cynosural/semantic-montecarlo-benchmark

Semantic Monte Carlo Benchmark

A synthetic benchmark of numeric research and forecasting questions for evaluating the semantic-montecarlo pipeline.

This release contains only benchmark inputs. Cached experiments, individual run artifacts, and aggregate results are intentionally excluded.

At a glance

Questions	Language	Splits	License
300	English	Validation and test	CC0 1.0

Dataset structure

The dataset has no training split:

Split	Source file	Rows	Intended use
validation	eval.csv	20	Parameter selection and development
test	test.csv	280	Final benchmark evaluation

Both splits are balanced across the ten confidence_mean levels. Validation contains two questions per level; test contains 28.

Fields

Field	Type	Description
id	integer	Identifier, unique within each split
confidence_mean	integer	Target expected-confidence level from 5 to 95, expressed as a percentage
domain	string	Topic category
question	string	Numeric question to research or forecast
answer_unit	string	Required unit for the numeric estimate

confidence_mean is the target used by the current benchmark score. It is not a model prediction, an observed frequency, or a guarantee that the answer is correct. The original assignment method was not recorded and remains a provenance limitation.

Data creation and provenance

The questions were generated with GPT-5.6 Sol and organized into validation and test splits by project contributors. The generation prompt and human-review procedure were not retained in this repository. The question set was first committed in July 2026.

The dataset contains no source documents or personal user records. Its questions cover public topics such as economics, companies, climate, sports, and long-range forecasts.

Benchmark protocol

The reference implementation is scripts/benchmark.py. For each test question, it:

1. Runs the pipeline with the question and answer_unit.
2. Converts the bootstrap-mean distribution into estimated confidence using norm_var_comp.
3. Converts confidence_mean to [0, 1] by dividing it by 100.
4. Reports mean squared error between expected and estimated confidence.

The score measures alignment with the benchmark's confidence targets. It does not measure numeric answer accuracy because resolved numeric answers are not included.

Comparable benchmark reports should record the dataset revision, code commit, run timestamp, model identifier, prompt and search configuration, paraphrase count, bootstrap resamples, random seed, token and search usage, failures, and retries.

Usage

from datasets import load_dataset

dataset = load_dataset("cynosural/semantic-montecarlo-benchmark")
validation = dataset["validation"]
test = dataset["test"]

Use validation while choosing parameters. Reserve test for the final reported evaluation.

Intended use

• Evaluate confidence distributions produced by web-enabled numeric research pipelines.
• Compare configurations under a fixed dataset revision and execution protocol.
• Study how question horizon and domain relate to distribution concentration and no-answer behavior.

This is not a factual answer key, a calibrated probability dataset, or training data for optimizing against the published test questions.

Limitations

• LLM-generated questions may contain ambiguities, incorrect premises, or generator biases.
• The provenance of the confidence_mean assignments is incomplete.
• Many questions are time-dependent or concern future events; available web evidence and pipeline outputs change with the execution date.
• The test questions are public. Repeated tuning against them invalidates claims of held-out evaluation and can cause benchmark contamination.
• Confidence-target MSE does not establish factual accuracy.

Citation

@dataset{cynosural_semantic_montecarlo_2026,
  title = {Semantic Monte Carlo Benchmark},
  author = {{Cynosural AI contributors}},
  year = {2026},
  url = {https://huggingface.co/datasets/cynosural/semantic-montecarlo-benchmark}
}

License

To the extent possible under law, the project contributors have dedicated this benchmark dataset to the public domain under CC0 1.0 Universal. It may be copied, modified, and redistributed for any purpose without conditions.