Prism English (prism-en)

English UPOS tagging, morphological features, and lemmatization with calibrated confidences, built for on-device, fully offline use. A compact 18.0 M-parameter model on a ModernBERT-lineage (Ettin) backbone — the same artifact contract and native runtimes as prism-no, now for English.

This repository mirrors the versioned release artifacts of the Prism project. Prism ships native runtimes for Swift, C++, C, and Java/Kotlin that read these artifacts directly. Requires the language-independent Prism runtimes, library 0.6.0 or later.

Which folder to use

Folder	Size	What it is
prism-en-0.1.0/	≈ 94 MB	fp32 reference model

This first English release ships a single precision (fp32). There is no int8 -fast variant yet: ModernBERT's rotary-embedding indexing does not lower cleanly through the int8 path, and a fast variant is planned.

An application bundles the folder as-is. It is everything a Prism runtime needs; point the tagger API at its local path:

let tagger = try PrismTagger(artifactURL: artifactFolder)                // Swift

prism::tagger::Tagger tagger("prism-en-0.1.0");                          // C++

try (var tagger = PrismTagger.load(Path.of("prism-en-0.1.0"))) {}        // Java

Quick starts for every binding: the project README. The artifact contract (programs, model.ptd weights, tokenizer and label schemas, checksums) is documented in docs/INTEGRATION.md.

Note: these are ExecuTorch programs with the decoding policy and calibration baked in — not transformers-loadable checkpoints. vocabulary.json is a standard Hugging Face tokenizer.json (the ModernBERT tokenizer) and loads with the tokenizers library.

Quality

Measured on the UD English-EWT test split with the official UD metrics, against UDPipe 2.17:

Test F1	prism-en 0.1.0	UDPipe 2.17
UPOS	97.29%	97.65%
UFeats	97.71%	98.16%
Lemmas	97.36%	97.99%

Competitive on EWT web text — within ~0.4–0.6 pp of UDPipe — from an 18 M-parameter model that runs fully offline on a laptop CPU. The Ettin-400m teacher beats UDPipe on all three metrics; distilling down to the 18 M student gives up ~0.5 pp, a deliberate later-optimization target (0.1.0 ships the current student so English is available now).

Model description

An 18.0 M-parameter encoder student on jhu-clsp/ettin-encoder-17m (a 7-layer ModernBERT-lineage encoder, hidden 256), distilled from the Ettin-encoder-400m teacher, with a character CNN feeding the morphology and lemma heads, a structured morphology decoder, and per-head temperature calibration. Label spaces: 18 UPOS · 21 morphology features · 1,632 lemma edit rules; 50,368 byte-level BPE subwords. Full technical reference: docs/ARCHITECTURE.md.

Training data and attribution

• Gold treebank: UD English-EWT (revision c5baffde, UD 2.17) by the Universal Dependencies contributors — CC BY-SA 4.0. (The GUM treebank is deliberately excluded: its CC BY-NC-SA license is incompatible with this release.)
• Silver corpora (unlabeled running text, labeled offline by the Prism teacher; the text itself is not redistributed): Project Gutenberg (public domain) and English Wikipedia (Wikimedia contributors — CC BY-SA 4.0).
• Backbone: jhu-clsp/ettin-encoder-17m and its Ettin-encoder-400m teacher by the JHU CLSP group — MIT.

Pinned revisions and checksums travel inside the artifact (manifest.json, LICENSES/).

License

Model weights: CC BY-SA 4.0 — the same share-alike license as the UD English-EWT gold annotations. Using or bundling the unmodified artifact — including commercially, in closed-source applications — is fine (keep the LICENSES/ attribution); redistributed modified weights must stay open. Prism source code is Apache 2.0.