SchemaForge

Papers

Every SchemaForge whitepaper, newest first — each version keeps its own citable URL rather than being silently overwritten by the next one.

SchemaForge V3 (draft, iterations 1–4)

A recipe sweep against V2's unsolved omission problem — narrower, not closed

Four iterations into targeting the failure mode V2 could not move: silent omission at 55–62% of failures. The best V3 configuration reaches 52.3% — a new low, and still above half. No V3 checkpoint yet beats the V2 release candidate on the comparable eval. Includes a teacher-generation deadlock recovered without re-running generation, by exploiting the pipeline's own determinism.

Read →

SchemaForge v1.5

Hybrid deterministic and distilled semantic structured extraction

Draft. The evaluation rebuilt around v1's weaknesses: corrupted documents instead of clean synthetic ones, held-out schemas, gated teacher labels, and scoring only on the fields regex cannot reach. Field F1 does not yet beat the un-distilled base model.

Read →

SchemaForge V2

A hybrid deterministic and distilled semantic extraction system, and the omission problem it did not solve

Hybrid field F1 0.6432–0.6827 against 0.2911 for rules alone and 0.4461 for the model alone, dominating both on every metric, on every checkpoint tested. The unsolved half: silent omission at 55–62% of every failure, immune to four independent fixes — and a non-deterministic teacher labeller that quietly invalidated part of the project's own early evidence before an audit caught it.

Read →

SchemaForge v1

Distilling Gemma-4 into a 1B edge model for enterprise JSON extraction

A 1.08B student matching its 31B teacher on the in-domain suite at 16x less VRAM and 5x the throughput — and a prompt-header change that collapses accuracy from 70% to 0%. Released on Hugging Face.

Read →