchmark \
  data/public/omnidocbench-smoke/manifests/omnidocbench-eval.jsonl \
  --benchmark opendatalab/OmniDocBench \
  --min-samples 5 \
  --out outputs/public-omnidocbench-smoke-audit

Result:

  • prepared samples: 5
  • audit status: passed
  • audit issues: none
  • slices: public=5, document_parsing=5, english=5, table=5
  • harder public slices included: table_hard=2, equation_hard=2, layout_hard=1

This proves the public benchmark conversion and provenance gate work on real OmniDocBench data. It is still only a smoke conversion, not a public benchmark accuracy result.

OmniDocBench Public Accuracy Smoke

Verified on k2 on 2026-06-14:

ocrtech run-experiment \
  --eval-manifest data/public/omnidocbench-smoke/manifests/omnidocbench-eval.jsonl \
  --out outputs/experiment-public-omnidocbench-smoke-v1 \
  --baselines candidate,surya \
  --candidate-model-config outputs/figure-caption-table-lexicon-v1/model-card.json \
  --skip-validation

ocrtech summarize-benchmark \
  --benchmark-report outputs/experiment-public-omnidocbench-smoke-v1/benchmark/report.json \
  --out outputs/summary-public-omnidocbench-smoke-v1 \
  --candidate candidate \
  --baseline surya \
  --metric cer \
  --metric table_cell_f1 \
  --metric table_exact_match \
  --metric table_teds_structure \
  --metric table_teds_text

ocrtech validate-claim \
  --benchmark-report outputs/experiment-public-omnidocbench-smoke-v1/benchmark/report.json \
  --config configs/validation-public-omnidocbench-surya.json \
  --eval-manifest data/public/omnidocbench-smoke/manifests/omnidocbench-eval.jsonl \
  --out outputs/validation-public-omnidocbench-smoke-v1

Result:

system samples CER table_cell_f1 table_exact_match table_teds_structure table_teds_text mean latency sec
candidate 5 0.481468 0.232540 0.000000 0.700358 0.465758 72.082501
Surya 5 0.480531 0.232540 0.000000 0.704062 0.465758 66.597825

Validation status: not_validated.

Failure reasons:

  • public sample count is 5, below the required 100
  • table sample count is 5, below the required 40
  • candidate CER is slightly worse than Surya: improvement -0.000936
  • candidate table TEDS structure is slightly worse than Surya: improvement -0.003704
  • candidate ties Surya on table cell F1 and text-sensitive table TEDS

This is an important negative result. The current figure_caption_refine_composite candidate is not a public OmniDocBench improvement over Surya. Its internal generated-pack table wins do not transfer to this public English table-heavy smoke. The next public-benchmark work should focus on better table matching, reference alignment, and page-level structure extraction before scaling to 100+ public samples.

OmniDocBench Public Rescore With Best-Match Tables

The initial public smoke exposed a validation weakness: table metrics compared reference table i to predicted table i. That is too brittle for public document parsing because engines may emit small table fragments before the main table. We changed table scoring to greedily match each reference table to the best unused predicted table by table similarity, then rescored the frozen OCR outputs without rerunning OCR:

ocrtech rescore-benchmark \
  --benchmark-report outputs/experiment-public-omnidocbench-smoke-v1/benchmark/report.json \
  --eval-manifest data/public/omnidocbench-smoke/manifests/omnidocbench-eval.jsonl \
  --out outputs/experiment-public-omnidocbench-smoke-v2-rescored/benchmark

ocrtech summarize-benchmark \
  --benchmark-report outputs/experiment-public-omnidocbench-smoke-v2-rescored/benchmark/report.json \
  --out outputs/summary-public-omnidocbench-smoke-v2-rescored \
  --candidate candidate \
  --baseline surya \
  --metric cer \
  --metric table_cell_f1 \
  --metric table_exact_match \
  --metric table_teds_structure \
  --metric table_teds_text

Corrected result:

system samples CER table_cell_f1 table_exact_match table_teds_structure table_teds_text mean latency sec
candidate 5 0.481468 0.296264 0.000000 0.740128 0.526327 72.082501
Surya 5 0.480531 0.296264 0.000000 0.740128 0.526327 66.597825

Validation remains not_validated:

  • public sample count is 5, below the required 100
  • table sample count is 5, below the required 40
  • candidate CER is still slightly worse than Surya: improvement -0.000936
  • candidate ties Surya on table cell F1, structure TEDS, and text TEDS

Table-shape diagnostics after the corrected rescore:

sample reference shape predicted shape exact cells missing cells diagnosis
omnidocbench-00142... 6x2 6x4 0 0 column segmentation mismatch
omnidocbench-00206... 6x3; 6x4 6x3; 2x7; 7x4 12 16 extra split table and row/column drift
omnidocbench-00208... 3x5 1x3; 16x5 1 12 table fragment emitted before oversized main table

This corrected evaluator is stricter and more defensible. It also narrows the model problem: the current candidate is mostly Surya with post-processing, so public failures are inherited Surya table segmentation failures. We need a real table reconstruction improvement, not just table-name text correction.

OmniDocBench Public Table-Normalization Candidate

We added a conservative runtime table cleanup to the figure_caption_refine_composite candidate:

  • drop entirely empty columns
  • drop separator-only columns such as —
  • drop exact duplicate adjacent columns

This is reference-free post-processing. It is not learned from OmniDocBench labels, and it does not inspect ground truth at runtime.

Verified on k2 on 2026-06-14 with model card:

  • outputs/figure-caption-table-lexicon-normalize-v2/model-card.json
  • model id: ocrtech-surya-tesseract-figure-table-lexicon-normalize-v2
  • model audit: outputs/audit-figure-caption-table-lexicon-normalize-v2-model/model-audit.json
  • audit status: passed

Public smoke comparison:

ocrtech run-experiment \
  --eval-manifest data/public/omnidocbench-smoke/manifests/omnidocbench-eval.jsonl \
  --out outputs/experiment-public-omnidocbench-smoke-normalize-v2-candidate \
  --baselines candidate \
  --candidate-model-config outputs/figure-caption-table-lexicon-normalize-v2/model-card.json \
  --skip-validation

We combined the new candidate rows with the unchanged Surya baseline rows from outputs/experiment-public-omnidocbench-smoke-v2-rescored/benchmark/report.json to avoid rerunning the same Surya baseline.

Result:

system samples CER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text
candidate normalize-v2 5 0.480080 0.496264 0.200000 0.759091 0.817547 0.603746
Surya 5 0.480531 0.296264 0.000000 0.573377 0.740128 0.526327

Cell analysis:

  • reference cells: 116
  • candidate exact cells: 56
  • Surya exact cells: 44
  • candidate better cells: 12
  • Surya better cells: 1
  • ties: 103

The main fixed sample is omnidocbench-00142..., where Surya emitted a 6x4 table with a separator column and duplicated ID column. The normalized candidate emits the expected 6x2 shape and reaches 12/12 exact cells on that table.

Validation status remains not_validated:

  • public sample count is 5, below the required 100
  • table sample count is 5, below the required 40
  • CER improvement is only 0.000452, below the required 0.005
  • bootstrap CI lower bounds are still 0.000000 on table gates because this is a tiny smoke and only one page materially improves

This is real progress over Surya on public table structure, but not enough evidence for a claim. The next step should be a larger public-table slice and a second table improvement aimed at row-fragment merging for pages like omnidocbench-00208....

Public OmniDocBench Row-Fragment Merge Candidate

We added a second reference-free table-grid postprocessor for sparse bullet-list row fragments. The rule is intentionally narrow: if a row has exactly one non-empty cell in column 0, that text starts with a bullet marker, and the previous kept row has a later bullet-list cell, append the fragment into that later cell and drop the sparse row. This targets failures where a multi-item cell is emitted as several one-cell rows.

Verified locally:

  • uv run pytest: 196 passed, 7 skipped

Verified on k2 on 2026-06-14 with model card:

  • outputs/figure-caption-table-lexicon-normalize-rows-v3/model-card.json
  • model id: ocrtech-surya-tesseract-figure-table-lexicon-normalize-rows-v3
  • model audit: outputs/audit-figure-caption-table-lexicon-normalize-rows-v3-model/model-audit.json
  • audit status: passed

Public smoke artifacts:

  • candidate-only run: outputs/experiment-public-omnidocbench-smoke-normalize-rows-v3-candidate/benchmark/report.json
  • combined candidate/Surya report: outputs/experiment-public-omnidocbench-smoke-normalize-rows-v3-combined/benchmark/report.json
  • summary: outputs/summary-public-omnidocbench-smoke-normalize-rows-v3-combined/summary.json
  • validation: outputs/validation-public-omnidocbench-smoke-normalize-rows-v3-combined/validation.json
  • cell analysis: outputs/table-cell-analysis-public-omnidocbench-smoke-normalize-rows-v3-combined/table-cell-analysis.json

Result:

system samples CER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text
candidate normalize-rows-v3 5 0.481629 0.564335 0.200000 0.934091 0.942836 0.675633
Surya 5 0.480531 0.296264 0.000000 0.573377 0.740128 0.526327

Compared with the v2 grid-normalization smoke, v3 mainly fixes the omnidocbench-00208... shape failure: the candidate table changes from the Surya-style 16x5 row-fragment table to 3x5, matching the reference shape. The combined report still has an extra 1x3 explanatory-text table on that page, and cell exactness for that sample remains poor (1/15 exact cells).

Cell analysis over the five public pages:

  • reference cells: 116
  • candidate exact cells: 56
  • Surya exact cells: 44
  • candidate better cells: 12
  • Surya better cells: 1
  • ties: 103

Validation status remains not_validated:

  • public sample count is 5, below the required 100
  • table sample count is 5, below the required 40
  • text CER regressed slightly against Surya on this smoke (0.481629 vs 0.480531)
  • table gates have positive mean improvements, but bootstrap CI lower bounds are still 0.000000 because only a few pages improve

This is a useful parser/postprocessing step for table structure, not a claim. The next useful experiment is a larger public table slice. If CER remains sensitive to Surya reruns or table text reconstruction, separate page-text CER from table-only metrics in the paper tables.

Public OmniDocBench Table40 Benchmark

We prepared and audited a larger public table-only slice on k2:

ocrtech prepare-public-benchmark opendatalab/OmniDocBench \
  --out data/public/omnidocbench-table40 \
  --limit 40 \
  --language english \
  --require-tables

ocrtech audit-public-benchmark \
  data/public/omnidocbench-table40/manifests/omnidocbench-eval.jsonl \
  --benchmark opendatalab/OmniDocBench \
  --min-samples 40 \
  --out outputs/public-omnidocbench-table40-audit

Audit result:

  • samples: 40
  • audit status: passed
  • slices: public=40, document_parsing=40, english=40, table=40, table_hard=32, figure=6
  • reference tables: 46
  • table-cell analysis reference cells: 10314

Benchmark command:

ocrtech run-experiment \
  --eval-manifest data/public/omnidocbench-table40/manifests/omnidocbench-eval.jsonl \
  --out outputs/experiment-public-omnidocbench-table40-normalize-rows-v3 \
  --baselines candidate,surya \
  --candidate-model-config outputs/figure-caption-table-lexicon-normalize-rows-v3/model-card.json \
  --skip-validation

Artifacts:

  • benchmark: outputs/experiment-public-omnidocbench-table40-normalize-rows-v3/benchmark/report.json
  • summary: outputs/summary-public-omnidocbench-table40-normalize-rows-v3/summary.json
  • validation: outputs/validation-public-omnidocbench-table40-normalize-rows-v3/validation.json
  • cell analysis: outputs/table-cell-analysis-public-omnidocbench-table40-normalize-rows-v3/table-cell-analysis.json

Result:

system samples CER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text mean latency sec
candidate normalize-rows-v3 40 0.713326 0.323122 0.058333 0.473233 0.520959 0.361666 67.947445
Surya 40 0.713960 0.280062 0.000000 0.414052 0.503603 0.358057 65.001688

Paired table metrics against Surya:

metric abs improvement wins losses ties validation note
CER 0.000634 1 1 38 fails: below margin, CI crosses zero, only 40/100 public docs
table_cell_f1 0.043059 7 9 24 fails: CI low -0.019152
table_structure_similarity 0.059181 8 7 25 diagnostic only, not in public config
table_teds_structure 0.017355 10 7 23 fails: CI low -0.038381
table_teds_text 0.003609 7 10 23 fails: below margin and CI low -0.040299
table_exact_match 0.058333 3 0 37 improves only three pages

Cell analysis:

  • candidate exact cells: 3440/10314
  • Surya exact cells: 3421/10314
  • candidate-better cells: 269
  • Surya-better cells: 224
  • ties: 9821
  • candidate missing cells: 7727
  • Surya missing cells: 7504

This is the most useful public result so far because it meets the table gate's 40 paired-page count. It still does not validate the candidate. The candidate has higher means on several table metrics, but the paired distribution is weak: more table-cell-F1 losses than wins, unstable TEDS text, and only marginal cell exactness improvement. The public-document CER gate also cannot pass on this slice because it requires 100 document pages and a 0.005 CER margin.

Paper framing: this is a negative/diagnostic public benchmark result. It shows that the current reference-free table postprocessors fix certain shape failures but do not generalize strongly enough across broad OmniDocBench tables. The next model work should target large/wide tables and table segmentation, not another small lexical or row-fragment heuristic.

Public OmniDocBench Table40 Width-Guard Rerun

The table40 analysis showed that v3 damaged wide sparse tables:

  • omnidocbench-00294...: normalized_table_grid_merged_rows=15, table-cell F1 collapsed to 0.025641 while Surya had 0.475978
  • omnidocbench-00216...: normalized_table_grid_dropped_columns=26, table-cell F1 fell to 0.079576 while Surya had 0.242276

We added a deliberately conservative guard: skip table-grid normalization entirely when the predicted table width is greater than 12. This keeps the small-table fixes while avoiding the wide sparse matrices where empty columns, duplicate-looking columns, and one-cell rows can be legitimate structure.

Verified locally:

  • uv run pytest: 198 passed, 7 skipped

Verified on k2 with model card:

  • outputs/figure-caption-table-lexicon-normalize-width-guard-v4/model-card.json
  • model id: ocrtech-surya-tesseract-figure-table-lexicon-normalize-width-guard-v4
  • audit: outputs/audit-figure-caption-table-lexicon-normalize-width-guard-v4-model/model-audit.json
  • audit status: passed

First, a targeted 5-page public subset confirmed the guard preserves the intended small-table fixes and repairs the known wide-table damage:

  • manifest: data/public/omnidocbench-width-guard-smoke/manifests/omnidocbench-eval.jsonl
  • benchmark: outputs/experiment-public-omnidocbench-width-guard-smoke-v4/benchmark/report.json
  • summary: outputs/summary-public-omnidocbench-width-guard-smoke-v4/summary.json
  • cell analysis: outputs/table-cell-analysis-public-omnidocbench-width-guard-smoke-v4/table-cell-analysis.json

Selected per-page checks:

sample v3 candidate F1 v4 candidate F1 v4 Surya F1 note
00142... 1.000000 1.000000 0.000000 compact separator/duplicate fix preserved
00208... 0.466667 0.466667 0.126316 compact row-fragment fix preserved
00216... 0.079576 0.200331 0.000000 wide-table damage reduced
00294... 0.025641 0.635692 0.469505 wide sparse-row damage repaired
00300... 1.000000 1.000000 0.000000 compact exact-table fix preserved

Then we ran v4 candidate-only on the full table40 slice and combined it with the already completed Surya rows from the previous table40 run:

  • candidate run: outputs/experiment-public-omnidocbench-table40-width-guard-v4-candidate/benchmark/report.json
  • combined report: outputs/experiment-public-omnidocbench-table40-width-guard-v4-combined/benchmark/report.json
  • summary: outputs/summary-public-omnidocbench-table40-width-guard-v4-combined/summary.json
  • validation: outputs/validation-public-omnidocbench-table40-width-guard-v4-combined/validation.json
  • cell analysis: outputs/table-cell-analysis-public-omnidocbench-table40-width-guard-v4-combined/table-cell-analysis.json

Full table40 result:

system samples CER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text mean latency sec
candidate width-guard-v4 40 0.707864 0.339319 0.058333 0.480787 0.545375 0.386013 71.991723
Surya 40 0.713960 0.280062 0.000000 0.414052 0.503603 0.358057 65.001688

Compared with v3 on the same table40 slice:

  • table-cell F1 improved from 0.323122 to 0.339319
  • table TEDS structure improved from 0.520959 to 0.545375
  • table TEDS text improved from 0.361666 to 0.386013
  • candidate exact cells improved from 3440/10314 to 3453/10314
  • candidate missing cells reduced from 7727 to 7533

Validation status remains not_validated, but the evidence is stronger:

gate status key numbers
public/document CER fail 40/100 pairs; improvement 0.006096, CI low 0.000000
table_cell_f1 pass improvement 0.059256, CI low 0.001433
table_teds_structure pass improvement 0.041772, CI low 0.004790
table_teds_text fail improvement 0.027957, CI low -0.000220

This is the first public OmniDocBench result where two configured public table gates pass on 40 pages. It is still not a SOTA claim: text-sensitive table quality is borderline, broad document CER is underpowered at 40/100 pages, and cell exactness is only marginally better than Surya (3453/10314 vs 3421/10314). The next paper-grade step is either a 100-page public run or a real table-text improvement that moves table_teds_text CI clearly above zero.

Public OmniDocBench Table40 Deterministic Postprocessor A/B

The v4 full run above compares a candidate rerun against Surya rows from an earlier run. That is useful deployment evidence, but it mixes two effects: postprocessing changes and Surya run-to-run OCR/layout variance. For paper analysis, we added a deterministic benchmark mode that derives the candidate from the exact same document.json emitted by an existing baseline and then rescored both outputs.

Command:

ocrtech derive-table-postprocess-benchmark \
  --benchmark-report outputs/experiment-public-omnidocbench-table40-normalize-rows-v3/benchmark/report.json \
  --source-baseline surya \
  --candidate-baseline candidate \
  --eval-manifest data/public/omnidocbench-table40/manifests/omnidocbench-eval.jsonl \
  --out outputs/experiment-public-omnidocbench-table40-derived-postprocess-v1/benchmark

Artifacts:

  • benchmark: outputs/experiment-public-omnidocbench-table40-derived-postprocess-v1/benchmark/report.json
  • summary: outputs/summary-public-omnidocbench-table40-derived-postprocess-v1/summary.json
  • validation: outputs/validation-public-omnidocbench-table40-derived-postprocess-v1/validation.json
  • cell analysis: outputs/table-cell-analysis-public-omnidocbench-table40-derived-postprocess-v1/table-cell-analysis.json

Aggregate deterministic A/B result:

system samples CER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text
derived candidate 40 0.713960 0.340092 0.058333 0.480343 0.547003 0.386181
fixed Surya source 40 0.713960 0.280062 0.000000 0.414052 0.503603 0.358057

Paired table gates on the same source OCR/layout:

gate status key numbers
public/document CER fail 40/100 pairs; improvement 0.000000; CER intentionally unchanged
table_cell_f1 pass improvement 0.060029; CI low 0.001601; wins/losses/ties 7/6/27
table_teds_structure pass improvement 0.043399; CI low 0.008255; wins/losses/ties 9/5/26
table_teds_text pass improvement 0.028124; CI low 0.001152; wins/losses/ties 7/7/26

Cell-level deterministic analysis:

  • candidate exact cells: 3495/10314
  • Surya exact cells: 3421/10314
  • candidate-better cells: 120
  • Surya-better cells: 37
  • ties: 10157
  • candidate exact / Surya not: 98
  • Surya exact / candidate not: 24
  • candidate missing cells: 7497
  • Surya missing cells: 7504

Paper framing: this is the cleanest evidence for the table postprocessor itself. It shows the reference-free table-grid and header-span repairs improve table structure and text-sensitive table similarity when the upstream OCR/layout is held fixed. It still does not validate an end-to-end OCR SOTA claim because CER cannot improve in this setup, the public document gate needs 100 pages, and the paired table improvements are concentrated in a minority of pages.

