maap

About maap

maap is a 4B parameter model designed to estimate numbers from unstructured text when no labeled data exists. It reads text records and returns calibrated probability distributions rather than single point estimates, including medians, quantiles at requested coverage levels, intervals fitted to held-out data, and probabilities for exceeding specified thresholds.

Unlike language models which generate text and provide unfounded confidence intervals, maap processes records in a single forward pass without text generation, answering in approximately 160 milliseconds. It explicitly reports whether targets are in-domain and whether intervals carry calibration guarantees.

The model can handle few-shot learning when supplied with at least eight labeled examples of new targets. It costs roughly two orders of magnitude less per prediction than frontier language models and includes 1,000 free estimates to try.

• api
• text-to-number
• calibration
• uncertainty quantification
• forecasting
• probability distribution
• estimation

FAQ

What does maap actually return?

A median, quantiles at requested levels, an interval at specified coverage levels, and the probability of exceeding any threshold you set. For count targets it can also return the probability of each whole number. Each response includes whether the target was in the training set and whether the interval carries a calibration guarantee.

Is maap just a language model under the hood?

It is built on a language model but does not generate text at inference. The record enters, the model reads a distribution over ordered bins in a single forward pass, and a conformal wrapper adjusts the spread, enabling responses in about 160 milliseconds rather than seconds.

What if my target is not one it trained on?

Send at least eight labeled examples of the same target and maap will estimate it few-shot, though results will be weaker than on trained targets. The response flags new targets so your code can fall back to another method.

What does it cost?

The first 1,000 estimates are free, with the count per estimate rather than per request since one request may contain up to 32 estimates. Pricing for additional usage is available by request.

How is this different from asking a language model for a number?

Two ways. A language model gives a point guess and an interval it has no basis for: on FermiEval, nominal 99% intervals from frontier models contained the real answer about 65% of the time. maap fits its intervals to held-out data per target and says so when that fitting failed. It also costs roughly two orders of magnitude less per prediction.