Hebrew text: model outputs a hallucinated token while Latin text on the same image is read correctly — is Hebrew supported?

Summary

On Hebrew-language documents the model produces no usable output: dense full pages return empty results at every size preset, and on a large-glyph crop it reads the Latin text correctly while answering the Hebrew text on the same image with an invented, niqqud-decorated token repeated several times. This looks like Hebrew (Hebr script) is outside the training distribution. Related: #3 (unanswered language-support question).

Environment

• baidu/Unlimited-OCR via transformers==4.55.0, trust_remote_code=True, torch_dtype=torch.bfloat16
• NVIDIA L4 (24 GB), CUDA build of torch 2.13, Linux (GCP Deep Learning VM)
• model.infer(tokenizer, prompt="<image>\nFree OCR.", image_file=..., output_path=..., save_results=True)

What we tested (controlled)

Input: a scanned Hebrew business invoice (A4, mixed Hebrew/Latin header), rendered at 200 DPI (1654×2339). Each run in a fresh process.

run	config	result
full page	base_size=1024, image_size=640, crop_mode=True (default/gundam)	empty output
full page	base_size=1024, image_size=1024, crop_mode=False	empty output (3–4 s)
full page	base_size=1280, image_size=1280, crop_mode=False	empty output (3–4 s)
top-30% header crop (large glyphs)	base_size=1280, image_size=1280, crop_mode=False	Latin tokens read correctly (Master-Coach, the invoice number 10707), but ALL Hebrew on the crop answered as תְבָרִי לַלְעֹן repeated ×3 — a token that appears nowhere in the image

The Latin success in the same inference rules out an image-quality or harness problem. An earlier full-page run with the <|grounding|> prompt degenerated into a repeated <|det|>text [102, 844, 190, 844]<|/det|>[Non-Text] loop until max_length.

Reproduction should work with any Hebrew document image (e.g. render a paragraph of Hebrew text plus one English word at ~1200 px height and compare which of the two comes back).

Questions

1. Is Hebrew (or RTL scripts generally) included in the training data / supported language set?
2. If not, could the README document the supported-language list? (#3 asks this too.)
3. If Hebrew is expected to work, is there a recommended config for RTL documents?