hagsmand1/laya-thai-decisions

Laya Thai Typed Decisions

A Thai dataset of typed decision questions. Each row is a piece of text (a chat message, product review, social post, news article, Wikipedia passage, or support ticket) plus one or more questions about it. Every question has a fixed type:

• choice: pick one option from a labelled set
• score: pick a level on an ordered scale
• noul: answer true or false

Each question comes with a gold label, the probability distribution an LLM teacher assigned to its options, and a soft target that blends the two. The format follows LocalLLaMA/typed-decisions. Laya (convaiinnovations/laya) and other models that score options given a state and an instruction can read it directly.

What the dataset provides:

• Thai decision data from real and synthetic sources:
  • intent and scenario (MASSIVE)
  • sentiment and question detection (wisesight)
  • review ratings (wongnai)
  • headline and topic (thaisum)
  • reading comprehension (iapp)
  • eight synthetic business workflows: support routing, urgency, frustration, churn, escalation, moderation
  • Thai NLI
• Soft labels: a teacher distribution for every Thai question in train. They are useful for distillation, label-noise analysis, and calibration studies.
• Non-Thai rows in the same format: English typed-decisions, and MASSIVE in zh-CN, ja, ko, ar, hi, and ru, for mixing with the Thai data.
• A calibration set of Thai gold questions, disjoint from train.
• 17 frozen eval suites covering Thai accuracy, unseen instruction phrasings, unseen MASSIVE scenarios, other languages, and behavioural probes.
• An unfiltered view of every question before quality filtering, with the reason each one was kept or dropped.

Size: train has 41,776 questions on 27,250 rows (71% Thai), calibration has 2,448 questions, and the eval suites have 9,161 cases.

This is a first release with known weaknesses. The audit lists them, and TASKS.md turns each one into a contributor-sized task.

Quick start

import json
from datasets import load_dataset

train = load_dataset("hagsmand1/laya-thai-decisions", split="train")
row = train[0]
state, questions = json.loads(row["state"]), json.loads(row["questions"])
target = json.loads(row["target"])          # {qid: {option_key: prob}}, gold blended with the teacher

calib = load_dataset("hagsmand1/laya-thai-decisions", split="calibration")
unfiltered = load_dataset("hagsmand1/laya-thai-decisions", "unfiltered", split="rows")
probe = load_dataset("hagsmand1/laya-thai-decisions", "eval", split="probe_noul_negation")

Nested fields are stored as JSON strings because their keys change from row to row (option keys are Thai words, English words, or letters). Decode them with json.loads.

Configs and splits

Config	Split	Rows	Questions	What it is
default	train	27,250	41,776	Questions that passed the quality filters. Each has a gold label, a teacher distribution, and a soft target. 71% Thai; the rest are English and multilingual rows in the same format
default	calibration	2,000	2,448	Thai gold questions from the five public sources, disjoint from train, with one-hot targets. Use it to measure or fit confidence calibration
unfiltered	rows	31,042	50,038	Every question before filtering, with the filter outcome for each. Use it to audit or change the filters
eval	17 splits	9,161 cases		Frozen test suites; see Evaluation suites

unfiltered rows have a role:

Role	Rows	Questions	Meaning
train	27,702	44,691	Source rows of the train split before filtering. 452 rows lost every question, so train has 27,250
calibration	2,000	2,448	The same rows as the calibration split
reshuffle_check	1,040	1,999	Copies of synthetic rows with the options reordered. A question was dropped when its teacher distribution changed with the order
eval_only	300	900	Source rows of synth_eval_domains.th (insurance claims, school admin) before filtering. Do not train on these

Data fields

default and unfiltered

Field	Type	Description
id	string	<source>:<upstream id>; unique
source	string	massive_th, wisesight, wongnai, thaisum, iapp, synth_<domain>, synth_nli, typed_decisions, massive_<lang>
lang	string	Language of the text (th, en, zh-CN, ...). Instructions can be Thai or English independently
text	string	The natural-language text the questions are about (for search, dedupe, and overlap checks)
state	JSON string	What the model reads: a plain string or an object such as {"ticket_id": ..., "message": ...}. Some states add realistic distractor fields
questions	JSON string	{qid: question}; see below
gold	JSON string	{qid: {"type", "label", "probabilities"}}. One-hot for public data; the generator's intended label for synthetic rows; upstream soft labels for typed_decisions
teacher	JSON string or null	Raw teacher distribution {qid: {key: prob}} before temperature. Null for the English and non-Thai MASSIVE rows
target	JSON string or null	Soft target {qid: {key: prob}}: gold blended with the teacher (see Soft targets and filtering). In calibration it equals the gold one-hot
role	string	unfiltered only; see above
decision	JSON string or null	unfiltered only: {qid: reason}, the filter rule that applied (table below)
teacher_adjudication	JSON string or null	unfiltered only: second teacher pass at higher effort on hard MASSIVE disagreements
teacher_reshuffle	JSON string or null	unfiltered only: teacher distribution with the options in a different order (synthetic rows)

Question types

Instructions refer to state fields in backticks, for example How urgent is `message`?.

