Practical overview
When it fits
Use TabFormer when you want research code and example data for experimenting with transformer training on multivariate tables, especially the included transaction data or the documented PRSA dataset.
Who it is for
- Researchers exploring transformer training on structured, time-indexed data.
- Developers prototyping models with transaction or air-quality tables.
- Practitioners investigating synthetic transaction sequences as input to separate fraud-detection experiments.
Documented uses
- Train tabular BERT on the included synthetic transaction dataset or the PRSA air-quality dataset.
- Train the documented GPT-2 path on credit-card transaction records selected by user ID.
- Explore hierarchical transformer components adapted for fields and masking in tabular data.
Requirements & limits
- The documented environment specifies Python 3.7, PyTorch 1.6.0, Hugging Face Transformers 3.2.0, scikit-learn 0.23.2 and Pandas 1.1.2.
- The synthetic transaction dataset is accessed through Git LFS or a README-listed direct download.
- The PRSA air-quality dataset is downloaded separately from Kaggle.
- The two training paths have different scope: BERT supports the transaction or PRSA dataset, while the documented GPT-2 command is for transaction records and accepts user IDs.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up TabFormer’s documented training environment
Create the Conda environment specified for the repository before choosing a dataset and model-training path.
Before you start
- Python 3.7
- The TabFormer repository and its setup.yml file
Steps
Obtain the TabFormer repository.
Create the environment from its YAML file with
conda env create -f setup.yml.Select a dataset and a documented BERT or GPT-2 training command for the next stage.
Expected outcome
A Conda environment is created from the repository’s setup file; model training requires proceeding to the relevant dataset and training instructions.
Guide limitations
- The README lists PyTorch 1.6.0, Hugging Face Transformers 3.2.0, scikit-learn 0.23.2 and Pandas 1.1.2 as requirements.
- This setup procedure does not itself download or prepare the PRSA dataset.
Prepare the included synthetic transaction data
Access the repository’s synthetic credit-card transaction dataset for a documented training path.
Before you start
- A local copy of the TabFormer repository
- Git LFS to retrieve the dataset through the repository, or access to the documented direct-download location
Steps
Obtain the repository’s
data/credit_carddataset using Git LFS.If Git LFS bandwidth is an issue, use the direct dataset link identified in the README.
Follow the README’s Tabular BERT or Tabular GPT-2 section for the corresponding training command.
Expected outcome
The synthetic transaction dataset is available for use with a documented model-training path.
Guide limitations
- The dataset is described as 24 million records with 12 fields.
- The GPT-2 command also requires a data path and user ID selection; the README does not describe a fraud-detection application or deployment workflow.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- IBM/TabFormer · Official homepage title · retained in its source language