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, tieu_de, ...)
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). The corpus is Vietnamese-only — there is no English replay stream.

Sources

Source	Train questions	License	Recast as
UIT-ViSFD (visolex/ViSFD)	85,646	research (UIT)	per-aspect 4-way choice + complaint noul
ViNLI (presencesw/vinli_4_label)	48,726	research (UIT)	relation choice(4) + entailment noul
VSFC (uitnlp/vietnamese_students_feedback)	22,790	research (UIT)	sentiment choice(3) + topic choice(4)
MASSIVE vi-VN (AmazonScience/massive v1.1)	22,069	CC BY 4.0	scenario choice(18) + intent choice(60)

Counts are unique questions after dedup. This release (v1.1) is the four-source scope; earlier revisions also carried ViGLUE, ViANLI, ViCTSD, VSMEC and UTS.

Deduplication

Rows are deduplicated by a canonical key over (normalized state, question type, criteria, instruction, gold label):

• the state is diacritic-stripped and lowercased, and criteria are compared as an unordered option set — so diacritic and option-order differences never produce two rows;
• the instruction is part of the key, because it carries the question's identity (e.g. ViSFD's per-aspect wording) — so distinct questions are not merged;
• dedup runs within each source, then again across the concatenated mixture, and any dev/test question whose key appears in train is dropped (leakage/decontamination);
• finally a state-level check drops any dev/test row whose state text (diacritic-stripped, lowercased) exactly or near-duplicates (MinHash, Jaccard >= 0.8) a train state, even under a different question.

The mixture is not materialized with augmentation. There are no paraphrase, option-shuffle, or diacritic-stripped duplicate rows on disk; diacritic stripping, option permutation, and instruction sampling belong at TypedDecisionDataset load time.

Splits

File	Cases	Size
data/train.jsonl	179,231	222 MB
data/dev.jsonl	25,098	32 MB
data/test.jsonl	42,062	53 MB
data/dev_stripped.jsonl, data/test_stripped.jsonl	same rows, diacritic-stripped state	31 / 51 MB

Evaluation reports both the accented and the stripped copy. Only choice (82%) and noul (18%) questions are present; option counts are 2, 3, 4, 18 and 60.

Intended use

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

Limitations

• Recast labels are one-hot; score-type data is deferred to v2.
• Narrow domains (student feedback, phone reviews, NLI, voice commands); short texts (median ~150 characters); some labels are skewed (e.g. entailment 75% false).
• MASSIVE intent descriptions are composed from label keys; a native translation pass is pending. MASSIVE ships the full intent set (K=60).
• Dedup is exact on the canonical key; semantically equivalent questions with different wording across sources are kept separate.

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 and processed data are regenerated from scripts and never committed to the code repo.