ThakiCloud/Qwen3.8-27B-Satoori-KO-Synth

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 primary distinction is its training methodology: it was trained entirely on synthetic data, meaning no actual source dialect text was used. Instead, colloquial Korean utterances generated by an in-house model were transformed into dialects using rule-based systems. This model was developed in parallel with a real-data counterpart (Qwen3.8-27B-Satoori-KO) to explore the effectiveness of synthetic data in specialized linguistic tasks.

Important Note: This is not a conversational model. Its training data consists solely of single-turn transformation and classification tasks, so it is not designed for or measured on conversational fluency in dialect.

Key Capabilities

• Korean Dialect Transformation: Converts standard Korean sentences into specific regional dialects (e.g., Gyeongsang, Jeolla, Jeju, Chungcheong, Gangwon).
• Dialect Identification: Accurately identifies the regional origin of a given dialect sentence.
• Synthetic Data Training: Demonstrates the potential of training specialized models without direct access to restricted real-world data, relying on rule-based transformations and generated utterances.
• Out-of-Domain Register Handling: Shows improved performance in differentiating dialects across formal registers (e.g., customer service) compared to its real-data counterpart, which tended to collapse on such inputs.

Performance Highlights

Evaluated on KoDialectBench v0, the model shows strong performance in dialect identification and generation, with an axis-balanced mean recovery of 75.8% compared to the real-data model. While comprehension recovery is lower (52.7%), its identification recovery is high at 91.2%. Notably, in out-of-domain register tests, the synthetic model achieved 97% region differentiation, outperforming the real-data model (77%).

Good for

• Linguistic Research: Ideal for researchers investigating the efficacy of synthetic data generation and rule-based transformation in specialized NLP tasks, particularly for low-resource or restricted-access linguistic phenomena.
• Dialect-Specific Applications: Useful for applications requiring the transformation of standard Korean into regional dialects or the identification of dialect regions, especially where conversational ability is not a primary requirement.
• Exploring Data Scarcity Solutions: Provides a case study for developing models when direct access to large, diverse, and annotated real-world datasets is limited or unavailable.