"""Escriba-OCR-1B: transcribe a handwritten essay page (diplomatic transcription). Recommended: vLLM (loads the checkpoint as-is). pip install "vllm>=0.34" pillow python inference_example.py page.png IMPORTANT: upscale the page so its longer side is 1200 px. The processor only downscales, and small scans (~550 px, like BRESSAY) give the model too few visual tokens. """ import base64 import io import sys from PIL import Image from vllm import LLM, SamplingParams MODEL = "tardellirs/Escriba-OCR-1B" LONG_SIDE = 1200 def to_data_url(path: str) -> str: im = Image.open(path).convert("RGB") f = LONG_SIDE / max(im.size) im = im.resize((round(im.width * f), round(im.height * f)), Image.LANCZOS) buf = io.BytesIO() im.save(buf, format="PNG") return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode() def main() -> None: llm = LLM(model=MODEL, max_model_len=8192, limit_mm_per_prompt={"image": 1}) params = SamplingParams(temperature=0.0, repetition_penalty=1.05, max_tokens=2048) # Empty prompt = transcription mode (image only), exactly as in training. messages = [[{"role": "user", "content": [{"type": "image_url", "image_url": {"url": to_data_url(p)}}]}] for p in sys.argv[1:]] for path, out in zip(sys.argv[1:], llm.chat(messages, params)): print(f"===== {path}\n{out.outputs[0].text}\n") if __name__ == "__main__": main()