Skip to main content
The shape is scikit-learn. What happens underneath is not.

​
fit does not train

No gradients are computed. No weights change. fit sends your labeled table to the model and returns a handle to it. The model was pretrained once, by us, on a very large number of tables. It did not learn any particular relationship between columns — it learned how to read a table and infer the relationships inside it. Your table becomes the context the model reads while it predicts. Closer to handing someone a reference sheet than making them study for the exam. The usual loop discards the table after training; Hollerith reads the same table on every call.

​
Nothing to tune

There is no learning rate, no tree depth, no regularization term, no early-stopping patience. There is no optimization loop for those knobs to sit on.

​
No world knowledge

Like a language model reading a prompt, Hollerith conditions on what you put in front of it. Unlike a language model, it has nothing stored to fall back on. It does not know that churn means a customer left, or that a price of -40 is probably a refund. It knows the shape of tables in general, and the one table you handed it. Two consequences:
  • Signal has to be in the columns. Ratios, aggregates and joins that encode domain knowledge help here for the same reason they help gradient boosting.
  • Column names carry no meaning. Renaming x7 to days_since_last_login will not change a prediction.

​
The cost moves to prediction time

Training cost does not disappear. It moves. Your table is read on every call, so its size is a cost you pay every time rather than once up front. Time and price both scale with how much table the model reads. Two things to internalize before building on this:
  • fit is a real job. It runs on a GPU, takes seconds to minutes, and bills the rows you send. “No training” is a claim about gradients, not about cost or latency.
  • More context is not better. Rows that carry no signal cost latency and money and buy nothing. The most informative table usually beats the largest one.

​
Fit once, predict many

fit returns a reusable artifact that lives on our side for 7 days. Each predict against it sends only the rows you want scored.
Expiry, resuming from another process, and the ways you can accidentally skip it are in The fitted context.

​
What the model reads

Numbers and categories are read as they are. Missing values are read as missing — they are information, not something to impute first. Free text and dates are accepted but read as categories. The model sees that two rows share a value, not what the value means. A signup_date tells it which rows signed up together, not which signed up earlier. The task is inferred from your target. A whole-number target with 20 or fewer distinct values is treated as classification — right for a 1–5 rating, wrong for a small count.
Row-wise predict has no notion of time. Time-indexed targets belong in Forecast.

​
Re-fit when your data moves

Predictions assume the rows you score look like the rows you fit on. A new product line or a pricing change breaks that, and the model has no way to tell you. Re-fitting costs one call. Do it often rather than building drift detection you do not need yet.

​
Next