{"type": "choice", "instructions": "`message` ควรอยู่ในคิวใด",
 "criteria": {"A": "เปลี่ยนหรืออัปเกรดแพ็กเกจ", "B": "บิล ค่าบริการ เติมเงิน การชำระเงิน", "C": "มือถือ เราเตอร์ ซิมการ์ด"}}

{"type": "score", "instructions": "How urgent is `message`?",
 "criteria": ["urgent or blocking", "should be handled soon", "routine, handle in the normal queue", "no time pressure"]}

{"type": "noul", "instructions": "Does the writer of `message` threaten to cancel or leave?",
 "criteria": {"false": "ไม่ได้ขู่ยกเลิก", "true": "ขู่ว่าจะยกเลิกหรือเลิกใช้"}, "labels": {"false": "no", "true": "yes"}}

• choice: option keys are letters, English slugs, or Thai words, and options are shuffled. Some questions add a "none of the other options fit" option.
• score: an ordered list of levels. The label is the level index as a string ("0", "1", ...), and the direction varies: sometimes level 0 is the worst, sometimes the best.
• noul: keys are always false / true; the optional labels map changes how they are displayed.

train by question type: 19,916 choice, 13,712 noul, 8,148 score.

eval

Field	Type	Description
suite	string	Original suite name, e.g. massive_intent.zh-CN
index	int	Case position in the frozen file
language	string	Language of the text
tags	list of strings	e.g. heldout-template, probe, regression
state, questions	JSON string	As above. Probe cases put two paired questions on one state
expected	JSON string	{qid: answer}: option index for choice and score, boolean for noul

eval/manifest.json records the sha256 of each original frozen file, so results stay comparable across versions. Suites are never edited in place; a changed suite gets a new versioned name.

How it was built

Sources

Source	Rows (train)	Questions (train)	Calibration rows	Tasks	License
MASSIVE th	3,455	4,852	513	intent / scenario choice, "asks for X" noul	CC BY 4.0
wisesight_sentiment	2,317	3,178	305	sentiment choice and score, "is a question" noul	CC0 1.0
wongnai_reviews	2,550	3,408	338	star score, sentiment score, "is positive" noul	LGPL-3.0
thaisum	3,976	3,976	524	headline choice, headline match noul, topic choice	MIT (articles: publisher copyright)
iapp_wiki_qa_squad	2,630	2,630	320	answer-span choice, answerable noul	MIT; passages CC BY-SA
Synthetic workflows (8 domains)	2,934	7,819		department, urgency, frustration, churn, needs-human, severity, ...	CC BY 4.0
Synthetic NLI	3,932	3,932		premise / hypothesis relation choice, "implies" noul	CC BY 4.0
typed-decisions (typed_decisions, en)	1,200	6,000		English workflow decisions: urgency, routing, risk, review flags	Apache-2.0
MASSIVE other languages (massive_<lang>: zh-CN, ja, ko, ar, hi, ru)	4,256	5,981		intent / scenario choice, "asks for X" noul	CC BY 4.0

Synthetic workflow domains: telco support, bank support, e-commerce support, content moderation, LINE chat, HR email, government service, and code-mixed Thai-English reviews.

• Public rows come from the upstream train splits.
• Any row whose normalised text matches a case in any eval suite (including suites not published here) was dropped.
• Six MASSIVE scenarios (cooking, email, lists, music, news, social) are absent from train and calibration. They appear only in massive_intent.th.heldout_scenarios.
• One instruction template per task is reserved for the *.sampler_eval suites, so those suites use phrasings that never occur in train.

Teacher labels

Every Thai public and synthetic question was labelled by an LLM teacher (gpt-5.6-terra). The teacher was asked for a calibrated probability over the options, with choice options shown in shuffled order and the output constrained by a JSON schema (low reasoning effort, four states per request). The same model generated the synthetic workflow and NLI states, each with an intended label.

