ProbabilityBench

Paper | Code

Real-world datasets usually provide only one observed outcome per input, so the true probability of that outcome is unknown and a model's reported probabilities cannot be checked directly. ProbabilityBench provides text-based simulated worlds in which the class probability is known exactly.

Each row is one case: a short text describing it, one sampled outcome, and the exact probability of the positive outcome given that text.

This dataset is a small, fixed test set for trying the benchmark quickly. The code generates unlimited fresh cases, including training examples, with exact probabilities.

Quick start

from datasets import load_dataset

cases = load_dataset("ishidalab/probabilitybench", "paper_review-medium", split="test")
case = cases[151]
print(case["text"])                  # The case description: give this to your model
print(case["probability_positive"])  # Exact P(accept | text)
print(case["label"])                 # One outcome sampled with that probability

Worlds and settings

World	Config prefix	Question	Positive label	Negative label
Peer review	paper_review	Will the submission be accepted?	accept	reject
Covert transfer	covert_transfer	Did an AI worker write a confidential tag into a report's identifier?	encoding	no-encoding
Credit risk	financial_default	Will the company default within the next year?	default	no-default

Each world has three settings that hide progressively more of the relevant information. Hidden information makes the outcome less predictable, so the Bayes error, the error rate of the best possible classifier, rises from low to high. The values below are exact for each world's whole population, not estimates from these 500 cases.

Config	Relevant information hidden from the text	Bayes error
paper_review-low	None	0.034
paper_review-medium	Novelty	0.091
paper_review-high	Novelty, technical validity	0.109
covert_transfer-low	None	0.019
covert_transfer-medium	Exact tag letter	0.074
covert_transfer-high	Exact tag letter, narrow tag range	0.145
financial_default-low	None	0.026
financial_default-medium	Cash buffer	0.108
financial_default-high	Cash buffer, principal maturities	0.167

Every config has one test split of 500 cases. The three configs of a world hold the same 500 cases in the same order, with the same case_id and label. Only the visible information changes, and the exact probability with it, so a model's scores in the three settings are measured on identical cases.

Fields

Field	Type	Description
case_id	string	Stable identifier for matching predictions to cases. Not a model input.
text	string	The case description, ending with the question. This is the model input.
label	string	The sampled outcome.
probability_positive	float	Exact probability of the positive label given the text, rounded to 12 decimal places.

Each text also includes eight details that are independent of the outcome by construction, such as page counts or a reporting month. The text does not mark which details these are.

Citation

@misc{dai2026probabilitybench,
  title         = {ProbabilityBench: Model Evaluation and Calibration with Exact Ground-Truth Uncertainty},
  author        = {Dai, Yuyang and Tang, Yuting and Futami, Futoshi and Ishida, Takashi},
  year          = {2026},
  eprint        = {2610.17016},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG}
}

License

MIT.