rlcd-decision-atlas — a typed-decision dataset

This is a dataset for training and evaluating models that make decisions rather than write prose. Every example is one situation, one question, and a short list of declared options — and exactly one of those options is correct. The model's whole job is to put a probability on each option and commit to one. Nothing here asks for free text.

It is assembled from 114 public sources plus one synthetic generator, each converted into the same tiny contract, then split so that a large in-distribution training corpus sits alongside a held-out set of sources the model never trains on — roughly 531k train / 74k validation / 74k test / 27.5k holdout rows, about 707k typed questions in total.

The name is RLCD — reinforcement learning for calibrated decisions. That is the training method the data was built for, but the data itself is method-agnostic: it is just labelled multiple-choice, and you can train on it with supervised learning, a bandit, or anything else.

Why this shape

Most classification corpora each speak their own dialect — different label names, different file layouts, a different idea of what an "answer" is. That makes it painful to train one model that decides across domains, and harder still to ask a fair question about calibration, because a model that is confident-and-wrong on banking intents looks nothing like one that is confident-and-wrong on entailment unless you first force both tasks into the same shape.

So everything is forced into the same shape. A sentiment rating, a yes/no reading question, a 77-way intent, a three-way entailment call — all of them become "here is the state, here is the question, pick the letter." Once the tasks share a contract, three things become possible that are awkward otherwise:

• One model, many decisions. A single policy can be trained and measured across dozens of task families without per-task heads or bespoke parsing.
• Calibration you can actually read. Because the output is a distribution over a small option set, you can compare a model's stated confidence to how often it is right, in the same units, on every source at once. Skewed sources are left skewed on purpose — if hate speech is genuinely rare, a well-calibrated model should say so, and rebalancing the classes would teach it a false prior.
• A real generalization test. A handful of sources are quarantined into a holdout split and appear nowhere in training, so you can ask how the model does on decision tasks it has never seen rather than on a fresh sample of ones it has.

What's in the box

The release is four line-delimited JSON files plus a stats file:

file	what it holds
train.jsonl	in-distribution training rows, shuffled across sources
val.jsonl	in-distribution validation, disjoint from train within each source
test.jsonl	in-distribution test, disjoint from train and val within each source
holdout.jsonl	out-of-distribution rows from sources that appear in no other file
stats.json	per-source and total counts, roles, licenses, and any skipped sources

DATA_CARD.md lists every source with its counts — the exhaustive companion to the summarized tables below.

The data contract

Every row is a JSON object with a handful of fields:

{
  "primitive": "choice",   // "choice" | "score" | "noul"
  "context":   "...",      // the state to decide on — non-empty, may be truncated
  "question":  "...",      // what is being asked
  "choices":   ["...", "..."],  // 2..26 bare, single-line, unique option strings
  "ordered":   false,
  "answer":    0,          // index of the correct choice in `choices` (or null if unlabelled)
  "source":    "ag_news",  // which source produced the row
  "id":        "ag_news-0" // stable per-row identifier
}

There are three primitives, and the difference between them is only how the options relate to each other:

• choice — an unordered set of options, from 2 up to 26. The answer is one index. Options are shuffled, so the correct answer lands in every position about equally often and the letter carries no signal on its own.
• score — an ordered scale, listed low to high (star ratings, sentiment levels). The order is meaningful, so these options are never shuffled.
• noul — a yes/no call, rendered as exactly ["true", "false"] with index 0 meaning the affirmative. Every yes/no recast was checked so that "true" means the thing the question actually asks.

A row becomes a prompt by listing the options as A) …, B) … and ending on a cue for the answer, so a model can answer by emitting a single letter — which is why options are single-line and there are at most 26 of them. Every row is guaranteed to hold 2–26 options, each non-empty and single-line, all unique after stripping, with an answer index that is either in range or null.

None of the above

Roughly a third of the choice and score rows carry a synthetic "None of the above" option. Half the time it replaces the true answer and becomes correct; half the time it is an extra distractor. Either way exactly one original option is dropped, so the option count never changes and the presence of the option leaks nothing about whether it is the answer — only the context can tell you. Yes/no rows are never touched.

