Mojicast Punct

Mojicast Punct is a Japanese punctuation restoration model provided in ONNX format. It is designed to restore punctuation marks, specifically the Japanese comma and period, in Japanese text at the character token level. The model is a conversion of a PyTorch model that combines the base from tohoku-nlp/bert-base-japanese-char-v3 and punctuation prediction weights from bobfromjapan/bert_japanese_punctuation. The ONNX version is exported with opset 17 and uses fp32 precision, packaged as a single file of approximately 364MB. The accompanying vocabulary file is identical to that used in the original BERT character model from tohoku-nlp. The model takes input in the form of input_ids and attention_mask, both as int64 character-level tokens, and outputs logits representing the likelihood of inserting punctuation after each token. The output logits can be processed with a sigmoid function and thresholding to determine punctuation placement. Mojicast Punct is used in the real-time captioning application Mojicast, where it assists in automatically restoring punctuation to Japanese text streams. This tool is licensed under the Apache-2.0 license, following the licenses of its component models. The evidence does not specify target users beyond its application in Mojicast, but its format and origin suggest it is intended for integration into Japanese language processing workflows that require punctuation restoration. Mojicast Punct is not deployed by any inference provider as of the evidence provided, and there is no information on pricing beyond the open-source license.