fullstop-coreml

oliverguhr/fullstop-punctuation-multilingual-base (XLM-RoBERTa token-classification, ~0.3 B) converted to CoreML (fp16) for on-device punctuation restoration — downloaded on first use and run locally through CoreML + swift-transformers.

This is a format conversion only — the weights are unchanged from the upstream PyTorch model. The model is a verbatim token classifier: it emits one punctuation label per token, so it can add punctuation (. , ? - :) but can never reword the input. Capitalization is handled separately (a heuristic restorer is typically chained after it).

Label map (id2label)

id	label	meaning
0	""	no punctuation
1	.	period
2	,	comma
3	?	question mark
4	-	dash
5	:	colon

Files (repo layout)

The repo root is the folder a loader consumes directly:

File	Purpose
Fullstop.mlpackage/	CoreML model bundle (compile before load)
tokenizer.json	XLM-R Unigram tokenizer (HuggingFace tokenizers)
tokenizer_config.json	tokenizer config
special_tokens_map.json	XLM-R special tokens (<s>, </s>, <pad>, <unk>, <mask>)
config.json	model config incl. id2label / label2id

Loading (Swift)

// 1. Snapshot the repo (swift-transformers Hub):
//    Hub.snapshot(from: Repo(id: "gregbarbosa/fullstop-coreml")) -> localDir
// 2. Compile + load the CoreML model, build the tokenizer from the same folder:
let compiled = try await MLModel.compileModel(at: localDir.appendingPathComponent("Fullstop.mlpackage"))
let model = try MLModel(contentsOf: compiled)
let tokenizer = try await AutoTokenizer.from(modelFolder: localDir)

Inference contract: feed input_ids (shape [1, seq], natural length — no padding) and attention_mask; the output logits are [1, seq, 6] → argmax per token yields the label id above. Attach each word's punctuation from the label on its final subword (token boundaries marked by the XLM-R ▁ U+2581 prefix); skip special tokens.

Provenance

• Source model: oliverguhr/fullstop-punctuation-multilingual-base (MIT)
• Conversion: PyTorch → ONNX → CoreML (fp16). Parity verified 15/15 labels identical to the PyTorch reference across the conversion fixtures; tokenizer parity re-verified on-device 2026-06-26.
• Conversion harness: restore-bench (standalone repo).

License

MIT — same as the upstream model. Conversion produced by Greg Barbosa.

Citation

@article{guhr-EtAl:2021:fullstop,
  title={FullStop: Multilingual Deep Models for Punctuation Prediction},
  author={Guhr, Oliver and Schumann, Anne-Kathrin and Bahrmann, Frank and B{\"o}hme, Hans Joachim},
  booktitle={Proceedings of the Swiss Text Analytics Conference 2021},
  month={June},
  year={2021},
  address={Winterthur, Switzerland},
  publisher={CEUR Workshop Proceedings},
  url={http://ceur-ws.org/Vol-2957/sepp_paper4.pdf}
}