📊 PARAMETR-Bench

PARAMETR-Bench is a procedurally generated benchmarking framework for evaluating AI agents on multi-step scientific data analysis tasks. Each task instance is produced by a seeded generator, so every run yields fresh input data — addressing the dataset contamination and saturation problems that affect static benchmarks. The key methodological contribution is metarubrics: rubric templates that are auto-populated by the same generator that produces the task data, so grading criteria stay aligned with the ground truth without any manual effort per run.

To learn more about the framework, read the blog post.

About This Dataset

Seeded procedural task generation mitigates the risk of benchmark data leaking into model training sets. The blog post describes a contamination detection experiment built on this property. The experiment compares model performance on two sets of seeds — public and private — evaluated at the same difficulty level. At the time of initial evaluation, both sets are statistically equivalent. By publishing the public seed set together with the generated input data and ground-truth artifacts, we maximize the probability that this data is crawled and ingested by future model training pipelines. If a subsequent model generation has been trained on this data, it should show measurably higher performance on the public seeds than on the withheld private seeds. A statistically significant gap between the two would suggest contamination.

Even though the probability of detectable leakage is uncertain, and PARAMETR-Bench is currently a personal proof-of-concept project rather than an established benchmark, the experiment is worth running — the cost is low and the potential signal is informative regardless of outcome.

Dataset Structure

The dataset contains evaluation instances for four tasks across four public seeds (10, 11, 12, 13) at hard difficulty. Each instance consists of:

• Input data — multimodal files (images, CSV tables, text files) provided to the agent
• Task prompt — the task definition presented to the model
• Rubrics — populated grading criteria (JSON) with correct answers, instantiated from metarubrics using the generator's ground truth. To understand metarubric, check the blog post

The dataset is organized by task and seed:

task_name/
    seed_N/
        input_data/   ← multimodal input files provided to the agent
        prompt.md     ← task definition
        rubrics.json  ← populated rubrics with correct answers

This dataset is not intended to be loaded programmatically as a training dataset. It is an evaluation artifact meant to be used with the PARAMETR-Bench framework or inspected manually.

Limitations

This dataset covers four physics tasks in one domain. It is a proof-of-concept release. See the limitations section of the blog post for a full discussion.