Input data overview

How to prepare the data to run Evidently Reports or Test Suites.

To run evaluations on your datasets with the Evidently Python library, you should prepare your data in a certain way. This section covers how to do that.

This applies to Evidently OSS, Evidently Cloud and Evidently Enterprise.

Looking for something else? Check Tracing to instrument your app. Check Datasets to work with datasets in the user interface. To run evaluations after you prepare the data, see Reports and Test Suites.

Input data format

Evidently works with Pandas DataFrames, with some metrics also supported on Spark.

Your input data should be in tabular format. All column names must be strings. The data can include any numerical, categorical, text, DateTime, and ID columns. You can pass embeddings as numerical features.

The structure is flexible. For example, you can pass:

• Any tabular dataset. You can run checks for data quality and drift for any dataset.

• Logs of generative LLM application. Include text inputs, outputs, and metadata.

• ML model inferences. You can analyze prediction logs that include model features (numerical, categorical, embeddings), predictions, and optional target values.

To run certain evaluations, you must include specific columns. For instance, to evaluate classification quality, you need columns with predicted and actual labels. These should be named "prediction" and "target", or you’ll need to point to the columns that contain them. This process is called Column Mapping.

Learn more in the next section:

Column mapping

Reference and current data

Usually, you evaluate a single dataset, which we call the current dataset. In some cases, you might also use a second dataset, known as the reference dataset. You pass them both when running an evaluation.

When you may need two datasets:

• Side-by-side comparison. If you want to compare model performance or data quality over two different periods or between model versions, you can do this inside one Report. Pass one dataset as current, and another as reference.

• Data drift detection. To detect distribution shifts, you compare two datasets using methods like distance metrics. You always need two datasets. Use your latest production batch as current, and choose a reference dataset to compare against, such as your validation data or an earlier production batch.

• Automatic Test generation. If you provide a reference dataset, Evidently can automatically set up Test conditions, like expected min-max values for specific columns. This way, you don’t have to write each test condition manually.

If you pass two datasets, the structure of both datasets should be identical.

Data volume

Running computationally intensive evaluations on large datasets can take time. This depends on the specific evaluation as well as your infrastructure.

In many cases, such as for probabilistic data drift detection, it’s more efficient to work with samples of your data. For instance, instead of running drift detection on millions of rows, you can apply random or stratified sampling and then compare samples of your data.

For datasets that don’t fit in memory, you can run calculations using Spark.