Please add a minimal reproducible example for fine-tuning DeepSeek-OCR on **Urdu** (dataset format, tokenizer, PEFT/LoRA example)

Hi — I want to fine-tune `deepseek-ai/DeepSeek-OCR` for **Urdu (Arabic script, RTL)** but the repo does not include a clear example. Could the authors please add (or point me to) a minimal reproducible example showing exactly what is required?

Specifically, please provide the following items (or a short example repository/PR that contains them):

1. **Exact dataset schema** the training scripts expect (JSONL/TFR/`datasets`-compatible). Include one tiny example file (3–10 samples) in the repo.
2. **Data loader example** or the exact script/`datasets` call to load image→text pairs used by training.
3. **Tokenizer behavior**: clarify whether targets should be plain UTF-8 text or use a special/compressed token format. If plain text, please show an example tokenization of an Urdu sample.
4. **Training examples** (one-line commands or scripts):

* Full fine-tune command (script name & exact args to reproduce).
* PEFT/LoRA (or QLoRA) command example (script name & args to reproduce).
5. **Evaluation example** showing how to compute CER (and optionally WER) including the exact normalization steps applied before scoring and an example `eval.py` or snippet.
6. **Decoding details** (any special control tokens, beam settings, postprocessing required for Urdu).
7. **One small example dataset** included under `examples/` (images + JSONL) that can be used to verify the training/eval pipeline works out of the box.
8. **Short note on RTL/Urdu gotchas** the authors consider important (e.g., ZWNJ handling, numerals, ligature/font issues) — one paragraph or checklist is enough.

A short example or a link to an existing example in the repo would be ideal. Thanks — I’m ready to test as soon as an example is added.