Xiaomi-OCR-0

INFERENCE MODES

One model, two parsing paths.

BEFORE YOU START

Use the matching checkpoint and environment.

If the Xiaomi-OCR Skill has already prepared the code, runtime, and model on your machine, you can skip this setup. Otherwise, start with the GitHub code and follow the installation guide: clone the code, choose and install SGLang or vLLM, then download the model checkpoint from Hugging Face. The inference examples below assume the local model service is already running.

• Use the repository setup instructions to install the code dependencies and your chosen inference backend; start an OpenAI-compatible endpoint at 127.0.0.1:8000 with model name SeerRay-Lab/Xiaomi-OCR-0.
• Download the Xiaomi-OCR-0 checkpoint from Hugging Face and configure your runtime to load its local model directory.
• For two-stage region parsing, also install PaddleX and download the PP-DocLayoutV3 layout weights as described in the repository guide. Full-page inference does not require this layout model.

Create images.jsonl with one image per line. Use absolute paths and unique filename stems, since output files are named after each input image.

{"images": ["/absolute/path/to/page-001.png"]}
{"images": ["/absolute/path/to/page-002.png"]}

RUN INFERENCE

Choose the command for your setup.

Install requirements.txt from the repository root, then use a fresh output directory for each configuration. Module invocation keeps the repository’s shared postprocessing imports available.

Full-page inference

python -m pipeline.e2e_img2md \
  --input-jsonl images.jsonl \
  --output_dir outputs/e2e \
  --served_model_name SeerRay-Lab/Xiaomi-OCR-0 \
  --port 8000 \
  --task document \
  --max_tokens 16384

Layout detection and region recognition

python -m pipeline.layout_detect_paddlex \
  --input_jsonl images.jsonl \
  --output_jsonl layout.jsonl \
  --gpu_ids 0

python -m pipeline.twostage_img2md \
  --input-jsonl images.jsonl \
  --output_dir outputs/two-stage \
  --layout-jsonl layout.jsonl \
  --served_model_name SeerRay-Lab/Xiaomi-OCR-0 \
  --port 8000 \
  --temperature 0.0 \
  --max_tokens 16384

If you already exported a PaddleX layout, skip the first command and pass its path with --layout-jsonl. Install a compatible PaddlePaddle wheel and requirements-paddlex.txt before layout detection; see the GitHub installation guide.

OUTPUTS

Keep the evidence behind the score.

postprocess/assemble_markdown.py can rebuild pages from saved region results. Preserve the assembly settings: text, formula, image, and reading-order handling can affect the evaluated output.

EVALUATION

Compare like with like.

The inspected repository provides parsing pipelines and the Agent runtime. It does not yet include a complete, pinned benchmark suite or the OCR-VQA evaluation commands. This guide documents the supported inference interface; exact score reproduction still needs those release artifacts.

TASK INSTRUCTIONS

Official prompts.

Extract all information from the main body of the document image and represent it in markdown format, ignoring headers and footers. Tables should be expressed in OTSL format, formulas in the document should be represented using LATEX format, and the parsing should be organized according to the reading order.

Extract key information in the image

Please output the key information in JSON format according to the following schema:
{
    "date": "",
    "seller_address": "",
    "seller_gst_id": "",
    "seller_name": "",
    "total_amount": "",
    "total_tax": ""
}