Definitions of terms on CloudFerro Cloud
When you move between My Portal, MLflow, and the Python examples, similar names can make it difficult to tell what is being created or saved. An instance, an experiment, and a run each refer to a different part of the setup.
The definitions below relate those names to the calculation and Sherlock examples. They also explain the inputs, measurements, and files saved with a run, so you can recognize the records in both the interface and the code.
Instance, experiment, and run
AI Experiments is the product name. In these articles, experiment on its own means an MLflow experiment, with the terms defined as follows:
Term |
Meaning |
|---|---|
Instance |
The hosted MLflow service you connect to. It has its own URL, membership, and storage configuration. The example instance is named experiment1. |
MLflow experiment |
A named group of related runs inside an instance. The calculation example uses first-experiment. |
Run |
A record of one execution or trial, with its own run ID. The calculation runs are named first-total and second-total. |
One instance can contain several experiments, and each experiment can contain many runs. In the Sherlock example, one run evaluates all six messages with one model and one prompt. The four model-and-prompt combinations produce four runs in the same experiment.
The portal page My Experiments lists service instances. The Experiments page inside MLflow lists experiment groups within the selected instance.
What a run records
The addition example uses these three record types:
Term |
Meaning and example |
|---|---|
Parameter |
An input setting recorded for a run. The numbers parameter contains [10, 20, 30]. A training program might record a learning rate instead. |
Metric |
A numerical measurement. The total metric is 60. A model evaluation might record accuracy or response time. |
Artifact |
A saved file. The calculation saves result.json with its inputs and total. Other programs might save a plot, dataset, or model file. |
A trace records application calls and their execution details. GenAI applications can save traces in an MLflow experiment, but a trace is a separate record type from a run. The Python examples in these guides explicitly log runs, parameters, metrics, and artifacts.
What the examples compare
The calculation exercise checks the tracking setup. Its two runs produce known totals, 60 and 100, from different input lists. Comparing the records checks that each run retains the expected inputs, result, and files. Neither calculation is better.
The Sherlock model comparison evaluates answers to the same messages. It asks whether changing a prompt improves classification accuracy, or whether another model gives more accurate answers or responds faster with the same prompt.
A statistical experiment is a study designed to investigate a question through controlled conditions and statistical analysis. MLflow can store records from such a study, but creating an MLflow experiment does not establish its design or statistical validity. The six-message Sherlock example demonstrates evaluation; broader conclusions require representative test data and repeated measurements.