trocr-vedic-cross-notation-rv-ms

The 15 OCR models of the main experiment in Transfer Between Accent Notations in Vedic OCR: Learning Visual Conventions from Parallel Text (PACLIC 2026).

Each model is yzk/trocr-large-printed-sanskrit-iso-preadapt fine-tuned (all parameters) on yzk/veda-ocr-rv-ms-noto-v2, which renders the same accented texts in two Vedic accent notations: the Rigveda (RV) convention and the Maitrāyaṇī Saṃhitā (MS) convention. Models differ only in the share of training lines shown in the RV notation versus the MS notation, and in the random seed. The output is accented ISO 15919 romanization.

Layout

One subfolder per condition: rv_{R}_ms_{M}/seed_{S}/, with R / M ∈ {100/0, 75/25, 50/50, 25/75, 0/100} and S ∈ {42, 43, 44}. Each subfolder holds the model, processor, tokenizer and generation config. run/ next to it holds that run's test predictions (rv_predictions.csv, ms_predictions.csv), metrics (results.json) and the training subset (subset_manifest.json, parallel_keys.txt).

from transformers import TrOCRProcessor, VisionEncoderDecoderModel

repo, sub = "yzk/trocr-vedic-cross-notation-rv-ms", "rv_50_ms_50/seed_42"
processor = TrOCRProcessor.from_pretrained(repo, subfolder=sub)
model = VisionEncoderDecoderModel.from_pretrained(repo, subfolder=sub)

Training

10,420 parallel training lines per model; 20 epochs; effective batch size 32; learning rate 2e-5 (linear, warmup ratio 0.1); weight decay 0.01; fp16. The checkpoint is selected on validation CER of the dominant notation (mean of both notations at 50/50). Decoding: beam 4, max length 256, no-repeat n-gram 3, length penalty 2.0. Code: https://github.com/Yuzki/vedic-ocr-cross-notation

Results

Test-set means over the three seeds (578 lines per notation). See the paper for uncertainty and the per-seed values in each run/results.json.

RV / MS training share	RV CER	MS CER	RV accent F1	MS accent F1
100 / 0	0.037	0.264	0.954	0.060
75 / 25	0.040	0.041	0.948	0.938
50 / 50	0.041	0.037	0.934	0.941
25 / 75	0.045	0.036	0.922	0.950
0 / 100	0.297	0.036	0.028	0.958

Limitations

Trained and tested on synthetic line images rendered in a single font (Noto Sans Devanagari). Performance on real scans of printed editions has not been evaluated in this release. The RV notation approximates printed RV editions and omits RV numerals; the MS grave accent is approximated with U+0952.

License

CC BY-NC-SA 4.0, inherited from the initialization; see the initialization's card for its training data and their licenses.

Citation

@inproceedings{TODO-paclic2026,
  title     = {Transfer Between Accent Notations in Vedic OCR: Learning Visual Conventions from Parallel Text},
  author    = {TODO},
  booktitle = {Proceedings of the 40th Pacific Asia Conference on Language, Information and Computation (PACLIC 2026)},
  year      = {2026}
}