---
license: other
language:
- th
task_categories:
- text-classification
tags:
- typed-decisions
- laya
- thai
- system-one
pretty_name: Laya Thai Trainset (typed decisions)
---

# Laya Thai Trainset

Training data for fine-tuning [`convaiinnovations/laya`](https://huggingface.co/convaiinnovations/laya)
on Thai typed decisions. Schema-compatible with the upstream
[`typed-decisions` training notebook](https://github.com/NandhaKishorM/laya/blob/main/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb).

- **`trainset.jsonl` — 37,091 rows** (current canonical round: C1, 2026-09-26 — v2-exact base 31,091 + Wongnai reviews 6,000; MASSIVE-th + oversample excluded after B1/B2 ablations), typed-decision format (`id`, `workflow`, `state`, `questions`, `gold`); all compound fields are JSON-encoded strings.
- Sources: Wisesight sentiment **train** split (21,628) + XNLI-th **train** stratified subsample (8,000, seed 42, de-segmented) + handwritten Thai intent cases (133 × 11 = 1,463) + Wongnai restaurant reviews **train** split (6,000 stratified subsample, seed 42: neg 2,000 / neu 2,000 / pos 2,000; star mapping 1-2→neg, 3→neu, 4-5→pos).
- Built reproducibly by `build_trainset.py` + `build_wongnai_reviews.py` (seed 42); question wording copied verbatim from the project's eval harness (train wording == gate wording).
- **No PII.** Derived from public datasets [`pythainlp/wisesight_sentiment`](https://huggingface.co/datasets/pythainlp/wisesight_sentiment), [`facebook/xnli`](https://huggingface.co/datasets/facebook/xnli) (th), [`Wongnai/wongnai_reviews`](https://huggingface.co/datasets/Wongnai/wongnai_reviews) (LGPL-3.0; used for research fine-tuning), plus handwritten items. Respect the source datasets' licenses.
- Leakage: disjoint from the project's frozen 100-item evaluation gate (0 overlap, verified at build time). Wisesight **test** / XNLI-th **validation** splits are excluded; Wongnai **test** split unused (unlabeled).