TEXR

TEXR is a collection of 36,115 synthetic tabular datasets generated for the TEXR project. It contains more than 72 million rows spanning diverse topics, feature schemas, generation models, and configurations.

Each dataset contains 2,000 rows and between 5 and 33 features. Every CSV is paired with a JSON metadata file describing its topic, features, value ranges, categorical values, and Bayesian-network structure.

Dataset structure

Files are organized by generation model and configuration:

<Model>/
└── <Generation configuration>/
    ├── synthetic_<topic>_2000.csv
    └── metadata_<topic>_2000.json

Example:

Meta-Llama-3.1-8B-Instruct/
└── 8B_fast_False_bn_llm/
    ├── synthetic_5G_cellular_networks_2000.csv
    └── metadata_5G_cellular_networks_2000.json

The metadata files contain:

• topic
• description
• n_samples
• n_features
• feature names, types, descriptions, categories, and ranges
• Bayesian-network nodes and edges

Generation models

TEXR includes datasets generated using nine models across 46 generation configurations:

• Llama 3.3 70B Instruct
• Llama 3.1 70B Instruct
• Llama 3.1 8B Instruct
• Mistral 7B Instruct v0.3
• QwQ 32B
• QwQ 32B Preview
• Qwen3 30B-A3B Thinking
• Qwen3 32B
• GPT-OSS 20B

Download and extraction

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="eddyliu-hf/texr",
    repo_type="dataset",
    local_dir="texr",
)

The datasets are distributed as model-level ZIP archives. Extract the archives before accessing individual CSV and metadata files:

from pathlib import Path
from zipfile import ZipFile

root = Path("texr")

for archive in root.glob("*.zip"):
    with ZipFile(archive) as zip_file:
        zip_file.extractall(root)

Reading a dataset

import json
from pathlib import Path

import pandas as pd

root = Path(
    "texr/Meta-Llama-3.1-8B-Instruct/8B_fast_False_bn_llm"
)

table = pd.read_csv(root / "synthetic_5G_cellular_networks_2000.csv")

with open(
    root / "metadata_5G_cellular_networks_2000.json",
    encoding="utf-8",
) as file:
    metadata = json.load(file)

print(table.head())
print(metadata["description"])
print(metadata["features"])

Dataset lookup

The companion TEXR Dataset Lookup provides name, path, model, configuration, and feature-based search across the collection.

Intended use and limitations

TEXR is intended for research involving synthetic tabular data, tabular representation learning, cross-table generalization, and dataset retrieval.

The tables are synthetically generated and should not be interpreted as verified real-world measurements. Generated content may contain unrealistic relationships, biases, inconsistencies, or domain inaccuracies. A small number of empty generation artifacts are present in the original archives and are excluded from the companion lookup index.

Citation

Citation information will be added with the TEXR paper release.