What "records trained" means

Three different record counts appear while you build a model, and they are supposed to differ. Here is which is which, and where each one is.

What it represents

Records trained is the labeled records the model actually fit on. Labeled means the record is old enough and complete enough to know whether it converted.

Scored records, shown on the score distribution, is the population the fitted model was applied to. This is normally the larger number, because scoring covers open records that have no outcome yet. That is the point of the model.

Features on the same card counts engineered feature columns, not the fields you selected. One categorical field with 40 values can produce many columns, so a 20-field configuration routinely trains on well over a hundred features.

How it’s calculated

Before fitting, ax1om splits the labeled records into a training partition and a held-out partition, 80/20 by default. The split is stratified, so both sides carry the same conversion rate, and it uses a fixed seed, so the same configuration reproduces the same split.

The model fits on the training partition only. AUC, lift, and precision are all computed on the held-out partition, which the model never saw during fitting. That is what makes those numbers an estimate of performance on new records rather than a report on memorized ones.

The records trained figure counts the training partition. It is smaller than your labeled dataset by exactly the held-out share.

Find each number, and set the split

1. Read merged records while you are still picking fields

   The footer of Selected Fields & Health on the Data step reads your field count and the merged record total, with the success count beside it once diagnostics have run. The same two counts head the diagnostics wall on Review and Train.

   This is your labeled dataset before any split. Nothing later can be larger than it, so if this number surprises you, stop here rather than reading on.

2. Set the train / test split, if the control is there

   Open the Filters & Success Criteria panel on the Data step, expand Success criteria, and the Train / test split slider sits at the bottom of it. It runs from 60 to 90 in steps of 5, and it starts at 80.

   Leave it at 80. Raising it to 90 to squeeze in more training data shrinks the held-out set, which makes every number on the model card noisier: you lose more in confidence than you gain in fit.

   When the model was created from a score, that whole section is hidden, because the score owns the success definition. The split is not exposed in the wizard in that case, and the model uses the default. The panel is headed Filters & Rules there instead.

3. Confirm the split on the review card before you train

   The summary card at the top of Review and Train lists Train/test split alongside your name, source, primary entity, field count, and success criteria. It is the last place the value is visible before it becomes part of a run.

   Read the whole card while you are there. Every value on it is a thing the model will be, and this is the cheapest moment to notice one is wrong.

4. Read Records trained on the model card after the run

   The Records trained card, last of the four KPI cards, carries the count, with the feature count on the line underneath it. This is the training partition, not your dataset.

   Expect it to be smaller than merged records by the held-out share and nothing more. A gap larger than that is not the split, it is the labeling.

5. Read the scored count on the score distribution

   The distribution below the KPI cards reports how many records were scored. This one is normally the biggest of the three, because it includes the open records the model exists to rank.

   If scored is not comfortably larger than records trained, the model is being applied only to history. Check the score's record population rather than the model.

6. When records trained looks far too small, check the right thing

   The usual cause is your date filter or your success criteria excluding records, not the split. Both are visible on the Data step, and their effect is visible in the two counts at the top of the diagnostics wall.

   Go to Dataset health first. Moving the split to recover records treats the symptom and costs you evaluation quality.

What it means for you

The three counts disagreeing is the healthy state. Records trained smaller than your dataset means an honest held-out evaluation exists. Scored larger than both means the model is doing its job on records that have no outcome yet.

The number to watch over time is merged records, because it is the one your CRM controls. The other two follow from it.

Check your understanding

Refresh reports 12,000 labeled records. The card reads 9,600 records trained and 143 features, and the score distribution reads 48,000 scored.

Nothing is missing. 2,400 records were held back for evaluation, which is exactly the 20% the slider was left on. The other 36,000 are open records with no outcome to learn from, which is why they were scored and not trained on. And 143 features from a 20-field selection is ordinary: the fields you ticked are not the columns the model sees.