On a 300-row pilot, the teacher agreed with gold 81% of the time (MASSIVE 91%, iapp 98%, thaisum 93%, wisesight 63%, wongnai 67%). Most of the low agreement on wisesight and wongnai traces back to noisy crowd labels. For the soft targets, the teacher distribution is first softened with a single temperature T = 1.447, fitted on that pilot. The teacher column holds the raw, unsoftened distribution.

Soft targets and filtering

target = floor(alpha * onehot(gold) + (1 - alpha) * teacher_T, 0.01), renormalised. The weight alpha on gold depends on how well gold and teacher agree. A question is dropped when they conflict and the gold label is unreliable. The decision column in unfiltered records which rule applied:

decision	Applies to	alpha	Questions
agree	teacher argmax == gold	0.5	15,694
soft_disagree	teacher p(gold) >= 0.2, or for score >= 0.5 of the mass within ±1 of gold	0.7	1,846
structural_gold	hard disagreement on thaisum / iapp, where gold is a real headline or span	0.9	272
adjudicated_keep / adjudicated_drop	hard MASSIVE disagreement, re-asked at higher effort; kept if p_adj(gold) >= 0.1	0.85 / dropped	232 / 200
noisy_drop	hard disagreement on wisesight / wongnai	dropped	516
kept	synthetic, teacher argmax == intended label; target is the floored teacher distribution	teacher only	11,751
generator_disagree	synthetic, teacher argmax != intended label	dropped	2,169
reshuffle_unstable	synthetic, teacher distribution moved by total variation > 0.3 when options were reordered	dropped	22
replay	typed_decisions: upstream soft target, floored; massive_<lang>: gold smoothed by 0.05	n/a	11,981

The calibration split is 2,000 Thai public rows, chosen so that every (question type, option count) bucket has at least 200 questions.

Evaluation suites

All suites come from upstream test splits or from synthetic domains absent from train. None of their texts appear in train or calibration.

Split	Cases	Measures
massive_intent_th, massive_intent_th_full	100, 2,974	Thai intent, upstream 20-option protocol
massive_intent_th_heldout_scenarios	796	Thai intents from the six scenarios absent from train
massive_intent_{ar,hi,ja,ko,ru,zh_CN}	100 each	Intent in six other languages, to check that non-Thai ability holds up
massive_th_sampler_eval, wisesight_sampler_eval, iapp_sampler_eval	1,092, 1,052, 647	Questions in the train format, phrased with instruction templates absent from train
probe_noul_negation	400 pairs	A claim and its negation on one state; predictions must flip
probe_noul_labels	400 pairs	Default false/true labels vs. relabelled A/B; predictions must agree
probe_score_orientation	400 pairs	A 3-level scale in normal and reversed order; predictions must follow the scale
probe_option_count	400	MASSIVE intent with 3 / 6 / 12 / 20 options
synth_eval_domains_th	300	Synthetic domains absent from train (insurance claims, school admin); see the caveat in H4

Suites built from XNLI, Belebele, SIB-200, XCOPA, thai_toxicity_tweet, thai_exam, and the typed-decisions test split are not included, because some are non-commercial or of unclear copyright. Build them from upstream if you need them.

Baseline

To show how hard each suite is, the table gives the zero-shot scores of the public multilingual checkpoint of convaiinnovations/laya, with no temperature scaling:

Suite	Metric	Score
massive_intent.th.full	accuracy	0.378
massive_intent.th.heldout_scenarios	accuracy	0.432
massive_intent.{ar,hi,ja,ko,ru,zh-CN}	mean accuracy	0.542
massive_th.sampler_eval	accuracy	0.584
wisesight.sampler_eval	accuracy	0.469
iapp.sampler_eval	accuracy	0.539
probe.noul_negation	flip consistency	0.118
probe.noul_labels	default / relabelled agreement	0.808
probe.score_orientation	orientation consistency	0.270
probe.option_count	accuracy at k = 3 / 20	0.825 / 0.363
synth_eval_domains.th	accuracy	0.553

Read the probe numbers together with H4 below: a constant predictor can score well on some of them.

Known issues

An audit of this release found the following. Severity reflects the expected effect on models trained with this data. Each item maps to tasks in TASKS.md.

