Eddy1919/openetruscan-classifier

OpenEtruscan Intelligence Suite V2 (v0.5.0)

This repository contains the official neural models for the OpenEtruscan platform. As of version 0.5.0, the models are optimized for both client-side ONNX inference and server-side Cloud Run autoscaling deployments.

Included Models

1. ByT5 Lacunae Restorer

• Base Model: google/byt5-small
• Task: Scholarly Span Corruption (LoRA)
• Purpose: Restoring damaged or missing characters in Etruscan inscriptions.
• Deployment: Now optimized for Google Cloud Run (/services/byt5-restorer).

2. Epigraphic Classifiers (ONNX)

• CNN Classifier (cnn.onnx): Ultra-lightweight (111KB) character-level CNN for client-side classification into 7 epigraphic classes.
• Transformer Classifier (transformer.onnx): High-accuracy (1.2MB) attention-based classifier. Achieves 99% Macro F1.

3. Prosopography Embeddings

• Supports Neural Entity Disambiguation (NED) via pgvector for Etruscan names and clans.

What's New in v0.5.0

• Cloud Run Deployment: The ByT5 restorer model is configured to be served as an independent API for autoscaling without blocking the main FastApi event loop.
• Improved Validation: Validated against the newly consolidated 200-query hybrid-search evaluation gate.

Usage

These models are seamlessly integrated into the OpenEtruscan Frontend and the openetruscan PyPI package.

For direct PyTorch usage:

from openetruscan.ml.neural import LacunaeRestorer

restorer = LacunaeRestorer("Eddy1919/openetruscan-classifier")
print(restorer.predict("mi ali[2]s"))