An experimental repository to demonstrate metadata extraction tools.
This repository is intended as a starting point for experimenting with metadata extraction tools, providing a Dockerfile with a working environment and sample scripts.
This is primarily indended for experimentation and inference across a variety of underlying architectures, using only CPU when possible rather than CUDA/GPU specific options.
- Docker installed on local machine
Clone repo to local machine, cd into repo directory:
git clone https://github.com/UCLALibrary/metadata-extraction-lab.git
cd metadata-extraction-lab
Build the container image:
docker compose up
The first time this is run, it may take a little time to download the tinyllama model (640MB).
JupyterLab will be accessible via port 9999, in the case you have a service already listening on that port.
In your terminal window, take note of the token assigned. It will look something like:
lab?token=1c5a66545jksfsdfsdfljk
In a web browser, navigate to localhost:9999. You will be prompted for the token above.
After approving, you should be able to navigate JupyterLab. A sample notebook in the notebooks directory contains a demo of the stock layout-parser model as a starting point, along with several other tools.
Example notebooks:
- Layout Parser
- Detectron2 : Object Detection and Instance Segmentation
- Amazon Science ReFinED NER
- PaddleOCR
- Ollama with initial tinyllama model
- Full ocr to metadata record pipeline example