Public OmniDocBench Table100 Claim Run

To close the 40/100 public document-page gap, we prepared and audited a claim-sized OmniDocBench table slice on k2:

ocrtech prepare-public-benchmark opendatalab/OmniDocBench \
  --out data/public/omnidocbench-table100 \
  --limit 100 \
  --language english \
  --require-tables

ocrtech audit-public-benchmark \
  data/public/omnidocbench-table100/manifests/omnidocbench-eval.jsonl \
  --benchmark opendatalab/OmniDocBench \
  --min-samples 100 \
  --out outputs/public-omnidocbench-table100-audit

Audit result:

  • samples: 100
  • audit status: passed
  • slices: public=100, document_parsing=100, english=100, table=100, table_hard=32, figure=22
  • artifact: outputs/public-omnidocbench-table100-audit/public-benchmark-audit.json

The long candidate-vs-Surya benchmark was launched on 2026-06-14 with the source-controlled runner:

uv run python scripts/run_public_table100_claim.py \
  --manifest data/public/omnidocbench-table100/manifests/omnidocbench-eval.jsonl \
  --candidate-model-config outputs/figure-caption-table-lexicon-normalize-width-guard-v4/model-card.json \
  --out outputs/experiment-public-omnidocbench-table100-width-guard-v4 \
  --sample-timeout-seconds 300

Run tracking:

  • remote PID at launch: 88058
  • current isolated-run PID after restart: 90501
  • log: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run.log
  • live status for future restarts: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json
  • PID file: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run.pid
  • expected benchmark report: outputs/experiment-public-omnidocbench-table100-width-guard-v4/benchmark/report.json
  • expected summary: outputs/experiment-public-omnidocbench-table100-width-guard-v4/summary/summary.json
  • expected validation: outputs/experiment-public-omnidocbench-table100-width-guard-v4/validation/validation.json
  • expected cell analysis: outputs/experiment-public-omnidocbench-table100-width-guard-v4/table-cell-analysis/table-cell-analysis.json

Operational note: the first all-or-nothing process reached three completed candidate rows and then stalled. We added benchmark checkpointing, --resume-existing, isolated per-page subprocesses, and --sample-timeout-seconds support, then restarted the run. The first measured timeout row was omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357, recorded as TimeoutExpired: timed out after 300s. A timeout is an explicit page error and should be counted against the candidate in validation.

Because PID 90501 was started before the runner's run-status.json write was loaded into that process, the active run's source of truth is the checkpointed benchmark/report.json. Future restarts with the current script should also write run-status.json.

Paper framing: this run is the first attempt at the full configured public OmniDocBench gate. It is not a result yet. Do not cite pass/fail numbers until the benchmark report, summary, validation, and cell-analysis artifacts exist and have been inspected.

2026-06-15 continuation status:

  • k2 time checked: 2026-06-15 00:03:12 +0545
  • active PID: 90501
  • checkpoint report rows: 15
  • rows by status: 14 ok, 1 error
  • rows by baseline: candidate=15
  • last completed sample: omnidocbench-00267-page-1cecd66d-3563-422a-a204-999bba0bb528
  • in-flight sample at the check: page-50ffe5d2-cd27-449a-a813-d4a8c44b12ea
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This is still a live run, not claim evidence. The important engineering outcome is that the isolated runner advanced beyond the measured timeout page and continued producing checkpoint rows.

2026-06-15 later checkpoint:

  • k2 time checked: 2026-06-15 00:06:16 +0545
  • active PID: 90501
  • checkpoint report rows: 17
  • rows by status: 16 ok, 1 error
  • rows by baseline: candidate=17
  • last completed sample: omnidocbench-00271-page-be2ebf31-5f24-4d3e-97f0-4a8df7e773b4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains live and should be allowed to continue. The candidate stage is still in progress; Surya baseline rows have not started yet.

2026-06-15 00:07 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 18
  • rows by status: 17 ok, 1 error
  • rows by baseline: candidate=18
  • last completed sample: omnidocbench-00273-page-035cb436-c01e-41db-b40e-8977678777eb
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:08 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 19
  • rows by status: 18 ok, 1 error
  • rows by baseline: candidate=19
  • last completed sample: omnidocbench-00274-page-49746735-505c-442d-b3f6-c55b1bf9e48b
  • in-flight sample at the check: page-9390721f-e06c-47ea-8b21-0f27c72df7b7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:11 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 21
  • rows by status: 20 ok, 1 error
  • rows by baseline: candidate=21
  • last completed sample: omnidocbench-00276-page-5d58cdf7-6d00-4ea5-97b2-358cbc0cab21
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:12 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 22
  • rows by status: 21 ok, 1 error
  • rows by baseline: candidate=22
  • last completed sample: omnidocbench-00278-page-f222e5f3-8628-47cd-a0b6-40f62123228f
  • in-flight sample at the check: page-c7771a62-5b6b-48f0-b591-6b657d26e125
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:13 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 23
  • rows by status: 22 ok, 1 error
  • rows by baseline: candidate=23
  • last completed sample: omnidocbench-00279-page-c7771a62-5b6b-48f0-b591-6b657d26e125
  • in-flight sample at the check: page-5cb4b2fe-c14d-450e-9b97-4b6e7c99493b
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:15 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 24
  • rows by status: 23 ok, 1 error
  • rows by baseline: candidate=24
  • last completed sample: omnidocbench-00280-page-5cb4b2fe-c14d-450e-9b97-4b6e7c99493b
  • in-flight sample at the check: page-eff3dc3a-8d44-4c68-9c3a-55df4aaf1e36
  • final summary, validation, and table-cell-analysis artifacts: not present yet

2026-06-15 00:17 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 26
  • rows by status: 25 ok, 1 error
  • rows by baseline: candidate=26
  • last completed sample: omnidocbench-00283-page-213e6a57-ba81-4afb-8184-2e1cf2a7664d
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains in the candidate stage. Surya baseline rows have not started yet, so no candidate-vs-Surya conclusion can be drawn from this checkpoint.

2026-06-15 00:18 checkpoint:

  • k2 time checked: 2026-06-15 00:18:13 +0545
  • active PID: 90501
  • checkpoint report rows: 27
  • rows by status: 26 ok, 1 error
  • rows by baseline: candidate=27
  • last completed sample: omnidocbench-00284-page-14a6b411-9097-4eec-86da-92075868d243
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This is still candidate-stage progress only. Surya baseline rows have not started yet.

2026-06-15 00:18:58 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 28
  • rows by status: 27 ok, 1 error
  • rows by baseline: candidate=28
  • last completed sample: omnidocbench-00285-page-0cbdcfa9-3248-4e54-8704-2bc73e6d29e7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run is still active and has not reached the Surya baseline phase.

2026-06-15 00:19 checkpoint:

  • k2 time checked: 2026-06-15 00:19:57 +0545
  • active PID: 90501
  • checkpoint report rows: 29
  • rows by status: 28 ok, 1 error
  • rows by baseline: candidate=29
  • last completed sample: omnidocbench-00286-page-458ab820-615f-42fe-af33-3a8fb15ad691
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains healthy and is still in the candidate phase.

2026-06-15 00:20 checkpoint:

  • k2 time checked: 2026-06-15 00:20:47 +0545
  • active PID: 90501
  • checkpoint report rows: 30
  • rows by status: 29 ok, 1 error
  • rows by baseline: candidate=30
  • last completed sample: omnidocbench-00287-page-b23ec9d4-e4aa-458f-8e65-c2c7f8e9dd46
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active. It still has not reached the Surya baseline phase.

2026-06-15 00:21 checkpoint:

  • k2 time checked: 2026-06-15 00:21:43 +0545
  • active PID: 90501
  • checkpoint report rows: 31
  • rows by status: 30 ok, 1 error
  • rows by baseline: candidate=31
  • last completed sample: omnidocbench-00289-page-f8c08102-a96d-4533-b3a1-dac6020513b5
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This remains candidate-only progress. No Surya rows have been written yet.

2026-06-15 00:22 checkpoint:

  • k2 time checked: 2026-06-15 00:22:28 +0545
  • active PID: 90501
  • checkpoint report rows: 32
  • rows by status: 31 ok, 1 error
  • rows by baseline: candidate=32
  • last completed sample: omnidocbench-00290-page-38fc7c6b-80ed-46be-bb53-29986a63ad3e
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run is still active and has not reached Surya rows.

2026-06-15 00:24 checkpoint:

  • k2 time checked: 2026-06-15 00:24:56 +0545
  • active PID: 90501
  • checkpoint report rows: 33
  • rows by status: 32 ok, 1 error
  • rows by baseline: candidate=33
  • last completed sample: omnidocbench-00291-page-f3e750d8-8201-4a22-9419-9776798b7a32
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This confirms the runner is still making progress after the temporary flat checkpoint at 32 rows.

2026-06-15 00:28 checkpoint:

  • k2 time checked: 2026-06-15 00:28:15 +0545
  • active PID: 90501
  • checkpoint report rows: 36
  • rows by status: 35 ok, 1 error
  • rows by baseline: candidate=36
  • last completed sample: omnidocbench-00300-yanbaopptmerge_PattPatelCh12.pdf_27
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The runner advanced through three more candidate pages after the flat 33-row check, so no restart was needed.

2026-06-15 00:29 checkpoint:

  • k2 time checked: 2026-06-15 00:29:49 +0545
  • active PID: 90501
  • checkpoint report rows: 37
  • rows by status: 36 ok, 1 error
  • rows by baseline: candidate=37
  • last completed sample: omnidocbench-00433-docstructbench_llm-raw-scihub-o.O-ceat.200600266.pdf_2
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:30 checkpoint:

  • k2 time checked: 2026-06-15 00:30:58 +0545
  • active PID: 90501
  • checkpoint report rows: 39
  • rows by status: 38 ok, 1 error
  • rows by baseline: candidate=39
  • last completed sample: omnidocbench-00446-docstructbench_llm-raw-scihub-o.O-hup.777.pdf_7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and has not reached the Surya baseline phase.

2026-06-15 00:32 checkpoint:

  • k2 time checked: 2026-06-15 00:32:13 +0545
  • active PID: 90501
  • checkpoint report rows: 40
  • rows by status: 39 ok, 1 error
  • rows by baseline: candidate=40
  • last completed sample: omnidocbench-00447-docstructbench_llm-raw-scihub-o.O-hup.908.pdf_5
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This reaches 40 candidate rows in the Table100 run, but it is not comparable to the earlier table40 validation yet because Surya rows for this run have not started.

2026-06-15 00:33 checkpoint:

  • k2 time checked: 2026-06-15 00:33:32 +0545
  • active PID: 90501
  • checkpoint report rows: 41
  • rows by status: 40 ok, 1 error
  • rows by baseline: candidate=41
  • last completed sample: omnidocbench-00451-docstructbench_llm-raw-scihub-o.O-ijc.22994.pdf_3
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:34 checkpoint:

  • k2 time checked: 2026-06-15 00:34:33 +0545
  • active PID: 90501
  • checkpoint report rows: 42
  • rows by status: 41 ok, 1 error
  • rows by baseline: candidate=42
  • last completed sample: omnidocbench-00465-docstructbench_llm-raw-scihub-o.O-j.chroma.2005.02.018.pdf_6
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and still has not started Surya rows.

2026-06-15 00:35 checkpoint:

  • k2 time checked: 2026-06-15 00:35:27 +0545
  • active PID: 90501
  • checkpoint report rows: 43
  • rows by status: 42 ok, 1 error
  • rows by baseline: candidate=43
  • last completed sample: omnidocbench-00466-docstructbench_llm-raw-scihub-o.O-j.chroma.2005.03.134.pdf_6
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:36 checkpoint:

  • k2 time checked: 2026-06-15 00:36:17 +0545
  • active PID: 90501
  • checkpoint report rows: 44
  • rows by status: 43 ok, 1 error
  • rows by baseline: candidate=44
  • last completed sample: omnidocbench-00467-docstructbench_llm-raw-scihub-o.O-j.chroma.2005.05.085.pdf_4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and has not started Surya rows.

2026-06-15 00:37 checkpoint:

  • k2 time checked: 2026-06-15 00:37:13 +0545
  • active PID: 90501
  • checkpoint report rows: 45
  • rows by status: 44 ok, 1 error
  • rows by baseline: candidate=45
  • last completed sample: omnidocbench-00468-docstructbench_llm-raw-scihub-o.O-j.chroma.2005.08.021.pdf_7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:37:56 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 46
  • rows by status: 45 ok, 1 error
  • rows by baseline: candidate=46
  • last completed sample: omnidocbench-00494-docstructbench_llm-raw-scihub-o.O-s00374-003-0589-2.pdf_6
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:38 checkpoint:

  • k2 time checked: 2026-06-15 00:38:53 +0545
  • active PID: 90501
  • checkpoint report rows: 47
  • rows by status: 46 ok, 1 error
  • rows by baseline: candidate=47
  • last completed sample: omnidocbench-00495-docstructbench_llm-raw-scihub-o.O-s00374-004-0768-9.pdf_2
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:40 checkpoint:

  • k2 time checked: 2026-06-15 00:40:09 +0545
  • active PID: 90501
  • checkpoint report rows: 48
  • rows by status: 47 ok, 1 error
  • rows by baseline: candidate=48
  • last completed sample: omnidocbench-00496-docstructbench_llm-raw-scihub-o.O-s00374-006-0103-8.pdf_11
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:41 checkpoint:

  • k2 time checked: 2026-06-15 00:41:10 +0545
  • active PID: 90501
  • checkpoint report rows: 50
  • rows by status: 49 ok, 1 error
  • rows by baseline: candidate=50
  • last completed sample: omnidocbench-00542-docstructbench_llm-raw-the-eye-o.O-TraneGuide.pdf_7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This is the halfway point of candidate rows for the Table100 run. It is still not a paired public result because Surya rows have not started.

2026-06-15 00:44 checkpoint:

  • k2 time checked: 2026-06-15 00:44:17 +0545
  • active PID: 90501
  • checkpoint report rows: 50
  • rows by status: 49 ok, 1 error
  • rows by baseline: candidate=50
  • last completed sample: omnidocbench-00542-docstructbench_llm-raw-the-eye-o.O-TraneGuide.pdf_7
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The report was flat relative to the 00:41 checkpoint, but the parent process was still alive and within the per-page timeout window. No restart was made.

2026-06-15 00:45 checkpoint:

  • k2 time checked: 2026-06-15 00:45:40 +0545
  • active PID: 90501
  • checkpoint report rows: 52
  • rows by status: 51 ok, 1 error
  • rows by baseline: candidate=52
  • last completed sample: omnidocbench-00545-yanbaopptmerge_3aa9b6677e17fd012b0a7a230b9f18db.pdf_1327
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The flat 50-row state resolved without a restart. The run remains active and candidate-only.

2026-06-15 00:47 checkpoint:

  • k2 time checked: 2026-06-15 00:47:03 +0545
  • active PID: 90501
  • checkpoint report rows: 53
  • rows by status: 52 ok, 1 error
  • rows by baseline: candidate=53
  • last completed sample: omnidocbench-00552-yanbaopptmerge_d346b889f1d85c61950a71d4b0ac2752.pdf_89
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:50 checkpoint:

  • k2 time checked: 2026-06-15 00:50:15 +0545
  • active PID: 90501
  • checkpoint report rows: 54
  • rows by status: 53 ok, 1 error
  • rows by baseline: candidate=54
  • last completed sample: omnidocbench-00553-yanbaopptmerge_d4fc7cba428625974e93183edfccea73.pdf_89
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The temporary flat 53-row state resolved without intervention. The run remains active and candidate-only.

2026-06-15 00:51 checkpoint:

  • k2 time checked: 2026-06-15 00:51:02 +0545
  • active PID: 90501
  • checkpoint report rows: 56
  • rows by status: 55 ok, 1 error
  • rows by baseline: candidate=56
  • last completed sample: omnidocbench-00577-yanbaor2_ey-global-fintech-adoption-index.pdf_30
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:52 checkpoint:

  • k2 time checked: 2026-06-15 00:52:01 +0545
  • active PID: 90501
  • checkpoint report rows: 58
  • rows by status: 57 ok, 1 error
  • rows by baseline: candidate=58
  • last completed sample: omnidocbench-00605-jiaocai_jiaocai_en_91
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:52:52 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 59
  • rows by status: 58 ok, 1 error
  • rows by baseline: candidate=59
  • last completed sample: omnidocbench-00623-docstructbench_enbook-zlib-o.O-17761417.pdf_894
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:55 checkpoint:

  • k2 time checked: 2026-06-15 00:55:39 +0545
  • active PID: 90501
  • checkpoint report rows: 61
  • rows by status: 60 ok, 1 error
  • rows by baseline: candidate=61
  • last completed sample: omnidocbench-00627-docstructbench_llm-raw-scihub-o.O-ceat.200407001.pdf_3
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The temporary flat 59-row state resolved without intervention. The run remains active and candidate-only.

2026-06-15 00:56 checkpoint:

  • k2 time checked: 2026-06-15 00:56:27 +0545
  • active PID: 90501
  • checkpoint report rows: 63
  • rows by status: 62 ok, 1 error
  • rows by baseline: candidate=63
  • last completed sample: omnidocbench-00629-docstructbench_llm-raw-scihub-o.O-chem.200700285.pdf_4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:57 checkpoint:

  • k2 time checked: 2026-06-15 00:57:32 +0545
  • active PID: 90501
  • checkpoint report rows: 64
  • rows by status: 63 ok, 1 error
  • rows by baseline: candidate=64
  • last completed sample: omnidocbench-00630-docstructbench_llm-raw-scihub-o.O-hup.777.pdf_8
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 00:58 checkpoint:

  • k2 time checked: 2026-06-15 00:58:47 +0545
  • active PID: 90501
  • checkpoint report rows: 65
  • rows by status: 64 ok, 1 error
  • rows by baseline: candidate=65
  • last completed sample: omnidocbench-00631-docstructbench_llm-raw-scihub-o.O-j.1444-2906.2002.00466.x.pdf_4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 01:00 checkpoint:

  • k2 time checked: 2026-06-15 01:00:05 +0545
  • active PID: 90501
  • checkpoint report rows: 67
  • rows by status: 66 ok, 1 error
  • rows by baseline: candidate=67
  • last completed sample: omnidocbench-00633-docstructbench_llm-raw-scihub-o.O-s00128-007-9171-1.pdf_3
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 01:00:58 checkpoint:

  • active PID: 90501
  • checkpoint report rows: 68
  • rows by status: 67 ok, 1 error
  • rows by baseline: candidate=68
  • last completed sample: omnidocbench-00634-docstructbench_llm-raw-scihub-o.O-s00128-008-9367-z.pdf_4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only.

