manurubo/CoCoGraph ? reverse-engineered prompt

Build me a Python project that can generate realistic synthetic molecules from an existing chemical dataset, and make sure the generated molecules are always chemically valid. I want a simple workflow where I can train the model, run a quick demo with a pre trained checkpoint, and then generate new molecules into an output folder.

It should also include a way to compare the generated molecules against the original set and against a couple of common benchmark methods, with simple plots and summary files so I can see how well it did. If possible, add support for doing molecule generation from a base model, a version that uses extra molecular fingerprints, and a basic inpainting mode where fragments can be attached to an existing molecule.

Please keep it easy to run from the command line, use the provided data files if they are there, and make the README clear enough that someone with only basic Python experience can follow the setup. If you need current package details or RDKit docs, look them up online.