---
license: mit
tags:
- relational-deep-learning
- relational-databases
- relational-data
- tabular
- tabular-classification
- tabular-regression
- foundation-model
- world-model
- in-context-learning
- few-shot
- pretraining
- synthetic-data
library_name: pytorch
---

# TabWM

TabWM (Tabular World Model) is a relational world model introduced in *[Relational Learning as World Modeling](https://arxiv.org/abs/2610.16763)*. Given a few observed executions from a relational system, it predicts the structured edits that a new action makes to the database, without updating its parameters. The model is pretrained on WorldRel, a generator of executable relational worlds, and its encoder is initialized from the [RT-PluRel](https://huggingface.co/stanford-star/rt-plurel) Relational Transformer.

- **Code:** [github.com/oooranz/TabWM](https://github.com/oooranz/TabWM)

Architecture: 24.7M parameters, float32, 12 Relational Transformer blocks, d_model 256, 8 heads, d_ff 1024, context length 4096. The repository contains `model.safetensors` (weights) and `config.json` (architecture).

## Installation

```bash
pip install git+https://github.com/oooranz/TabWM
```

## Usage

```python
from tabwm import TabWM

model = TabWM.from_pretrained("oooranz/TabWM")
```

Examples for building behavioral context from WorldRel and running inference are provided in the [GitHub repository](https://github.com/oooranz/TabWM).

## License

MIT. The weights are partly initialized from the PluRel paper checkpoint, whose MIT notice is included in `LICENSE`.