Tw Instruct 500K 2511

tw-instruct-500k-2511 is a 2025 November version of a synthetic dialogue dataset for training Taiwanese Mandarin conversational models. It combines reference-based and reference-free generation methods to produce instruction-response pairs aligned with Taiwanese context. The dataset was created by lianghsun and updated on HuggingFace in May 2026.

Use Cases

• Supervised fine-tuning (SFT) of Taiwanese Mandarin dialogue models based on the dataset's stated purpose.
• Training models on culturally contextual responses based on reference-based generation from Taiwanese texts.
• Generating reference-free instruction-response pairs based on the Self-Instruct methodology mentioned.
• Cleaning and standardizing dialogue data based on the described process of removing overly long, short, or duplicate samples.
• Adopting OpenAI messages schema for dialogue formatting based on the dataset's unified structure.

Strengths

• Dataset format is unified to the OpenAI messages schema, suggesting a standardized structure.
• Data has been cleaned to remove overly long, short, and duplicate samples.
• Includes both reference-based and reference-free instruction samples, indicating a multi-source generation approach.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

lianghsun on HuggingFace.

Collection Method

Synthetic generation via reference-based (using Taiwanese texts) and reference-free (using Self-Instruct) methods.

Freshness

Last updated 2026-05-03 21:19:02; freshness should be verified.

Geography

Taiwan (based on the focus on Taiwanese context and texts).