ThakiCloud/Qwen3.8-27B-Satoori-KO-Synth
ThakiCloud/Qwen3.8-27B-Satoori-KO-Synth is a 27 billion parameter Qwen3.8-based model developed by ThakiCloud, specifically designed for Korean regional dialect transformation. This model was uniquely trained using only synthetic colloquial Korean utterances converted to dialect via rule-based transformers, without any direct source corpus text. It excels at transforming standard Korean into various regional dialects and identifying dialect regions, serving as an experimental counterpart to real-data trained models.
Loading preview...
Model Overview
ThakiCloud/Qwen3.8-27B-Satoori-KO-Synth is a 27 billion parameter model built on Qwen3.8, focused on Korean regional dialect transformation. Its unique characteristic is that it was trained exclusively on synthetic data, meaning no actual source dialect text was used in its training corpus. Instead, colloquial Korean utterances generated by another Qwen3.8-27B model were converted to dialect using rule-based transformers.
Key Capabilities
- Korean Dialect Transformation: Converts standard Korean sentences into regional dialects (e.g., Gyeongsang, Jeolla, Jeju, Chungcheong, Gangwon).
- Dialect Identification: Identifies the region of a given dialect sentence.
- Synthetic Data Training: Demonstrates the potential of achieving dialect transformation capabilities without direct access to restricted real-world dialect corpora.
Performance Insights
This model's performance is evaluated against a real-data trained sibling model, Qwen3.8-27B-Satoori-KO, to assess how much can be achieved without restricted data. On the KoDialectBench v0, it achieves 74.6% accuracy in dialect identification and a generation recovery of 72.4% (chrF against human reference), with an axis-balanced mean recovery of 75.8% compared to the real-data model's gain. Notably, it shows improved region differentiation (97%) in out-of-domain registers compared to the real-data model (77%).
Limitations
It is not a conversational model; its training was limited to single-turn transformation and classification. Regional variance is worse than the real-data model, particularly for Gangwon identification, and comprehension recovery sits at approximately 52-54%.