Download ocr-tech/paper-buildkit/ARTIFACTS.md from prosperious/ocr-projects-archive: direct link, hf CLI and curl.
- Browser
- Download file 20.1 kB
-
https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/paper-buildkit/ARTIFACTS.md
- Command line
-
hf download hf://prosperious/ocr-projects-archive/ocr-tech/paper-buildkit/ARTIFACTS.md
-
curl -L -o ARTIFACTS.md https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/paper-buildkit/ARTIFACTS.md
Limbu paper claim-to-artifact ledger
“Verified” below means the value was read from a surviving artifact and is
checked by paper/verify_claims.py. It does not make machine-converted silver
references equivalent to human ground truth.
A. Sirijonga real-print source and split
Verified values: 16 source dates; 4,872 source line pairs; 38 cross-split exact-label rows removed; 4,834 retained rows; train/dev/test counts of 2,711/1,221/902 over 9/4/3 dates.
Artifact:
outputs/limbu-sirijonga-silver-pack-v2-2026-07-05/pack-summary.json
Identity: pack content SHA-256
e9c7851851c9613794cf0c0e5d72659186077ddebee650b7443c03969fe12e9a.
Boundary: gt_status=pdf_text_layer_silver and
claim_evidence_eligible=false. Text is converted from a Namdhinggo legacy
layer through data/mappings/limbu-legacy/Limbu.map, not transcribed by a
human reviewer.
B. ft-v2 training payload and configuration
Verified values: 25,065 retained synthetic rows; 3,090 missing synthetic rows excluded; 2,711 real training lines replayed eight times; 46,753 training rows; 1,221 development rows; 40 configured epochs; batch size 64; learning rate $10^{-4}$; input 3x48x320; maximum text length 128.
Artifacts:
outputs/limbu-ft-v2-payload-2026-07-05/limbu-ft-v2-payload.tgz, memberslimbu-ft/payload-summary.jsonandlimbu-ft/limbu-sirijonga-ft-v2.yml;outputs/limbu-ft-v2-run-2026-07-06/checkpoints/config.yml; andoutputs/limbu-ft-v2-run-2026-07-06/train.log.
Recorded history: RESEARCH.md records approximately 8.5 L4 GPU-hours and
best epoch 39. The configuration and selected checkpoint survive; the paper
build does not repeat training.
C. Date-disjoint line recognition
Verified values:
| System | Lines | Codepoint CER | Grapheme CER | Exact-line accuracy |
|---|---|---|---|---|
| Previous | 902 | 0.013787556730051128 | 0.028400954653937948 | 0.6374722838137472 |
| ft-v2 | 902 | 0.004739472625955076 | 0.009785202863961813 | 0.8592017738359202 |
Artifacts:
outputs/limbu-ft-v2-run-2026-07-06/score-champion-frozen-test/metrics.jsonoutputs/limbu-ft-v2-run-2026-07-06/score-ftv2-frozen-test/metrics.json
Derived: codepoint CER relative reduction 65.625%; grapheme CER relative reduction 65.546%; exact-line accuracy gain 22.173 percentage points.
Boundary: the split is frozen and date-separated for this diagnostic, but the labels are unreviewed silver. The previous recognizer's older training history may include some of these dates; ft-v2's dates are disjoint.
D. Non-oracle page diagnostic on the line-test dates
The page subset is deterministically selected by
paper/build_heldout_page_subset.py from dates 2082-03-24, 2082-08-12, and
2082-12-25. It reuses already-produced document JSON and does not rerun OCR.
Verified ft-v2 values: 3 pages; 907 reference lines; 904 predictions; 877 matches; detection F1 0.9685256764218665; page codepoint CER 0.025746508829373408; page grapheme CER 0.03360869073214892; matched-line codepoint CER 0.01012004871542075; matched-line exact accuracy 0.7616875712656784.
