Training data
Training data are the examples used to fit a machine learning model. They may contain text, images, measurements, audio or combinations of these. In supervised learning, labels specify the target output. In self-supervised learning, the training process derives a target from the input data.
Choosing examples
Begin with the intended use. A model trained on formal letters may handle short, misspelled support messages poorly. A dataset with many records can still omit the languages or situations that users encounter.
Document how people collected the data, what each field means, who supplied labels and which cases they excluded. The Datasheets for Datasets proposal provides a framework covering motivation, composition, collection, recommended uses and maintenance.[1]
Check the conditions under which you can use and retain the material. Public access to a document does not, by itself, settle copyright, personal-data or contractual questions. Keep provenance and usage restrictions with the dataset so that later teams can assess them.
Separating training from evaluation
Training examples fit the model. Validation examples support development choices. A held-out test set provides a final assessment after those choices. Fitting preprocessing steps on test data creates leakage and can make results look better than they will be in use.[2]
Consider a collection of maintenance reports. If several reports concern the same machine, a random split may put near-duplicates on both sides. A split by machine, site or time may provide a more realistic test, depending on where the team plans to deploy the service.
Quality work before another training run
Inspect duplicates, conflicting labels and missing cases. Keep a record of dataset versions and changes so that a performance difference can be investigated. Removing records without examining their role can also remove uncommon but important cases.
For Sustainable IT, data preparation is an opportunity to avoid repeated, poorly targeted training. More data can improve a model, but collecting or generating additional examples should address an observed gap. Measure the result instead of treating dataset size as a quality score.
See also
References
- ↑ Gebru et al., Datasheets for Datasets.
- ↑ scikit-learn, Common pitfalls and recommended practices.
