Create a Custom Model with a Training Dataset
6 mins read - Updated on Oct 28, 2026You can create a custom model in Kimola from a training dataset: a file of customer feedback your team has already labeled, with one piece of text and its label on every row. Kimola trains a classifier that learns from these examples and then tags new feedback with the same labels in Reports and Funnels. This article explains how to prepare the file and walks through the four steps of the By Training Set wizard: Training Set, Mapping, Labels and Review.
A training dataset suits teams that already have labeled data, such as a coded survey archive or support tickets sorted into categories over the years. If you don't have labeled data, you can create a custom model with a prompt by describing its labels, or create one from a Report you have already analyzed. What is a Custom Model? compares the three ways.
Custom models are available in the paid packages, not in Starter. You need a file of feedback that is already labeled, with one label on every row. Sign in to Kimola and open Models from the main menu.
Prepare Your Training Set
A training set is a table with two columns that matter: a text column that holds the feedback and a label column that holds the label each piece of feedback should get. Every row is one example, and the model learns what a label means from the rows that carry it. Put the column names in the first row, because the wizard lists them by those names. Other columns can stay in the file; the wizard reads only the two you choose. Save the file in .xls, .xlsx, .csv or .tsv format, up to 30 MB.

A hotel chain that has coded its guest survey comments by hand could export them as a training set. One row holds "The room smelled of smoke and the window wouldn't open" with the label Room Condition, another "We waited 40 minutes to check in" with Check-in. Other rows carry Breakfast, Staff and Noise.
The model can only be as good as these examples. The three sections below cover what to check before you upload the file.
Choose Clear Labels
A training set needs at least 2 and at most 25 labels; with fewer or more, the wizard doesn't let you continue. Every row gets one label, written in the label column.
Use short, descriptive and neutral names that say what the feedback is about, not how the customer feels about it: Delivery rather than Bad Delivery, Pricing rather than Too Expensive. Keep the names in one style, for example all singular nouns with the same capitalization, and spell each label the same way on every row, so that one label is always one value in the label column.
Choose labels that don't overlap. If two labels describe the same thing, the model can't tell them apart, and neither can the people reading the results. Every sector has its own topics: if there is something you want to track, such as Product Returns or Repeat Purchases, make it a label of its own and apply it to every row it fits.
Clean the Text
Remove the rows that teach the model nothing: empty rows, duplicates, and rows that hold only emojis, punctuation marks or a link. Delete links, numbers and names from the remaining rows too, so the model learns from what customers say rather than from details that belong to one customer. In Google Sheets, the REGEXREPLACE function can remove them from a whole column at once.
Balance the Labels
Give every label enough examples, and keep the counts of the labels close to each other. A label with far more rows than the others pulls the model toward it, and a label with only a few rows is hard for the model to learn. The wizard points out both in step 3, so you can correct the file before the model is trained.
Step 1: Upload the Training Set
On the Models page, click Create AI Model and choose By Training Set. The wizard opens on Training Set, the first of its four steps. Drag your file onto the box that says Upload a dataset to train your custom AI model, or click the box to choose the file.

The training set can be one .xls, .xlsx, .csv or .tsv file of up to 30 MB.
If your package doesn't include custom models, the window says so and offers to upgrade your package or contact sales. If you have reached the custom model limit of your package, it offers to upgrade or to delete a model you don't actively use. See kimola.com/pricing for the packages.
Step 2: Choose the Text and Label Columns
The Sheets & Columns screen shows a preview of your file, with the number of rows in total and the file's name and size under it. If an Excel file has more than one sheet, choose the sheet that holds your training set from the sheet names under the preview.
Above the preview, choose two columns:
- Text Column is the column that holds the labeled content, the feedback itself.
- Label Column is the column that holds the label of each row.
A column chosen for one of them isn't offered for the other. Click Next to continue, or Back to upload another file.

Step 3: Check the Label Distribution
The Labels screen lists every label found in your label column, with its Count (the number of rows that carry it), its Percentage of all rows and its Status. A sentence above the list sums up the training set.
When every label is in the expected range, the sentence says your training set is well-balanced. When one or more labels are flagged, it says the training set "has some balancing issues but suitable for model building", and you can still continue with Next. When the file has fewer than 2 or more than 25 labels, it says "Your training set is invalid." and Next is not available until you upload a file with the right number of labels.

The Status column marks the labels that need attention. Above the overrepresented threshold means the label has far more rows than the others, and Below the underrepresented threshold means it has too few. To correct them, remove some rows of an overrepresented label or add rows to an underrepresented one, then go Back and upload the corrected file.
In the hotel chain's training set, Room Condition might hold most of the rows, because guests write about their rooms more than anything else, and be marked above the overrepresented threshold. Noise, with only a handful of rows, might be marked below the underrepresented threshold.
Step 4: Name and Review the Model
The Review screen sums up the model before it is trained. Fill in the fields and check the rest:
- Model Name is the name the model gets in Kimola, up to 100 characters (e.g., Guest Survey Topics). The name must be different from the names of your other models.
- Category sets the kind of classifier: Content Classification sorts feedback into topics, and Sentiment Classification detects how customers feel. Choose the one that matches your labels.
- Dataset shows the number of rows, the sheet and the file the model is trained on.
- Required Query shows the queries that training will use, next to your package and its usage.
Click Complete! to start training.

Training uses queries by the rule queries = rows × models × Dimensions. A training set is one model with no Dimensions, so every row of the file is one query. If your package has no queries left, the Required Query line says so. See How to Calculate Queries.
Training uses one query per row of the training set.
What Happens Next
After you click Complete!, Kimola uploads the training set and shows how many records were uploaded out of the total. Then the window says Getting Ready while Kimola optimizes the model for the best accuracy. This may take a little while, and you don't need to keep the window open: on the Models list, the model shows Training... until it is ready.
When training is done, the window says All Set! and Go to Model! opens the model's page. On the Models list, the model shows its number of records and its consistency, the percentage of semantically consistent pairs in its records. See What is Model Consistency?.
From the model's page you can test the model on your own text or on a file, and explore its records to add, delete or download them. If the results aren't what you expected, see Optimize Dataset-Based Custom Models.
Once the model works as you want, choose it when you create a Report using a custom model, or add it to a Funnel so that every Space using the Funnel classifies its files with it. Funnels and Spaces are available in the Enterprise package.