2026-06-15 01:03 checkpoint:

  • k2 time checked: 2026-06-15 01:03:27 +0545
  • active PID: 90501
  • checkpoint report rows: 71
  • rows by status: 70 ok, 1 error
  • rows by baseline: candidate=71
  • last completed sample: omnidocbench-00637-docstructbench_llm-raw-scihub-o.O-s003950050126.pdf_6
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only. No public Table100 candidate-vs-Surya claim can be drawn until Surya rows and final validation artifacts exist.

2026-06-15 01:04 checkpoint:

  • k2 time checked: 2026-06-15 01:04:18 +0545
  • active PID: 90501
  • checkpoint report rows: 72
  • rows by status: 71 ok, 1 error
  • rows by baseline: candidate=72
  • last completed sample: omnidocbench-00638-docstructbench_llm-raw-scihub-o.O-s00396-003-0935-9.pdf_2
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The run remains active and candidate-only. The next meaningful paper milestone is completion of the Surya baseline phase plus final validation and cell analysis artifacts.

2026-06-15 01:05 checkpoint:

  • k2 time checked: 2026-06-15 01:05:46 +0545
  • active PID: 90501
  • checkpoint report rows: 73
  • rows by status: 72 ok, 1 error
  • rows by baseline: candidate=73
  • last completed sample: omnidocbench-00639-docstructbench_llm-raw-scihub-o.O-s0167-8655%2800%2900107-0.pdf_6
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

This is still candidate-stage progress only. It is useful operational evidence that the isolated runner continues past slow pages, but it is not a benchmark result suitable for a paper claim.

2026-06-15 01:07 checkpoint:

  • k2 time checked: 2026-06-15 01:07:30 +0545
  • active PID: 90501
  • checkpoint report rows: 75
  • rows by status: 74 ok, 1 error
  • rows by baseline: candidate=75
  • last completed sample: omnidocbench-00641-docstructbench_llm-raw-scihub-o.O-s0362-546x%2897%2900070-9.pdf_7
  • error rows still only include: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 with TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

We added a guarded --status-only mode to scripts/run_public_table100_claim.py. It writes run-status.json from the current checkpoint without running OCR, summary, validation, or cell analysis, and marks paper_claim_ready=false until all expected baseline rows and final artifacts exist. This is intended for safe interim monitoring only; it must not be cited as validation evidence.

2026-06-15 status-only artifact:

  • command wrote: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json
  • checkpoint rows at write time: 79
  • expected rows for candidate plus Surya: 200
  • rows by baseline: candidate=79
  • row statuses: 78 ok, 1 error
  • next expected row: candidate / omnidocbench-00649-docstructbench_llm-raw-the-eye-o.O-Player%27s%20Options%20-%20Halflings.pdf_11
  • final artifacts: summary=false, validation=false, table_cell_analysis=false
  • paper_claim_ready=false

This status artifact is useful because it gives future paper notes a machine-readable source for incomplete-run state. It also guards against the bad habit of treating a candidate-only checkpoint as validation.

2026-06-15 01:11 checkpoint:

  • k2 time checked: 2026-06-15 01:11:57 +0545
  • active PID: 90501
  • live checkpoint report rows: 82
  • rows by status: 81 ok, 1 error
  • rows by baseline: candidate=82
  • last completed sample: omnidocbench-00654-jiaocaineedrop_2005_QP.pdf_5
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The live report had advanced beyond the previously written run-status.json (79 rows). We therefore extended --status-only so future status artifacts also include last_result, error_count, and recent_errors. The purpose is to make incomplete-run monitoring auditable without hand-parsing the benchmark report.

2026-06-15 refreshed status artifact:

  • command wrote: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json
  • checkpoint rows at write time: 85
  • expected rows for candidate plus Surya: 200
  • rows by baseline: candidate=85
  • row statuses: 84 ok, 1 error
  • last result: candidate / omnidocbench-00676-jiaocaineedrop_ced6f3c05e1dfefa8455fa60aacdd2d76ffd79489b631c2dbde22a74e927544b.pdf_84 / ok
  • recent error: candidate / omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 / TimeoutExpired: timed out after 300s
  • next expected row: candidate / omnidocbench-00677-jiaocaineedrop_chap02.pdf_16
  • paper_claim_ready=false

This remains candidate-only evidence. The useful change is that the status artifact now carries enough detail for audit logs without promoting the run to validation evidence.

2026-06-15 01:14 checkpoint:

  • k2 time checked: 2026-06-15 01:14:57 +0545
  • active PID: 90501
  • live checkpoint report rows: 87
  • rows by status: 86 ok, 1 error
  • rows by baseline: candidate=87
  • last completed sample: omnidocbench-00680-jiaocaineedrop_chap5.pdf_4
  • final summary, validation, and table-cell-analysis artifacts: not present yet

We extended --status-only again to include baseline_progress and current_phase. The status artifact now reports expected, completed, ok, error, and missing counts per baseline. This makes the transition from candidate rows to Surya rows machine-visible instead of relying on manual row counts.

The same command now writes both run-status.json and run-status.md. The Markdown file is for human-readable paper audit notes; the JSON remains the source for scripted checks.

2026-06-15 phase-aware status artifact:

  • command wrote: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json
  • checkpoint rows at write time: 90
  • expected rows for candidate plus Surya: 200
  • current phase: candidate
  • candidate progress: 90/100 completed, 89 ok, 1 error, 10 missing
  • Surya progress: 0/100 completed, 0 ok, 0 error, 100 missing
  • last result: candidate / omnidocbench-00742-jiaocaineedrop_jiaocai_needrop_en_1898 / ok
  • next expected row: candidate / omnidocbench-00853-jiaocaineedrop_jiaocai_needrop_en_854
  • paper_claim_ready=false

The candidate stage is close to completion, but this still has zero Surya rows. It remains unusable as a paired public benchmark claim.

2026-06-15 Markdown status artifact:

  • command wrote both: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json and outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.md
  • checkpoint rows at write time: 93
  • expected rows for candidate plus Surya: 200
  • current phase: candidate
  • candidate progress: 93/100 completed, 92 ok, 1 error, 7 missing
  • Surya progress: 0/100 completed, 0 ok, 0 error, 100 missing
  • paper_claim_ready=false

The Markdown status file is intentionally labeled incomplete and is suitable for paper audit notes, not for results tables.

2026-06-15 refreshed candidate-phase status:

  • command wrote both: outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.json and outputs/experiment-public-omnidocbench-table100-width-guard-v4/run-status.md
  • checkpoint rows at write time: 94
  • expected rows for candidate plus Surya: 200
  • current phase: candidate
  • candidate progress: 94/100 completed, 93 ok, 1 error, 6 missing
  • Surya progress: 0/100 completed, 0 ok, 0 error, 100 missing
  • last result: candidate / omnidocbench-01035-newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_18 / ok
  • next expected row: candidate / omnidocbench-01036-newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_2
  • paper_claim_ready=false

This is still pre-paired evidence. The run is close to Surya transition but has not yet produced any candidate-vs-Surya comparison rows.

2026-06-15 Surya phase transition:

  • k2 time checked: 2026-06-15 01:27:57 +0545
  • active PID: 90501
  • live checkpoint report rows: 101
  • rows by status: 100 ok, 1 error
  • rows by baseline: candidate=100, surya=1
  • candidate stage: complete with 99 ok, 1 error
  • Surya stage: started with 1 ok, 0 error, 99 missing
  • first Surya sample: omnidocbench-00010-page-f9583127-1277-4424-be9d-35430e0aedd6
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The refreshed status artifacts now report current_phase=surya and paper_claim_ready=false. This is the first paired-run milestone, but it is still not validation evidence because only one Surya row exists and the configured public gates require 100 candidate-vs-Surya document pairs plus 40 table pairs.

2026-06-15 Surya phase checkpoint:

  • k2 time checked: 2026-06-15 01:29:42 +0545
  • active PID: 90501
  • live checkpoint report rows: 103
  • rows by status: 102 ok, 1 error
  • rows by baseline: candidate=100, surya=3
  • refreshed status artifact rows: 103
  • current phase: surya
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 3/100 completed, 3 ok, 0 error, 97 missing
  • last result: surya / omnidocbench-00142-page-52e82aa9-7bd7-4683-8adc-520cae0291e0 / ok
  • next expected row: surya / omnidocbench-00206-page-83587fc3-fe55-4e51-a3c5-f0664edba743
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This confirms the run is now in the paired baseline stage. It still cannot support a claim until the remaining Surya rows and final validation artifacts exist.

2026-06-15 Surya slow-page checkpoint:

  • k2 time checked: 2026-06-15 01:34:45 +0545
  • active PID: 90501
  • live checkpoint report rows: 105
  • rows by status: 104 ok, 1 error
  • rows by baseline: candidate=100, surya=5
  • candidate stage: complete with 99 ok, 1 timeout error
  • Surya stage: 5 ok, 0 error, 95 missing
  • last result: surya / omnidocbench-00208-page-28c45f5f-7e0d-464a-89ec-8de3a4abb927 / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The apparent stall on Surya sample omnidocbench-00206-page-83587fc3-fe55-4e51-a3c5-f0664edba743 resolved without manual intervention. The isolated child process completed within the configured timeout policy, and the runner continued to subsequent Surya rows.

2026-06-15 paired timeout checkpoint:

  • k2 time checked: 2026-06-15 01:38:59 +0545
  • active PID: 90501
  • live checkpoint report rows: 107
  • rows by status: 105 ok, 2 error
  • rows by baseline: candidate=100, surya=7
  • Surya progress: 7/100 completed, 6 ok, 1 timeout error, 93 missing
  • last result: surya / omnidocbench-00255-page-2deb83b3-abe2-442c-a544-0b6a8db464f1 / ok
  • timeout rows:
    • candidate / omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 / TimeoutExpired: timed out after 300s
    • surya / omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 / TimeoutExpired: timed out after 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet

The page that timed out under the candidate also timed out under raw Surya and was recorded as an explicit page-level baseline error. This is useful failure evidence, but the public validation gate still cannot run until Surya completes the remaining rows.

2026-06-15 Surya 9-row checkpoint:

  • k2 time checked: 2026-06-15 01:40:37 +0545
  • refreshed status artifact rows: 109
  • rows by baseline: candidate=100, surya=9
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 9/100 completed, 8 ok, 1 error, 91 missing
  • last result: surya / omnidocbench-00259-page-232d386f-6c9d-4b50-9463-89dedce995e5 / ok
  • next expected row: surya / omnidocbench-00260-page-cb30bac7-d4bd-47d0-ac84-e0b46d051fe9
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains in the Surya baseline phase. There are not enough Surya rows for any public candidate-vs-Surya conclusion.

2026-06-15 Surya 12-row checkpoint:

  • k2 time checked: 2026-06-15 01:42:33 +0545
  • refreshed status artifact rows: 112
  • rows by baseline: candidate=100, surya=12
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 12/100 completed, 11 ok, 1 error, 88 missing
  • last result: surya / omnidocbench-00262-page-ddb9d794-5230-4b43-b1ec-273b9e7ebc11 / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Surya is progressing, but the result remains far below the configured 100 paired document rows required for the public validation gate.

2026-06-15 Surya 14-row checkpoint:

  • k2 time checked: 2026-06-15 01:44:19 +0545
  • refreshed status artifact rows: 114
  • rows by baseline: candidate=100, surya=14
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 14/100 completed, 13 ok, 1 error, 86 missing
  • last result: surya / omnidocbench-00265-page-d2cdac80-318c-4157-ac6e-bdab333f66bd / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains in the Surya baseline phase. The validation gate cannot pass or fail yet because the required paired coverage is incomplete.

2026-06-15 Surya 16-row checkpoint:

  • k2 time checked: 2026-06-15 01:46:11 +0545
  • refreshed status artifact rows: 116
  • rows by baseline: candidate=100, surya=16
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 16/100 completed, 15 ok, 1 error, 84 missing
  • last result: surya / omnidocbench-00270-page-50ffe5d2-cd27-449a-a813-d4a8c44b12ea / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This is still an in-progress baseline run. No aggregate or gate outcome should be reported from this checkpoint.

2026-06-15 Surya 18-row checkpoint:

  • k2 time checked: 2026-06-15 01:48:03 +0545
  • refreshed status artifact rows: 118
  • rows by baseline: candidate=100, surya=18
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 18/100 completed, 17 ok, 1 error, 82 missing
  • last result: surya / omnidocbench-00273-page-035cb436-c01e-41db-b40e-8977678777eb / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is still below the public validation gate requirements. Treat this as operational progress only.

2026-06-15 Surya 19-row checkpoint:

  • k2 time checked: 2026-06-15 01:49:58 +0545
  • refreshed status artifact rows: 119
  • rows by baseline: candidate=100, surya=19
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 19/100 completed, 18 ok, 1 error, 81 missing
  • last result: surya / omnidocbench-00274-page-49746735-505c-442d-b3f6-c55b1bf9e48b / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This remains an incomplete public baseline. No SOTA or public win/loss claim is supported at this point.

2026-06-15 Surya 21-row checkpoint:

  • k2 time checked: 2026-06-15 01:51:50 +0545
  • refreshed status artifact rows: 121
  • rows by baseline: candidate=100, surya=21
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 21/100 completed, 20 ok, 1 error, 79 missing
  • last result: surya / omnidocbench-00276-page-5d58cdf7-6d00-4ea5-97b2-358cbc0cab21 / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The Surya baseline is making progress but remains far from the 100 paired document rows required by the public validation config.

2026-06-15 Surya 22-row checkpoint:

  • k2 time checked: 2026-06-15 01:53:49 +0545
  • refreshed status artifact rows: 122
  • rows by baseline: candidate=100, surya=22
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 22/100 completed, 21 ok, 1 error, 78 missing
  • last result: surya / omnidocbench-00278-page-f222e5f3-8628-47cd-a0b6-40f62123228f / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains incomplete. The useful observation is that the paired baseline continues to advance under timeout guardrails.

2026-06-15 Surya quarter checkpoint:

  • k2 time checked: 2026-06-15 01:57:32 +0545
  • refreshed status artifact rows: 126
  • rows by baseline: candidate=100, surya=26
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 26/100 completed, 25 ok, 1 error, 74 missing
  • last result: surya / omnidocbench-00283-page-213e6a57-ba81-4afb-8184-2e1cf2a7664d / ok
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Surya has crossed the first quarter of the public baseline run. This is still only run-progress evidence; the public gates require the completed run.

2026-06-15 Surya 30-row checkpoint:

  • k2 time checked: 2026-06-15 02:01:18 +0545
  • refreshed status artifact rows: 130
  • rows by baseline: candidate=100, surya=30
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 30/100 completed, 29 ok, 1 error, 70 missing
  • last result: surya / omnidocbench-00287-page-b23ec9d4-e4aa-458f-8e65-c2c7f8e9dd46 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains in the Surya baseline phase. This checkpoint is useful for auditability and paper chronology, but it is still not a result.

2026-06-15 Surya 32-row checkpoint:

  • k2 time checked: 2026-06-15 02:03:26 +0545
  • refreshed status artifact rows: 132
  • rows by baseline: candidate=100, surya=32
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 32/100 completed, 31 ok, 1 error, 68 missing
  • last result: surya / omnidocbench-00290-page-38fc7c6b-80ed-46be-bb53-29986a63ad3e / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This continues to be a live-run checkpoint only. A paper can cite this as reproducibility bookkeeping, not as a public result.

2026-06-15 Surya 33-row checkpoint:

  • k2 time checked: 2026-06-15 02:05:19 +0545
  • refreshed status artifact rows: 133
  • rows by baseline: candidate=100, surya=33
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 33/100 completed, 32 ok, 1 error, 67 missing
  • last result: surya / omnidocbench-00291-page-f3e750d8-8201-4a22-9419-9776798b7a32 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The public Table100 paired run is now one third of the way through the Surya baseline. This is still operational progress, not validation evidence.

2026-06-15 Surya 34-row checkpoint:

  • k2 time checked: 2026-06-15 02:08:24 +0545
  • refreshed status artifact rows: 134
  • rows by baseline: candidate=100, surya=34
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 34/100 completed, 33 ok, 1 error, 66 missing
  • last result: surya / omnidocbench-00292-page-ab4f331b-e8f3-41fb-ba6b-8da5ade54983 / ok
  • health note: this page was observed as an active isolated Surya subprocess near the 300s timeout boundary and then completed successfully; the parent runner remained healthy
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run continues to make forward progress under per-page isolation. This is useful engineering evidence for the benchmark harness, but the public result still requires the remaining Surya rows and final validation artifacts.

2026-06-15 Surya 36-row checkpoint:

  • k2 time checked: 2026-06-15 02:09:16 +0545
  • refreshed status artifact rows: 136
  • rows by baseline: candidate=100, surya=36
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 36/100 completed, 35 ok, 1 error, 64 missing
  • last result: surya / omnidocbench-00300-yanbaopptmerge_PattPatelCh12.pdf_27 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is making steady progress through the Surya phase. The conclusion is unchanged: no public Table100 claim until all paired rows and final validation files exist.

2026-06-15 Surya 37-row checkpoint:

  • k2 time checked: 2026-06-15 02:10:52 +0545
  • refreshed status artifact rows: 137
  • rows by baseline: candidate=100, surya=37
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 37/100 completed, 36 ok, 1 error, 63 missing
  • last result: surya / omnidocbench-00433-docstructbench_llm-raw-scihub-o.O-ceat.200600266.pdf_2 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The benchmark remains healthy and incomplete. This checkpoint should be read as run provenance only, not as a paper result.

2026-06-15 Surya 40-row checkpoint:

  • k2 time checked: 2026-06-15 02:12:53 +0545
  • refreshed status artifact rows: 140
  • rows by baseline: candidate=100, surya=40
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 40/100 completed, 39 ok, 1 error, 60 missing
  • last result: surya / omnidocbench-00447-docstructbench_llm-raw-scihub-o.O-hup.908.pdf_5 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This reaches 40% of the Surya side of the public Table100 run. It improves the audit trail, but still cannot support a benchmark claim.

2026-06-15 Surya 41-row checkpoint:

  • k2 time checked: 2026-06-15 02:14:42 +0545
  • refreshed status artifact rows: 141
  • rows by baseline: candidate=100, surya=41
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 41/100 completed, 40 ok, 1 error, 59 missing
  • last result: surya / omnidocbench-00451-docstructbench_llm-raw-scihub-o.O-ijc.22994.pdf_3 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains healthy and incomplete. The next material milestone is Surya completion, not another partial-row interpretation.

2026-06-15 Surya 43-row checkpoint:

  • k2 time checked: 2026-06-15 02:16:44 +0545
  • refreshed status artifact rows: 143
  • rows by baseline: candidate=100, surya=43
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 43/100 completed, 42 ok, 1 error, 57 missing
  • last result: surya / omnidocbench-00466-docstructbench_llm-raw-scihub-o.O-j.chroma.2005.03.134.pdf_6 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

This is still a partial public run. We should not derive table, OCR, or SOTA claims from it before the Surya phase and final validators finish.

2026-06-15 Surya 46-row checkpoint:

  • k2 time checked: 2026-06-15 02:18:37 +0545
  • refreshed status artifact rows: 146
  • rows by baseline: candidate=100, surya=46
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 46/100 completed, 45 ok, 1 error, 54 missing
  • last result: surya / omnidocbench-00494-docstructbench_llm-raw-scihub-o.O-s00374-003-0589-2.pdf_6 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is approaching the halfway point of the Surya phase. This is still progress monitoring only; the acceptance gates cannot be evaluated yet.

