BDRC/tibetan-ocr

Tibetan OCR — Yigdzin 1

<!-- Model card for BDRC/tibetan-ocr ("Yigdzin 1"). Weights: checkpoint eliev8coarsegrow26ep2 (becmixedeliev8 mix, evaluation benchmark pages held out of training via imagesexcludefromtrain.csv). -->

A vision-first OCR model for Tibetan pecha — modern publications, woodblock prints, and manuscripts (uchen and u-med), including pecha layout and orthographic shorthands. Built by the Buddhist Digital Resource Center (BDRC) and trained on the bec_mixed_elie_v8 mix of curated real + synthetic pages, with the evaluation-benchmark pages held out of training.

• —Repository: BDRC/tibetan-ocr (display name Yigdzin 1)
• —Base model: `PaddlePaddle/PaddleOCR-VL-1.6` (Apache-2.0)
• —Live demo: ocr.bdrc.io · Desktop app: buda-base/tibetan-ocr-app

TL;DR

On a 1,070-page hand-transcribed benchmark (production serving: vLLM + sequential regime + DRY guard + temperature-retry), median page CER is ~1.4% and clean-page (non-catastrophic) mean CER is ~4.0%, with 0 hard repetition loops. The benchmark pages are held out of training. It is a specialized OCR model, not a general chat VLM.

Model details

Intended use

• —OCR of Tibetan pecha pages: uchen and u-med (dbu med) scripts across modern print, woodblock, and manuscript sources.
• —Batch OCR at corpus scale (BDRC is applying it to ~25M scanned pages).

Out of scope: general vision-language chat / VQA; non-Tibetan scripts; layout analysis of illustrations, tables, or diagrams; line/region detection (this model transcribes page or line crops it is given).

How to use

Prompt (used at training and inference), rendered through the repo's chat_template.jinja:

Extract all Tibetan text. Preserve line breaks.

This checkpoint uses the native PaddleOCR-VL architecture (model_type: paddleocr_vl, PaddleOCRVLForConditionalGeneration) — no `trust_remote_code` needed, but it requires transformers ≥ 5.15 (native paddleocr_vl support) or vLLM ≥ 0.26. Decode greedy (temperature=0).

pip install "transformers>=5.15" torch torchvision accelerate pillow
# torchvision is required by the PaddleOCR-VL image processor.

⚠ Position regime — required for correct output. This model was trained in the sequential image-token M-RoPE regime (1-D positions), not the default "grid" regime. If the processor emits a non-zero mm_token_type_ids, structured pages loop badly (some pages regress from ~0.10 to ~0.62 CER). Serve it sequential: - HF: zero the mask before the forward — inputs["mm_token_type_ids"] = torch.zeros_like(inputs["input_ids"]). - vLLM: install the vllm_paddleocr_seqpos plugin and set OCR_VLLM_IMAGE_TOKEN_POSITIONS=sequential.

HF transformers (reference / accuracy baseline)

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "BDRC/tibetan-ocr"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="cuda")

image = Image.open("page.jpg").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Extract all Tibetan text. Preserve line breaks."}]}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=[prompt], images=[image], return_tensors="pt").to("cuda")
inputs["mm_token_type_ids"] = torch.zeros_like(inputs["input_ids"])  # sequential regime (required)
out = model.generate(**inputs, do_sample=False, max_new_tokens=4096)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

To add the DRY anti-loop guard under HF too (optional; dry_logits_processor.py ships in this repo):

from transformers import LogitsProcessorList
from dry_logits_processor import make_hf_dry_processor

prompt_len = inputs["input_ids"].shape[1]
out = model.generate(**inputs, do_sample=False, max_new_tokens=4096,
    logits_processor=LogitsProcessorList([make_hf_dry_processor(
        prompt_len=prompt_len, multiplier=0.8, base=1.75, allowed_length=12)]))

vLLM (production)

Serve with vLLM ≥ 0.26. Two things matter for production quality:

1. Sequential image-token positions (required). Stock vLLM serves PaddleOCR-VL in the grid regime and never reads mm_token_type_ids; this model needs the sequential regime. A tiny vLLM plugin ships in this repo under vllm_paddleocr_seqpos/:

pip install "git+https://huggingface.co/BDRC/tibetan-ocr#subdirectory=vllm_paddleocr_seqpos"
export OCR_VLLM_IMAGE_TOKEN_POSITIONS=sequential   # 'grid' / unset = no-op

Without it, structured (book/list) pages skip or merge lines and CER regresses badly.

