SpectraNet

SpectraNet is a multimodal spectroscopy benchmark for evaluating scientific reasoning in foundation models.

The benchmark contains curated experimental spectra across Raman, X-ray diffraction (XRD), Fourier-transform infrared spectroscopy (FTIR), and mass spectrometry (MS), together with metadata, spectrum images, processed spectral arrays, peak annotations, and evaluation-related outputs.

SpectraNet is designed to evaluate whether foundation models can perceive plotted experimental spectra, extract characteristic peaks, and support downstream scientific reasoning over spectroscopic evidence.

Dataset Structure

The dataset is released in sharded form because the full image directory contains a large number of files.

spectranet/
├── public_shards/
│   ├── public_part_0000.zip
│   ├── public_part_0001.zip
│   └── public_part_0002.zip
├── splits/
│   └── data.csv
└── manifest.json

After extracting all ZIP files in public_shards/, the original public dataset structure will be restored as:

public/
├── data/
│   └── shards/
│       ├── 00.parquet
│       ├── 01.parquet
│       └── ...
└── images/
    ├── *.png
    └── ...

Files

public_shards/

The full public/ folder is provided as multiple ZIP shards. Each ZIP file preserves the original folder structure, including:

public/data/
public/images/

To reconstruct the full dataset, download all ZIP files under public_shards/ and extract them into the same directory.

public/data/

This folder contains Parquet shards storing the processed spectral arrays and compact metadata.

Each Parquet record includes fields such as:

sample_id
mid
modality
subkey
x
y
x_unit
y_unit
x_min
x_max
y_min
y_max
n_points

The x and y fields store the processed spectral coordinates and intensities.

public/images/

This folder contains PNG spectrum previews generated from the processed spectral arrays.

The image filenames correspond to the sample_id field used in the metadata table.

splits/data.csv

data.csv is the main post-processed metadata table. It contains one row per processed sample and includes both metadata fields and file pointers.

Important columns include:

sample_id
mid
modality
subkey
image_path
data_path
x_min
x_max
y_min
y_max
n_points
x_unit
y_unit
method
measurement_condition
crystal_system
material_name
phase_label
classification
preferred_chemical_formula
source_format
file_path

The image_path column points to the corresponding PNG image under public/images/.

The data_path column points to the corresponding Parquet shard under public/data/shards/.

Reconstructing the Dataset

Download all ZIP files in public_shards/, then extract them into the same output directory:

from pathlib import Path
import zipfile

root = Path("spectranet")
shard_dir = root / "public_shards"
out_dir = root / "extracted"

out_dir.mkdir(parents=True, exist_ok=True)

for zip_path in sorted(shard_dir.glob("public_part_*.zip")):
    print("Extracting", zip_path.name)
    with zipfile.ZipFile(zip_path, "r") as zf:
        zf.extractall(out_dir)

print("Done.")

After extraction, the reconstructed dataset should contain:

extracted/public/data/
extracted/public/images/

Loading Metadata

import pandas as pd

df = pd.read_csv("splits/data.csv")
print(df.head())

Loading a Parquet Shard

import pandas as pd

parquet_path = "public/data/shards/00.parquet"
data = pd.read_parquet(parquet_path)
print(data.head())

Linking Images and Data

Each row in splits/data.csv contains an image_path and a data_path.

import pandas as pd
from pathlib import Path

df = pd.read_csv("splits/data.csv")

row = df.iloc[0]
image_path = Path(row["image_path"])
data_path = Path(row["data_path"])

print("Image:", image_path)
print("Parquet shard:", data_path)

License

This dataset is released for research use.

Please refer to the original data sources for source-specific licensing terms.

Citation

If you use SpectraNet, please cite the associated paper:

@article{spectranet2026,
  title={SpectraNet: A Multimodal Spectroscopy Benchmark for Evaluating Scientific Reasoning in Foundation Models},
  author={Gu, Yijun and Yang, Jingyun and Liu, Yongtao and Wang, Haozhe},
  year={2026}
}

Notes

This repository is currently under active release preparation.

Dataset files are uploaded in segmented batches because of repository file-count and upload-rate limitations.