GitHub repository
View the README, source code, and examples.
Docision-1.0 model
View and download the model weights on Hugging Face.
Quickstart
This example requires Python 3.10 or later and uses CUDA GPU0. Download the model and run inference with the input included in the repository:
Read the results
examples/multiple.json asks several questions in Chinese and English about the same table image. One asks for the RR for severely malnourished patients. After running the command, find that question by id in outputs/document-decisions.json. This excerpt from the project’s recorded output shows scores rounded to eight decimal places. Sample image · Input JSON · Full recorded output
decision shows that the model chose A (3.4). Use probability to compare candidate scores for the same question. Review decisions when the scores are close.
Use your own document image
Save the following JSON to a file and point--input to it. Set image to a path relative to the input file’s directory. In questions, list the questions and their candidate answers.
stable inference and prefix caching to reuse the page context. Package configuration guide
Use cases
You can use the same approach to ask questions with defined options about engineering drawings and scientific charts. For example, ask which experimental group each of two curves represents based on the chart legend.
Ways to use Docision
- Local Playground. Upload a document, enter your questions and options, and view the results in the browser.
- Python or HTTP API. Add page checks to your application. When you submit a PDF, Docision returns results by page. Package documentation
- MCP. Connect to a running Docision service to inspect local documents in Codex or Claude Code and export decisions as JSON or CSV. MCP setup guide
Model and public benchmark
Docision-1.0 is based on Qwen3-VL-4B-Instruct. The publicdoc-decision-bench dataset contains 608 document images and 2,617 Chinese–English question pairs, for a total of 5,234 language records. It covers document understanding and decisions as well as classification across industries. Dataset card
The project reports 96.18% accuracy for Docision, 86.78% for the base Qwen3-VL-4B model, and 95.72% for JEV-27B-VL on this benchmark. Official benchmark results
The project uses the Apache License 2.0.
