---
task_categories:
- text-classification
language:
- en
pretty_name: Big Decisions
size_categories:
- 1M<n<10M
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: train.parquet
  - split: dev
    path: dev.parquet
  - split: test
    path: test.parquet
---

# Big Decisions

Big Decisions brings together more than four million examples from 74 datasets, covering classification, language understanding, reasoning, policy, tool use, preference, and other decision tasks. Each example pairs a context with a question, possible answers, and a target answer.

## Dataset splits

| Split | Examples |
|---|---:|
| Train | 3,632,283 |
| Dev | 326,539 |
| Test | 86,931 |
| **Total** | **4,045,753** |

## Load the data

```python
from datasets import load_dataset
import json

data = load_dataset("ELBennett/big-decisions")
row = data["train"][0]
questions = json.loads(row["questions_json"])
```

Each row's `state` is the context. Its `questions_json` contains one or more questions. A question has a consistent choice-based structure:

```json
{
  "type": "choice",
  "instructions": "Which category does this text belong to?",
  "criteria": ["Album", "Animal", "Artist", "Company"],
  "label": 3
}
```

`criteria` lists the possible answers. `label` is a zero-based index into that list, so `3` selects `"Company"`. Some examples include `label_probs`, a probability for each answer, or `choice_ranks`, the rank of each answer in ordinal tasks. Binary and ordinal decisions use the same choice format.

Rows also include a task `domain` and source identifiers. The data is tokenizer-neutral; use the tokenizer appropriate for your model.

## Included source datasets

Counts below are examples retained in this release after deduplication.