2026-06-15 Surya 48-row checkpoint:

  • k2 time checked: 2026-06-15 02:20:34 +0545
  • refreshed status artifact rows: 148
  • rows by baseline: candidate=100, surya=48
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 48/100 completed, 47 ok, 1 error, 52 missing
  • last result: surya / omnidocbench-00496-docstructbench_llm-raw-scihub-o.O-s00374-006-0103-8.pdf_11 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The Surya phase is now just short of halfway. This remains bookkeeping, not evidence for a model-quality claim.

2026-06-15 Surya halfway checkpoint:

  • k2 time checked: 2026-06-15 02:21:47 +0545
  • refreshed status artifact rows: 150
  • rows by baseline: candidate=100, surya=50
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 50/100 completed, 49 ok, 1 error, 50 missing
  • last result: surya / omnidocbench-00542-docstructbench_llm-raw-the-eye-o.O-TraneGuide.pdf_7 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Surya has reached the halfway point of the public baseline phase. The benchmark harness is doing what we need, but this is still not a result: no validation claim is allowed until the remaining 50 Surya rows and final artifacts finish.

2026-06-15 Surya 52-row checkpoint:

  • k2 time checked: 2026-06-15 02:26:58 +0545
  • refreshed status artifact rows: 152
  • rows by baseline: candidate=100, surya=52
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 52/100 completed, 51 ok, 1 error, 48 missing
  • last result: surya / omnidocbench-00545-yanbaopptmerge_3aa9b6677e17fd012b0a7a230b9f18db.pdf_1327 / ok
  • health note: omnidocbench-00543-docstructbench_llm-raw-the-eye-o.O-chap08.pdf_12 was observed as an active isolated Surya subprocess and then resolved without adding a new timeout error
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The second half of the Surya phase has started and the runner remains healthy. No model-quality conclusion is available yet.

2026-06-15 Surya 53-row checkpoint:

  • k2 time checked: 2026-06-15 02:29:02 +0545
  • refreshed status artifact rows: 153
  • rows by baseline: candidate=100, surya=53
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 53/100 completed, 52 ok, 1 error, 47 missing
  • last result: surya / omnidocbench-00552-yanbaopptmerge_d346b889f1d85c61950a71d4b0ac2752.pdf_89 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains active and past the halfway mark. This should be treated as benchmark provenance only until the final validators run.

2026-06-15 Surya 58-row checkpoint:

  • k2 time checked: 2026-06-15 02:31:27 +0545
  • refreshed status artifact rows: 158
  • rows by baseline: candidate=100, surya=58
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 58/100 completed, 57 ok, 1 error, 42 missing
  • last result: surya / omnidocbench-00605-jiaocai_jiaocai_en_91 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The Surya baseline phase continues to advance. This checkpoint is useful for reproducibility, but it does not change the claim boundary.

2026-06-15 Surya 60-row checkpoint:

  • k2 time checked: 2026-06-15 02:33:35 +0545
  • refreshed status artifact rows: 160
  • rows by baseline: candidate=100, surya=60
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 60/100 completed, 59 ok, 1 error, 40 missing
  • last result: surya / omnidocbench-00626-docstructbench_llm-raw-scihub-o.O-bf03175144.pdf_4 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The public run is now 60% through the Surya baseline phase. The result remains uninterpretable for paper claims until the final artifacts exist.

2026-06-15 Surya 63-row checkpoint:

  • k2 time checked: 2026-06-15 02:35:53 +0545
  • refreshed status artifact rows: 163
  • rows by baseline: candidate=100, surya=63
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 63/100 completed, 62 ok, 1 error, 37 missing
  • last result: surya / omnidocbench-00629-docstructbench_llm-raw-scihub-o.O-chem.200700285.pdf_4 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run remains healthy and incomplete. The only honest interpretation remains progress toward a public benchmark, not a benchmark result.

2026-06-15 Surya 66-row checkpoint:

  • k2 time checked: 2026-06-15 02:38:03 +0545
  • refreshed status artifact rows: 166
  • rows by baseline: candidate=100, surya=66
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 66/100 completed, 65 ok, 1 error, 34 missing
  • last result: surya / omnidocbench-00632-docstructbench_llm-raw-scihub-o.O-j.eswa.2008.01.070.pdf_5 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The public run is two thirds of the way through the Surya baseline phase. We still cannot evaluate the configured gates until the run finishes.

2026-06-15 Surya 68-row checkpoint:

  • k2 time checked: 2026-06-15 02:40:17 +0545
  • refreshed status artifact rows: 168
  • rows by baseline: candidate=100, surya=68
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 68/100 completed, 67 ok, 1 error, 32 missing
  • last result: surya / omnidocbench-00634-docstructbench_llm-raw-scihub-o.O-s00128-008-9367-z.pdf_4 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The remaining Surya tail is shrinking, but the claim boundary is unchanged: final metrics and validators are mandatory.

2026-06-15 Surya 71-row checkpoint:

  • k2 time checked: 2026-06-15 02:42:40 +0545
  • refreshed status artifact rows: 171
  • rows by baseline: candidate=100, surya=71
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 71/100 completed, 70 ok, 1 error, 29 missing
  • last result: surya / omnidocbench-00637-docstructbench_llm-raw-scihub-o.O-s003950050126.pdf_6 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The public run has entered the final third of the Surya phase. No benchmark claim is available until the remaining rows and final artifacts finish.

2026-06-15 Surya three-quarter checkpoint:

  • k2 time checked: 2026-06-15 02:45:07 +0545
  • refreshed status artifact rows: 175
  • rows by baseline: candidate=100, surya=75
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 75/100 completed, 74 ok, 1 error, 25 missing
  • last result: surya / omnidocbench-00641-docstructbench_llm-raw-scihub-o.O-s0362-546x%2897%2900070-9.pdf_7 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Surya has reached three quarters of its public baseline phase. The benchmark is close enough that the next important work is finishing and validating, not interpreting partial metrics.

2026-06-15 Surya 77-row checkpoint:

  • k2 time checked: 2026-06-15 02:47:35 +0545
  • refreshed status artifact rows: 177
  • rows by baseline: candidate=100, surya=77
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 77/100 completed, 76 ok, 1 error, 23 missing
  • last result: surya / omnidocbench-00643-docstructbench_llm-raw-scihub-o.O-s10653-010-9369-8.pdf_11 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is in the final quarter of the Surya baseline. We should continue monitoring, then inspect the final validators before writing any claim.

2026-06-15 Surya 82-row checkpoint:

  • k2 time checked: 2026-06-15 02:50:07 +0545
  • refreshed status artifact rows: 182
  • rows by baseline: candidate=100, surya=82
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 82/100 completed, 81 ok, 1 error, 18 missing
  • last result: surya / omnidocbench-00654-jiaocaineedrop_2005_QP.pdf_5 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The public baseline is close to completion. The remaining work is to finish the last Surya rows and then inspect the generated validators.

2026-06-15 Surya 87-row checkpoint:

  • k2 time checked: 2026-06-15 02:52:42 +0545
  • refreshed status artifact rows: 187
  • rows by baseline: candidate=100, surya=87
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 87/100 completed, 86 ok, 1 error, 13 missing
  • last result: surya / omnidocbench-00680-jiaocaineedrop_chap5.pdf_4 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is close to the final validation stage. We still should not write any paper claim until the runner emits summary, validation, and table-cell analysis.

2026-06-15 Surya 91-row checkpoint:

  • k2 time checked: 2026-06-15 02:55:23 +0545
  • refreshed status artifact rows: 191
  • rows by baseline: candidate=100, surya=91
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 91/100 completed, 90 ok, 1 error, 9 missing
  • last result: surya / omnidocbench-00853-jiaocaineedrop_jiaocai_needrop_en_854 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is inside the last ten Surya rows. Do not stop at row completion: the final report, validation, and cell-analysis files are the evidence that matters for the paper.

2026-06-15 Surya 93-row checkpoint:

  • k2 time checked: 2026-06-15 02:57:39 +0545
  • refreshed status artifact rows: 193
  • rows by baseline: candidate=100, surya=93
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 93/100 completed, 92 ok, 1 error, 7 missing
  • last result: surya / omnidocbench-01026-newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_1 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

The run is very close to row completion. The next paper-relevant step is inspection of the final artifacts, not another partial metric read.

2026-06-15 Surya 95-row checkpoint:

  • k2 time checked: 2026-06-15 02:59:29 +0545
  • refreshed status artifact rows: 195
  • rows by baseline: candidate=100, surya=95
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 95/100 completed, 94 ok, 1 error, 5 missing
  • last result: surya / omnidocbench-01036-newspaper_2a6b4fa088699701a6fa9ccecfb5c25d_2 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Only five Surya rows remain. The next checkpoint should prioritize final artifact inspection over more progress logging.

2026-06-15 Surya 97-row checkpoint:

  • k2 time checked: 2026-06-15 03:01:37 +0545
  • refreshed status artifact rows: 197
  • rows by baseline: candidate=100, surya=97
  • candidate progress: 100/100 completed, 99 ok, 1 error, 0 missing
  • Surya progress: 97/100 completed, 96 ok, 1 error, 3 missing
  • last result: surya / omnidocbench-01117-newspaper_fe5ed29024932fad071afc53807b16ba_2 / ok
  • paired timeout so far: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s
  • final summary, validation, and table-cell-analysis artifacts: not present yet
  • paper_claim_ready=false

Only three Surya rows remain. The next useful checkpoint should be final artifact inspection, not more partial-run commentary unless the runner stalls.

Blunt Summary

The repo now supports a paper about comparative OCR/document evaluation and routed-system tradeoffs for Nepali/English documents.

It does not yet support a paper claiming the best OCR system.

2026-06-14 Hard15 Milestone

The strongest current hard15 candidate is figure_caption_refine_composite with:

  • Surya primary OCR/layout
  • Tesseract fallback only for better figure-caption lines
  • guarded structural-label correction for वार्षिक प्रतिवेदन, दैनिक हिसाब, चित्र, and निष्कर्ष
  • the small curated table-name lexicon at configs/nepali-table-name-lexicon-v1.txt

Result on outputs/experiment-figure-caption-table-lexicon-hard15-v3/benchmark/report.json:

System CER figure_caption_cer table_cell_f1 reading_order_pair_accuracy
candidate 0.010671 0.008746 1.000000 0.520000
Surya 0.236839 0.092323 0.979167 0.520000
Tesseract 1.266364 0.051020 0.000000 0.613333

Validation is still not_validated; the only failed gate is table-cell F1 bootstrap CI low 0.000000. Cell analysis says candidate gets 96/96 exact cells and Surya gets 94/96, but the table slice has only 8 pages and only 2 improving pages. That is not enough statistical evidence for a table SOTA claim.

Also record the metric-contract lesson: hard15 table-page text is only the title दैनिक हिसाब. Table contents are evaluated through structured table references. We verified this before rejecting the idea of pushing table grids into Document.text.

2026-06-14 Targeted Table Validation

We generated data/eval-pack-table-ledger-claim-v1 to resolve the hard15 table-cell CI problem without moving the gate:

  • 60 table-ledger image pages
  • 720 reference cells
  • offset 300
  • degradations: clean, scan, blur, low_contrast, skew

Validation with configs/validation-table-ledger-surya.json passed after adding a structure non-regression guard:

Metric Candidate Surya CI low Status
table_cell_f1 1.000000 0.979167 0.012500 pass
table_exact_match 1.000000 0.750000 0.150000 pass
table_structure_similarity 1.000000 1.000000 0.000000 pass, non-regression

Cell exactness:

  • candidate: 720/720
  • Surya: 705/720
  • Tesseract full-page parser: 0/720

This gives us a clean paper subsection about targeted table-ledger cell-content extraction with grid-structure non-regression. It still does not solve the broad SOTA claim. The obvious reviewer criticism is valid: generated tables are simple, the candidate is lexicon-assisted, and the table names come from a small known pool. The scoring path now accepts structured table references with rowspans/colspans, but the table-ledger pack does not exercise them. The next paper-grade step is a real scanned-table set with unseen names, merged-cell/header variety, and true TEDS/TEDS-IOU.

Follow-up added after this result: the internal eval-pack generator now has complex-table-ledger, which writes span-aware references with a merged header while keeping rendering simple. Use it to run a targeted complex-table benchmark; do not cite complex-table behavior from the simple table-ledger pack.

2026-06-14 Complex-Table Benchmark

We ran data/eval-pack-complex-table-v1 on k2:

  • 40 generated complex-table pages
  • span-aware references with a merged header
  • 600 reference cells
  • candidate/Surya/Tesseract comparison

Validation with configs/validation-complex-table-surya.json is validated, but the contract is intentionally modest:

Metric Candidate Surya CI low Status
table_cell_f1 0.928333 0.915000 0.005000 pass
table_exact_match 0.000000 0.000000 0.000000 pass, non-regression
table_structure_similarity 0.875000 0.875000 0.000000 pass, non-regression

This is a useful negative result. The candidate improves some cell text because the lexicon correction now works below grouped headers, but neither candidate nor Surya preserves the merged-header span. Do not claim complex-table SOTA from this. The next table branch should either preserve span metadata from HTML/table modules or infer spans from repeated/merged header geometry before scoring against real complex tables.

2026-06-14 Complex-Table Span-Inference Rerun

We added a narrow table postprocessor that collapses adjacent identical cells in the first header rows into a single colspan cell when a following detail row has distinct child columns. This is deliberately heuristic and should be described as parser postprocessing, not as a newly trained recognizer.

Rerun on k2:

  • benchmark: outputs/experiment-complex-table-span-infer-v1/benchmark/report.json
  • validation: outputs/experiment-complex-table-span-infer-v1/validation/validation.md
  • summary: outputs/summary-complex-table-span-infer-v1/summary.md
  • cell analysis: outputs/table-cell-analysis-complex-table-span-infer-v1/table-cell-analysis.md

Validation against Surya on data/eval-pack-complex-table-v1 is now stronger:

Metric Candidate Surya CI low Status
table_cell_f1 0.995000 0.915000 0.071667 pass
table_exact_match 0.925000 0.000000 0.825000 pass
table_structure_similarity 1.000000 0.875000 0.125000 pass

Cell analysis:

  • candidate: 597/600 exact cells, mean cell CER 0.001000
  • Surya: 549/600 exact cells, mean cell CER 0.071000
  • Tesseract full-page parser: 40/600 exact cells, mean cell CER 0.933333

This is no longer a negative result for this generated merged-header pattern. It is still not a broad complex-table SOTA claim: the pack has one repeated-header span pattern, table text is generated, and page-level text CER/reading-order metrics remain poor on this pack because the parser/reference text contract for table pages is still diagnostic rather than publication-grade.

2026-06-14 Table Tree Metrics

We added deterministic ordered table-tree edit metrics:

  • table_teds_structure: table/tr/td tree shape plus rowspan/colspan
  • table_teds_text: same tree with normalized cell text included in td labels

The stricter complex-table rerun is:

  • benchmark: outputs/experiment-complex-table-teds-v1/benchmark/report.json
  • validation: outputs/experiment-complex-table-teds-v1/validation/validation.md
  • summary: outputs/summary-complex-table-teds-v1/summary.md

Validated gates:

Metric Candidate Surya CI low Status
table_teds_structure 1.000000 0.904762 0.095238 pass
table_teds_text 0.996250 0.891667 0.098690 pass

This is a better paper metric than the earlier tokenized span score, but it is still internal. Before calling this SOTA, run a public table benchmark or a real scanned-table set using an accepted implementation or a clearly specified reproducible scorer.

2026-06-15 Public OmniDocBench Table100 Final

The public Table100 run finished on k2 under:

outputs/experiment-public-omnidocbench-table100-width-guard-v4

Final artifacts:

  • benchmark report: outputs/experiment-public-omnidocbench-table100-width-guard-v4/benchmark/report.json
  • summary: outputs/experiment-public-omnidocbench-table100-width-guard-v4/summary/summary.json
  • validation: outputs/experiment-public-omnidocbench-table100-width-guard-v4/validation/validation.json
  • table cell analysis: outputs/experiment-public-omnidocbench-table100-width-guard-v4/table-cell-analysis/table-cell-analysis.json

Run status after refreshing --status-only at k2 time 2026-06-15 03:06:36 +0545:

  • rows: 200/200
  • candidate: 100/100 completed, 99 ok, 1 error
  • Surya: 100/100 completed, 99 ok, 1 error
  • final artifacts: present
  • shared timeout: omnidocbench-00216-page-dca64e05-1ce6-49ee-857a-9b5d05a87357 timed out for both candidate and Surya at 300s

Important status-contract correction: the runner now distinguishes evidence_package_ready from paper_claim_ready. The completed Table100 benchmark package is ready to inspect, but the paper claim is not ready unless the validation report itself passes. This change is covered by tests/test_public_table100_runner.py.

Public Table100 aggregate metrics over the 99 paired non-timeout samples:

System samples CER WER table_cell_f1 table_exact_match table_structure_similarity table_teds_structure table_teds_text latency_s
candidate width-guard-v4 99 0.613201 0.666187 0.484364 0.077441 0.646758 0.720160 0.565284 57.362883
Surya 99 0.612623 0.665433 0.406694 0.037037 0.587030 0.687550 0.543987 54.583374

Candidate minus Surya:

  • CER: +0.000578 worse
  • WER: +0.000755 worse
  • table_cell_f1: +0.077670
  • table_cell_precision: +0.085841
  • table_cell_recall: +0.071341
  • table_exact_match: +0.040404
  • table_structure_similarity: +0.059727
  • table_teds_structure: +0.032609
  • table_teds_text: +0.021297
  • latency: +2.779509s slower

Validation status: not_validated.

Gate details:

Gate Candidate Surya Delta CI low pairs Status
CER, lower is better 0.613201 0.612623 -0.000578 -0.002648 99 fail
table_cell_f1 0.484364 0.406694 +0.077670 +0.031407 99 pass
table_teds_structure 0.720160 0.687550 +0.032609 +0.012948 99 pass
table_teds_text 0.565284 0.543987 +0.021297 +0.003130 99 pass

The failed validation requirement is real and should not be softened:

  • paired samples are 99, below the configured 100, because the same page timed out for both systems
  • candidate OCR CER is slightly worse than Surya
  • the CER bootstrap confidence interval crosses zero

Table-cell analysis is still useful for the paper:

  • candidate exact cells: 6732/15992
  • Surya exact cells: 6508/15992
  • candidate mean cell CER: 0.618767
  • Surya mean cell CER: 0.649101
  • candidate-better cells: 445
  • Surya-better cells: 208
  • ties: 15339
  • candidate exact / Surya not: 363
  • Surya exact / candidate not: 139

Conclusion for the paper: this is evidence for a table-structure and table-cell extraction improvement over Surya on public OmniDocBench Table100, not evidence for a broad OCR SOTA claim. The OCR recognizer path still needs a real gain over Surya/stock recognizers, especially for Devanagari plus English, before the headline claim is defensible.

2026-06-15 Recognizer Training Direction Reset

We should be explicit about the model state:

  • The public Table100 candidate is not a newly trained recognizer. It is mostly Surya plus table/caption postprocessing.
  • We did train HF/TrOCR recognizers earlier, but those were negative results: the full-page table/figure dev run and crop-level run both lost badly to Surya/Tesseract.
  • Therefore the engine architecture remains worth pushing, but the recognizer core is still missing.

