Product model (research preview)

This page describes Upright's research preview methodology for producing net impact data. For information on the classic product model. see this page.

Inputs

The main inputs for the product model include:

• Research papers found relevant by the teacher model when teaching the specialized model (discussed below)

• Base model weights (current base model is Ministral 3), which are distilled from usual data sources used to train Large Language Models.

Algorithm

Conceptually, the product-level impact size estimates are produced in three phases:

1. Training a specialized model for high-performance impact size inference

2. Using the specialized model to pairwise compare the sizes of millions of impacts (thousands per product)

3. Running a solver to derive coherent implied impact sizes estimates from the pairwise comparison results

The impacts to be sized are defined in Upright's universal impact structure, a hierarchical list of around ~800 impact topics across Environment, Society, Knowledge, and Health.

The training and motivation for the approach is further discussed in this blog post.