| Task area | Dataset | Examples | Original source |
|---|---|---:|---|
| Aspect and sentiment classification | ACOS (`acos`) | 11,971 | NUSTM/ACOS |
| Exam | AGIEval (`agieval`) | 7,082 | hails/agieval-* |
| Search relevance | Amazon ESCI (`amazon_esci`) | 47,658 | tasksource/esci or amazon-science/esci-data |
| Semantics and inference | ANLI (`anli`) | 169,072 | facebook/anli |
| Tool use | API-Bank (`api_bank`) | 10,908 | liminghao1630/API-Bank |
| Exam | ARC-Challenge (`arc_challenge`) | 3,710 | allenai/ai2_arc:ARC-Challenge |
| Exam | ARC-Easy (`arc_easy`) | 2,251 | allenai/ai2_arc:ARC-Easy |
| Intent classification | BANKING77 (`banking77`) | 13,071 | PolyAI/banking77 |
| Reasoning benchmark | BBH (`bbh`) | 5,499 | lukaemon/bbh |
| Decision-making | BEV Decision Mix (`bev_decision`) | 501,927 | avbiswas/bev-decision |
| Tool use | BFCL (`bfcl`) | 1,906 | gorilla-llm/gorilla berkeley-function-call-leaderboard |
| Reasoning benchmark | BIG-Bench (`bigbench`) | 451,563 | tasksource/bigbench |
| Reading comprehension | BoolQ (`boolq`) | 9,427 | google/boolq |
| Humor and language | BPoMP (`bpomp`) | 23,532 | zenodo.org/records/7299879 |
| Search relevance | BRIGHT (`bright`) | 937 | xlangai/BRIGHT |
| Causal reasoning | CLadder (`cladder`) | 8,917 | causalNLP/cladder (HF mirror or GitHub release) |
| Intent classification | CLINC150+OOS (`clinc150`) | 28,266 | clinc/clinc_oos |
| Semantics and inference | ContractNLI (`contract_nli`) | 7,837 | tasksource/contract-nli:contractnli_a |
| Code reasoning | CRUXEval (`cruxeval`) | 800 | cruxeval-org/cruxeval |
| Topic classification | DBpedia-14 (`dbpedia14`) | 99,988 | fancyzhx/dbpedia_14 |
| Policy | Domain decisions (`domain_decisions`) | 122,438 | altslate/certo-decisions-v2:domain_generator |
| Policy | Executable policies (`executable_policies`) | 96,318 | altslate/certo-decisions-v2:own_policy_generator |
| Financial sentiment | FinEntity (`finentity`) | 1,985 | yixuantt/FinEntity |
| Forecasting | ForecastBench (`forecastbench`) | 4,562 | forecastingresearch/forecastbench-datasets |
| Grammar and acceptability | GLUE CoLA (`glue_cola`) | 9,537 | nyu-mll/glue:cola |
| Semantics and inference | GLUE MNLI (`glue_mnli`) | 412,129 | nyu-mll/glue:mnli |
| Semantics and inference | GLUE MRPC (`glue_mrpc`) | 4,076 | nyu-mll/glue:mrpc |
| Semantics and inference | GLUE QNLI (`glue_qnli`) | 110,184 | nyu-mll/glue:qnli |
| Semantic similarity | GLUE QQP (`glue_qqp`) | 403,904 | nyu-mll/glue:qqp |
| Semantics and inference | GLUE RTE (`glue_rte`) | 2,765 | nyu-mll/glue:rte |
| Sentiment classification | GLUE SST-2 (`glue_sst2`) | 67,845 | nyu-mll/glue:sst2 |
| Semantic similarity | GLUE STS-B (`glue_stsb`) | 7,197 | nyu-mll/glue:stsb |
| Semantics and inference | GLUE WNLI (`glue_wnli`) | 706 | nyu-mll/glue:wnli |
| Emotion classification | GoEmotions (`go_emotions`) | 428,892 | google-research-datasets/go_emotions |
| Exam | GPQA (`gpqa`) | 439 | hendrydong/gpqa_main_mc + gpqa_diamond_mc |
| Mathematical reasoning | GSM8K (`gsm8k`) | 8,792 | openai/gsm8k |
| Group preference | Habermas Machine (`habermas`) | 7,710 | google-deepmind/habermas_machine |
| Commonsense reasoning | HellaSwag (`hellaswag`) | 49,947 | Rowan/hellaswag |
| Tool use | Home appliance simulator (`home_appliance`) | 851 | apolinario/decision-index home_appliance.py |
| Fact verification | HoVer (`hover`) | 22,005 | vincentkoc/hover-parquet |
| Humor and language | Humicroedit (`humicroedit`) | 8,325 | cs.rochester.edu humicroedit SemEval-2020 task-7 |
| Sarcasm detection | iSarcasmEval (`isarcasm`) | 4,844 | iabufarha/iSarcasmEval |
| Probability and belief | Known posterior (`known_posterior`) | 120,000 | altslate/certo-decisions-v2:genworld |
| Exam | LogiQA (`logiqa`) | 16,429 | lgw863/LogiQA-dataset |
| Exam | MMLU (remaining) (`mmlu_full`) | 11,956 | cais/mmlu |
| Exam | MMLU high school (`mmlu_high_school`) | 3,850 | cais/mmlu |
| Exam | MMLU-Pro (`mmlu_pro`) | 11,972 | TIGER-Lab/MMLU-Pro |
| Semantics and inference | MultiNLI (`multinli`) | 239,992 | nyu-mll/multi_nli |
| Multi-step reasoning | MuSR (`musr`) | 828 | TAUR-Lab/MuSR |
| Humor and language | New Yorker (`new_yorker`) | 11,674 | jmhessel/newyorker_caption_contest |
| Semantics and inference | NLI4CT (`nli4ct`) | 7,376 | ai-systems/Task-2-SemEval-2024 |
| Security classification | PhishNChips (`phishnchips`) | 2,013 | AreLit/PhishNChips |
| Music understanding | POP909-CL (`pop909_cl`) | 24,000 | AndyWeasley2004/POP909-CL-Dataset |
| Reading comprehension | RACE (`race`) | 97,734 | ehovy/race |
| Factuality detection | RAGTruth (`ragtruth`) | 17,722 | wandb/RAGTruth-processed |
| Exam | ReClor (`reclor`) | 5,123 | metaeval/reclor |
| Model routing | RouterBench (`routerbench`) | 36,483 | withmartian/routerbench |
| Exam | SAT analogies (`sat_analogies`) | 374 | willwhitney/6.863-final-project |
| Multi-step reasoning | SATA-Bench (`sata_bench`) | 1,605 | sata-bench/sata-bench |
| Semantics and inference | SuperGLUE AX-b (`sg_axb`) | 1,100 | aps/super_glue:axb |
| Semantics and inference | SuperGLUE AX-g (`sg_axg`) | 356 | aps/super_glue:axg |
| Reading comprehension | SuperGLUE BoolQ (`sg_boolq`) | 12,697 | aps/super_glue:boolq |
| Semantics and inference | SuperGLUE CommitmentBank (`sg_cb`) | 306 | aps/super_glue:cb |
| Causal reasoning | SuperGLUE COPA (`sg_copa`) | 500 | aps/super_glue:copa |
| Reading comprehension | SuperGLUE MultiRC (`sg_multirc`) | 32,028 | aps/super_glue:multirc |
| Reading comprehension | SuperGLUE ReCoRD (`sg_record`) | 110,656 | aps/super_glue:record |
| Semantics and inference | SuperGLUE RTE (`sg_rte`) | 76 | aps/super_glue:rte |
| Lexical semantics | SuperGLUE WiC (`sg_wic`) | 6,066 | aps/super_glue:wic |
| Coreference resolution | SuperGLUE WSC (`sg_wsc`) | 606 | aps/super_glue:wsc.fixed |
| Dialogue intent classification | SGD/SGD-X (`sgd`) | 10,890 | google-research-datasets/dstc8-schema-guided-dialogue |
| Tool use | ToolRet (`toolret`) | 7,460 | mangopy/ToolRet-Queries |
| Stance classification | VAST (`vast`) | 12,048 | emilyallaway/zero-shot-stance |
| Tool use | When2Call (`when2call`) | 18,401 | nvidia/When2Call |
| Commonsense reasoning | WinoGrande (`winogrande`) | 41,659 | allenai/winogrande |

## Data quality note

925 malformed source records with only one possible answer were excluded because they do not form a choice task. No alternative answers were invented.