Semilattice trained 420 models on data from groups of small-business merchants to predict different responses within the same audience

Semilattice trained 420 models on data from groups of small-business merchants to predict different responses within the same audience

On 4 August, Semilattice introduced Anacreon. The system analyzes public posts by small-business merchants who accept card payments in person, then builds a separate model for each group of similar authors.

An average score does not show how responses are distributed within an audience. Everyone in one group may choose the midpoint of the scale, while in another group half may answer “for” and half “against.” The average treats these cases as equivalent, although they call for different decisions about a pricing plan, service, or campaign.

In the Semilattice report, Anacreon groups public posts according to similarities among their authors. The system trains a LoRA adapter for each group. This is a small module that connects to a language model. The authors trained 420 such modules and evaluated 400 after selection. Each model represents its own group of similar merchants.

For evaluation, the authors set aside the most recent 20% of records based on time, then asked the models questions that appeared later. They varied the order of the answer choices and reversed the polarity of the wording. For the median group, the authors report 67,9% exact matches on multiple-choice questions. For scale questions, their metric applies a larger penalty when a response is farther from the actual rating.

In the study by Joon Sung Park and colleagues, the researchers created a separate agent for each of 1 052 participants using a two-hour interview and questionnaires. Park’s agent models the responses of one person, while Anacreon models the distribution of positions within a group of similar people.