High

• H1. The synthetic agreement filter skews class balance. Synthetic questions are kept only when the teacher agrees with the generator, and the teacher rejects middle levels far more often than extremes. Moderation severity "medium" keeps 15%, telco urgency "no pressure" 27%, while "urgent" keeps 100%; after filtering, "urgent" makes up 52% of telco urgency (27% before). A model trained on train is pushed away from middle levels.
• H2. thaisum topic is treated as structural gold. Its gold is a keyword heuristic over the article category, not a real headline or span, yet 223 topic disagreements are forced onto it with alpha 0.9.
• H3. Few instruction templates. Most tasks have 2–3 instruction strings per language, and one is reserved for eval, so many tasks have only one or two phrasings in train.
• H4. Some eval suites cannot detect what they target.
  • probe.noul_labels is 97.8% false, so a constant predictor scores 100% agreement.
  • probe.score_orientation is 54% neutral, which reversal leaves unchanged.
  • The n = 100 non-Thai MASSIVE suites carry a ±9.8 pp 95% CI.
  • synth_eval_domains.th keeps only questions that the same teacher (the one that labelled train) agrees with, so it measures teacher agreement rather than correctness. Its 100-row human spot-check (annotation/spot_check_synth_eval.csv) has not been filled in yet.
• H5. The calibration split leaks passages and uses noisy gold. It splits by row, so 464 train iapp rows share a passage with a calibration row. It also includes wisesight and wongnai one-hot gold, which the teacher agrees with only 51–70% of the time.

Medium

• M1. The "none of the other options fit" option is always last (1,429 questions), so the last position is gold more often than chance.
• M2. Negatives are too easy. iapp unanswerable passages are random, and thaisum headline distractors are random; the teacher scores 98–99.5% on them.
• M3. Some synthetic questions are vaguely defined. Teacher and generator agree only 50–66% on needs_human, urgency, moderation severity, frustration, and LINE-chat intent.
• M4. Synthetic NLI has low diversity: a distinct 4-gram ratio of 0.117, and 11% near-duplicates.
• M5. Passages are clipped at 600–700 characters, and every thaisum state ends in "…".
• M6. A single teacher temperature is used for every question type.
• M7. No held-out suite exists for wongnai star scoring or thaisum topic / headline.
• M8. Non-Thai rows make up 28.7% of train questions, and no other share has been tested. Also, 994 non-Thai MASSIVE questions use Thai option keys.

Low

• L1. Overlap filtering strips Thai vowels and tone marks, so เสือ and เสื่อ normalise to the same string and some rows were dropped as false-positive overlaps.
• L2. Small bookkeeping gaps in the eval manifest: massive_intent.th has no language or tags.
• L3. synth_gov_service has 320 rows against 375 for the other domains, and 8 synthetic questions have no teacher label.
• L4. 217 wisesight and 204 massive_th rows have under 10 characters of content; some are noise (555, one emoji).

Areas to improve

TASKS.md has 30 tasks with a problem statement, what to do, and a measurable "done when" for each. Most of the analysis and re-filtering tasks can be done with this repo alone, because unfiltered keeps the teacher distribution, gold, and decision for every question.

Good places to start:

Task	Effort	Needs
T02 Route thaisum topic disagreements away from structural gold	S	this repo
T07 Macro-F1, balanced accuracy, and majority baselines for every suite	S	this repo
T09 Keep synthetic score questions within ±1 of the intended level	S	this repo
T10 Per-class retention report for the synthetic filter	S	this repo
T16 Fill in the 100-row human spot-check	S	native Thai reader
T19 Group-level calibration split	S	this repo
T21 Per-question-type teacher temperature	M	this repo
T30 Reproducible audit script for every number in "Known issues"	M	this repo

To contribute, open a Discussion to claim a task, then a pull request with the before/after numbers that the task's Done when asks for. New sources need a license entry in LICENSE.md, and nothing with a non-commercial license may enter train.

Licensing and attribution

There is no single license; each row follows its source. See LICENSE.md for the per-source table.

• Synthetic rows and all teacher distributions were generated for this project and are released under CC BY 4.0. They are model outputs, so the teacher model provider's usage terms may also apply to how you use them.
• thaisum articles are scraped Thai news. The dataset card says MIT, but copyright in the article text stays with the original publishers. Drop source == "thaisum" if that matters for your use.
• iapp passages come from Thai Wikipedia under CC BY-SA, so redistributed passages must stay share-alike.
• wongnai reviews carry LGPL-3.0, a code license applied to data by the upstream repo.
• Reviews and social-media posts (wisesight, wongnai) are user-generated text and may contain names or offensive language.

Please cite the upstream datasets when you use their rows: MASSIVE (FitzGerald et al., 2022), wisesight_sentiment (Suriyawongkul et al., 2019), Wongnai reviews, ThaiSum (Chumpolsathien, 2020), iapp_wiki_qa_squad (iApp Technology), and typed-decisions (LocalLLaMA).