Miyauel/meari-dialect-lora

Meari dialect LoRA — standard Korean → regional dialect (LMT-60 1.7B)

LoRA adapter in GGUF format (for llama.cpp --lora) that rewrites standard Korean as one of five regional dialects: 경상 Gyeongsang, 전라 Jeolla, 충청 Chungcheong, 강원 Gangwon, 제주 Jeju. One adapter covers all five regions. Used by Meari (local speech engine, MIT) to read text in a regional dialect.

Base model: NiuTrans/LMT-60-1.7B (Apache-2.0). Meari runs the GGUF mradermacher/LMT-60-1.7B-GGUF (Q4_K_M) with this adapter.

Prompt

The same chat message as in training (thinking off). Region names: Gyeongsang, Jeolla, Chungcheong, Gangwon, Jeju.

Translate the following text from Korean into Gyeongsang dialect Korean.
Korean: <standard Korean sentence>
Gyeongsang dialect Korean:

With the llama.cpp server build Meari uses, the adapter is applied to every request that does not set lora — send "lora": [{"id": 0, "scale": 0}] with normal translation requests and the region's scale from dialect.json for dialect requests.

Strength per region (dialect.json)

At scale 1.0 the adapter changes only about half as many sentences as real speakers do, so dialect.json sets a scale per region (missing regions: 1.0). Chosen by measurement on the same 1,500 sentences: the scale whose share of changed sentences is closest to real speakers, as long as chrF++ on the sentences real speakers changed does not fall more than 1 point below scale 1.0. Each cell is "share of sentences changed · chrF++ on the sentences real speakers changed".

Region	Real speakers	Scale 1.0	1.25	1.5	1.75	Chosen
Gyeongsang	59%	30% · 79.2	63% · 78.8	82% · 71.6	92% · 62.3	1.25
Jeolla	59%	28% · 83.2	46% · 83.3	67% · 77.4	80% · 68.3	1.25
Chungcheong	55%	28% · 84.4	47% · 84.0	67% · 80.1	80% · 73.0	1.25
Gangwon	35%	17% · 84.5	28% · 83.4	54% · 79.0	72% · 67.1	1.0
Jeju	97%	96% · 68.5	99% · 64.0	100% · 46.8	100% · 28.3	1.0

Results

1,500 held-out sentences (300 per region), chrF++ against the real dialect transcript, at scale 1.0. "Unchanged" keeps the standard sentence; "word list" replaces words with a dictionary built from the same training data.

Region	Unchanged	Word list	This adapter
Gyeongsang	83.3	84.0	85.4
Jeolla	85.0	86.6	88.7
Chungcheong	86.0	88.4	90.5
Gangwon	89.2	91.1	92.8
Jeju	40.6	59.2	69.5

Training: LoRA rank 16 (alpha 32, dropout 0.05) on all attention and MLP projections, learning rate 2e-4, batch 8, 10,000 steps on CPU, sentence pairs of standard and dialect transcripts.

Data notice

이 모델은 AI 허브(aihub.or.kr)의 「중·노년층 한국어 방언 데이터(강원도, 경상도)」·「중·노년층 한국어 방언 데이터(충청도, 전라도, 제주도)」의 전사로 학습했습니다. 본 결과물은 한국지능정보사회진흥원의 사업결과임.

Trained on transcripts of the AI Hub datasets "Korean dialect speech of middle-aged and elderly speakers". This is a result of a project of the National Information Society Agency of Korea (NIA). No transcripts of the dataset are included.

License

MIT for this adapter. The base model is Apache-2.0.