float32/README.en.md

[日本語](README.md)

# SBV2 Core ML shared model: original FP32 (float32)

This folder contains the **original FP32 (float32)** shared BERT and Open JTalk dictionary, approximately 1.52 GB.
BERT uses the original unquantized FP32 weights. BERT computation and tensor interfaces remain FP32.
Supply a separate compatible voice model. Both shared variants use the same voice models.

## Download and use

Choose this variant's `download.json` from the [model selection table](https://huggingface.co/AILogDev/sbv2-coreml-common).
If downloading folders, select this `float32/` root in the sample's **モデル設定 → 共通モデル**, then press **モデルを準備**.
For the SDK, pass its `bert/` and `dictionary/` subfolders and a separate voice folder.
Compilation caches require additional storage.

## Attribution and license

BERT derives from Kyoto University's NLP group's [Japanese DeBERTa](https://huggingface.co/ku-nlp/deberta-v2-large-japanese-char-wwm).
See this folder's `LICENSE.md` for CC BY-SA 4.0 terms and `dictionary/COPYING` for separate dictionary terms.
Provenance is in `provenance.json`; file hashes are in `checksums.json`.
This unofficial conversion is not endorsed by the original authors.

This is the original converted Core ML BERT without the INT8 weight quantization, not the PyTorch checkpoint.