audreyt/sayit-archive-tw

Audrey Tang Transcript Corpus

Public transcripts of Audrey Tang (唐鳳) — 2025 Right Livelihood Laureate, civic hacker, and Taiwan's Cyber Ambassador.

Tang is co-author of Plurality: The Future of Collaborative Technology and Democracy and an inaugural Senior Accelerator Fellow at the Oxford Institute for Ethics in AI. She served as Taiwan's first Digital Minister (2016–2024) and the world's first nonbinary cabinet minister, awarded the Right Livelihood Award for "advancing the social use of digital technology to empower citizens, renew democracy and heal divides."

A child prodigy who practiced Taoism to manage a congenital heart condition, Tang left formal schooling at 14. By 24, she was a leader in the free and open-source software communities, revitalizing the Haskell and Perl languages, and a computational linguistics consultant to Apple. She was instrumental in the creation of g0v (gov-zero) and played a pivotal role in the 2014 Sunflower Movement, facilitating digital consensus-building during the occupation of Taiwan's legislature.

As Digital Minister, Tang implemented radical transparency and participatory democracy platforms like vTaiwan and Join.gov.tw. Her tenure saw public trust in Taiwan's government rise from single digits to over 70%, driven by innovations such as the real-time Mask Map during COVID-19 and pioneering defenses against cyber interference in the 2024 elections. She describes her philosophy as becoming a "good enough ancestor," striving to leave future generations a wider canvas for democratic possibility.

Dataset Summary

Statistic	Value
Transcripts parsed	1,931 (from 1,992 markdown files)
Date range	2015-07-22 to 2026-02-20
Total speaker turns	213,366
Audrey Tang turns	82,011
SFT training pairs	59,518 (en: 27,141 / zh: 32,377)
RAG chunks	85,094
Bilingual terms	765
Languages	English (642 files), Traditional Chinese (1,289 files)

Topics Covered

Radical transparency, digital democracy, civic tech, AI governance, participatory democracy, vTaiwan, Join.gov.tw, COVID-19 digital response (Mask Map, 1922 SMS, prebunking), plurality, techno-communitarianism, the 6-Pack of Care, g0v, Sunflower Movement, open-source software, international relations, cyber defense, deepfake regulation, alignment assemblies, policy coordination.

Dataset Configurations

turns

Individual speaker turns from all transcripts.

Field	Type	Description
id	string	Unique turn ID ({transcript}/{index})
date	string	ISO date of the transcript
title	string	Transcript title
source_file	string	Source markdown filename
speaker	string	Speaker name
text	string	Turn text content
turn_index	int	Position within the transcript
is_audrey	bool	Whether the speaker is Audrey Tang
language	string	en or zh
stage_directions	list[string]	Stage directions / annotations
total_turns	int	Total turns in the source transcript

conversations

Full transcripts structured as multi-turn conversations with human/assistant roles (Audrey Tang as assistant).

Field	Type	Description
id	string	Transcript ID
date	string	ISO date
title	string	Transcript title
language	string	en or zh
participants	list[string]	All speakers in the conversation
messages	list[object]	{role, speaker, content} message objects

sft_pairs

Instruction-response pairs for supervised fine-tuning, where the instruction is the preceding conversational context and the response is Audrey Tang's reply.

Field	Type	Description
id	string	Unique pair ID
date	string	ISO date
title	string	Transcript title
language	string	en or zh
instruction	string	Conversational context (preceding turns)
response	string	Audrey Tang's response
context_turns	int	Number of context turns in the instruction
source_file	string	Source markdown filename

chunks

RAG-optimized chunks of Audrey Tang's responses with surrounding context, suitable for embedding and retrieval.

Field	Type	Description
id	string	Unique chunk ID
date	string	ISO date
title	string	Transcript title
language	string	en or zh
speaker	string	Speaker name (唐鳳 or Audrey Tang)
question	string	Preceding question or prompt
text	string	Audrey Tang's response text
context_before	string	Preceding conversational context
token_estimate	int	Estimated token count
source_file	string	Source markdown filename

lexicon

Bilingual EN↔ZH terminology across AI, governance, tech, and society domains.

Field	Type	Description
id	int	Term ID
en	string	English term
zh	string	Traditional Chinese term
domain	string	Domain category

Source and License

The transcripts are sourced from sayit.archive.tw, published as records of a public office holder's official activities under CC0 1.0 Universal — no rights reserved. This dataset inherits the same dedication to the public domain.

Intended Uses

• Fine-tuning grounded chatbots — Use the sft_pairs config with RAG retrieval from chunks to build conversational agents grounded in Audrey Tang's policy positions and communication style. The TAIDE model (Gemma-3-TAIDE-12b-Chat-2602) is the recommended base model for fine-tuning.
• RAG retrieval — The chunks config provides pre-segmented, context-rich passages for embedding and retrieval-augmented generation.
• Digital democracy research — Study civic tech policy language, open government discourse, and participatory democracy practices across a decade of public service.
• Bilingual NLP — The corpus spans English and Traditional Chinese with parallel terminology in the lexicon config.

Citation

@dataset{audrey_tang_transcripts,
  title={Audrey Tang Transcript Corpus},
  author={Tang, Audrey},
  year={2026},
  url={https://sayit.archive.tw},
  license={CC0-1.0}
}