OCeanpaRk-small-v1

Vietnamese text-line recognition by ANZ Innovation, based on the PP-OCRv6 medium recognition architecture from PaddleOCR. The name “small” is the release name; this is not the official PP-OCRv6 tiny architecture.

The default weights now include handwriting adaptation. This release uses the best mixed-validation checkpoint saved during the latest 10-epoch adaptation, initialized from our Vietnamese epoch-12 checkpoint. It recognizes both printed text and handwriting. The previous printed-document epoch-30 model is preserved under epoch30/, and epoch 12 remains under epoch12/.

On the official 1,000-example VietHTR-Line test split, the new dynamic checkpoint achieves 5.76% CER / 18.08% WER, compared with 15.53% / 41.61% before handwriting adaptation. This is a 62.9% relative CER reduction. On the same internal 1,900-crop document test used previously, it achieves 2.51% CER / 4.24% WER. Document accuracy regresses slightly versus epoch 12 and is lower than the older epoch-30 release. This release trades a small document-recognition regression for substantially better handwriting recognition.

Handwriting comparison

Percentages; lower CER/WER is better. Both checkpoints use identical samples, preprocessing and dictionary. “Before adaptation” is our epoch-12 checkpoint, not the original tieubaoca model.

The original tieubaoca and official PP-OCRv6 baselines have not been evaluated on this handwriting test in this release; no handwriting score is claimed for either baseline.

Same-test comparison with HTR-ConvText

We also evaluated the released Vietnamese checkpoint from DAIR-Group/HTR-ConvText on the exact same 1,000 VietHTR-Line test crops and reference strings. On this target-domain test, OCeanpaRk-small-v1 reduces CER by 84.8% and WER by 75.2% relative to that checkpoint.

OCeanpaRk has fewer character errors on 988/1,000 lines; HTR-ConvText has fewer on 7 lines, and 5 tie. Counts and reproducibility metadata are in metrics/htr-convtext-vietHTR-test.json.

Comparison scope: HTR uses its released VNOnDB-trained checkpoint, while OCeanpaRk was adapted using the VietHTR training split. This measures the released checkpoints' performance on our target domain; it is not a comparison of architectures trained with identical data, nor evidence of superiority on VNOnDB or handwriting in general. HTR's author-reported 3.45% CER is on VNOnDB, a different benchmark. Upstream HTR training-source overlap with VietHTR has not been audited.

HTR inference uses state_dict_ema, strict loading with all keys/shapes matched, its original VNOnDB vocabulary, and the upstream preprocessing functions: grayscale, height 64, capped width 512 and white padding. Decoding is FP32 greedy CTC. Metrics use the same NFC micro CER and whitespace WER as OCeanpaRk, without punctuation/case normalization; all 1,000 examples are retained. OCeanpaRk uses the previously verified dynamic-checkpoint predictions on those identical samples. HTR revision: 8ed2ad5390d67f44828c528d9254cc72deeab423.

HTR's vocabulary does not cover all reference characters in 79 samples. Even on the 921 fully supported references, HTR has 37.69% CER, versus 5.48% for OCeanpaRk. This diagnostic subset does not replace the full-test score. Native preprocessing differs between models, and 929 HTR inputs are compressed horizontally to fit width 512; these differences may affect results.

Comparison on the original document test

All models below were evaluated on the same 1,900 reference crops. Earlier baseline/epoch-30 scores are preserved from the previous evaluation; the new model was evaluated after adaptation. This split includes documents, invoices, tables and mixed printed/handwritten forms; it is not a pure printed-text benchmark.

The official baseline uses its original dictionary. It cannot represent at least one reference character in 1,423/1,900 crops, so its gap reflects vocabulary coverage as well as adaptation. This is not an architecture-only comparison. Private benchmark images/references are not distributed, so these document scores cannot be independently reproduced from this repository alone.

On the 3,023-crop document validation split, the new checkpoint achieves 3.83% CER / 5.92% WER, compared with 3.85% / 5.62% for epoch 12. Choose epoch30/ if your workload primarily resembles the older document benchmark; evaluate the new default for handwriting or mixed workloads.

