AI Model Artifact Detection

Overview

Mend AI detects various types of AI model artifacts.

• To minimize noise and increase accuracy, Mend AI automatically filters out irrelevant file types (e.g., .txt, .tar, .zip) and employs smart content validation to handle ambiguous extensions such as .pkl, .npy, etc.

• Mend AI only detects files equal to or larger than 10MB in size.

Testing Artifact Detection with Git LFS

To properly test the AI Model Artifact Detection capability, it is highly recommended to use Git LFS (Large File Storage), since model artifacts are typically larger than 10MB.

1. Clone a Hugging Face repository.

2. Run git lfs.

3. Verify the files are on your file system.

Prerequisites

Before testing, ensure you have Git LFS installed. If not, install it following the Hugging Face requirements guide.

Testing Steps

1. Clone a repository with model artifacts from Hugging Face:

   CODE

   git clone https://huggingface.co/[model-name]

   Example repositories you can use for testing:

   • https://huggingface.co/bert-base-uncased

   • https://huggingface.co/gpt2

2. Pull the LFS files:

   CODE

   cd [repository-name]
   git lfs pull

3. Verify the files are present on your filesystem:

   CODE

   ls -lh

   Look for files larger than 10MB (typically .bin, .safetensors, or .ckpt files). These should now show their actual size rather than being LFS pointer files.

4. Run the Mend artifact scanner:

   Now you can test the artifact detection capabilities against these actual model files.

Common Issues

• Issue: Scanner doesn't detect artifacts

  • Cause: Git LFS files weren't pulled, only pointer files exist

  • Solution: Run git lfs pull to download the actual artifacts

• Issue: Files appear to be only a few KBs in size

  • Cause: These are LFS pointer files, not the actual artifacts

  • Solution: Ensure Git LFS is installed and run git lfs pull

AI Model Artifact File Types

    "bin", "safetensors", "safetensorsc", "ggmlv3", "trt", "tdict", "safetens", "argosmodel", "pt",
    "v2", "pth", "onnx", "ckpt", "gguf", "guff", "wv", "model", "weight", "weights",
    "caffe", "caffemodel", "nemo", "pdmodel", "neuron", "skops", "pkl", "index", "npz", "npy",
    "pb", "pickle", "qweight", "tfrecord", "engine", "pbmm", "scorer", "tflite", "data", "binary",
    "llamafile", "dat", "mlmodel", "keras", "ggml", "safetensor", "tfbson", "tensors", "gguff", "mil",
    "torch", "safetensorsa", "savetensors", "savetensor", "msgpack", "h5", "h5ad", "pyth", "wandb",
    "onnx_data", "pdparams", "cleanrl_model", "qzeros", "safetesors",
    "mar", "joblib", "cbm", "mlpackage", "plan", "dlc", "ubj", "pmml"