DranxX/typed-decisions-v1

Typed Decision Dataset (EN + ID)

458,093 typed decisions across 17 sub-datasets in two languages, where every answer is a full probability distribution over discrete options (noul), category sets (choice), or ordered levels (score) — usable as soft targets for training decision/classification models.

Tree structure

text/
├── en/
│   ├── aita-moral-judgment/          # moral judgment, 5-way verdict + severity
│   ├── aita-moral-judgment-d1-ultra/ # same states, alternate label set
│   ├── civil-comments-moderation/    # toxicity level + threat/insult noul
│   └── github-issues-triage/         # issue type + severity + priority
└── id/
    ├── mc4-id-moderation/            # gambling/spam/scam moderation (obfuscated fonts)
    ├── id-newspapers-2018/           # news category + bias + honest-title
    ├── campus-complaints-triage/     # complaint action + urgency + emotion
    ├── smsa-sentiment-multiview/     # sentiment + utterance type + anger
    ├── id-emotion-6way/              # 6-way emotion + ideal response
    ├── id-hate-speech-offensiveness/ # offensiveness + identity targeting
    └── id-clickbait-editorial/       # clickbait + sensationalism + editorial action
image/
└── en/
    ├── figureqa-chart-noul/          # chart yes/no reasoning
    ├── hateful-memes-noul/           # meme + caption hatefulness
    ├── nlvr2-visual-entailment-noul/ # two-image statement verification
    ├── rvl-cdip-document-routing/    # 16-way document routing
    ├── mvtec-defect-inspection/      # industrial defect detection
    └── chestxray14-finding-noul/     # 4 radiology findings per X-ray

Each dataset folder contains train.jsonl (one decision per line) and instructions.json (question definitions — type, instructions, options). Image folders additionally contain images/.

Composition

path	rows	source dataset
text/en/aita-moral-judgment	51,770	OsamaBsher/AITA-Reddit-Dataset
text/en/aita-moral-judgment-d1-ultra	40,000	OsamaBsher/AITA-Reddit-Dataset
text/en/civil-comments-moderation	30,000	google/civil_comments
text/en/github-issues-triage	29,989	sharjeelyunus/github-issues-dataset
text/id/mc4-id-moderation	60,003	indonesian-nlp/mc4-id
text/id/id-newspapers-2018	100,000	indonesian-nlp/id_newspapers_2018
text/id/campus-complaints-triage	20,000	YogaSWD/indonesian-campus-complaints
text/id/smsa-sentiment-multiview	30,000	carant-ai/indonesian_sentiment_dataset
text/id/id-emotion-6way	30,000	talithaolga/Indonesian-Emotion-Classification
text/id/id-hate-speech-offensiveness	12,703	haipradana/indonesian-twitter-hate-speech-cleaned
text/id/id-clickbait-editorial	2,048	mteb/IndonesianIdClickbaitClassification
image/en/figureqa-chart-noul	36,000	vikhyatk/figureqa
image/en/hateful-memes-noul	7,000	limjiayi/hateful_memes_expanded
image/en/nlvr2-visual-entailment-noul	2,490	lmms-lab/NLVR2
image/en/rvl-cdip-document-routing	2,690	chainyo/rvl-cdip
image/en/mvtec-defect-inspection	2,200	Voxel51/mvtec-ad
image/en/chestxray14-finding-noul	1,200	arudaev/chest-xray-14
Total	458,093

Dataset statistics

Measure	Value
Total rows (decisions)	458,093
Sub-datasets	17
Languages	id, en
Total tokens (Qwen3 BPE)	249,727,865
State characters	933,083,428

Token counts use the Qwen/Qwen3-0.6B byte-level BPE tokenizer, measured over per-row state content plus each unique question template counted once.

Column structure

{
  "id": "mc4-000042",
  "state": "the text the questions refer to",
  "images": ["images/00/nlvr-000405.jpg"],
  "questions": {
    "message_type": { "type": "choice", "instructions": "...", "criteria": { "normal": "...", "online_gambling": "..." } },
    "block_action": { "type": "noul", "instructions": "..." }
  },
  "gold":     { "message_type": null, "block_action": null },
  "answers": {
    "message_type": { "type": "choice", "choice": "online_gambling", "confidence": 0.99,
                      "probabilities": { "normal": 0.004, "online_gambling": 0.99 } },
    "block_action": { "type": "noul", "noul": 0.97 },
  }
}

For choice: choice + confidence + probabilities per option. For noul: a single noul probability (0–1). For score: score (expected value) + probabilities per level.

gold holds human reference labels where the source dataset provides them (AITA, GitHub issues, civil comments, SMSA, emotion, clickbait, hate speech, and the image datasets); it is null on fields that have no annotation.

Usage

from datasets import load_dataset

# a single sub-dataset
ds = load_dataset("DranxX/typed-decisions-v1", "mc4-id-moderation-id", split="train")

# list all sub-datasets
from huggingface_hub import get_dataset_config_names
get_dataset_config_names("DranxX/typed-decisions-v1")

License & attribution

Built on the public datasets cited above — review each source license before commercial use.