---
license: apache-2.0
task_categories:
- image-segmentation
- image-to-image
language:
- en
tags:
- lithography
- semiconductor
- EUV
- OPC
- mask-optimization
- inverse-lithography
size_categories:
- n<1K
pretty_name: Litho-Tiny-100
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# Litho-Tiny-100

A tiny placeholder lithography dataset shipped by
[OpenLithoHub](https://github.com/OpenLithoHub/OpenLithoHub) — **100 deterministic
(mask, aerial-image) pairs** generated from the project's rule-based PDK-aware
synthesizer and Gaussian-PSF forward model.

> ⚠️ **This dataset is intentionally tiny.** It exists to nail down the schema
> and prove `load_dataset("OpenLithoHub/litho-tiny")` works end-to-end. If you
> need a real research-scale dataset, see the strategic plan's discussion of the
> upcoming `Litho-1M` pre-training set, or use one of the upstream sources
> referenced in OpenLithoHub's data layer.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("OpenLithoHub/litho-tiny", split="train")
print(ds)
print(ds.column_names)

# Decode the first mask
import io, numpy as np
from PIL import Image
mask = np.array(Image.open(io.BytesIO(ds[0]["mask"])))    # (512, 512) uint8 in {0, 255}
aerial = np.array(Image.open(io.BytesIO(ds[0]["aerial"])))  # (512, 512) uint8

# Or use the raw float32 bytes for full precision
mask_f32 = np.frombuffer(ds[0]["mask_npy"], dtype=np.float32).reshape(512, 512)
aerial_f32 = np.frombuffer(ds[0]["aerial_npy"], dtype=np.float32).reshape(512, 512)
```

## Schema

| Column        | Type         | Description                                                    |
|---------------|--------------|----------------------------------------------------------------|
| `id`          | string       | Stable identifier `<pattern>-<index>`, e.g. `sram-007`.        |
| `pattern`     | string       | One of `sram`, `contact_array`, `random_logic`.                |
| `pdk`         | string       | PDK preset (always `freepdk45` for this tiny version).         |
| `size_px`     | int32        | Mask edge length (512).                                        |
| `seed`        | int32        | PRNG seed used by the synthesizer.                             |
| `mask`        | bytes (PNG)  | Binary mask, 8-bit grayscale, `{0, 255}`.                      |
| `aerial`      | bytes (PNG)  | Aerial-image intensity normalized to `[0, 255]` (clipped at 1).|
| `mask_npy`    | bytes        | Raw `float32` mask, row-major `(H, W)`.                        |
| `aerial_npy`  | bytes        | Raw `float32` aerial intensity (un-clipped), row-major `(H, W)`. |

Counts: 34 SRAM + 33 contact-array + 33 random-logic = **100 rows**, all in `train`.

## Reproducing

```bash
git clone https://github.com/OpenLithoHub/OpenLithoHub.git
cd OpenLithoHub
pip install -e '.[data]'
python scripts/build_litho_tiny.py --out out/litho-tiny
```

The script is fully deterministic; identical commits produce identical bytes.

## Forward model

The aerial images come from `openlithohub._utils.forward_model.simulate_aerial_image`
— a Gaussian-PSF approximation (`sigma_px=4.0`, `dose=1.0`) of the Hopkins forward
model, with circular padding. For a research-grade SOCS Hopkins simulation, see
`openlithohub._utils.hopkins.simulate_aerial_image_hopkins`.

## License

Apache-2.0 — same as OpenLithoHub itself. Patterns are synthetic and free of any
real fab IP.

## Citation

If this dataset helped your work, please ⭐ the
[OpenLithoHub repo](https://github.com/OpenLithoHub/OpenLithoHub) and cite the project.