Luigi/ivod-zhtw-10min-maps

IVOD zh-TW 10-min Extractive Meeting Summaries (MAP)

10-minute periodic meeting summaries in Traditional Chinese (Taiwan), built as MAP targets for a map-reduce meeting summarizer: MAP (this dataset, one bounded summary per 10-min window, ≤512 tokens) → REDUCE (cloud model over map outputs at meeting end).

Source

Speech→transcript base: OpenFormosa/parliament (Taiwan Legislative Yuan IVOD, ~1,286 h, embedded opus audio + transcripts). Gazette metadata from openfun/taiwan-legislator-transcript for speaker enrichment.

Target format (extractive-first, standing design)

• —問→答 minutes skeleton; every bullet cites segment ids [S..]
• —answers may only paraphrase their cited segments
• —unanswered questions are first-class: A: （本段未見回應）
• —關鍵詞： salvage line: every term must occur verbatim in the transcript
• —<|summary_zh_tw|>…</|summary_zh_tw|> marker span

Validation (deterministic, per row)

citations (bigram containment: mis-cite vs hallucination) · roles (only ungrounded titles flagged; quoting is allowed) · self-contained (no 如前所述/承上) · false-「未見回應」 check · truncation (missing 關鍵詞 line) · TW-terms/script · length ≤512 · entity gates (recall ≥0.6 on orgs/numbers/dates; titles reported, weight 0 — IVOD segments carry no inline speaker labels, verified 1/32).

Provenance

target_source: local:gemma4-26b-moe (Gemma-4-26B-A4B NVFP4 GGUF, host, thinking OFF) or teacher:muse-spark-1.3-contributor (repair-only rows). Validators + judge version stored per row; train_eligible is the gate.

Intended use

LoRA targets for a streaming ASR decoder (VibeVoice-ASR-Streaming-7B family) to emit periodic same-language summaries: mixed batches (~4:1 ASR-only), gated on WER+DER+summary jointly. Teacher rows are ablatable via target_source.

Limitations — read before training

• —Transcript text here is the human reference, not your own streaming ASR output (asr_source field). Regenerate text with your ASR before LoRA, or the summarizer trains on cleaner inputs than inference.
• —Single register: formal legislative 質詢. Add ~5-10% conversational zh-TW.
• —IVOD segments have UNKNOWN speakers; role attribution beyond verbatim quotes is unverified.
• —Teacher rows came from a Meta contributor-tier model (prompts trainable).
• —quarantined split documents failures; do not train on it.

Splits

train / validation / control (never train) split by meeting (seed 7); quarantined = ineligible rows with validator reports.