To keep the next fine-tune honest, we added and verified:

ocrtech inspect-hf-dataset <dataset-id> ...

This records actual Hugging Face dataset structure before conversion. Live k2 inspection artifacts:

  • outputs/hf-inspect-training-v1/gauravgiri__nepali-ocr-dataset-train-inspect.json
  • outputs/hf-inspect-training-v1/prashant0919__nepali-synthetic-ocr-lines-train-inspect.json
  • outputs/hf-inspect-training-v1/nvidia__OCR-Synthetic-Multilingual-v1__en-test-inspect.json

All three passed on 2026-06-15:

dataset verified schema
gauravgiri/nepali-ocr-dataset image + text via datasets.load_dataset
prashant0919/nepali-synthetic-ocr-lines repo metadata.jsonl with file_name + text
nvidia/OCR-Synthetic-Multilingual-v1 HDF5 images + annotations.line_bboxes[].bbox/text

Next credible branch:

  1. Build a larger line/crop recognizer corpus from these verified sources.
  2. Prefer a recognizer architecture suited to line OCR, not full-page TrOCR.
  3. If using PaddleOCR, train on a supported Paddle training host rather than pretending k2/macOS arm64 is official Paddle training infrastructure.
  4. Package the trained recognizer into the existing document parser and rerun the public/held-out gates.

2026-06-15 Mixed Recognizer Recipe v2

We hardened the PaddleOCR recognizer recipe writer before launching any serious training. Two bugs/risks were fixed:

  • ocrtech train-recognizer --backend paddleocr now actually propagates --train-batch-size, --eval-batch-size, --learning-rate, and --epochs into paddle_recognizer.yml.
  • Paddle label serialization now collapses tabs/newlines inside labels so labels/train.txt remains one physical line per manifest entry.

The second fix matters: before it, the mixed-script training manifest had 1700 entries but the Paddle label file expanded to 1760 physical lines because some labels contained embedded newlines. That would corrupt PaddleOCR training.

Regenerated k2 artifact:

runs/paddle-recognizer-mixed-v2/

Recipe inputs:

  • train manifest: data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl
  • eval manifest: data/splits-mixed-v1/eval-with-refs.jsonl
  • base model: PaddlePaddle/devanagari_PP-OCRv3_mobile_rec
  • batch sizes: train 16, eval 16
  • learning rate: 0.0005
  • epochs: 10

Audited counts after regeneration:

  • train manifest entries: 1700
  • train label lines: 1700
  • eval manifest entries: 180
  • eval label lines: 180
  • dictionary characters: 178
  • train/eval leakage check: pass

Train slices:

  • nepali=850
  • english=850
  • real=428
  • synthetic=1272
  • line_crop=850

Eval slices:

  • nepali=30
  • english=150
  • real=15
  • synthetic=165
  • line_crop=150

This is now a usable PaddleOCR line-recognition recipe artifact. Do not call it a trained model. Also do not describe k2/macOS arm64 as an official PaddleOCR training host; use k2 for corpus/recipe generation and HF-compatible runs, or move this recipe to a supported Paddle training machine.

Portable bundle for supported Paddle training host:

runs/paddle-recognizer-mixed-v2-bundle/paddle-recognizer-mixed-v2-training.tar.gz

The bundle materializes all referenced images under bundle_assets/images and rewrites manifests plus Paddle label files to those portable paths. This matters because some source images live in the Hugging Face cache outside ~/ocr-tech.

Bundle audit on k2:

  • files in archive: 1887
  • materialized images: 1880
  • uncompressed bytes: 79133479
  • compressed bytes: 74061239
  • required recipe/config/manifests present in tar: yes

Training handoff command after extraction:

python PaddleOCR/tools/train.py -c runs/paddle-recognizer-mixed-v2/paddle_recognizer.yml

2026-06-15 HF Mixed Recognizer v2 Full Run Started

Because k2 is not an official PaddleOCR training host, we started the HF-compatible recognizer path on k2 while the Paddle bundle remains the handoff artifact for a supported Paddle machine.

Active k2 run:

runs/hf-recognizer-mixed-v2-full

Launch command:

uv run ocrtech train-recognizer \
  --backend hf-vision-encoder-decoder \
  --base-model microsoft/trocr-small-printed \
  --train-manifest data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl \
  --eval-manifest data/splits-mixed-v1/eval-with-refs.jsonl \
  --out runs/hf-recognizer-mixed-v2-full \
  --train-batch-size 4 \
  --eval-batch-size 4 \
  --learning-rate 0.00005 \
  --epochs 3 \
  --run

Status checkpoint:

  • k2 time: 2026-06-15 03:32:05 +0545
  • PID: 24642
  • process: running
  • status artifact: runs/hf-recognizer-mixed-v2-full/train-status.json
  • log: runs/hf-recognizer-mixed-v2-full/train.log
  • expected trainer steps: 1275
  • early log shows active training on MPS with the usual TrOCR warnings about newly initialized encoder pooler weights and tokenizer/generation token alignment

Follow-up checkpoint:

  • k2 time: 2026-06-15 03:35:55 +0545
  • PID: 24642
  • process: still running
  • latest observed log progress: 474/1275 trainer steps
  • training_metrics.json: not present
  • final-model/: not present

This is a real fine-tune run, but it is not evidence of a better recognizer until it finishes, is packaged, and beats Tesseract/Surya/stock recognizer baselines on the held-out eval. Previous TrOCR recognizer attempts were negative, so treat this as a measured experiment, not a likely win.

2026-06-15 AWS G6 PaddleOCR Recognizer Training Result

We used the playaround AWS account to launch a supported CUDA training host:

  • instance: i-044a50b7737e8a067
  • type: g6.xlarge
  • GPU: NVIDIA L4, 23034 MiB
  • AMI: Deep Learning Base OSS Nvidia Driver GPU AMI, Ubuntu 22.04
  • region: us-east-1
  • local artifact bundle: outputs/aws-g6-paddle-mixed-v2-cuda/ocrtech-paddle-mixed-v2-cuda-artifacts.tar.gz

The original generated PaddleOCR recipe was not runnable because it lacked the real PaddleOCR training sections (Architecture, Loss, PostProcess, Metric). We fixed the local recipe writer to emit a PP-OCRv3 mobile Devanagari recognizer config matching the upstream PaddleOCR training shape.

The CUDA run completed:

  • config: outputs/aws-g6-paddle-mixed-v2-cuda/runs/paddle-recognizer-mixed-v2-cuda/paddle_recognizer.yml
  • train log: outputs/aws-g6-paddle-mixed-v2-cuda/runs/paddle-recognizer-mixed-v2-cuda/train.log
  • best training metric: acc=0.1888888783950623
  • best normalized edit distance: 0.6377056438836495
  • best epoch: 10

The model exported, packaged, and audited:

  • model card: outputs/aws-g6-paddle-mixed-v2-cuda/runs/paddle-recognizer-mixed-v2-cuda/package/model-card.json
  • audit: outputs/aws-g6-paddle-mixed-v2-cuda/runs/paddle-recognizer-mixed-v2-cuda/audit-v4/model-audit.json
  • audit status: pass

Packaging fixes found during runtime validation:

  • PaddleOCR 3.x accepts text_recognition_model_dir, not the old rec_model_dir when both are present.
  • lang="devanagari" is invalid for the installed package; the usable language key is ne, though it is ignored once explicit model directories are set.
  • the runtime must be given text_recognition_model_name="devanagari_PP-OCRv3_mobile_rec" or it defaults to PP-OCRv6_medium_rec and rejects our exported config.
  • rec_char_dict_path is not accepted by the PaddleOCR 3.x constructor; the exported inference.yml carries the character dictionary.
  • paddlex[ocr-core]==3.7.1, aiohttp, and Tesseract eng+nep language data were required for the benchmark host.

Held-out benchmark result after materializing references on AWS:

  • report: outputs/aws-g6-paddle-mixed-v2-cuda/outputs/experiment-paddle-mixed-v2-cuda-vs-tesseract-v2/rescored-with-refs/report.json
  • summary: outputs/aws-g6-paddle-mixed-v2-cuda/outputs/experiment-paddle-mixed-v2-cuda-vs-tesseract-v2/rescored-with-refs-summary/summary.md
  • rows: 360, all ok (180 candidate, 180 Tesseract)
  • candidate overall CER: 0.690203
  • Tesseract overall CER: 0.364915
  • candidate overall WER: 0.944234
  • Tesseract overall WER: 0.521763
  • candidate mean latency: 0.077315s
  • Tesseract mean latency: 0.177510s

Slice detail:

  • English CER: candidate 0.630908, Tesseract 0.288197
  • Nepali CER: candidate 0.986676, Tesseract 0.748507
  • real Nepali slice CER: candidate 0.973353, Tesseract 1.175661

Interpretation: this is a real trained PaddleOCR recognizer, but it is a negative result for the main claim. It is faster than Tesseract and wins one narrow real-Nepali slice, but it loses the full held-out mixed eval and loses English badly. Do not cite this as an engine win. Cite it as evidence that the training/runtime path is now operational and that the current dataset/recipe is not sufficient.

One benchmark hygiene issue remains: PaddleOCR logged unsupported-extension warnings for .img files in the synthetic subset. The rows still completed, but a claim-grade rerun should normalize those images to supported suffixes before comparing systems.

Follow-up implementation: added ocrtech normalize-manifest-images to make that hygiene step reproducible. The command copies supported image files, converts unsupported suffixes to PNG, updates image hashes, and preserves source-image provenance in manifest metadata. Use it before the next PaddleOCR/Tesseract claim run, then materialize references from the rewritten manifest:

ocrtech normalize-manifest-images \
  data/splits-mixed-v1/eval-with-refs.jsonl \
  --out data/splits-mixed-v1/eval-normalized-images

ocrtech materialize-references \
  data/splits-mixed-v1/eval-normalized-images/manifest-images-normalized.jsonl \
  --out data/references-mixed-v1-normalized \
  --rewritten-manifest data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl

2026-06-15 k2 Normalized Mixed Eval Preflight

We synced normalize-manifest-images and the preflight suffix guard to k2 and ran them against the canonical mixed eval manifest. The canonical k2 manifest already used supported .png paths, unlike the AWS portable bundle that emitted .img warnings.

Artifacts copied locally:

  • bundle: outputs/k2-mixed-v1-normalized-preflight/k2-mixed-v1-normalized-preflight.tar.gz
  • image normalization report: outputs/k2-mixed-v1-normalized-preflight/data/splits-mixed-v1/eval-normalized-images/image-normalization.json
  • rewritten manifest with references: outputs/k2-mixed-v1-normalized-preflight/data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl
  • manifest audit: outputs/k2-mixed-v1-normalized-preflight/outputs/manifest-audit-mixed-v1-eval-normalized/manifest-audit.json
  • preflight: outputs/k2-mixed-v1-normalized-preflight/outputs/preflight-mixed-v1-normalized-script-router-path/preflight.json

Normalization result:

  • samples: 180
  • copied: 180
  • converted: 0
  • unsupported suffixes: none
  • manifest audit: pass, with required nepali and english slices present

Preflight command used the existing script-router candidate and component baselines:

PATH=/opt/homebrew/bin:$PATH ocrtech preflight-claim \
  --eval-manifest data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl \
  --out outputs/preflight-mixed-v1-normalized-script-router-path \
  --baselines candidate,tesseract,surya \
  --candidate-model-config outputs/calibration-surya-script-router/model-card.json \
  --train-manifest data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl \
  --allow-smoke-eval-pack

Preflight result:

  • overall: pass
  • eval-input-suffixes: pass
  • references: pass for all 180 samples
  • candidate model card: pass
  • component baselines present: pass
  • calibration separation: pass
  • leakage: warning only, due duplicate text hashes across train/eval
  • eval-pack: warning, because this mixed line-crop pack is not a full document-claim pack with tables, figures, forms, scans, skew, blur, and reading-order slices

This is useful evidence for OCR-line validation hygiene. It is not a document-understanding SOTA preflight because the pack lacks the document slices required by the objective.

2026-06-15 k2 Normalized Mixed Eval Benchmark

After adding manifest-driven benchmark inputs, we ran the cleaned mixed eval checkpoint directly from the normalized manifest:

PATH=/opt/homebrew/bin:$PATH ocrtech benchmark \
  --inputs-from-manifest data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl \
  --out outputs/live-k2-mixed-v1-normalized-script-router-benchmark \
  --baselines candidate,tesseract,surya \
  --candidate-model-config outputs/calibration-surya-script-router/model-card.json \
  --resume-existing \
  --sample-timeout-seconds 180

Local artifact bundle:

outputs/k2-mixed-v1-normalized-script-router-benchmark/k2-mixed-v1-normalized-script-router-benchmark.tgz

Checkpoint status:

  • report rows: 428
  • candidate rows: 180/180, all ok
  • Tesseract rows: 180/180, all ok
  • Surya rows: 68/180, all ok before the run stopped making progress
  • the Python signal timeout did not turn the blocked Surya call into an error, so the process was stopped manually and the checkpoint was preserved

Complete candidate-vs-Tesseract summary:

  • summary: outputs/k2-mixed-v1-normalized-script-router-benchmark/outputs/summary-mixed-v1-normalized-script-router-vs-tesseract/summary.json
  • candidate overall CER/WER: 0.354222 / 0.493555
  • Tesseract overall CER/WER: 0.329012 / 0.486485
  • candidate English CER/WER: 0.272420 / 0.409552
  • Tesseract English CER/WER: 0.243379 / 0.370552
  • candidate Nepali CER/WER: 0.763234 / 0.913571
  • Tesseract Nepali CER/WER: 0.757175 / 1.066151

Interpretation: the cleaned line-style eval does not rescue the script-router candidate. It loses to Tesseract on aggregate CER/WER and English CER/WER, and only shows a Nepali WER improvement. This is a useful negative result for the paper because it confirms that component routing is not the engine path. The Surya leg is diagnostic only until rerun under a hard backend watchdog.

2026-06-15 k2 Normalized Mixed Eval Hard-Timeout Rerun

After the Surya leg stalled at 68/180, the benchmark runner was changed so --sample-timeout-seconds uses a process-backed watchdog on fork platforms. That turns backend stalls into row-level TimeoutError entries and lets the benchmark finish.

Rerun command:

PATH=/opt/homebrew/bin:$PATH ocrtech benchmark \
  --inputs-from-manifest data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl \
  --out outputs/live-k2-mixed-v1-normalized-script-router-benchmark-hard-timeout \
  --baselines candidate,tesseract,surya \
  --candidate-model-config outputs/calibration-surya-script-router/model-card.json \
  --resume-existing \
  --sample-timeout-seconds 45

Local artifact bundle:

outputs/k2-mixed-v1-normalized-script-router-hard-timeout/k2-mixed-v1-normalized-script-router-hard-timeout.tgz

Run status:

  • rows: 540
  • candidate: 180/180, all ok
  • Tesseract: 180/180, all ok
  • Surya: 180/180 attempted, 178 ok, 2 timed out at 45s

Summary:

  • summary: outputs/k2-mixed-v1-normalized-script-router-hard-timeout/outputs/summary-mixed-v1-normalized-script-router-hard-timeout/summary.json
  • candidate overall CER/WER: 0.354840 / 0.469944
  • Tesseract overall CER/WER: 0.329012 / 0.486485
  • Surya scored-row CER/WER: 3.157317 / 2.428888
  • candidate vs Tesseract paired CER: 17 wins, 4 losses, 159 ties, but worse aggregate mean
  • candidate vs Tesseract paired WER: 16 wins, 4 losses, 160 ties, and slightly better aggregate mean

Interpretation: the benchmark infrastructure is now healthier, but the engine story is still negative. The script-router candidate is not a stronger OCR engine than Tesseract on this cleaned line-style eval because CER regresses, and the Surya line-mode baseline is not competitive on this pack.

2026-06-15 PaddleOCR Recognizer Failure Slice

We converted the negative AWS G6 PaddleOCR recognizer result into a concrete next-eval target instead of another broad retraining guess.

Commands:

ocrtech benchmark-error-analysis \
  --benchmark-report outputs/aws-g6-paddle-mixed-v2-cuda/outputs/experiment-paddle-mixed-v2-cuda-vs-tesseract-v2/rescored-with-refs/report.json \
  --out outputs/paddle-mixed-v2-cuda-error-analysis \
  --system candidate \
  --baseline tesseract \
  --metric cer \
  --top-n 50

ocrtech export-failure-manifest \
  --source-manifest data/splits-mixed-v1/eval-normalized-images-with-refs.jsonl \
  --top-failures outputs/paddle-mixed-v2-cuda-error-analysis/top-failures.json \
  --out data/splits-mixed-v1/eval-paddle-mixed-v2-cuda-hard50-normalized.jsonl \
  --split paddle_hard_eval

The source manifest in the second command is the normalized k2 eval manifest, not the AWS bundle manifest. That matters because the AWS manifest uses bundle_assets/... paths, while the k2-normalized manifest resolves to runnable PNG paths with materialized references.

Artifacts:

  • error analysis: outputs/paddle-mixed-v2-cuda-error-analysis/error-analysis.json
  • top failures: outputs/paddle-mixed-v2-cuda-error-analysis/top-failures.json
  • k2 hard-50 bundle: outputs/k2-paddle-mixed-v2-hard50-normalized/k2-paddle-mixed-v2-hard50-normalized.tgz
  • k2 audit: outputs/k2-paddle-mixed-v2-hard50-normalized/outputs/manifest-audit-paddle-mixed-v2-cuda-hard50-normalized/manifest-audit.json

Failure profile:

  • top failure rows: 50
  • synthetic: 50
  • english/en/line_crop: 38
  • nepali: 12
  • k2 audit: pass, with required paddle_hard_eval, english, and nepali slices present

Interpretation: do not train on this hard-50 manifest. It is a held-out diagnostic gate for the next recognizer. The failed PaddleOCR branch had already seen a balanced 850 English / 850 Nepali training mix on k2, so the failure is not just class-count imbalance. The next recognizer branch must explicitly preserve English line accuracy while improving Nepali, likely by changing the base recognizer or using a script-aware local recognizer stack rather than continuing the same Devanagari PP-OCRv3 recipe unchanged.

Follow-up validation config:

configs/validation-paddle-hard50-tesseract.json

Negative-control validation against the failed AWS G6 recognizer:

ocrtech validate-claim \
  --benchmark-report outputs/aws-g6-paddle-mixed-v2-cuda/outputs/experiment-paddle-mixed-v2-cuda-vs-tesseract-v2/rescored-with-refs/report.json \
  --config configs/validation-paddle-hard50-tesseract.json \
  --eval-manifest data/splits-mixed-v1/eval-paddle-mixed-v2-cuda-hard50.jsonl \
  --out outputs/validation-paddle-mixed-v2-cuda-hard50-tesseract-negative

Result:

  • status: not_validated
  • hard-50 CER: candidate 0.887557, Tesseract 0.080249
  • hard English line-crop CER: candidate 0.852048, Tesseract 0.031244
  • hard English line-crop WER: candidate 0.995614, Tesseract 0.114464
  • hard Nepali WER: candidate 1.000000, Tesseract 0.716369

This is the next recognizer admission gate, not the final document claim gate. Any new local recognizer that still fails this hard slice is not worth promoting to full document-pack or public-benchmark evaluation.

