A manifest file records metadata or associations that let downstream systems interpret a dataset or model artifact. Its contents depend on the workflow: dataset manifests map files to labels, while model manifests combine runtime metadata with prediction-class mappings. section 2.2.7 section 7.1.4
Common production roles include:
- Define training data by associating image filenames with numeric label IDs and mapping those IDs to label names. section 2.2.7
- Supply model-serving metadata such as algorithm, framework, framework version, model version, and class names. section 7.1.4
- Support prediction by translating an integer model result into a meaningful class name. section 7.1.4
>> examples.csv: <image filename>,<label id> “imageA.jpg”, 0 “imageB.jpg”, 1 “imageC.jpg”, 0 >> labels.csv: <label id>,<label name> 0, cat 1, dog >> examples/ - folder imageA.jpg imageB.jpg imageC.jpg
Inspect a compact image-classification manifest pair: examples.csv links image files to label IDs, labels.csv resolves IDs to names, and examples/ stores the images. section 2.2.7
├── manifest.json #1
├── model.pth #2
└── vocab.pth #3
// A sample manifest.json file
{
"Algorithm": "intent-classification", #4
"Framework": "Pytorch",
"FrameworkVersion": "1.9.0",
"ModelName": "intent",
"CodeVersion": "80bf0da",
"ModelVersion": "1.0",
"classes": { #5
"0": "cancel",
"1": "ingredients_list",
"2": "nutrition_info",
"3": "greeting",
.. .. ..
}
Inspect a model directory containing manifest.json beside model.pth and vocab.pth, with metadata and class mappings in the manifest. section 7.1.4