Splits

In-distribution rows are split per source and grouped by context, so every question that shares a state string lands in the same split — a passage with five questions never has some questions in train and others in test. Validation and test target about 1,000 rows per source each (scaled down for small sources), and train takes the rest up to a cap of 8,000. Across the whole corpus that lands near a 90 / 5 / 5 split.

The holdout is different in kind. Whole sources are set aside — currently MMLU, MedMCQA, ANLI, TruthfulQA, Climate-FEVER, GPQA, and all of the BBQ bias categories — and their rows go only to holdout.jsonl, roughly 2,000 per source. Nothing from a holdout source ever reaches train, validation, or test, so the holdout measures decisions on unseen task families rather than unseen examples of seen ones.

Each source is generated from its own deterministic stream keyed by the source name, so the release is reproducible and each source's rows and splits are fixed independently of the others — the mixture is stable rather than a lucky draw.

The numbers

In this release:

split	rows	primitives (rows)	"none of the above"	median context
train	531,088	choice 311,981 · noul 192,440 · score 26,667	~32%	122 chars
validation	73,864	choice 43,072 · noul 27,459 · score 3,333	~31%	119 chars
test	73,864	choice 43,072 · noul 27,459 · score 3,333	~31%	120 chars
holdout	27,548	choice 27,548	~50%	203 chars

• 114 sources are included — 98 feed the in-distribution splits and 16 are holdout-only. One further source — GPQA — is gated on the Hub and is not included here; it is recorded as skipped in stats.json.
• Option counts span the full range. In train: 210k rows are yes/no (K=2), 222k sit at 3–5 options, 44k at 6–10, 45k at 11–20, and 9k use 21–26. Sources drawn from very large label sets (banking intents, CLINC, MASSIVE, language ID) sample a fresh subset of 4–26 real labels per row, so their K varies row to row.
• Yes/no answers run about 45% true to 55% false in aggregate, but individual sources keep their natural rates — some are near even, some are heavily one-sided, and that is intentional.
• Contexts are short by default — a median of ~120 characters — with passages, tables, and dialogues truncated to 1,500 characters. A small tail of exam questions (a few dozen rows from MMLU-Pro, MMLU, and MedQA) runs longer, up to ~4,700 characters, because the question itself is that long and clipping it would throw away what is being asked.
• Licenses are recorded, not filtered. Every source carries its license string in stats.json as metadata only. The spread runs from permissive (MIT, Apache-2.0, the various CC-BY variants) through unstated to a few copyleft and non-commercial terms. Treat the license column as a starting point for your own review before redistributing anything — a source's presence here is not clearance to publish it.

The sources

Grouped by the kind of decision they ask for. The K column shows a representative option count; large-label sources sample a subset per row, so their K varies. The rows column is the training count, or the holdout count for holdout-only sources.

Commonsense and reasoning

source	Hugging Face id	type	K	split role	license	rows
social_iqa	allenai/social_i_qa	choice	3	train	CC-BY-4.0	8,000
cosmos_qa	allenai/cosmos_qa	choice	4	train	CC-BY-4.0	8,000
swag	allenai/swag (regular)	choice	4	train	unstated	8,000
quartz	allenai/quartz	choice	2	train	CC-BY-4.0	1,864
winogrande	allenai/winogrande (winogrande_xl)	choice	2	train	unstated	8,000
logiqa	lucasmccabe/logiqa	choice	4	train	unstated	6,634
reclor	metaeval/reclor	choice	4	train	research	3,138
codah	jaredfern/codah (codah)	choice	4	train	ODC-BY	2,219
hellaswag	Rowan/hellaswag	choice	4	train	MIT	8,000
piqa	ybisk/piqa	choice	2	train	AFL-3.0	8,000

Science and knowledge exams

source	Hugging Face id	type	K	split role	license	rows
mmlu	cais/mmlu (all)	choice		holdout	MIT	2,000
mmlu_pro	TIGER-Lab/MMLU-Pro	choice	10	train	MIT	8,000
arc	allenai/ai2_arc (Challenge+Easy)	choice	4	train	CC-BY-SA-4.0	5,787
openbookqa	allenai/openbookqa (main)	choice	4	train	Apache-2.0	3,954
commonsense_qa	tau/commonsense_qa	choice	5	train	MIT	8,000
sciq	allenai/sciq	choice	4	train	CC-BY-NC-3.0	8,000
qasc	allenai/qasc	choice	8	train	CC-BY-4.0	7,060
medmcqa	openlifescienceai/medmcqa	choice		holdout	Apache-2.0	2,000
medqa	GBaker/MedQA-USMLE-4-options	choice	4	train	unstated	8,000
head_qa	dvilares/head_qa	choice	5	train	MIT	8,000
gpqa	Idavidrein/gpqa (gpqa_main)	choice		holdout (gated, skipped)	gated/unstated	0

