kora-v0.1

📓 Notebook: kora_inference.ipynb (download). It loads the model once and OCRs images or PDF pages, showing each page next to its Markdown. It runs on Google Colab (upload it with File → Upload notebook and pick a GPU runtime) or locally (NVIDIA GPU, Apple GPU or CPU).

kora-v0.1 is an OCR model for Hindi and English document pages. It turns a page image into Markdown: headings, lists and bold text as Markdown, tables as HTML, math as LaTeX, pictures as ![image](image_N.png) placeholders. Headers, footers and page numbers are kept, and digits are written in the script they're printed in.

Quick start

import torch
from PIL import Image
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "shubhdotai/kora-v0.1"
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.float32 if device == "cpu" else torch.bfloat16).to(device).eval()

img = Image.open("page.png").convert("RGB")
img.thumbnail((1540, 1540))                                  # longest edge at most 1,540 px
prompt = processor.apply_chat_template([{"role": "user", "content": [{"type": "image"}]}],
                                       add_generation_prompt=True, tokenize=False, enable_thinking=False)
inputs = processor(text=[prompt], images=[img], return_tensors="pt").to(device)
end = processor.tokenizer.convert_tokens_to_ids("<|im_end|>")
out = model.generate(**inputs, max_new_tokens=6144, do_sample=False, eos_token_id=end)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Requires transformers 5.5.4 or newer.

Tips

• Input: one page per image, with no text prompt. Render PDFs at 200 DPI and keep the longest edge at 1,540 px or less (the notebook does both).
• Repetition: if a page's output starts repeating and runs to the length cap, run it again with repetition_penalty=1.05.
• Precision: bf16 on recent GPUs and Apple silicon; use fp32 on a T4 or a CPU.

License

Apache-2.0. Based on lightonai/LightOnOCR-3-0.8B (Apache-2.0).