Sys1Cal-v1

PAPER: Jev thinks "I don't know", but doesn't say it: Introducing Sys1Cal-v1 Dataset for Probability Calibration

Sys1Cal-v1 is a synthetic benchmark for probability calibration in Jev-like typed decision models. Each item is a True/False question about a proposition A where the exact probability P(A) is known by construction. The benchmark evaluates whether a model returns probabilities with the right numerical meaning, not only whether it selects the right label.

Dataset Summary

Sys1Cal-v1 procedurally constructs probability problems with exact gold distributions, renders equivalent forms of the same latent item, and queries the proposition through Jev-style primitives:

• noul: one scalar probability that the proposition is true.
• choice: a categorical distribution over False and True.
• score: a distribution over ordered truth levels, evaluated by expected normalized truth value.

The primary metric is total variation distance from the exact gold distribution. For binary True/False questions, this is exactly absolute probability error.

Splits and Configs

Both configs expose a single test split.

Config	File	Rows	Use
parallel_primitives	data/v0.1.0_tiny_cleanvars_parallel_primitives.jsonl	365	Main Noul/Choice/Score benchmark contract
binary	data/v0.1.0_tiny_cleanvars_binary.jsonl	365	Binary Choice-like probability recovery

Load

from datasets import load_dataset

ds = load_dataset("RiccardoPorcedda/Sys1Cal-v1", "parallel_primitives", split="test")
print(ds[0])

For local use before upload:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="data/v0.1.0_tiny_cleanvars_parallel_primitives.jsonl",
    split="train",
)

Row Schema

Important fields:

• instance_id: unique rendered benchmark item id.
• latent_instance_id: shared id for equivalent representations of the same latent probability problem.
• family: generator family.
• representation: rendering form, such as direct, counts, ratio, table, prose, or distractor.
• state: sufficient information for the model.
• proposition and prompt: proposition being evaluated.
• queries: Noul/Choice/Score query objects for the parallel primitive config.
• outcomes: usually ["True", "False"].
• gold_distribution: exact target distribution.
• gold_argmax: exact argmax label.
• probability_band: bin for the true probability.

Minimal Example

{
  "instance_id": "explicit_probability__seed_000042__repr_direct__parallel_primitives",
  "latent_instance_id": "explicit_probability__seed_000042",
  "family": "explicit_probability",
  "representation": "direct",
  "proposition": "Event A is true.",
  "outcomes": ["True", "False"],
  "gold_distribution": {"False": 1.0, "True": 0.0},
  "queries": {
    "noul": {"proposition": "Event A is true."},
    "choice": {
      "question": "What is the truth status of the following proposition?",
      "proposition": "Event A is true.",
      "options": ["False", "True"]
    },
    "score": {
      "question": "To what degree is the following proposition true?",
      "proposition": "Event A is true."
    }
  }
}

Evaluation

This repository includes a standalone evaluator:

python scripts/evaluate_sys1cal.py \
  --dataset data/v0.1.0_tiny_cleanvars_parallel_primitives.jsonl \
  --predictions examples/predictions_minimal.jsonl \
  --output evaluation_summary.json

Prediction rows must include instance_id and one of:

• predicted_distribution, distribution, prediction, or probabilities, with labels such as True and False;
• probability_true or p_true, for binary True/False submissions.

The evaluator reports accuracy, pointwise probability fidelity (1 - TV), total variation distance, and Brier regret overall and by model/group, family, representation, and probability band.

How to cite

@misc{porcedda2026sys1cal,
  title = {Jev thinks "I don't know", but doesn't say it: Introducing Sys1Cal-v1 Dataset for Probability Calibration},
  author = {Riccardo Porcedda},
  year = {2026},
  eprint = {2609.35342},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2609.35342},
  url = {https://arxiv.org/abs/2609.35342}
}

GitHub Repo

https://github.com/little-g-ai/Sys1Cal-v1