Verified previous-system values: 3 pages; 907 reference lines; 903 predictions; 877 matches; detection F1 0.969060773480663; page codepoint CER 0.03957126465731956; page grapheme CER 0.06280411904492475; matched-line codepoint CER 0.025691584991010846; matched-line exact accuracy 0.45952109464082097.
Artifacts:
paper/evidence/limbu-page-test-dates/eval.jsonlpaper/evidence/limbu-previous-heldout-page-score/metrics.jsonpaper/evidence/limbu-ftv2-heldout-page-score/metrics.json- each score directory's
page-eval-run.json,claim-safety.md, line alignment, text pairs, and review files.
Boundary: date-disjoint, non-oracle, and real-print, but only three pages;
the references are silver, the subset is not a claim pack, and the scorer
correctly reports claim_safe=false.
E. Sixteen-page engineering gate
Verified ft-v2 values: 16 pages; page codepoint CER 0.027222632149080803; page grapheme CER 0.034681104813771124; detection F1 0.964986138207208; matched-line codepoint CER 0.0089806695569971; matched-line exact accuracy 0.7927218557139817.
Verified previous-system values under shared column-major scoring: page codepoint CER 0.04328827626628151; page grapheme CER 0.06713527319549546; detection F1 0.9656872816930347; matched-line codepoint CER 0.026572279207232267; matched-line exact accuracy 0.4340425531914894.
Artifacts:
outputs/limbu-ftv2-page-gate-2026-07-06/score-v2/metrics.jsonoutputs/limbu-column-major-2026-07-05-score-v2/metrics.json
Boundary: 13 of the 16 dates participate in ft-v2 training or development. This is a deployment regression gate, not independent evaluation.
F. Page-linearization correction
Verified values: historical native serialization has codepoint CER 0.4134103646953546; shared column-major serialization has codepoint CER 0.04328827626628151. Both use 4,839 predictions and 4,700 matched lines from the earlier recognizer.
Artifacts:
outputs/limbu-native-regression-check-2026-07-05/metrics.jsonoutputs/limbu-column-major-2026-07-05-score-v2/metrics.json
Boundary: this is a metric-definition correction. The values are not two model runs and cannot be presented as model improvement.
G. Source-pixel and representation verification
paper/LIMBU_SOURCE_VERIFICATION.md records exact paths, hashes, commands, and
bounded manual observations for:
- a real six-column Gorkhapatra Sirijonga page used in the page test; and
- CDC mathematics PDF page 8, whose pixels are visually Sirijonga while its curriculum reference contains Devanagari Unicode.
The CDC finding excludes the historical 31-crop result from the verified Devanagari OCR tables. Its raw score directory was lost; only the packaged model and handoff ledger survive. The result is properly described as cross-script transcription evidence.
H. Runtime artifact identity
Artifacts:
spaces/limbu-ocr/models/limbu-sirijonga-ft-v2/model-card.jsonspaces/limbu-ocr/models/limbu-sirijonga-ft-v2/artifacts/spaces/limbu-ocr/app.py
The model card binds the required inference JSON, parameters, YAML, and dictionary by SHA-256. The application points to ft-v2 and deva-v2 and exposes line, Paddle page, and Paddle book paths. This verifies the local payload, not the current network status of a hosted service.
I. Figures
paper/build_figures.py reads the recorded artifacts in this ledger and writes
PDF, SVG, and 300-DPI PNG versions of:
research-program;system-architecture;limbu-source-evidence;limbu-recognizer-results;limbu-page-results;limbu-linearization-diagnostic; andlanguage-pack-status.
No publisher or curriculum pixels are embedded. The first two figures frame the delivered platform, human-review loop, scientific completion gates, and longer research horizon; the remaining figures are generated from recorded evidence.
J. Local application packaging repair
The application uses models/limbu-sirijonga-ft-v2/artifacts and
models/deva-v2/artifacts. During the paper audit, its static artifact preflight
was found to expect the obsolete Devanagari filename nep_eng_dict.txt even
though the current model card binds devanagari-v2.txt. The check and Space
documentation were corrected. paper/verify_claims.py now requires all eight
model files named by the two current bundles to exist and requires the corrected
dictionary path in app.py.