Math (multiple choice)

source	Hugging Face id	type	K	split role	license	rows
aqua_rat	deepmind/aqua_rat (raw)	choice	5	train	Apache-2.0	8,000
math_qa	allenai/math_qa	choice	5	train	Apache-2.0	8,000

Reading comprehension

source	Hugging Face id	type	K	split role	license	rows
race	ehovy/race (all)	choice	4	train	research	8,000
quail	textmachinelab/quail	choice	4	train	CC-BY-NC-SA-4.0	8,000
dream	dataset-org/dream	choice	3	train	unstated	8,000
boolq	google/boolq	noul	2	train	CC-BY-SA-3.0	8,000

Natural language inference

source	Hugging Face id	type	K	split role	license	rows
mnli	nyu-mll/multi_nli	choice	3	train	unstated	8,000
snli	stanfordnlp/snli	choice	3	train	CC-BY-SA-4.0	8,000
anli	facebook/anli	choice		holdout	CC-BY-NC-4.0	2,000
scitail	allenai/scitail (tsv_format)	noul	2	train	Apache-2.0	8,000
folio	tasksource/folio	choice	3	train	CC/MIT	964
glue_qnli	nyu-mll/glue (qnli)	noul	2	train	permissive	3,462
glue_rte	nyu-mll/glue (rte)	noul	2	train	permissive	223
superglue_cb	aps/super_glue (cb)	choice	3	train	permissive	46
superglue_copa	aps/super_glue (copa)	choice	2	train	permissive	80

Fact verification and truthfulness

source	Hugging Face id	type	K	split role	license	rows
vitaminc	tals/vitaminc	choice	3	train	CC-BY-SA-3.0	8,000
fever	fever/fever (v1.0)	choice	3	train	CC-BY-SA-3.0	8,000
health_fact	ImperialCollegeLondon/health_fact	choice	4	train	MIT	8,000
pubmedqa	qiaojin/PubMedQA (pqa_labeled)	choice	3	train	MIT	800
tab_fact	ibm-research/tab_fact (tab_fact)	noul	2	train	CC-BY-4.0	8,000
halueval	pminervini/HaluEval (qa)	noul	2	train	unstated	8,000
squad_v2	rajpurkar/squad_v2	noul	2	train	CC-BY-SA-4.0	8,000
climate_fever	tdiggelm/climate_fever	choice		holdout	unstated	1,535
truthfulqa	truthfulqa/truthful_qa (multiple_choice)	choice		holdout	Apache-2.0	817

Paraphrase

source	Hugging Face id	type	K	split role	license	rows
paws	google-research-datasets/paws (labeled_final)	noul	2	train	free-to-use	8,000
glue_mrpc	nyu-mll/glue (mrpc)	noul	2	train	permissive	2,076
glue_qqp	nyu-mll/glue (qqp)	noul	2	train	permissive	8,000
med_q_pairs	curaihealth/medical_questions_pairs	noul	2	train	unstated	1,048

Safety, ethics, and bias

source	Hugging Face id	type	K	split role	license	rows
ethics_commonsense	hendrycks/ethics (commonsense)	noul	2	train	MIT	8,000
ethics_justice	hendrycks/ethics (justice)	noul	2	train	MIT	8,000
ethics_deontology	hendrycks/ethics (deontology)	noul	2	train	MIT	8,000
ethics_virtue	hendrycks/ethics (virtue)	noul	2	train	MIT	8,000
moral_stories	demelin/moral_stories (full)	noul	2	train	MIT	8,000
ethos_binary	iamollas/ethos (binary)	noul	2	train	AGPL/GPL	800
tweets_hate	tweets-hate-speech-detection/tweets_hate_speech_detection	noul	2	train	GPL-3.0	8,000

