---
base_model: meta-llama/Llama-3.2-1B-Instruct
datasets:
- AvaLovelace/StableText2Brick
language:
- en
library_name: peft
license: mit
pipeline_tag: text-to-3d
---

# Model Card for BrickGPT

These are the model weights for BrickGPT, the first approach for generating physically stable toy brick models from text prompts, as described in [Generating Physically Stable and Buildable LEGO® Designs from Text](https://huggingface.co/papers/2505.05469).
This model was fine-tuned from [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct).

## Model Details

### Model Description

- **Developed by:** [Carnegie Mellon University Generative Intelligence Lab](https://www.cs.cmu.edu/~generative-intelligence-lab/)
- **Funded by:** This work is partly supported by the Packard Foundation, Cisco Research Grant, and Amazon Faculty Award. This work is also in part supported by the Manufacturing Futures Institute, Carnegie Mellon University, through a grant from the Richard King Mellon Foundation. KD is supported by the Microsoft Research PhD Fellowship.
- **Model type:** Autoregressive
- **Language(s):** English
- **License:** MIT
- **Finetuned from model:** [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct)
- **Project page:** https://avalovelace1.github.io/BrickGPT/

### Model Sources

- **Repository:** [AvaLovelace1/BrickGPT](https://github.com/AvaLovelace1/BrickGPT)
- **Paper:** [Generating Physically Stable and Buildable LEGO® Designs from Text](https://huggingface.co/papers/2505.05469)
- **Demo:** [cmu-gil/BrickGPT-Demo](https://huggingface.co/spaces/cmu-gil/BrickGPT-Demo)

## Limitations

The model is restricted to creating structures made of 1-unit-tall cuboid bricks on a 20x20x20 grid. It was trained on a dataset of 21 object categories: *basket, bed, bench, birdhouse, bookshelf, bottle, bowl, bus, camera, car, chair, guitar, jar, mug, piano, pot, sofa, table, tower, train, vessel*. Performance on prompts from outside these categories may be limited.

## How to Get Started with the Model

See the [GitHub repo](https://github.com/AvaLovelace1/BrickGPT) for usage examples and an interactive CLI demo.

## Training Details

### Training Data

BrickGPT was trained using [StableText2Brick](https://huggingface.co/datasets/AvaLovelace/StableText2Brick), a dataset of 47k toy brick structures.

### Training Procedure

The model was fine-tuned using LoRA applied to the `q_proj` and `v_proj` matrices. We used AdamW optimization. The learning rate followed a cosine decay with warmup.

#### Training Hyperparameters

- **Training regime:** bf16 mixed precision
- **Epochs:** 3
- **Global batch size:** 64
- **Max learning rate:** 0.002
- **Learning rate warmup steps:** 100
- **LoRA rank:** 32
- **LoRA alpha:** 16
- **LoRA dropout:** 0.05

## Evaluation

See the [paper](https://huggingface.co/papers/2505.05469) for detailed evaluations.

## Environmental Impact

- **Hardware Type:** 8x NVIDIA RTX A6000 (48 GB)
- **Hours used:** 0.5

## Citation

If you find this model useful for your research, please cite the following work.

```bibtex
@article{pun2025brickgpt,
    title   = {Generating Physically Stable and Buildable LEGO Designs from Text},
    author  = {Pun, Ava and Deng, Kangle and Liu, Ruixuan and Ramanan, Deva and Liu, Changliu and Zhu, Jun-Yan},
    journal = {arXiv preprint arXiv:2505.05469},
    year    = {2025}
}
```

## Model Card Contact

Ava Pun (apun@andrew.cmu.edu)

### Framework versions

- PEFT 0.15.0