JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Implementation guide

Serve structured reads from DiffusionGemma

A technical guide to serving bounded label decisions and span extraction from DiffusionGemma behind a `/v1/systemone` interface, including constrained reads and a reproducible span battery.

Original by mmastracExtractionGitHub README and serverSource reviewed

Before you dive in

What you’ll find in the original

  1. Send state and typed questions to the documented `/v1/systemone` route instead of parsing generated JSON.
  2. Constrained mode runs unembedding, sampling, and self-conditioning only over declared labels, changing how probabilities are normalized.
  3. Exercise span fields, lists, and mixed schemas with the included battery before treating extraction behavior as compatible.
Worth knowing

Requires a DiffusionGemma-capable inference environment. Constrained probabilities are normalized over labels and are not directly equivalent to unconstrained label mass; repository results were not independently benchmarked.