README.md

metadata

pretty_name: OpenJevData-140k
language:
  - en
  - zh
license: mit
size_categories:
  - 100K<n<1M
task_categories:
  - question-answering
tags:
  - decision-making
  - reasoning
  - supervised-fine-tuning
  - reinforcement-learning
configs:
  - config_name: default
    data_files:
      - split: SFT
        path: data/SFT.parquet
      - split: RL
        path: data/RL.parquet

Dataset Card for OpenJevData-140k

Dataset Summary

OpenJevData-140k is a curated release of the data collection used to train OpenJev-4B. It contains 146,738 decision-making examples across 19 task categories, organized into SFT and RL splits.

Each example presents a state, a question, and a request-specific set of natural-language options. The data include hard answers and soft probability distributions, with candidate sets ranging from 2 to 77 options. Languages are English and Chinese.

Data Splits

The SFT split contains 100,345 examples for supervised fine-tuning: 86,023 hard and 14,322 soft examples. The RL split contains 46,393 examples for reinforcement learning: 28,667 hard and 17,726 soft examples. Together they provide 146,738 examples.

Task Distribution

• Probabilistic reasoning: Infer outcome probabilities from uncertainty, conditional information, sampling, or noisy evidence.
• Deductive reasoning: Derive conclusions from stated premises, rules, or natural-language entailment.
• Commonsense context: Use everyday knowledge to answer a question or select a plausible continuation.
• Knowledge QA: Answer factual and subject-specific questions across multiple domains.
• Policy exceptions: Apply policies while resolving exceptions, rule precedence, and conflicting conditions.
• Multi-hop reasoning: Connect facts across passages, documents, or entities to reach an answer.
• Temporal / numeric: Apply rules involving dates, durations, numeric thresholds, or time-dependent conditions.
• Quantitative reasoning: Solve numerical questions involving arithmetic, word problems, financial information, or tables.
• Intent routing: Identify the user intent and select an appropriate tool, route, or action.
• Evidence validation: Check whether evidence supports a claim or whether a response meets stated semantic and format requirements.
• Reading comprehension: Answer questions using information stated or implied in a supplied passage.
• Answer preference: Choose the response that best satisfies the request among candidate answers.
• Option logic: Jointly interpret logical statements and dependencies distributed across the candidate options.
• Score quality: Assign a rating for qualities such as helpfulness, coherence, clarity, or preference strength.
• Score rubrics: Assign a score by applying explicit criteria for completeness, coverage, compliance, or task completion.
• Constraint planning: Choose an action or plan that satisfies interacting constraints and trade-offs.
• Code / database: Reason about code behavior or select the database query that answers a request.
• Emotion recognition: Identify the emotion conveyed by a piece of text.
• Score intensity: Rate the strength of a tone or the severity of an impact on an ordered scale.

Dataset Creation

The source benchmarks include ARC, BANKING77, BoolQ, CodeContests, CommonsenseQA, ConditionalQA, ContractNLI, FinQA, GoEmotions, HellaSwag, HelpSteer2, HoVer, LogiQA 2.0, MASSIVE, MathQA, MMLU auxiliary data, MultiDoc2Dial, MuSiQue, OpenBookQA, OR-ShARC, PRM800K, QASC, RAGTruth, RuleTaker, SNLI, Spider, TAT-QA, TempReason, TimeQA, ToolBench, WANLI, and When2Call, together with instruction-following constraint datasets and human-annotated ratings. These sources are converted into a shared decision format and complemented by program-generated tasks and LLM-assisted synthesis. Labels come from source annotations, executable checks, exact probability calculations, or reviewed judgments, depending on the task.

Synthetic data generation uses the input token-length characteristics of JevBench Public as a reference to broaden input-length coverage. The collection includes both short, straightforward questions and longer, more demanding decision problems. The harder tasks include multi-step reasoning, interacting rules and exceptions, cross-option logic, temporal and quantitative constraints, and decisions under uncertainty.

Data Format

Each example contains the following fields:

• state: context, evidence, rules, or a scenario.
• question: the requested decision.
• options: an ordered list of {id, name, criteria} objects. Nullable criteria may contain essential conditions and is part of the model input.
• category: the broad task category.
• task: choice, score, or noul; each uses the supplied candidate options.
• score_values: numeric values aligned with the options for scoring tasks, otherwise null.
• answer: {kind, probabilities}, where kind is hard or soft. The probabilities align by position with options and sum to one. A hard answer is represented by a one-hot distribution.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("shenjunhao/OpenJevData-140k")
sft = dataset["SFT"]
rl = dataset["RL"]

print(len(sft), len(rl))  # 100345 46393
example = sft[0]
options = example["options"]
gold_probabilities = example["answer"]["probabilities"]

Licensing Information

MIT applies to the original contributions in this release. Adapted source material retains its upstream license and attribution requirements; see SOURCE_NOTICES.md.