Annotations¶

You can annotate your Jobs with aibom.io/* keys to provide experiment metadata. Annotations are optional: without them, the AIBOM is still generated from auto-detected data (hardware discovery, dataset detection, telemetry).

Annotation	AIBOM field
`aibom.io/experiment-intent`	`experiment_intent` (`training`, `sft`, `inference`)
`aibom.io/experiment-name`	`experiment_name`
`aibom.io/model-name`	`model.name`
`aibom.io/model-framework`	`model.framework`
`aibom.io/git-repository`	`source_code.git_repository`
`aibom.io/git-commit`	`source_code.git_commit`
`aibom.io/git-branch`	`source_code.git_branch`
`aibom.io/dataset-name`	`dataset.declared.name`
`aibom.io/dataset-source`	`dataset.declared.source`
`aibom.io/dataset-version`	`dataset.declared.version`
`aibom.io/dataset-license`	`dataset.declared.license`
`aibom.io/optimizer`	`training.optimizer`
`aibom.io/batch-size`	`training.batch_size`
`aibom.io/epochs`	`training.epochs`
`aibom.io/learning-rate`	`training.learning_rate`
`aibom.io/top-k`	`inference.top_k`

Precedence¶

Auto-detected values are used as defaults, and a corresponding annotation overrides them.

The exceptions are aibom.io/learning-rate, aibom.io/batch-size, aibom.io/epochs and aibom.io/random-seed. For these, the runtime hooks win, then the command-line arguments, and the annotation is used only when neither produced a value.

For every field the pipeline can populate, see Detected Fields.