2026-06-15 Line-Level PaddleOCR Routing Admission Test

We implemented a line-level script-select composite to test the engine shape directly:

  • primary: Tesseract nep+eng
  • secondary: AWS G6 trained PaddleOCR recognizer
  • routing: only replace primary lines whose detected text is Devanagari-heavy
  • model card: outputs/line-script-paddle-mixed-v2-cuda/model-card.json

The first k2 hard-50 run crashed all candidate rows with:

BenchmarkError: timeout worker exited without a result while running candidate ...; exit_code=-11

This was a benchmark runtime bug. The hard timeout path created a reusable PaddleOCR engine in the parent process and then forked per sample. PaddleOCR / PaddlePaddle runtime state is not safe to inherit that way. We changed run_benchmark so timeout mode creates the OCR engine inside the watchdog child process. The no-timeout path still reuses engines for speed.

Regression coverage:

uv run pytest \
  tests/test_training_benchmark.py::test_benchmark_timeout_builds_engine_inside_watchdog_worker \
  tests/test_training_benchmark.py::test_benchmark_sample_timeout_uses_process_watchdog_when_available \
  tests/test_models_candidate.py::test_candidate_engine_supports_script_select_composite_line_routing \
  tests/test_models_candidate.py::test_nested_model_kwargs_resolve_relative_paths_from_model_card \
  -q

k2 rerun:

ocrtech benchmark \
  --inputs-from-manifest data/splits-mixed-v1/eval-paddle-mixed-v2-cuda-hard50-normalized.jsonl \
  --out outputs/experiment-line-script-paddle-hard50-v2 \
  --baselines candidate,tesseract \
  --candidate-model-config outputs/line-script-paddle-mixed-v2-cuda/model-card.json \
  --resume-existing \
  --sample-timeout-seconds 120

Artifacts:

  • outputs/experiment-line-script-paddle-hard50-v2/report.json
  • outputs/summary-line-script-paddle-hard50-v2/summary.md
  • outputs/validation-line-script-paddle-hard50-v2/validation.md

Validation status: not_validated.

slice metric candidate Tesseract result
all hard-50 CER 0.276489 0.120106 fail
all hard-50 WER 0.340517 0.271302 fail
English line-crop CER 0.060429 0.060429 tie/pass
English line-crop WER 0.132259 0.132259 tie/pass
Nepali hard CER 0.960679 0.309083 fail
Nepali hard WER 1.000000 0.711607 fail

Interpretation:

  • The engine shape is useful: English preservation works because English lines stay on the Tesseract primary path.
  • The available trained PaddleOCR recognizer is not an admissible Devanagari override. It makes the Nepali hard slice much worse.
  • The next AWS/CUDA work should be a better recognizer training run or a different Devanagari recognizer base, not more routing around this model.
  • Do not spend GPU on this exact composite again unless the secondary recognizer first passes the hard-50 admission gate.

2026-06-15 AWS G6 PaddleOCR V3 Training Launch

We created a fresh g6.xlarge through the playaround AWS profile:

  • instance: i-0b8604ffbcbce25c7
  • public IP at launch: 44.202.252.0
  • GPU: NVIDIA L4, 23034 MiB
  • AMI: Deep Learning Base OSS Nvidia Driver GPU AMI Ubuntu 22.04, ami-01011b868ec560823

The host passed paddle.utils.run_check() with paddlepaddle-gpu==3.3.1.

Before training could start, the G6 exposed two recipe/runtime bugs:

  1. Integral epoch counts were serialized as floats (20.0). PaddleOCR's cosine scheduler passes that value to CosineAnnealingDecay(T_max=...), which requires an integer. We patched the recipe generator to emit 20 for integral epoch values and added a regression test.
  2. The default PaddleOCR pretrained_model value used the Hugging Face model id PaddlePaddle/devanagari_PP-OCRv3_mobile_rec. PaddleOCR training expects a local path or URL resolving to .pdparams. We changed the default to the official Paddle pretrained URL: https://paddle-model-ecology.bj.bcebos.com/paddlex/official_pretrained_model/devanagari_PP-OCRv3_mobile_rec_pretrained.pdparams.

V3 run:

  • run dir: /opt/ocrtech/runs/paddle-recognizer-mixed-v3-cuda
  • train/eval: 1700 / 180, same verified split as v2
  • LR: 0.0001
  • epochs: 20
  • process: PID 4071

Initial status:

  • dataloaders built (106 train iters, 12 eval iters)
  • pretrained weights downloaded and loaded
  • expanded-dictionary head weights were shape-mismatched and skipped, as expected
  • first logged step: epoch 1/20, global step 10, loss 137.718468, ETA about 44 minutes

Final status:

  • training completed all 20 epochs
  • best checkpoint: epoch 17
  • best reported eval accuracy: 0.09444443919753115
  • best norm edit distance: 0.5833825886055732
  • export succeeded from checkpoints/best_accuracy.pdparams
  • package audit passed for ocrtech-paddle-mixed-v3-cuda

We had to materialize references on the G6 before scoring. The first hard-50 benchmark run had OCR outputs but no CER/WER because reference files were not available to load_reference. We copied the 50 normalized hard-50 PNGs from k2, materialized reference files, and rescored the existing outputs.

Hard-50 validation result: not_validated.

slice metric candidate Tesseract result
all hard-50 CER 0.921416 0.080249 fail
all hard-50 WER 1.010476 0.258921 fail
English line-crop CER 0.917511 0.031244 fail
English line-crop WER 0.996241 0.114464 fail
Nepali hard CER 0.933782 0.235433 fail
Nepali hard WER 1.055556 0.716369 fail

Artifacts:

  • outputs/aws-g6-paddle-mixed-v3-cuda/ocrtech-paddle-mixed-v3-cuda-artifacts.tgz
  • outputs/aws-g6-paddle-mixed-v3-cuda/outputs/validation-paddle-mixed-v3-cuda-hard50-rescored/validation.md
  • outputs/aws-g6-paddle-mixed-v3-cuda/runs/paddle-recognizer-mixed-v3-cuda/train.log

Conclusion: the second PaddleOCR CUDA branch fixed real infrastructure issues, but it does not fix the engine. This PP-OCRv3 Devanagari recipe family should not receive more GPU time without a new model hypothesis.

2026-06-15 SOTA Claim-Review Metric Coverage Gate

After the failed PaddleOCR v3 branch, we tightened the portfolio-level SOTA review instead of pretending the current engine is publishable.

Change:

  • review-claim now supports required_metrics per required experiment slot.
  • Each required metric must appear as a finite value on at least one successful benchmark row for every required baseline in that slot.
  • configs/claim-review-sota-v1.json now requires internal document evidence for OCR (cer, wer), tables (table_cell_f1), reading order (reading_order_pair_accuracy), and figures/captions (figure_detection_f1, figure_caption_cer).
  • Public slots now declare their own required metric coverage.
  • review-claim now rejects stale experiment evidence with a domain ClaimReviewError when a referenced manifest or benchmark report path is missing, instead of leaking raw filesystem errors or guessing a replacement.

This matters for the paper because a Devanagari+English document-understanding claim cannot be supported by a single CER table. The claim package must prove that the same candidate model was evaluated across the required document capabilities and public/held-out suites, with intact provenance for manifests, references, model cards, benchmark reports, and validation reports.

Follow-up: candidate model provenance is now hash-checked even when review-claim is run without an explicit --model-config. The reviewer loads the single candidate model path recorded by the experiments, recomputes the model-card SHA-256, and compares it with every recorded candidate_model_config_sha256 before audit. This matches the fallback-model behavior and prevents stale experiment reports from silently reviewing a different candidate card than the one benchmarked.

Follow-up: required experiment error counts are now derived from benchmark/report.json rows during review instead of trusted from experiment.json benchmark_status_counts. This prevents a hand-edited or stale experiment summary from hiding failed benchmark rows in a claim package.

Follow-up: required experiment eval manifests are now hash-bound during portfolio review. review-claim compares the current eval manifest SHA-256 against experiment.json environment.eval_manifest_sha256 and fails the required slot on mismatch. This prevents a claim package from silently swapping the dataset behind an existing experiment report path.

Follow-up: successful benchmark rows are now bound back to the reviewed eval manifest. For rows with a sample_id, review-claim verifies that the sample exists in the eval manifest and that the row input_path matches the manifest entry. This prevents a benchmark report from satisfying required metrics with rows from outside the declared evaluation set.

Follow-up: row slice labels are now bound to the eval manifest too. Review fails if a successful benchmark row reports slices that differ from the manifest entry for its sample_id. This prevents a report from relabeling an easier sample as table, figure, or another claim-critical slice.

Verification:

  • uv run pytest tests/test_claim_review.py -q

Candidate fallback component provenance gate

Follow-up hardening for the selective-fallback claim: candidate model cards can now record provenance.fallback_model_card through create-model-card --fallback-model-card or write_model_card(..., fallback_model_card=...). review-claim can require this with require_candidate_fallback_provenance: true, and the SOTA profile enables it.

This closes a reproducibility gap in the hybrid-engine story. Experiment reports already record fallback_model_config and its SHA-256, but that only proves what was supplied at benchmark time. The candidate card itself now has to declare the fallback component it was designed to use. Review hashes that declared fallback card, compares it with every reviewed experiment fallback hash, and audits the fallback card before a SOTA claim can pass.

Paper implication: the engine definition is now an auditable composite artifact, not a loose primary-card plus runtime flag. A reported result can be traced to both the local primary component and the selective fallback component.

Verification:

  • uv run pytest tests/test_models_candidate.py tests/test_claim_review.py tests/test_claim_review_sota_config.py tests/test_cli.py -q

Paired win-rate validation gates

Follow-up hardening for SOTA validation: ClaimGate now supports min_pair_win_rate, defined as wins divided by all paired samples for that gate. This is intentionally stricter than the existing reported win_rate, which is wins divided by non-tied outcomes. The distinction closes an important paper-review loophole: a candidate could have one win, many ties, no losses, and a reported non-tie win rate of 1.0, while only improving a tiny fraction of documents.

The SOTA validation configs now require min_pair_win_rate >= 0.55 on gates that claim positive improvement. No-regression/tie-acceptable gates, such as English preservation checks with zero required margin, remain exempt. This keeps the claim standard aligned with the objective: broad Devanagari+English document-engine improvement, not a narrow average shift.

Verification:

  • uv run pytest tests/test_references_validation.py tests/test_claim_review_sota_config.py -q

Context-aware validation preflight

Follow-up hardening for remote/public runs: preflight-claim now checks an explicit --validation-config against the actual requested baselines and eval manifest before OCR starts. It fails when:

  • the validation candidate label is not present in --baselines
  • a gate baseline is not scheduled for the run
  • min_samples exceeds the eval manifest size
  • required slices are absent or undersized
  • a gate's min_pairs exceeds the number of rows matching the gate slice intersection

If --validation-config is omitted, preflight now warns instead of pretending the default validation gates are claim-ready for the run. This matters for the paper workflow because a four-hour GPU or public benchmark window should not be spent producing evidence that could never satisfy the configured claim gates.

Verification:

  • uv run pytest tests/test_preflight.py::test_preflight_accepts_context_compatible_validation_config tests/test_preflight.py::test_preflight_rejects_validation_config_that_cannot_match_run_context -q

Preflight fallback provenance consistency

Follow-up hardening for hybrid engine runs: preflight-claim now checks that a configured --fallback-model-config matches the candidate model card's provenance.fallback_model_card. The check resolves the candidate-declared fallback path, verifies any declared SHA-256, hashes the configured fallback card, and fails if the two cards differ.

This is the pre-run version of the portfolio claim-review fallback provenance gate. It prevents a benchmark from accidentally evaluating one fallback component while the candidate model card describes another composite engine. That matters for reproducibility: the paper artifact must define the whole engine, including the selective fallback component, before metrics are run.

Verification:

  • uv run pytest tests/test_preflight.py::test_preflight_audits_fallback_model_config tests/test_preflight.py::test_preflight_requires_candidate_fallback_provenance_for_configured_fallback tests/test_preflight.py::test_preflight_rejects_candidate_fallback_provenance_mismatch -q

PaddleOCR dictionary coverage gate

Follow-up hardening before another recognizer run: train-recognizer now validates custom PaddleOCR dictionaries against the exact normalized train/eval labels that will be written into labels/train.txt and labels/eval.txt. Recipe generation fails if any label contains a character missing from the dictionary.

This prevents a wasted GPU job where the recognizer head cannot emit characters that appear in the supervised labels. It also makes dictionary coverage a launch-time invariant rather than something inferred after a failed training run. Auto-generated dictionaries still use the train/eval labels, but custom dictionaries must now prove coverage before handoff.

Verification:

  • uv run pytest tests/test_training_benchmark.py::test_train_recognizer_accepts_custom_dictionary_covering_labels tests/test_training_benchmark.py::test_train_recognizer_rejects_custom_dictionary_missing_label_characters -q

HF recognizer launch provenance summary

Follow-up for the non-Paddle recognizer path: train-recognizer --backend hf-vision-encoder-decoder now writes recognizer-training-summary.json beside hf_recognizer_config.json, matching the PaddleOCR recipe provenance pattern. The summary records:

  • train/eval manifest paths and SHA-256 hashes
  • sample, dataset, slice, and script coverage
  • generated config hash
  • base model and hyperparameters
  • aggregate character coverage, including ASCII and Devanagari presence

This does not change the interpretation of previous TrOCR runs; those remain negative results. It makes future HF-compatible recognizer attempts auditable under the same trained-recognizer provenance gate as PaddleOCR packages.

Verification:

  • uv run pytest tests/test_training_benchmark.py::test_train_hf_recognizer_writes_recipe tests/test_training_benchmark.py::test_package_hf_recognizer_model_writes_auditable_model_card -q

Recognizer corpus audit before the next GPU run

The failed PaddleOCR mixed v4 run showed that a nominal mixed corpus is not enough. The next recognizer branch needs an explicit corpus admission step before training, otherwise we can waste GPU time on a dataset that cannot preserve English or does not contain enough Devanagari/mixed-script coverage.

Added ocrtech audit-recognizer-corpus, which reads the actual train and eval manifests and writes:

  • recognizer-corpus-audit.json
  • recognizer-corpus-audit.md

The audit checks:

  • train/eval sample counts
  • Latin-script sample coverage for English preservation
  • Devanagari sample coverage for Nepali recognition
  • mixed Devanagari+Latin samples for real document text
  • real and synthetic training coverage
  • train/eval leakage using the existing leakage detector

This is not proof of model quality. It is a pre-training gate that prevents another branch from being trained without the corpus properties required by the admission validation gates.

Follow-up: bundle-recognizer-training --require-corpus-audit now runs the same audit during bundle creation and refuses to create a remote/GPU training bundle if the audit fails. The bundle manifest records recognizer_corpus_audit with whether the check was required, present, and passed, plus the JSON audit path. This makes the pre-training gate enforceable at the point where we package a job for a GPU host.

Verification:

  • uv run pytest tests/test_datasets.py tests/test_cli.py -q
  • uv run pytest tests/test_training_benchmark.py tests/test_datasets.py tests/test_cli.py -q
  • uv run pytest -q

Follow-up corpus check on the available table/figure development split:

uv run ocrtech audit-recognizer-corpus \
  --train-manifest data/splits-table-figure-dev-v1/train.jsonl \
  --eval-manifest data/splits-table-figure-dev-v1/eval.jsonl \
  --out outputs/recognizer-corpus-audit-table-figure-dev-v1

Result:

  • artifact: outputs/recognizer-corpus-audit-table-figure-dev-v1/recognizer-corpus-audit.json
  • markdown: outputs/recognizer-corpus-audit-table-figure-dev-v1/recognizer-corpus-audit.md
  • exit code: 1
  • train/eval size: 60 / 20, below the default 1000 / 100 gate
  • train/eval mixed-script samples: 60 / 20
  • train/eval Latin-only samples: 0 / 0
  • train/eval Devanagari-only samples: 0 / 0
  • real/synthetic training slices: 0 / 0

This confirms that the current local table/figure dev split is useful for parser and structure experiments, but it is not an admissible recognizer training corpus. The next recognizer corpus needs isolated English-only and Nepali-only line crops in addition to mixed-script rows.

Follow-up implementation: ocrtech build-recognizer-corpus now creates an explicit training manifest from verified source manifests and upsampling targets. It measures the same script/source buckets that the audit cares about:

  • latin_only
  • devanagari_only
  • mixed_devanagari_latin
  • real
  • synthetic

The command writes {manifest-stem}-recognizer-corpus-build.json and {manifest-stem}-recognizer-corpus-build.md next to the output manifest. It annotates duplicated rows with recognizer_corpus_source_sample_id, recognizer_corpus_bucket, and recognizer_corpus_duplicate, so any training run can distinguish real source size from upsampled training size.

The important guardrail is negative: if a requested bucket is absent, the builder fails instead of fabricating a balanced-looking manifest. This keeps the next CUDA recognizer attempt tied to actual English-only, Nepali-only, mixed-script, real, and synthetic evidence rather than reusing the mixed-page table/figure split as a fake recognizer corpus.

Verification:

  • uv run pytest tests/test_datasets.py tests/test_cli.py -q

k2 recognizer-next-v1 corpus handoff

After syncing the updated corpus-audit/builder code to k2, we built a new PaddleOCR recognizer handoff corpus from actual split-safe sources instead of using the table/figure crop split alone.

The failed crop-only audit was useful:

  • train/eval: 480 / 64
  • leakage: pass
  • train mixed-script crops: 68
  • eval Latin-only samples: 0
  • train real/synthetic slices: 0 / 0
  • result: fail

The passing recognizer-next-v1 corpus combines:

  • data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl
  • data/splits-crops-table-figure-dev-v1/train.jsonl
  • eval merge of data/splits-mixed-v1/eval-with-refs.jsonl and data/splits-crops-table-figure-dev-v1/eval.jsonl

Build command on k2:

~/.local/bin/uv run ocrtech build-recognizer-corpus \
  data/splits-mixed-v1/train-rebalanced-nep-eng.jsonl \
  data/splits-crops-table-figure-dev-v1/train.jsonl \
  --out data/splits-recognizer-next-v1/train.jsonl \
  --target-latin-only 850 \
  --target-devanagari-only 850 \
  --target-mixed 100 \
  --target-real 400 \
  --target-synthetic 1000

Audit result:

  • artifact mirror: outputs/k2-recognizer-next-v1/recognizer-corpus-audit.json
  • train/eval: 2212 / 244
  • train Latin/Devanagari/mixed: 1084 / 1174 / 100
  • eval Latin/Devanagari/mixed: 172 / 90 / 28
  • train real/synthetic: 428 / 1272
  • leakage: pass
  • result: pass

Generated PaddleOCR recipe and CUDA handoff bundle on k2:

  • recipe: runs/paddle-recognizer-next-v1/paddle_recognizer.yml
  • training summary mirror: outputs/k2-recognizer-next-v1/recognizer-training-summary.json
  • bundle: runs/paddle-recognizer-next-v1-bundle/paddle-recognizer-next-v1-training.tar.gz
  • bundle manifest mirror: outputs/k2-recognizer-next-v1/recognizer-training-bundle.json
  • bundle corpus audit: required and passed
  • files in bundle: 2464
  • uncompressed bytes: 82095047

This is still pre-training evidence. It proves that the next CUDA run has a better admitted corpus and portable bundle. It does not prove model quality. Quality remains unproven until the trained recognizer is exported, packaged, benchmarked against Tesseract/Surya/stock Paddle, and admitted through the validation gates.

