editorial-v1-modernbert

Emberdyn Editorial V1 — ModernBERT (Core ML)

The runtime resource for Emberdyn Editorial Intelligence V1, used on-device by EmberdynAI inside CutScript. This repository exists so the model can be downloaded on demand rather than embedded in an application bundle.

This is not a general-purpose model. It is a packaged runtime artifact: a Core ML conversion of ModernBERT-base, the upstream tokenizer, and an Emberdyn-trained classification head, with digests so a truncated or substituted download is rejected rather than silently producing different editorial output.

What is in the archive

editorial-v1-modernbert-base-coreml-fp32-v1.tar.gz

Base model

answerdotai/ModernBERT-base, revision 8949b909ec900327062f0ebf497f51aef5e6f0c8.

What Emberdyn changed

1. Core ML conversion. The upstream distribution has no Core ML artifact; the weights were converted for on-device Apple inference. Architecture and learned parameters are unchanged by that step — it is a format conversion, not a retrain or a quantisation.
2. A trained head. v1_head.bin is Emberdyn's own work. It consumes ModernBERT embeddings at runtime and contains no upstream parameters.

Answer.AI did not create, review or endorse Emberdyn Editorial V1. The base model is theirs; the conversion, the head and the editorial behaviour are Emberdyn's.

Licence

Apache License 2.0, matching upstream. The licence text is included in the archive and in this repository, and retains the attribution it carries (Copyright 2022 MosaicML Examples authors — ModernBERT builds on MosaicBERT).

Upstream publishes no NOTICE file, so none is included here.

Integrity

EmberdynAI pins an immutable revision of this repository and verifies the archive digest before installing. A mismatch is reported as not-ready rather than used.

Citation

ModernBERT's authors ask to be cited; see the base model card.