Code/Data to LLM-Assisted Interpretation of Graphene Image Analysis in Serverless Workflows

«LLM-Assisted Interpretation of Graphene Image Analysis in Serverless Workflows»
Code and data artefact to WoAIS/DEBS 2026 paper

Preparation
-----------

- run local Ollama with the model qwen2.5:0.5b-instruct
- place microscopy images into the directory data/cv_annotations/ of the following format:
 PNG image data, 907 x 907, 8-bit grayscale, non-interlaced

Running
-------

- follow the instructions from th ecompose file, with slight modification as follows:

docker compose -f docker-compose.eval.yml build

- to build the container image separately, use:

docker build -f Dockerfile.eval -t tdf --progress=plain .

The build process might take around 10 minutes with the optimisation of bypassing pip by installing system packages. It would otherwise take around 25 minutes due to the b
ehaviour of PyPI.

- run the experiments, e.g.:

docker compose -f docker-compose.eval.yml run --rm eval-handler python scripts/handler.py --strategy A

The output should be something like:

...
=== DESCRIPTOR DATASET SUMMARY ===
Total images: 3
...
✓ Results saved to /app/results/handler_output.json
(which is mapped to results/handler_output.json on the host)