AWS G6 PaddleOCR recognizer-next-v1 final result

We restarted the existing g6.xlarge L4 host i-0fa19dff7b233a353, opened SSH only to the current workstation /32, trained paddle-recognizer-next-v1, exported the best checkpoint, packaged the inference model, benchmarked it on hard-50, and stopped the instance again.

Training summary:

  • host: AWS g6.xlarge, NVIDIA L4
  • recipe: runs/paddle-recognizer-next-v1/paddle_recognizer.yml
  • base: Devanagari PP-OCRv3 mobile recognizer
  • epochs: 20
  • batch size: 16
  • learning rate: 0.00005
  • best epoch: 19
  • best training-log eval acc: 0.016393441951088446
  • best training-log normalized edit distance: 0.37240666956041846

Packaged artifact:

  • model card: runs/paddle-recognizer-next-v1/package/model-card.json
  • local slim mirror: outputs/aws-g6-paddle-recognizer-next-v1/ocrtech-paddle-recognizer-next-v1-slim-artifacts.tgz
  • slim mirror sha256: 3558cb36fc63e463650d229cd21e1b91955c208cfcd483e419e89d3e86b309ce
  • full checkpoint archive retained on AWS at run time: /tmp/ocrtech-paddle-recognizer-next-v1-artifacts.tgz
  • full archive sha256: 4ed465ece824a2865fd8e93ca1845af64a5ae0ffd707bb58e6d4c2c639d8447b

Hard-50 benchmark against Tesseract:

  • report: outputs/experiment-paddle-recognizer-next-v1-hard50/benchmark/report.json
  • summary mirror: outputs/aws-g6-paddle-recognizer-next-v1/slim/outputs/summary-paddle-recognizer-next-v1-hard50/summary.md
  • candidate all CER/WER: 0.964227 / 1.000000
  • Tesseract all CER/WER: 0.080249 / 0.258921
  • English line-crop CER: candidate 0.965341, Tesseract 0.031244
  • Nepali CER: candidate 0.960698, Tesseract 0.235433
  • paired CER wins/losses/ties vs Tesseract: 0 / 50 / 0
  • candidate low-confidence line rate: 1.000000

Admission decision:

  • validation: outputs/validation-paddle-recognizer-next-v1-hard50/validation.json
  • admission: outputs/recognizer-admission-paddle-recognizer-next-v1-hard50/recognizer-admission-decision.json
  • decision: reject
  • claim status: not_validated

Interpretation: this is another negative PaddleOCR recognizer result. The better admitted corpus did not fix the core recognizer failure. This branch must not be promoted into the document engine, and another single Devanagari PP-OCRv3 fine-tune is not a credible next step without changing the recognizer design itself.

PaddleOCR Devanagari PP-OCRv3 family stop guardrail

Follow-up implementation: added ocrtech recognizer-failure-review, which aggregates recognizer admission decisions or validation reports across runs and emits:

  • recognizer-failure-review.json
  • recognizer-failure-review.md

The guardrail groups equivalent Devanagari PP-OCRv3 mobile base identifiers even when one model card records the Hugging Face-style name and another records the PaddleX pretrained URL. It recommends stop_family only when a family has at least three rejected runs and zero paired wins across failed admission gates.

Current review artifact:

  • review: outputs/recognizer-failure-review-paddle-ppocrv3/recognizer-failure-review.json
  • markdown: outputs/recognizer-failure-review-paddle-ppocrv3/recognizer-failure-review.md
  • runs included: PaddleOCR mixed v2, v3, v4, and recognizer-next-v1
  • family: paddleocr::PaddlePaddle/devanagari_PP-OCRv3_mobile_rec
  • rejected runs: 4
  • failed gates: 16
  • paired gate wins/losses/ties: 0 / 545 / 7
  • failed slices: english, line_crop, nepali, paddle_hard_eval
  • recommendation: stop_family

This is not a claim that local OCR is impossible. It is a specific ablation result: the repeated Devanagari PP-OCRv3 mobile fine-tune family is not the engine core. The next recognizer-side work must change the base architecture, the objective, or move to a document-native model; otherwise the GPU spend is not justified by the existing admission evidence.

Follow-up enforcement: train-recognizer and bundle-recognizer-training now accept --failure-review. When supplied, they reject a backend/base-model pair whose family has recommendation: stop_family in the review artifact. The bundle command reads the actual recognizer-training-summary.json from the recipe directory, so remote GPU handoffs are checked against the recorded training provenance rather than a manually repeated flag.

This converts the PP-OCRv3 negative result from a note into an operational guardrail: future AWS/G6 jobs must either use a different recognizer family or carry a new failure-review artifact that does not stop the proposed family.

2026-06-15 Document Reconstruction Metric

The benchmark scorer now treats document reconstruction as a first-class metric instead of only exporting document.md as an artifact.

Reference JSON can provide one of:

  • markdown
  • document_markdown
  • reconstruction_markdown

Plain Markdown reference files such as document.md are also treated as reconstruction references, so materialized reference packs can emit this metric without switching to JSON references.

When present, the scorer renders the predicted Document back to Markdown and reports:

  • document_reconstruction_cer
  • document_reconstruction_wer
  • document_reconstruction_similarity

document_reconstruction_similarity is 1 - document_reconstruction_cer, clamped at zero, so validation and claim-review configs can use it as a higher-is-better metric. The SOTA claim config now requires this metric for the internal document pack and public document reconstruction slots.

This plugs a paper-methods gap: CER/WER can improve while Markdown structure, tables, figures, or reading order regress. The reconstruction metric gives us a single deterministic score over the final user-facing Markdown output, while the table, reading-order, and figure metrics remain the diagnostic breakdown.

2026-06-15 Auditable Selective Fallback Metadata

The parser's selective fallback path is now auditable. When parse-doc is run with --fallback-engine, document.json records metadata.fallback with:

  • primary engine
  • fallback engine
  • primary average confidence
  • threshold
  • trigger reason
  • outcome (not_triggered, success, or error)

Benchmark scoring also surfaces fallback diagnostics:

  • fallback_triggered
  • fallback_succeeded
  • fallback_failed
  • fallback_primary_average_confidence
  • fallback_threshold

This matters for the intended local-engine-plus-selective-VLM design. A paper claim must be able to say whether the local path handled the document or whether the fallback carried it. These fields make fallback usage measurable instead of implicit.

2026-06-15 Stricter SOTA Gate Coverage

The SOTA validation contract now treats named claim dimensions as gates, not just reported metrics.

Config changes:

  • configs/validation-internal-sota-v1.json now gates Nepali WER against Tesseract and English WER against stock PaddleOCR.
  • configs/validation-internal-sota-v1.json now gates figure_detection_f1 and figure_caption_cer on the figure slice against Surya.
  • configs/validation-public-real5-sota-v1.json now gates WER against Surya and GLM-OCR on the public slice.
  • configs/claim-review-sota-v1.json now requires matching passing validation gates for those WER and figure/caption metrics.
  • tests/test_claim_review_sota_config.py pins that every required review gate is covered by its validation config.

This closes a validation loophole: a portfolio can no longer pass the SOTA review merely because WER, figure detection, or figure-caption metrics are present in benchmark rows. They must be paired against the named baseline and pass the configured validation gate.

2026-06-15 Structured Artifact Evidence Gate

The SOTA review now checks that candidate benchmark rows produce the actual structured artifacts named by the engine claim, not just scalar metrics.

Implementation:

  • review-claim accepts required_output_artifacts and required_output_artifact_slices per required experiment slot.
  • Artifact paths are inferred from the benchmark report directory plus candidate/<input-stem>/.
  • Supported artifact checks:
    • document_json: document.json
    • document_md: document.md
    • quality_json: quality.json
    • table_csv_html: at least one tables/*.csv and one tables/*.html
    • figure_metadata: figures/metadata.json
    • figure_images: at least one exported figure image

The checked-in SOTA profile now requires:

  • all required slots: document_json, document_md, quality_json
  • internal table slice: table_csv_html
  • internal figure slice: figure_metadata, figure_images
  • public OmniDocBench table slice: table_csv_html

This closes another paper-methods loophole. A run cannot support a structured document-understanding claim if the report has metrics but the reconstructed Markdown, JSON, table files, or figure files are missing.

Follow-up: the artifact gate now validates minimal artifact content rather than only checking filenames. document.json and quality.json must be valid JSON, document.json must expose the expected structured-document keys, document.md, table CSV/HTML exports, and figure image files must be nonempty, and figure metadata must be a nonempty JSON list. This prevents empty or malformed placeholders from satisfying the structured-output requirement.

2026-06-15 Validation Config Binding Gate

The SOTA portfolio now binds each required experiment slot to the exact checked-in validation config used for the paper claim.

Implementation:

  • review-claim supports required_validation_config per required experiment.
  • During review, the configured validation file is hashed and compared against experiment.json environment.validation_config_sha256.
  • configs/claim-review-sota-v1.json binds:
    • internal pack: configs/validation-internal-sota-v1.json
    • public OmniDocBench: configs/validation-public-omnidocbench-sota-v1.json
    • public Real5/olmOCR slot: configs/validation-public-real5-sota-v1.json
  • tests/test_claim_review_sota_config.py now checks that the review config points at those files and that their gates cover the required review gates.

This closes a methods loophole: a run cannot pass the SOTA review by using a locally weakened validation config with smaller sample counts, easier slice thresholds, or omitted gates while preserving the same experiment name.

2026-06-15 Trained Recognizer Provenance Gate

The SOTA claim review now has an explicit trained-recognizer requirement.

Change:

  • configs/claim-review-sota-v1.json sets require_trained_recognizer: true.
  • Direct recognizer candidates (paddleocr, hf_vision_encoder_decoder) must include recorded artifacts plus train/eval manifest and training-summary provenance with SHA-256 hashes.
  • Composite candidates can satisfy the requirement only by pointing at a nested trained recognizer model card through provenance.trained_recognizer_model_card.

This closes a loophole in the claim story. A pure Tesseract/Surya/router/VLM wrapper may be a useful engineering baseline, but it does not satisfy the user objective of a trained local Devanagari+English recognizer with document capabilities. The paper claim now has to prove both: trained recognizer provenance and benchmark superiority.

Follow-up: preflight-claim now has --require-trained-recognizer, using the same model-card provenance rule. This moves the trained-recognizer requirement from end-of-portfolio review into the pre-benchmark checklist, so we fail before spending remote/GPU time on a claim run whose candidate is only a wrapper.

Follow-up: the trained-recognizer provenance gate now also requires provenance.training_summary. This makes the launch artifact mandatory for direct trained recognizers, not just a nice-to-have package file. A recognizer model card with weights and train/eval hashes but no recipe/config summary no longer satisfies the SOTA trained-recognizer requirement.

Follow-up: trained-recognizer provenance is now hash-checked, not just field-checked. The gate resolves provenance.train_manifest, provenance.eval_manifest, and provenance.training_summary from the model card and recomputes SHA-256 before admitting a direct PaddleOCR or HF recognizer. This blocks stale or manually edited training evidence from passing the SOTA claim checklist.

2026-06-15 Model Admission Validation Gate

Trained provenance is necessary but not sufficient. The failed PaddleOCR v2/v3 branches had artifacts and train/eval manifests, yet the hard-50 validation rejected them. Model cards can now record:

{
  "metrics": {
    "admission_validation": {
      "path": "outputs/validation-next-recognizer-hard50/validation.json",
      "sha256": "...",
      "required_status": "validated"
    }
  }
}

preflight-claim --require-model-admission and SOTA review-claim can require that admission report to have claim_status=validated and passed=true. This prevents us from promoting a trained recognizer just because it was trained; it must also pass the branch admission gate before entering the full document SOTA portfolio.

Follow-up: recognizer packaging now accepts --admission-validation-report for PaddleOCR packages, Hugging Face recognizer packages, and finalized Hugging Face recognizer runs. This means admission evidence is attached when the candidate model card is created, instead of being left as a manual model-card edit before claim review.

2026-06-15 Metric-Slice Coverage Gate

The portfolio claim review now supports required_metric_slices. This closes a subtle validation loophole: previously an experiment could contain a figure slice and a finite figure_caption_cer metric somewhere else in the report, and the review would treat the metric family as covered. The SOTA config now requires metric evidence on the relevant rows:

  • cer and wer on both Nepali and English rows.
  • table_cell_f1 on table rows.
  • reading_order_pair_accuracy on reading-order rows.
  • figure_detection_f1 and figure_caption_cer on figure rows.
  • document reconstruction similarity on table, reading-order, and figure rows for the internal document pack.

This matters for the paper because the claim is not "some metric was emitted"; it is that the engine handles the target document capabilities on the examples designed to test them.

2026-06-15 Validation Artifact Cross-Check

Portfolio review now reads the referenced validation.json whenever a required experiment slot declares require_validation_status. The copied validation_status in experiment.json must agree with the validation artifact's claim_status, and a validated requirement also requires passed: true in that artifact.

This prevents stale or manually edited experiment summaries from satisfying the paper claim review when the authoritative validation report is actually not_validated.

Follow-up: validate-claim now writes provenance into validation.json for the benchmark report, validation config, eval manifest, and train manifests. review-claim cross-checks those SHA-256 values against the experiment report. This prevents a passing validation artifact from one run being reused against a different benchmark, eval manifest, validation config, or training-manifest set in the final portfolio review.

2026-06-15 Public Slot Validation Tightening

The SOTA claim-review profile now requires require_validation_status: validated for every required experiment slot, including public benchmark slots. Public slots also require non-candidate baselines instead of accepting a candidate-only run with emitted metrics.

The internal Nepali/English document slot now explicitly requires Tesseract, stock PaddleOCR, GLM-OCR, and PaddleOCR-VL alongside candidate evidence. This matches the paper objective: the claim is comparative against strong existing systems, not merely an existence proof that the candidate can produce Markdown, JSON, tables, and figure metrics.

2026-06-15 Required Passing Validation Gates

The claim-review config now supports required_validation_gates. This closes the gap between metric coverage and actual superiority: a slot can require that validation.json contain a passed gate for a specific baseline, metric, and slice set.

The SOTA profile uses this to require passing gates against Tesseract, stock PaddleOCR, GLM-OCR, Surya, and PaddleOCR-VL where relevant. This is stricter than checking that those baselines ran or that a metric exists; it requires the configured validation test to say the candidate won.

Follow-up: claim-review config loading now rejects contradictory required_validation_gates. A gate cannot be declared unless the experiment slot requires validation status, and its baseline/metric must also be listed in require_baselines and required_metrics. This keeps the SOTA profile internally auditable instead of relying on later, ambiguous review failures.

Follow-up: added claim-grade validation configs for the three SOTA review slots:

  • configs/validation-internal-sota-v1.json
  • configs/validation-public-omnidocbench-sota-v1.json
  • configs/validation-public-real5-sota-v1.json

Regression coverage now checks that these validation configs cover every required_validation_gates entry in configs/claim-review-sota-v1.json. This turns the paper validation plan into executable configuration rather than a manual checklist.

Follow-up: portfolio review now re-checks the thresholds embedded in matched validation gates. A gate marked passed: true is not enough; its paired count, margins, pair-win rate, p-value, and CI fields must satisfy the gate payload itself. This catches hand-edited validation reports that flip pass status without satisfying min_pairs or the statistical checks required by the SOTA configs.

2026-06-15 Model-Card Selective Fallback

parse-doc now accepts --fallback-model-config in addition to --fallback-engine. When primary OCR confidence falls below the configured threshold, the parser can instantiate the fallback as a candidate engine from that model card.

This matters for the engine design because a selective VLM/document-model fallback should be packaged and audited like any other candidate component, not hidden behind an engine name. metadata.fallback now records the fallback model card path when used, so benchmark diagnostics can distinguish local-core output from model-card-backed fallback output.

Follow-up: benchmark and run-experiment now expose the same fallback knobs for candidate/ours rows:

  • --fallback-engine
  • --fallback-model-config
  • --low-confidence-threshold

This is important for the paper because the benchmarked candidate should be the actual hybrid engine we intend to ship, not only the primary recognizer path. The fallback remains candidate-only, so Tesseract, stock PaddleOCR, Surya, GLM-OCR, and PaddleOCR-VL baseline rows are not modified by our routing policy. experiment.json records the fallback engine, fallback model-card path, fallback model-card SHA-256, and confidence threshold for reproducibility.

Follow-up: portfolio review can now constrain fallback dependency with:

  • require_candidate_fallback_metrics
  • max_candidate_fallback_trigger_rate
  • max_candidate_fallback_failed_rate

The checked-in SOTA profile requires fallback diagnostics on every candidate row, caps fallback trigger rate at 0.35, and requires fallback failure rate 0.0 for each required experiment slot. This prevents a paper claim from being silently converted into "the fallback model handled the benchmark" while still allowing selective fallback on hard pages.

Follow-up: benchmark report.md now includes per-row fallback columns: fallback and fallback_failed. The JSON report already carried these metrics, and summary.md aggregates them automatically, but the human run log now shows fallback dependency without requiring manual JSON inspection.

Follow-up: review-claim now audits fallback model cards when an experiment records fallback_model_config. The review checks that the fallback card path exists, its SHA-256 matches experiment.json, and audit-model passes. The SOTA profile also requires fallback provenance to exist and requires a shared fallback model across required slots. This keeps the hybrid-engine claim honest: the fallback component is part of the system under review, not an unaudited sidecar.

Follow-up hardening: review-claim now supports require_fallback_model_present, and configs/claim-review-sota-v1.json sets it to true. This closes a loophole where fallback metrics could be required while the portfolio omitted fallback_model_config entirely. For the SOTA profile, every candidate experiment must now record the fallback model-card path and SHA-256 in experiment.json, and validation provenance must bind back to the same fallback card when validation is required.

Follow-up: validate-claim now records candidate and fallback model-card provenance when those paths are supplied, and run-experiment passes them automatically. review-claim cross-checks those hashes against experiment.json alongside the existing benchmark, eval manifest, validation config, and train-manifest provenance. This prevents a validated benchmark artifact from being reused after swapping either the primary recognizer card or the fallback model card.

Follow-up: preflight-claim now accepts --fallback-model-config and audits that fallback card before the benchmark starts. This moves another failure mode from end-of-portfolio review into the cheap pre-run checklist, which matters for remote or GPU-backed benchmark runs.

2026-06-15 Recognizer Training Recipe Provenance

train-recognizer --backend paddleocr now writes recognizer-training-summary.json next to paddle_recognizer.yml. The summary captures train/eval manifest paths and hashes, sample counts, dataset counts, slice counts, dictionary character coverage, generated label/config hashes, and launch hyperparameters.

This does not train a better model by itself. It closes a paper/reproducibility gap for the next CUDA recognizer attempt: the training bundle can now prove what data and dictionary went into the run before we evaluate the exported model.

Follow-up: bundle-recognizer-training now records the summary path and SHA-256 under training_summary in recognizer-training-bundle.json. This makes the CUDA handoff auditable even before the archive is extracted.

Follow-up: packaged recognizer model cards now accept --training-summary and record its path and SHA-256 under provenance.training_summary. The final candidate card can now point back to the exact recipe summary used for training, instead of relying on adjacent files or runbook prose.

