Datasetpublic

oroikono/SIGS-symbolic-expressions

SIGS symbolic expression latent corpus This dataset contains 23,695 grammar-generated symbolic expressions used by the SIGS Grammar-VAE, together with their 32-dimensional latent statistics and variable-based mathematical classes. Dataset structure Each row contains: id: stable row index; expression: symbolic expression generated by the SIGS grammar; math_class: one of CONSTANT, TEMPORAL_1D, SPATIAL_1D, SPATIAL_2D, SPATIOTEMPORAL_2D, or SPATIOTEMPORAL_3D; has_x… See the full description on the dataset page: https://huggingface.co/datasets/oroikono/SIGS-symbolic-expressions.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

1likes40downloads

Dataset Card

SIGS symbolic expression latent corpus

This dataset contains 23,695 grammar-generated symbolic expressions used by the SIGS Grammar-VAE, together with their 32-dimensional latent statistics and variable-based mathematical classes.

Dataset structure

Each row contains:

—id: stable row index;
—expression: symbolic expression generated by the SIGS grammar;
—math_class: one of CONSTANT, TEMPORAL_1D, SPATIAL_1D, SPATIAL_2D, SPATIOTEMPORAL_2D, or SPATIOTEMPORAL_3D;
—has_x, has_t, has_y: variables present in the expression;
—latent_mu: 32-dimensional encoder mean;
—latent_log_sigma: 32-dimensional encoder log-scale statistic stored by the original pipeline.

The data is exported to Parquet so it can be inspected with the Hugging Face dataset viewer and loaded without executing pickle code.

Usage

python

from datasets import load_dataset

dataset = load_dataset("oroikono/SIGS-symbolic-expressions", split="train")
print(dataset[0]["expression"])
print(dataset[0]["math_class"])

Source and generation

Expressions were generated from the context-free grammar in the SIGS repository. The latent statistics were produced by the trained SIGS Grammar-VAE. Mathematical classes are derived deterministically from whether x, y, and t occur as variables.

The included build_dataset.py script converts the original trusted SIGS artifact into a typed Parquet file. Because Python pickle can execute code, do not run the converter on artifacts from an untrusted source.

Intended use

This dataset supports research on grammar-guided representation learning, symbolic regression, and analytical differential-equation solution discovery. Latent proximity does not establish mathematical equivalence or physical validity.

Limitations