PhoLaya-base — Vietnamese typed-decision mixture

Typed-decision training data for PhoLaya, a Vietnamese choice / score / noul model fine-tuned from convaiinnovations/laya-multilingual (mmBERT-base + Laya decision head).

Not affiliated with VinAI (PhoBERT) or Convai Innovations (Laya).

Each row is one question about a state, in the Laya schema:

Column	Contents
state	JSON string; Vietnamese keys (binh_luan, doan_van, cau_noi, ...)
questions	JSON string {qid: {type, instructions, criteria}}; ASCII snake_case option keys, Vietnamese descriptions
gold	JSON string {qid: {label, probabilities}}

type is choice / score / noul. probabilities is one-hot for recast sources (soft targets are a v2 item).

Sources

Source	Rows (base questions)	License	Used here
UIT-ViSFD (visolex/ViSFD, LuongPhan/UIT-ViSFD)	122,342	Research-only (UIT)	per-aspect 4-way choice + complaint noul
UIT-VSFC (uitnlp/vietnamese_students_feedback)	32,350	Research-only (UIT)	sentiment choice(3) + topic choice(4)
ViNLI (presencesw/vinli_4_label)	60,752	Research-only (UIT)	relation choice(4) + entailment noul
MASSIVE vi-VN (AmazonScience/massive, v1.1)	33,042	CC BY 4.0	scenario choice(18) + intent choice(60)

The corpus is Vietnamese-only; there is no English replay stream. The base corpus is ~250k questions. The released data/*.jsonl is the augmented mixture (see below), ~1.8 GB across the three splits.

Recast rules

• ViSFD: for each of 10 aspects, a 4-way choice {positive, negative, neutral, none} (none = aspect not mentioned); plus a noul for "does this comment contain a complaint?".
• VSFC: sentiment {negative, neutral, positive} and topic {lecturer, program, facility, others} over the same comment.
• ViNLI: relation {entailment, contradiction, neutral, other} over doan_van (premise) + cau_noi (hypothesis); plus an entailment noul.
• MASSIVE: scenario over 18 domains and intent over 60 labels; option descriptions are Vietnamese (scenario name + action suffix).

Augmentation (this release)

The mixture materializes, per base question:

• up to 3 Vietnamese instruction paraphrases (template-based — not LLM-generated);
• diacritic-stripped state copies for ~30% of train questions (test is scored on both accented and stripped copies);
• 2 label-preserving option-order permutations (train split only).

Mixture ratios (by base question count): ViSFD 0.30 / ViNLI 0.37 / VSFC 0.18 / MASSIVE 0.15. Under-sized sources are oversampled to hold the ratio.

Per split (augmented):

Split	Rows	Size
train	1,622,502	1.59 GB
validation	76,446	76 MB
test	128,181	128 MB

Decontamination

Records carry a group id at build time (comment id / premise SHA-1 / utterance id); the plan requires that no test state hash (accented or stripped) appears in train/calibration. The base shards under sources/ retain canonical order for re-checking; the augmented rows are derived.

Intended use

Training and evaluating Vietnamese typed-decision models and studying calibration / budget ablations. Not for production decisions.

Limitations

• Recast labels are one-hot; score data is deferred to v2.
• Instruction paraphrases are templates, not natural paraphrases.
• MASSIVE intent descriptions are composed from label keys; a native translation pass is pending.
• The mixture duplicates each base question across augmentation variants — it is a view of a ~250k-question corpus, not 1.6M distinct questions.

License & attribution

license: other (research-use-only): the mixture contains derivatives of the UIT corpora (ViSFD, VSFC, ViNLI), which are research-licensed. Do not use commercially without contacting the dataset authors. MASSIVE vi-VN is CC BY 4.0. Attribute each source per its own license.

Reproduce with scripts/data/download.py + scripts/data/prepare.py; raw data is never committed to the code repo.