Yoda6131027/twscholar-lm-dataset

Dataset Card — twscholar-lm training data

Summary

A Traditional Chinese (Taiwan) academic-writing-polish dataset: colloquial draft sentences paired with polished, journal-register rewrites, plus a small general-conversation supplement for diversity. Built for supervised fine-tuning of a writing-assistant model on consumer hardware.

476 training examples, merged from three provenance classes.

Provenance breakdown (this is the honest part)

Count	Share	Source	What it means
199	41.8%	human_polished	Claude generated a rough colloquial draft; the human author wrote every polished (academic) target — the actual writing-skill demonstration
150	31.5%	external_licensed	Quality-filtered subset of yentinglin/TaiwanChat (single-turn, 20–220 char, no code, Traditional-purity checked) for general-conversation diversity
127	26.7%	ai_drafted_human_reviewed	Claude drafted both draft and polished; the human reviewed and edited each pair

Why this split is stated plainly: the polished academic targets in the 199 human_polished pairs are the author's own writing, drawn from real domains they work in (sports/volleyball analytics, graph neural networks, eye-tracking, rs-fMRI/neuroimaging). The AI-assisted portions are labeled as such rather than presented as fully hand-written.

License

⚠️ cc-by-nc-4.0 (non-commercial) — because the external_licensed subset derives from TaiwanChat (cc-by-nc-4.0), the merged dataset inherits its non-commercial restriction. The human_polished and ai_drafted_human_reviewed portions alone would be freely relicensable; they are kept in separate files (data/raw/*.jsonl) so a commercial-safe subset can be reconstructed by excluding external_taiwanchat.jsonl.

Categories

The polish pairs span 25 academic-writing situations: abstract polishing, introduction/motivation, literature review, methods, results, figure/table description, discussion/limitations, conclusion, contributions, hypotheses, operational definitions, research questions, cover letters, reviewer responses, advisor emails, acknowledgements, grant-proposal language, colloquial→academic register shift, concision, objective-tone conversion, abbreviation conventions, statistical reporting, and research ethics.

Editorial principles applied to the polish targets

1. No fabrication — polishing changes how something is said, never what is claimed; no adjectives/claims/numbers absent from the draft.
2. Register consistency — third-person academic voice throughout (本研究 / 本文), not mixed with first-person 我們.
3. Reserved words — 顯著 (statistically significant) is used only when the draft actually reports a statistical test; otherwise 明顯/大幅 etc.
4. Traditional Chinese, Taiwan usage — e.g. 藉助/運用 not 借助.
5. Concision over inflation — polishing often shortens; length is not padded to seem formal.

Held-out evaluation set

data/eval/holdout_prompts.jsonl — 50 draft prompts, verified to have zero overlap with the training drafts, for the blind A/B evaluation (M3).

Building the training set

python scripts/build_training_set.py   # -> data/processed/sft_train.jsonl

Pairs are wrapped into messages with a randomly chosen instruction template (5 variants) so the model learns the task, not one fixed prompt.

Known limitations

• Scale — 476 examples is a small SFT set; sufficient for demonstrating the pipeline and the task, not a production-grade corpus.
• rejected/failure coverage (for the DPO stage) is narrow — see the companion dpo-lab investigation.
• Single task — the model is a focused polisher, not a general assistant; the external subset only lightly offsets this.
• AI assistance — ~59% of examples involved Claude in drafting; labeled per-row via source.