K. AI-detector receipt
The old detector receipt was bound to a superseded manuscript with the wrong
scope and is invalid for this paper. paper/AI_INTEGRITY_CHECK.md remains red
until the corrected PDF is built and checked on the requested remote host.
L. Newar-Devanagari normalized pack splits
Verified values: train/dev/test contain 4,096/1,528/1,139 rows.
Artifact:
outputs/newar-pack-v1-normalized-2026-07-15/stats.json
(train.total_lines, dev.total_lines, and test.total_lines).
Boundary: these are inventory counts for the recovered normalized pack, not evidence that its labels, provenance, or split isolation are claim-grade. The independent admission decision in section P applies to all three splits.
M. Newar ft-v1 line diagnostic
Verified values: 1,139 lines; codepoint CER 0.04655477596980231; grapheme CER 0.09569411931616656; exact-line accuracy 0.24495171202809482.
Artifact:
outputs/newar-ft-v1-verify-omac-2026-07-07/metrics.json, under metrics.
Boundary: this is a diagnostic receipt, not claim-grade CER. The receipt
records a host-local label path but no label-file hash, and the audited source
pack is NOT_ADMITTED; the number therefore cannot override the boundary in
section P.
N. deva-v2 Newar line baseline
Verified values: 1,139 lines; codepoint CER 0.062229425689364695; grapheme CER 0.13119812726111937; exact-line accuracy 0.15803336259877085. The verifier also requires its recorded codepoint and grapheme reference lengths to equal the ft-v1 receipt's reference lengths.
Artifact:
outputs/deva-v2-reexport-2026-07-06/score-newar-test/metrics.json, under
metrics.
Boundary: this is an internal diagnostic baseline. The two metric receipts name different absolute label paths and neither supplies a label hash, so their matching row and reference-length fields do not establish a cryptographically bound independent benchmark.
O. Newar non-oracle page diagnostic
Verified values: 19 pages; page codepoint CER 0.19918426853185692; page grapheme CER 0.27885891823188813; line-detection F1 0.911629205640095; matched-line codepoint CER 0.09052417830487239.
Artifact:
outputs/newar-page-eval-v2-2026-07-05/newar-ft-v1/score-v2/metrics.json.
Boundary: the same artifact records zero claim-eligible rows and classifies
all 19 references as nayanepal_font_mapping_silver_unreviewed. These are page
diagnostics against unreviewed silver, not claim-grade accuracy. Its leakage
block also records checked=false with no training manifests supplied.
P. Newar claim-boundary and leakage audit
Verified values: train versus dev+test contains 18 exact-overlap left rows and 575 left rows sharing a substring of at least 16 codepoints, including the exact overlaps.
Artifact:
outputs/newar-devanagari-normalized-claim-boundary-audit-2026-07-15-r12/claim-boundary-audit.json.
Boundary: the audit reports leakage status FAIL, overall status
NOT_ADMITTED, claim_grade=false, train_permitted=false,
eval_permitted=false, and cer_permitted=false. It also records missing
row-level source provenance, Gorkhapatra ledger status RISK, absent qualified
native review, and unreconstructable date/edition isolation. Sections L--O must
not be promoted from diagnostics to publication claims.
Q. Videha Tirhuta real-print silver
Verified value: 10,696 kept lines; the crop archive also records 10,696 file entries and verified archive/label parity.
Artifact:
outputs/videha-tirhuta-silver-v2-2026-07-17/summary.json, under totals,
artifacts.crop_archive, and verification.
Boundary: this is mechanically converted silver substrate. The count and archive parity do not make the labels native-reviewed ground truth or establish recognizer accuracy.
R. Tirhuta fine-tune payload
Verified values: status READY; 9,765 real training rows; 844 real
development rows. The progress receipt also says training was not launched.
Artifact:
outputs/tirhuta-v2-ft-payload-2026-07-17/PROGRESS.md.