Sentiment, emotion, and rating

source	Hugging Face id	type	K	split role	license	rows
sst5	SetFit/sst5	score	5	train	unstated	8,000
yelp	Yelp/yelp_review_full	score	5	train	unstated	8,000
imdb	stanfordnlp/imdb	noul	2	train	permissive	8,000
rotten_tomatoes	cornell-movie-review-data/rotten_tomatoes	noul	2	train	unstated	8,000
amazon_polarity	fancyzhx/amazon_polarity	noul	2	train	Apache-2.0	8,000
app_reviews	sealuzh/app_reviews	score	5	train	unstated	8,000
emotion	dair-ai/emotion (split)	choice	6	train	unstated	8,000
financial_phrasebank	takala/financial_phrasebank (sentences_allagree)	choice	3	train	CC-BY-NC-SA-3.0	1,812

Topic and intent

source	Hugging Face id	type	K	split role	license	rows
ag_news	fancyzhx/ag_news	choice	4	train	unstated	8,000
dbpedia_14	fancyzhx/dbpedia_14	choice	14	train	CC-BY-SA-3.0	8,000
yahoo_topics	community-datasets/yahoo_answers_topics	choice	10	train	unstated	8,000
newsgroups	SetFit/20_newsgroups	choice	20	train	unstated	8,000
trec_coarse	CogComp/trec (coarse_label)	choice	6	train	unstated	3,952
trec_fine	CogComp/trec (fine_label)	choice	5	train	unstated	3,952
clinc_oos	clinc/clinc_oos (plus)	choice	4	train	CC-BY-3.0	8,000
massive_intent	AmazonScience/massive (en-US)	choice	18	train	CC-BY-4.0	8,000
massive_scenario	AmazonScience/massive (en-US)	choice	18	train	CC-BY-4.0	8,000
subj	SetFit/subj	noul	2	train	unstated	8,000
sms_spam	ucirvine/sms_spam	noul	2	train	unstated	3,574
lang_id	papluca/language-identification	choice	12	train	derived	8,000
banking77	legacy-datasets/banking77	choice	5	train	CC-BY-4.0	8,000
bitext	bitext/Bitext-customer-support-llm-chatbot-training-dataset	choice	9	train	unstated	8,000

Synthetic decisions

source	Hugging Face id	type	K	split role	license	rows
triage	(synthetic)	choice	4	train	synthetic	8,000

Config bundles (each config is its own source; see DATA_CARD.md for the full list)

family	Hugging Face id	type	split role	configs	rows
tweet_eval_*	cardiffnlp/tweet_eval	choice / noul	train	5	29,653
babi_nli_*	tasksource/babi_nli	noul	train	20	23,990
bbq_*	heegyu/bbq	choice	holdout	11	19,196

Loading it

Each line of every file is one JSON object, so read it with whatever you already use:

import json

with open("train.jsonl") as f:
    rows = [json.loads(line) for line in f]

q = rows[0]
print(q["source"], q["primitive"], q["choices"], q["answer"])

or with the datasets library:

from datasets import load_dataset

ds = load_dataset("goutam/rlcd-decision-atlas")   # train, validation, test, holdout

A row turns into a prompt by listing the options as lettered lines and ending on a cue for the answer letter — one natural rendering is:

User: Context:
<context>

Question: <question>
Options:
A) <choice 0>
B) <choice 1>
...
Answer with the letter only.
Assistant: The answer is

The holdout split has the same schema as the others and is meant to be read exactly like test — it is simply a stricter test, since none of its sources appear in the other splits.

Good to know

• No invented answers. A source is only included if it has a real fixed option set or can be recast as a clean yes/no. Distractors are never fabricated — the only sampling that happens is drawing a subset from a source's genuine label set when it has more than 26 labels.
• Positions carry no signal. Unordered options are permuted per row, so a model cannot learn "the answer is usually B." Ordered scales and yes/no rows keep their fixed order because the order is the point.
• Natural class balance is preserved. Skewed sources stay skewed. For a model meant to be calibrated, the real base rate is information, not noise.
• Reproducible and auditable. The release is built deterministically, and stats.json records exactly what went in — per-source counts, split roles, licenses, and any source that was skipped.
• Licenses are metadata, not clearance. Each source's license string is recorded for reference; check it yourself before redistributing any subset.