Analyzing Ethereum Smart Contracts by fintuning CodeBert on it to detect vurnabilities
https://huggingface.co/datasets/mwritescode/slither-audited-smart-contracts
For training the TU Wien Jupyter Hub Servers where used with GPU support. You need 40GB of RAM to train these models.
QLoRA was used for training: https://github.com/artidoro/qlora
https://github.com/microsoft/CodeBERT
In the Notebook, data and model checkpoints are made. Thats why the disk requirement is so high.
This is a screenshot of the wandb (Weights and Biases)
Here you can see the usage of the last run