Data, checkpoint selection and evaluation

• Handwriting data: 5CD-AI/VietHTR-Line, revision a43e22a5076f4b62ecfcd00ef5a56058f64acd1e, accessed with authorization. The adaptation sampled 50,000 handwriting lines and 50,000 selected old-domain lines per epoch. No dataset images, annotations or private documents are uploaded here.
• Handwriting validation was split from the dataset's training split by source image ID, with rotated variants grouped where identifiable. Writer IDs were unavailable; writer-independent generalization is not established. The official 1,000-example test split remained held out. One detected train/test source-ID overlap was removed from training.
• Training/handwriting-validation filtering retained supported dictionary characters and labels up to 98 characters, without truncation. The official handwriting test retained all 1,000 examples, including 15 references with unsupported characters and 13 references over 98 characters.
• The default is best mixed-validation exact-line accuracy during epoch 10, not the final epoch-10 snapshot and not a checkpoint selected by test CER/WER. Weight SHA-256: 66a4deb98c63d8e4424ce1599f3239138be153fdece6eebb44d40ca11133de4c.
• The original document benchmark excluded 9/1,909 test and 11/3,034 validation crops with unsupported reference characters before the earlier comparison. All models receive the same remaining 1,900/3,023 crops.
• Recognition uses fixed supplied text crops, BGR preprocessing, aspect-ratio resize to height 48 and padding to width 1280 (overwide lines are compressed). Text is NFC-normalized without accent stripping or case folding.
• CER is total character edit distance / total reference characters. WER uses whitespace-separated tokens. Exact accuracy requires the whole line to match. Diacritic accuracy counts correctly recovered diacritic-bearing reference characters under edit alignment.
• Dataset review found upside-down images, multi-line crops with single-line labels, inconsistent transcriptions and mixed print/handwriting in both nominal splits. Reported numbers retain the original evaluation labels; they are not manually cleaned benchmark scores.
• Metrics describe recognition on line crops. They exclude detection, reading order, page layout, PDF rendering and table reconstruction. This model does not directly produce structured HTML/Markdown.
• Broad multilingual robustness and writer-independent handwriting performance have not been established. Confidence is not a calibrated probability of correctness.

Aggregate metrics and evaluation hashes are in metrics/. No per-example predictions or dataset contents are published.

Export verification

The comparison tables above report dynamic-checkpoint FP32 evaluation. The new exported Paddle Inference artifacts were also evaluated on all 1,000 handwriting and 1,900 document test crops:

Exported versus dynamic inference differs on 10 predictions in each split, with small aggregate metric differences. The checks used PaddlePaddle 3.3.1, native Paddle Inference on GPU and width 1280. See export_verification.json. The earlier epoch-30 export verification is preserved in epoch30/export_verification.json.

Usage

Install PaddlePaddle for your device using its official installation instructions, then install paddleocr and huggingface_hub. The following uses PaddleOCR's high-level recognition API; this release's export checks used native Paddle Inference.

from huggingface_hub import snapshot_download
from paddleocr import TextRecognition

model_dir = snapshot_download(
    repo_id="ANZ-Innovation/OCeanpaRk-small-v1",
    allow_patterns=["inference.*", "ppocr_keys.txt"],
)
model = TextRecognition(
    model_name="PP-OCRv6_medium_rec",
    model_dir=model_dir,
    device="gpu:0",  # or "cpu"
    input_shape=(3, 48, 1280),
)
for result in model.predict(input="cropped_line.png", batch_size=1):
    result.print()

For an earlier model, download epoch12/* or epoch30/* and pass that subdirectory as model_dir. Full images/PDF pages require a separate detector to produce line crops. See the PaddleOCR recognition documentation.

Files

Training code, training configurations, optimization scripts, optimizer state and dataset contents are not distributed. inference.yml is the public inference preprocessing/decoding configuration, not the training recipe.

Attribution and licensing

Derived from PaddleOCR and the tieubaoca Vietnamese checkpoint; handwriting adaptation used VietHTR-Line. PaddleOCR source is Apache-2.0. The upstream tieubaoca checkpoint and VietHTR-Line model/dataset cards do not declare a license in their metadata at the time of this release. This card does not grant new rights to inherited weights or dataset content. Consult the upstream terms for your intended use.