Boundary: READY is a payload-readiness state under the recorded Videha
license caveat, not a trained-model or evaluation result.
S. Tirhuta synthetic render archive
Verified value: 48,960 PNG members. The verification receipt has
passed=true, an equal expected count, and an empty issue list.
Artifact:
outputs/tirhuta-synth-v1-prep-2026-07-17/audits/render-archive-verification.json.
Boundary: this is a synthetic pretraining inventory and archive-integrity result. It says nothing about real-print recognition quality.
T. Ol Chiki real evaluation pack v3
Verified values: 895 candidate lines; 123 excluded blank renders; 772 clean evaluation lines. The verifier requires the clean and excluded counts to reconcile exactly to the candidate total.
Artifacts:
data/olchiki-real-eval-pack-v3/summary.json; anddocs/DATA_LICENSING.mdfor the source-specific rights boundary.
Boundary: the pack summary is an inventory artifact and contains no claim-eligibility field. It combines a public-marked issue with the September-2023 issue whose license is recorded as unspecified, so the full pack remains diagnostic rather than claim-grade.
U. Ol Chiki v2 line diagnostic
Verified values: 772 lines; codepoint CER 0.08263840048226666; exact-line accuracy 0.2033678756476684.
Artifact:
outputs/olchiki-v2-run-2026-07-07/score-real-print-v3pack/metrics.json, under
metrics.
Boundary: this is explicitly diagnostic-only. The mixed-rights evaluation boundary in section T and the recorded pending crop, label, punctuation, leakage, and asset-independence review prevent a claim-grade interpretation.
V. Ol Chiki v3 fine-tune payload
Verified values: 5,307 total images: 242 real training, 5,000 synthetic training, and 65 real development images.
Artifact:
outputs/olchiki-v3-ft-payload-2026-07-17/olchiki-v3-ft-payload.tgz, exact
member metadata/assembly-manifest.json, under payload_counts.
Boundary: the manifest records training_launched=false; these are staged
payload counts, not v3 model or accuracy evidence. Its license boundary keeps
the unspecified-license September issue evaluation-only.
W. Sikkim Herald silver-pack yields
Verified values: 9,743 clean Sunuwar lines; 3,699 clean Kirat Rai lines; zero clean Lepcha lines.
Artifacts:
outputs/sikkim-silver-packs-2026-07-16/sunuwar/summary.json;outputs/sikkim-silver-packs-2026-07-16/kiratrai/summary.json; andoutputs/sikkim-silver-packs-2026-07-16/lepcha/summary.json.
Boundary: the summaries' readiness verdict is mechanical converter/archive readiness. The rows are silver, not native-reviewed labels; the separate legal gate in section X still controls training.
X. Sunuwar and Kirat Rai fine-tune payloads
Verified values: Sunuwar has 8,422 real training and 514 real development rows; Kirat Rai has 2,865 real training and 193 real development rows.
Artifacts:
outputs/sunuwar-v1-ft-payload-2026-07-17/PROGRESS.md; andoutputs/kiratrai-v1-ft-payload-2026-07-17/PROGRESS.md.
Boundary: both receipts report BLOCKED-PENDING-LEGAL and
Training launched: false. These payloads are legal-blocked staging artifacts,
not training or evaluation claims.
Y. Lepcha real-evaluation anchor
Verified value: 132 line crops, stored as lines=132.
Artifact: data/lepcha-real-eval-pack/README.json.
Boundary: the artifact encodes claim_evidence_eligible=false inside its
label_status string rather than as a Boolean field. The labels are
machine-converted and human verification is pending.
Z. Lepcha zero-yield root cause
Verified artifact:
outputs/lepcha-zero-yield-diagnosis-2026-07-17/REPORT.md exists and records
that the article bodies are rasterized as image XObjects.
Boundary: this is a source-format diagnosis. It supports a raster/OCR or transcription route, not a converter yield, label, model, or accuracy claim.
AA. Newa CTC-T200 preparation
Verified values: exact and long-substring filtering removed 430 unique
training rows and left 1,961; the patched train and evaluation paths both emit
T=200.
