ThakiCloud/Qwen3.8-27B-Satoori-KO
ThakiCloud/Qwen3.8-27B-Satoori-KO is a 27 billion parameter language model, based on the Qwen3.8 architecture, specifically designed for Korean regional dialect transformation. It was trained directly on real dialect corpora from five regions (Gangwon, Gyeongsang, Jeolla, Chungcheong, and Jeju) and serves as a real-data reference for dialect transformation tasks. This model excels at single-turn tasks including dialect comprehension (converting dialect to standard Korean), dialect identification, and dialect generation (converting standard Korean to a specified regional dialect). It is not a conversational model.
Loading preview...
Overview
ThakiCloud/Qwen3.8-27B-Satoori-KO is a 27 billion parameter model, built on the Qwen3.8 architecture, specialized in Korean regional dialect transformation. It covers five distinct regions: Gangwon, Gyeongsang, Jeolla, Chungcheong, and Jeju. This model was trained directly on authentic dialect corpora, positioning it as a crucial real-data reference arm for dialect transformation, particularly when compared against its synthetic-data counterpart, Qwen3.8-27B-Satoori-KO-Synth.
Key Capabilities
- Dialect Comprehension: Transforms dialect sentences into standard Korean.
- Dialect Identification: Determines the specific region of a given dialect sentence.
- Dialect Generation: Converts standard Korean sentences into a specified regional dialect.
- High Performance: Achieves strong results on KoDialectBench v0, with 80.1 chrF for comprehension and 79.4% accuracy for identification.
Good For
- Research and Benchmarking: Ideal for researchers studying Korean dialects and for benchmarking synthetic dialect models against real-world data.
- Single-Turn Dialect Tasks: Suited for applications requiring direct, non-conversational dialect transformations.
- Linguistic Analysis: Useful for analyzing and understanding the nuances of Korean regional dialects.
Limitations
It is important to note that this is not a conversational model; its training data is exclusively single-turn transformation and classification. The model's dialectness metric is a lower bound, and it processes only transcribed text, lacking pronunciation-level information. Performance is measured on Qwen3.8-27B, and results at other scales are unverified. For more details, refer to the associated paper: Metric-Construction Coupling Inflates Measured Synthetic Dialect Recovery.