eg2-ko-reader / NOTICE
topabaem's picture
EG2-KO Reader v1.3 (source 1c047ba75986)
a3adeb1 verified
Raw History Blame Contribute Delete
1.69 kB
EG2-KO Reader
Copyright (c) 2026 Haverbex.co
The reader (reader/, coreai/) and eg2_ko_reader.py are licensed under the EG2-KO Reader
License 1.0 (LICENSE).
Encoder
-------
encoder/ holds an unmodified copy of files from google/embeddinggemma-2 (EmbeddingGemma 2),
revision 914f7f89142e33e77833254d9c9b90c3cef7303b, licensed under the Apache License,
Version 2.0. See encoder/LICENSE and encoder/NOTICE.
The Core AI reader bundle expects the encoder bundle macos/eg2_audio_fp16_T3000.aimodel of
mlboydaisuke/EmbeddingGemma-2-Transcribe-CoreAI (Apache-2.0). That bundle is not included.
Training data
-------------
The reader was trained on EG2 encoder outputs of these data sets. No audio and no transcript
is included in this repository.
* Zeroth-Korean (OpenSLR SLR40), Lucas Jo and Wonkyum Lee, via the Hugging Face dataset
kresnik/zeroth_korean. License: Creative Commons Attribution 4.0 International (CC BY 4.0).
* FLEURS (Conneau et al., 2022), Korean (ko_kr) subset, via the Hugging Face dataset
google/fleurs. License: CC BY 4.0.
* YODAS (Li et al., 2023), Korean shards, via the Hugging Face dataset espnet/yodas2: audio
from YouTube videos that were published under the Creative Commons Attribution license.
License: Creative Commons Attribution 3.0 (CC BY 3.0).
* Korean ASR pseudo-labels for YODAS2, via the Hugging Face dataset nhatminh/korean-asr:
segment boundaries and transcripts generated with Qwen/Qwen3-Omni-30B-A3B-Instruct.
License: CC BY 3.0.
Changes to the data: Zeroth-Korean and FLEURS were split again by sentence. Of the YODAS
segments, only those whose pseudo-label agrees with the original caption (character error
rate at most 10 %) were used.