dunguasli/detecode-model-v1

DeteCode Model v1

Fine-tuned CodeBERT checkpoint for the optional AI engine in DeteCode.

This model is experimental and should be used as an assistive semantic layer, not as the only source of truth. DeteCode's local rules and taint analysis are recommended for stable CLI scanning.

Usage

Download or clone this repository into:

models/codebert-webvuln

Then run:

python -m detecode scan .\tests\samples --engine hybrid --model-path .\models\codebert-webvuln --format table

Training Data

The model was fine-tuned using the DeteCode training script with PHP/JavaScript vulnerability data, primarily CrossVul (hitoshura25/crossvul).