2. DRY anti-loop guard + temperature retry (recommended). Greedy decoding runs away into repetition on a small fraction of pages. The surgical fix is the DRY ("Don't Repeat Yourself") penalty: it ties an n-gram ban on corpus CER, zeroes hard loops, and — unlike repetition_penalty / no_repeat_ngram_size — barely touches clean pages or legitimate repetitive scripture (mantras, litanies). It ships here as the self-contained dry_logits_processor.py (torch-only):

from vllm import LLM, SamplingParams
from dry_logits_processor import DRYLogitsProcessor   # ships in this repo

llm = LLM(model="BDRC/tibetan-ocr", logits_processors=[DRYLogitsProcessor])  # + seqpos env above
params = SamplingParams(temperature=0, max_tokens=4096, extra_args={
    "dry_multiplier": 0.8, "dry_base": 1.75, "dry_allowed_length": 12,  # production config, no breakers
})

For the pages that still loop under greedy+DRY, re-decode at temperature and keep the cleanest sample. The strong signal is how often DRY fired: the measured knee is ≥ 100 fires — re-decode only those pages at temperature=0.4, n=3 (DRY still on) and pick the sample with the lowest leftover repetition. On this benchmark that is ~2% of pages and clears the residual hard loops. dry_logits_processor.py writes per-request fire counts (pass dry_stats_id / dry_stats_path in extra_args, read them back with load_dry_stats_dir); deploy/fast_inference/bench.py in the training repo is the reference implementation of the full greedy → fire-count gate → temperature-retry loop.

Do not reach for aggressive repetition_penalty / no_repeat_ngram_size: they "correct away" legitimate repeated scripture. DRY with the config above is the tested, surgical choice.

Training data

Curated mix (bec_mixed_elie_v8), assembled from BDRC alignment collections + BoCorpus-rendered synthetic pages, with an easy→hard curriculum, u-med over-weighting, and repetition-aware filtering. The evaluation-benchmark pages are held out of training (images_exclude_from_train.csv). Released component datasets:

• —Real transcriptions: ALL-BDRC (ACIP Sungbum), PalriParkhang, Berkeley, MonlamAI-transcriptions, MonlamAI-handwritten, Stok, TibSchol (gated).
• —Synthetic: Tibetan OCR synthetic v5.

Evaluation

Evaluated on the BDRC hand-transcribed benchmark 20260315 (1,070 pages) in the production configuration (vLLM + vllm_paddleocr_seqpos sequential regime + DRY guard mult=0.8 base=1.75 allowed_length=12 + temperature-retry temp=0.4 n=3 on pages with ≥100 DRY fires). Metrics from the benchmark's compute_cer.py (botok normalization, whitespace stripping, tsheg folding, placeholder removal); pages with CER > 50% are counted as catastrophic and reported separately.

Full method + the multi-system leaderboard: [BDRC Tibetan OCR benchmark](https://huggingface.co/datasets/BDRC/tibetan-ocr-benchmark) and the [leaderboard](https://huggingface.co/spaces/BDRC/tibetan-ocr-leaderboard).

Limitations & recommendations

• —Residual repetition loops on a small fraction of pages. Production uses the DRY guard + temperature retry (see How to use), which clears them (0 hard loops on the benchmark) while leaving legitimate repetitive scripture (mantras) intact; keep any anti-loop mitigation this gentle.
• —Line-break joins on some dense (6-line) woodblock pecha.
• —Struggles on illustrations, non-Tibetan text, and illegible/damaged pages.

Citation

A comprehensive article is forthcoming (Springer Language Resources and Evaluation). Until then, please cite the model and BDRC:

@misc{bdrc_tibetan_ocr_2026,
  title  = {Tibetan OCR (Yigdzin 1)},
  author = {Roux, Elie and Werner, Eric},
  year   = {2026},
  howpublished = {Buddhist Digital Resource Center, Hugging Face},
  note   = {https://huggingface.co/BDRC/tibetan-ocr}
}

Acknowledgements

Built on PaddleOCR-VL (Apache-2.0). Funded by the Khyentse Foundation. Thanks to the BDRC, Dharmaduta, and collaborating transcription teams.