Tw Instruct 500K: 500,000 Synthetic Task-Oriented Dialogues for Taiwan Mandarin

500,000 synthetic task-oriented dialogues in Traditional Chinese, tailored for Taiwan's social context. The dataset was created by author lianghsun, combining reference-based and reference-free generation methods using LLMs, with a subset derived from the lianghsun/tw-instruct collection. It was last updated on January 10, 2025.

Use Cases

• Fine-tuning large language models for task completion based on the described synthetic dialogue structure.
• Training conversational AI assistants for Taiwan-specific domains and social contexts mentioned in the description.
• Benchmarking model performance on generating or understanding Traditional Chinese task-oriented dialogues.
• Studying the characteristics of LLM-generated synthetic data for instruction tuning.

Strengths

• Contains 500,000 dialogue instances, providing a substantial volume for model training.
• Specifically targets Taiwan's social context and common tasks, offering regional linguistic relevance.
• Combines both reference-based and reference-free generation methods, as described, which may enhance diversity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the source texts and generation methods used.

Provenance

Source

huggingface

Collection Method

Synthetic generation by LLMs, combining reference-based (using texts from training lianghsun/Llama-3.2-Taiwan-3B) and reference-free methods.

Freshness

Last updated 2025-01-10 05:03:52; freshness should be verified.

Geography

Taiwan