OSIRIS: Bridging Analog Layout Circuit Design and Machine Learning with Scalable Dataset Generation

OSIRIS generates and evaluates analog circuit layouts. This repository contains a browsable table of the generated layout variants, the full circuit artifacts, and the pipeline code released with the ICLR 2026 paper. The ICLR 2026 paper.

Browse the dataset

The Dataset Viewer displays one row per top-level layout GDS file in data/Osiris_Dataset.zip. There are 89,108 distinct rows and layout files, covering the five circuit families below.

Circuit family	Layouts (#)
ahuja_ota	14,617
five_transistors_ota	13,694
miller_ota	14,870
ota_ff	23,028
ota_lpf	22,899
Total	89,108

Column	Meaning
id	Unique path-like variant identifier: circuit family, instance, and variant.
circuit_family, circuit_instance, variant	Identifiers from the release directory structure.
area	Numeric value from the variant's area.npy.
pex_score	Numeric array from the variant's pex_score.npy.
layout_member	GDS layout path inside data/Osiris_Dataset.zip.
netlist_member, moves_member	Associated netlist and move metadata paths inside the same archive.
simulation_pre_member, simulation_post_member	Pre- and post-extraction simulation CSV paths inside the same archive.

Load the Viewer table with:

from datasets import load_dataset

variants = load_dataset("hardware-fab/osiris", "default", split="full")
print(variants[0])

Repository files

Path	Contents
viewer/variants.parquet	The typed, searchable variant table used by the Dataset Viewer.
data/Osiris_Dataset.zip	Complete layouts, netlists, metadata, metrics, and simulations. Extracts to Osiris_Dataset/.
code/osiris_code.zip	Pipeline code and supporting files. Extracts to osiris/.

The Parquet table is an index and summary of the original artifacts. The GDS layouts and simulation files remain in the ZIP archive.

After extracting the dataset archive, its main directories are:

Osiris_Dataset/
├── data/
│   ├── layouts/
│   ├── metadata/
│   ├── metrics/
│   └── simulations/
└── netlists/

Each data category is organized by circuit family, circuit instance, and variant. The *_member values in the table are exact paths relative to the root of this archive.

Download an artifact

from zipfile import ZipFile
from datasets import load_dataset
from huggingface_hub import hf_hub_download

row = load_dataset("hardware-fab/osiris", "default", split="full")[0]
archive = hf_hub_download(
    repo_id="hardware-fab/osiris",
    repo_type="dataset",
    filename="data/Osiris_Dataset.zip",
)
with ZipFile(archive) as z:
    z.extract(row["layout_member"], path=".")

Download code/osiris_code.zip in the same way if you need the pipeline. Inside that ZIP, osiris/README.md describes installation and usage; it appears at that path after extraction. The dataset archive contains about 24 GB of uncompressed file contents and millions of entries, so allow additional disk space and time for full extraction.