Artifact: outputs/newa-ctc200-prep-2026-07-15/PROGRESS.md.
Boundary: the receipt states that no GPU launch is authorized. The sealed Cambridge evaluation labels still require native pixel-level review, so this is recipe and payload evidence rather than a trained-model result.
AB. Gurung Khema mechanical eligibility audit
Verified values: 759 mechanical candidates; 749 mechanically eligible; 10 mechanically rejected.
Artifact:
outputs/sikkim-herald-gurung-khema-r13-adversarial-bundle-audit-2026-07-15-r13/adversarial-audit.json.
Boundary: both top-level and result status are
PASS_READY_NOT_ADMITTED; admission is NOT_ADMITTED, claim_grade=false,
and labels_present=0. “Eligible” means mechanically convertible, not reviewed
or admitted training data.
AC. Gorkhapatra monthly backfill
Verified values: the exact glob
outputs/gorkhapatra-backfill-*/silver-line-packs/silver-pack-summary.json
matches 17 files, whose total_kept_lines sum to 515,586.
Artifacts: those 17 monthly silver-pack-summary.json files, spanning the
surviving March-2024 through July-2025 backfill directories.
Boundary: this is converted silver inventory, not language-verified or
claim-grade evidence. The summaries have no claim-status field, and
Gorkhapatra training, numerical-publication, and redistribution rights remain
RISK in docs/DATA_LICENSING.md.
AD. Per-language silver rebuild receipts
Verified values: the last line of batch 1 records kept=122882; the last
line of batch 2 records kept=184089; their sum is 306,971.
Artifacts:
outputs/silver-lines-rebuild-b1-2026-07-05.log; andoutputs/silver-lines-rebuild-b2-2026-07-05.log.
Boundary: these terse completion logs establish generated silver-row counts only. They do not establish per-language native review, rights, split isolation, or accuracy.
AE. Nepal language-pipeline registry
Verified value: languages is a JSON array of length 50.
Artifact: configs/nepal-language-pack-pipeline-v1.json.
Boundary: a 50-language configuration describes registry breadth, not 50 trained, evaluated, rights-cleared, or claim-grade language systems.
AF. Manuscript value and scope binding
paper/verify_claims.py requires paper/DRAFT.md to contain the formatted
values backed by sections L--AE, including the 50-language, 515,586, 306,971,
T=200, payload, diagnostic, and silver-pack tokens. It rejects the known-wrong
tokens 0.0441, 299,611, 0.0961, and 5,330, while retaining the existing
Limbu boundary phrases cross-script transcription, operational research platform, and review burden.
Boundary: this is a manuscript-consistency gate, not independent evidence. Every number remains subject to its source artifact and boundary above.
AG. Language-pack status figure
paper/build_figures.py writes non-empty PDF, SVG, and PNG versions of
paper/figures/language-pack-status. paper/verify_claims.py requires all
three alongside the six existing figure stems.
Boundary: file presence and nonzero size verify build output only. The underlying values and status labels remain gated by their source checks and do not turn silver or diagnostic substrate into reviewed ground truth.
Publication gate
Before the diagnostic values become submission claims, the project needs:
- a frozen, native-reviewed Sirijonga reference set;
- at least 30 independent page units across source groups;
- an explicit near-duplicate and train/evaluation leakage audit;
- external baselines on identical inputs;
- pack-wide resolution of the Sikkim Herald legal gate, including source-host
eligibility, third-party screening, required attribution, and clearance of
the
BLOCKED-PENDING-LEGALpayloads; - qualified native pixel-label review, frozen separately for every script whose rows or metrics would appear as a claim;
- resolution of the remaining Gorkhapatra redistribution question. The
numerical-publication question was resolved by a recorded owner
determination (2026-07-22,
docs/DATA_LICENSING.md): derived metrics and counts may be published; pixel and crop redistribution remains withheld pending a formal license statement; and - a separate benchmark of visibly Devanagari-written Limbu pages.