2026-06-15 AWS G6 V4 Provisioning

Created a fresh CUDA host through the playaround AWS profile before the SSO window expired:

  • instance: i-0fa19dff7b233a353
  • type: g6.xlarge
  • region/AZ: us-east-1 / us-east-1a
  • public IP at launch: 3.231.208.121
  • AMI: Deep Learning Base OSS Nvidia Driver GPU AMI Ubuntu 22.04, ami-01011b868ec560823
  • SSH key: ocrtech-g6-training-20260615
  • security group: sg-05ad6744a62dcabdd, SSH restricted to 27.34.64.200/32
  • root volume: 200 GiB gp3, delete on termination

Purpose: use the instance for the next PaddleOCR recognizer branch only after the training bundle and model-card summary provenance are in place.

Bootstrap status:

  • installed repo on Python 3.11.15 via uv
  • installed paddlepaddle-gpu==3.3.1
  • checked out PaddleOCR commit c166448875bcecb8d3b7628fd697ac1c28f8705b
  • paddle.utils.run_check() passed on one NVIDIA L4
  • copied and extracted the mixed-v2 training bundle from k2

First v4 launch found another recipe bug before training consumed GPU time: YAML parsed learning_rate: 5e-05 as a string, and Paddle rejected it with TypeError: The type of learning rate must be float. The generator now emits small learning rates as plain decimals, for example 0.00005, and has a regression test for this exact case.

Follow-up: a bad source sync copied the local macOS .venv onto the Linux G6 host and broke remote console scripts with invalid interpreter paths. The host venv was repaired with uv venv --python 3.11, repo extras, paddlepaddle-gpu==3.3.1, PaddleOCR requirements, and another successful paddle.utils.run_check(). The remote-training runbook now explicitly warns not to sync .venv and records the exact v4 recovery/relaunch commands.

Local verification after the YAML scalar, export checkpoint-prefix, and package-relative model-card hash fixes:

  • focused training/CLI tests: 78 passed
  • full suite: 270 passed, 7 skipped

AWS G6 PaddleOCR mixed v4 final result

The v4 CUDA run completed on the g6.xlarge L4 instance after the YAML float and remote .venv recovery fixes. This branch used the official Devanagari PP-OCRv3 recognizer checkpoint, the verified mixed-v1 train split, a balanced 850 English / 850 Nepali training manifest, batch size 16, learning rate 0.00005, and 20 epochs.

Training produced an exportable PaddleOCR inference package, but the admission gate failed decisively:

  • artifact bundle: outputs/aws-g6-paddle-mixed-v4-cuda/ocrtech-paddle-mixed-v4-cuda-artifacts.tgz
  • run directory inside bundle: runs/paddle-recognizer-mixed-v4-cuda
  • package card inside bundle: runs/paddle-recognizer-mixed-v4-cuda/package/model-card.json
  • experiment summary inside bundle: outputs/experiment-paddle-mixed-v4-cuda-hard50/summary/summary.md
  • hard-50 manifest: data/splits-mixed-v1/eval-paddle-mixed-v4-hard50-with-refs.jsonl
  • validation status: not_validated
  • candidate all CER/WER: 0.940500 / 1.028667
  • Tesseract all CER/WER: 0.080249 / 0.258921
  • English line-crop CER: candidate 0.950730, Tesseract 0.031244
  • Nepali CER: candidate 0.908108, Tesseract 0.235433
  • paired CER wins/losses/ties vs Tesseract: 0 / 50 / 0

The paper conclusion for this branch is negative and useful: lowering the learning rate fixed neither the English collapse nor Nepali quality. The same single Devanagari PP-OCRv3 mixed-script fine-tuning path has now failed v2, v3, and v4 hard gates. Treat it as evidence against making this recognizer the core engine. Future engine work should admit recognizers behind measurable gates rather than assume that a single local mixed-script recognizer will dominate.

Admission rejection details

Follow-up implementation: model_admission_status now extracts failed requirements and failed-gate deltas from a referenced admission validation.json. Preflight and claim review already use this status object, so a rejected recognizer now reports details like:

  • missing_requirement=...
  • failed_gate=cer baseline=tesseract slice=nepali
  • candidate_mean=...
  • baseline_mean=...
  • absolute_improvement=...
  • pairs=... win_count=... loss_count=... tie_count=...

This matters after the v4 CUDA failure because the next engine iteration should reject bad recognizers at the admission boundary with explicit evidence, not bury the failure inside a tarball or require manual JSON inspection.

Verification:

  • uv run pytest tests/test_models_candidate.py tests/test_preflight.py tests/test_claim_review.py -q
  • uv run pytest -q

Admission provenance binding

Follow-up hardening: a passing admission report is no longer accepted only because it says claim_status=validated. model_admission_status now requires the validation report to include provenance.candidate_model_config, loads that model card, and compares stable identity fields with the current candidate:

  • model_id
  • backend
  • base_model
  • artifact path, hash, kind, and required flag

This intentionally avoids requiring the final model-card SHA-256 to match, because embedding the admission report changes the model card hash. The stable identity comparison still closes the important loophole: a passing admission report from another recognizer cannot be attached to a new model card and used for SOTA review.

Follow-up: model admission now also requires quantitative passing gate evidence. claim_status=validated and passed=true are not sufficient unless the admission report contains passed gates with paired counts, finite means, improvement statistics, win/loss/tie counts, and thresholds that satisfy the gate payload. This brings recognizer admission in line with portfolio claim-review gate evidence.

Follow-up: recognizer admission and portfolio claim review now use the same gate_has_quantitative_evidence implementation. This prevents the two claim paths from drifting on what counts as a valid passed gate.

Verification:

  • uv run pytest tests/test_models_candidate.py tests/test_preflight.py tests/test_claim_review.py tests/test_training_benchmark.py -q

Required validation gate evidence

Follow-up hardening: review-claim no longer treats a passed: true validation gate as enough evidence for a required portfolio gate. A required gate match must now include quantitative paired evidence:

  • pairs > 0
  • finite candidate_mean
  • finite baseline_mean
  • finite absolute_improvement
  • numeric win_count, loss_count, and tie_count
  • win_count + loss_count + tie_count == pairs

This closes a paper-review loophole where a hand-edited validation report could declare a gate as passed without carrying the paired comparison data needed to audit the claim.

Verification:

  • uv run pytest tests/test_claim_review.py tests/test_claim_review_sota_config.py -q

Recognizer admission decision artifact

Follow-up implementation: added ocrtech recognizer-admission-decision and the underlying decide_recognizer_admission helper. The command reads an admission validation.json and writes:

  • recognizer-admission-decision.json
  • recognizer-admission-decision.md

The artifact makes the branch decision explicit:

  • admit only when claim_status=validated and passed=true
  • reject otherwise, preserving missing requirements and failed gate metrics
  • next-action hints based on failed slices and metrics, including English preservation failures and no-win admission gates

This turns failed recognizer evidence, like the PaddleOCR v4 hard-50 result, into a reusable decision artifact instead of only a narrative note. The command returns exit code 0 for admit and 1 for reject, so remote training scripts can stop before packaging or promotion when an admission gate fails.

Verification:

  • uv run pytest tests/test_error_analysis.py tests/test_cli.py -q

Applied to the AWS G6 PaddleOCR mixed v4 run:

uv run ocrtech recognizer-admission-decision \
  --validation-report outputs/aws-g6-paddle-mixed-v4-cuda/extracted-validation/outputs/experiment-paddle-mixed-v4-cuda-hard50/validation/validation.json \
  --out outputs/aws-g6-paddle-mixed-v4-cuda/recognizer-admission-decision

Result:

  • decision artifact: outputs/aws-g6-paddle-mixed-v4-cuda/recognizer-admission-decision/recognizer-admission-decision.json
  • markdown artifact: outputs/aws-g6-paddle-mixed-v4-cuda/recognizer-admission-decision/recognizer-admission-decision.md
  • decision: reject
  • failed gates include hard-50 CER 0.940500 vs Tesseract 0.080249, English line-crop CER 0.950730 vs 0.031244, and zero CER wins.
  • next action recorded by the artifact: do not train another Devanagari-only branch until English preservation is explicitly designed and gated.

Verification after applying the real v4 decision:

  • uv run pytest -q

Figure artifact export from bboxes

The parser now produces inspectable figure image artifacts when the engine detects a figure but does not provide a separate figure.image_path.

Behavior:

  • if figure.image_path exists, export_document copies it into figures/
  • otherwise, for ordinary image inputs and page-bundle directories, it crops the source page image using the figure bbox into figures/{figure_id}.png when Pillow is available
  • for page bundles, figure.page_index selects the source page image
  • if the crop cannot be produced, metadata.export_warnings records the reason

This matters for the paper because figure detection should be auditable as a visual artifact, not only as a JSON bbox. It also keeps failure explicit: missing Pillow, unsupported source formats, bad bboxes, or missing figure paths become metadata warnings instead of silent omissions.

Verification:

  • uv run pytest tests/test_pipeline.py -q
  • uv run pytest -q

Portfolio review now requires paired metric coverage

review-claim now rejects a required experiment slot when candidate and baseline metrics are present but not paired on the same sample ids. This is separate from metric presence and slice coverage:

  • required_metrics proves each required baseline emitted the metric
  • required_metric_slices proves the metric exists on the relevant document type
  • paired coverage proves candidate and external baselines were measured on the same samples for that metric and slice

For unsliced requirements, the paired coverage check uses all successful rows. For metrics listed in required_metric_slices, it checks the configured slices. This closes a portfolio-review loophole where a report could contain candidate metrics on one set of documents and baseline metrics on another set, then appear complete despite never making a same-document comparison.

Follow-up: portfolio review now also requires full sample coverage for every require_baselines entry. Each required baseline must have a successful row for every sample id in the reviewed eval manifest. This closes the adjacent loophole where a report could point at a hard eval pack but only benchmark an easier subset.

Verification:

  • uv run pytest tests/test_claim_review.py -q

Current training status and engine artifact traceability

As of this note, no recognizer training job is active. The last AWS G6 PaddleOCR mixed v4 candidate was evaluated and rejected by admission gates, so it should be treated as a negative result, not a publishable trained model. The current engineering work is on the document engine and evidence pipeline: ensuring every structured output can be traced to concrete artifacts before the next training run is launched.

The parser's table export path now has explicit regression coverage that document.json records the generated tables/{table_id}.csv and tables/{table_id}.html paths, plus the HTML table body. This matters for paper reproducibility because table accuracy claims must point to the exact CSV/HTML artifacts used for scoring, not only to an in-memory structure.

Verification:

  • uv run pytest tests/test_pipeline.py -q

Full eval-manifest coverage for SOTA validation

validate-claim now supports require_full_eval_manifest_coverage in the validation config. When enabled, validation fails unless the candidate and every baseline named by configured gates have successful rows for every sample in the provided eval manifest.

The three SOTA validation profiles now enable this:

  • configs/validation-internal-sota-v1.json
  • configs/validation-public-omnidocbench-sota-v1.json
  • configs/validation-public-real5-sota-v1.json

This closes a methods loophole: a run can no longer validate by scoring only an easy subset of a held-out or public manifest while preserving enough samples to satisfy min_samples, slice counts, and gate min_pairs. The validation artifact must cover the full manifest it cites.

Verification:

  • uv run pytest tests/test_references_validation.py::test_validate_claim_can_require_full_eval_manifest_coverage -q
  • uv run pytest tests/test_claim_review_sota_config.py -q

Embedded preflight for guarded experiment runs

run-experiment now supports an embedded preflight phase. Passing --preflight through the CLI, or run_preflight=True through the Python API, runs the same claim-readiness checks before benchmark execution and stops before writing a benchmark report if preflight fails.

This matters for remote/GPU runs: the command can now reject an impossible claim setup before spending hours on OCR. The embedded report is written under <experiment>/preflight/preflight.json, and successful experiments record preflight_report plus preflight_passed in experiment.json.

The intended claim-grade command now includes:

ocrtech run-experiment \
  --eval-manifest data/eval-pack/manifests/ocrtech-eval-pack.jsonl \
  --references-dir data/eval-pack/references \
  --baselines candidate,tesseract,stock-paddle,glm-ocr,paddleocr-vl \
  --candidate-model-config runs/recognizer/package/model-card.json \
  --validation-config validation.json \
  --train-manifest data/splits/train.jsonl \
  --preflight \
  --require-trained-recognizer \
  --require-model-admission \
  --out outputs/claim-experiment \
  --require-validated

Verification:

  • uv run pytest tests/test_experiments.py -q

Next recognizer branch moves away from stopped PaddleOCR family

The PaddleOCR Devanagari PP-OCRv3 recognizer family is stopped by admission evidence:

  • failure review: outputs/recognizer-failure-review-paddle-ppocrv3/recognizer-failure-review.json
  • rejected runs: 4
  • paired wins/losses/ties across failed admission gates: 0 / 545 / 7
  • failed slices: English, line-crop English, Nepali, and hard admission rows

The next local recognizer candidate is therefore a materially different branch: Hugging Face vision-encoder-decoder from microsoft/trocr-small-printed. The actual audited manifests were copied back from k2 and re-audited locally:

  • train manifest: data/splits-recognizer-next-v1/train.jsonl, 2212 samples
  • eval manifest: data/splits-recognizer-next-v1/eval.jsonl, 244 samples
  • local corpus audit: outputs/recognizer-corpus-audit-next-v1-local/recognizer-corpus-audit.json
  • audit status: passed

Generated recipe:

  • config: runs/hf-recognizer-next-v1/hf_recognizer_config.json
  • summary: runs/hf-recognizer-next-v1/recognizer-training-summary.json
  • backend: hf_vision_encoder_decoder
  • base model: microsoft/trocr-small-printed
  • max target length: 128
  • batch size: train/eval 8 / 8
  • learning rate: 0.00005
  • epochs: 5

This is not a trained model yet. It is the next GPU-ready recognizer branch, selected specifically because the prior PaddleOCR family is stopped.

Follow-up implementation: the generated HF recipe README now renders the full reproducible launch command from the actual config, including base model, target length, batch sizes, learning rate, and epochs. This prevents handoff docs from drifting away from the recorded recipe.

Verification:

  • uv run ocrtech audit-recognizer-corpus --train-manifest data/splits-recognizer-next-v1/train.jsonl --eval-manifest data/splits-recognizer-next-v1/eval.jsonl --out outputs/recognizer-corpus-audit-next-v1-local
  • uv run ocrtech train-recognizer --backend hf-vision-encoder-decoder --base-model microsoft/trocr-small-printed --train-manifest data/splits-recognizer-next-v1/train.jsonl --eval-manifest data/splits-recognizer-next-v1/eval.jsonl --out runs/hf-recognizer-next-v1 --max-target-length 128 --train-batch-size 8 --eval-batch-size 8 --learning-rate 0.00005 --epochs 5 --failure-review outputs/recognizer-failure-review-paddle-ppocrv3/recognizer-failure-review.json

Follow-up bundle handoff from k2: the HF recognizer recipe has a portable training bundle with all referenced image assets staged under bundle_assets/images.

  • local archive mirror: outputs/k2-hf-recognizer-next-v1/hf-recognizer-next-v1-training.tar.gz
  • archive SHA-256: aff0b277b9cfdd353ed82178c65586353c3511d6db06bd20099cbc62df113873
  • bundle manifest: outputs/k2-hf-recognizer-next-v1/recognizer-training-bundle.json
  • bundle backend: hf_vision_encoder_decoder
  • bundle files: 2463
  • uncompressed bundle bytes: 82091601
  • corpus audit in bundle: required, present, passed
  • run command recorded by bundle: ocrtech train-recognizer --backend hf-vision-encoder-decoder --base-model microsoft/trocr-small-printed --train-manifest data/splits-recognizer-next-v1/train.jsonl --eval-manifest data/splits-recognizer-next-v1/eval.jsonl --out runs/hf-recognizer-next-v1 --max-target-length 128 --train-batch-size 8 --eval-batch-size 8 --learning-rate 5e-05 --epochs 5 --run

Implementation detail: bundle-recognizer-training now records recipe_backend and run_command in recognizer-training-bundle.json. Its README no longer assumes PaddleOCR; HF bundles now show the HF training command. The PaddleOCR label helper also creates its parent directory so non-Paddle recipes can still include portable label files for inspection.

Verification:

  • uv run pytest tests/test_training_benchmark.py::test_bundle_recognizer_training_writes_hf_run_command tests/test_training_benchmark.py::test_train_hf_recognizer_writes_recipe -q

Follow-up AWS launch on 2026-06-15:

  • instance: i-0fa19dff7b233a353
  • instance type: g6.xlarge
  • GPU: NVIDIA L4, 23034 MiB
  • public IP at launch: 3.91.25.53
  • remote workspace: /opt/ocrtech
  • archive staged at: /opt/ocrtech/incoming/hf-recognizer-next-v1-training.tar.gz
  • archive SHA-256: aff0b277b9cfdd353ed82178c65586353c3511d6db06bd20099cbc62df113873
  • torch runtime verified before launch: 2.12.0+cu130, torch.cuda.is_available() == True, device NVIDIA L4
  • remote PID: 1862
  • remote log: /opt/ocrtech/runs/hf-recognizer-next-v1/train.log
  • remote PID file: /opt/ocrtech/runs/hf-recognizer-next-v1/train.pid

Launch command:

ocrtech train-recognizer \
  --backend hf-vision-encoder-decoder \
  --base-model microsoft/trocr-small-printed \
  --train-manifest data/splits-recognizer-next-v1/train.jsonl \
  --eval-manifest data/splits-recognizer-next-v1/eval.jsonl \
  --out runs/hf-recognizer-next-v1 \
  --max-target-length 128 \
  --train-batch-size 8 \
  --eval-batch-size 8 \
  --learning-rate 5e-05 \
  --epochs 5 \
  --run

Status at launch check:

  • process alive
  • GPU memory in use: about 6500 MiB
  • GPU utilization observed between 46% and 98%
  • trainer progress observed past 200 / 1385 steps

This is not a quality result and not a SOTA claim. It is a live training run for the next non-Paddle recognizer branch. The prior HF model failed against Tesseract, so this branch must still pass held-out admission gates before it can be treated as an engine component.

Final training status:

  • training completed all 5 epochs on AWS G6.
  • final model copied locally to: outputs/aws-g6-hf-recognizer-next-v1/runs-hf-recognizer-next-v1/final-model
  • local package: outputs/aws-g6-hf-recognizer-next-v1/package/model-card.json
  • local model audit: outputs/aws-g6-hf-recognizer-next-v1/audit/model-audit.json
  • audit status: passed

Final training/eval telemetry on the recipe eval manifest:

  • eval_loss: 1.0254589319229126
  • eval_cer: 0.18840631911585276
  • eval_wer: 0.36464600972797695
  • eval_exact_line_accuracy: 0.4344262295081967
  • eval_runtime: 8.3352
  • epoch: 5.0

Operational note: after artifact retrieval, the instance was shut down through SSH with sudo shutdown -h now because the local AWS SSO session expired before an API stop call could be issued. SSH to the last public IP timed out after the shutdown command. Confirm final EC2 state through AWS after refreshing SSO.

This is still only a trained and auditable recognizer artifact. The next required evidence is a benchmark against Tesseract, Surya, stock PaddleOCR, and other configured baselines on held-out samples that were not used for training or tuning.