---
license: mit
tags:
- biology
- proteomics
- mass-spectrometry
- benchmark
configs:
  - config_name: massive_kb_native
    data_files:
      - split: train
        path: data/massive_kb_native/train.parquet
      - split: validation
        path: data/massive_kb_native/val.parquet
      - split: test
        path: data/massive_kb_native/test.parquet
  - config_name: massive_kb_canonical
    data_files:
      - split: train
        path: data/massive_kb_canonical/train.parquet
      - split: validation
        path: data/massive_kb_canonical/val.parquet
      - split: test
        path: data/massive_kb_canonical/test.parquet
  - config_name: prospect_native
    data_files:
      - split: train
        path: data/prospect_native/train.parquet
      - split: validation
        path: data/prospect_native/val.parquet
      - split: test
        path: data/prospect_native/test.parquet
  - config_name: prospect_canonical
    data_files:
      - split: train
        path: data/prospect_canonical/train.parquet
      - split: validation
        path: data/prospect_canonical/val.parquet
      - split: test
        path: data/prospect_canonical/test.parquet
  - config_name: ood_native
    data_files:
      - split: test
        path: data/ood_native/*.parquet
  - config_name: ood_canonical
    data_files:
      - split: test
        path: data/ood_canonical/*.parquet
  - config_name: sample_data
    data_files:
      - split: train
        path: data/sample_data/*.parquet
---

# Overview

PepSpecBench is a unified benchmark dataset designed for peptide MS/MS spectrum prediction.

## Dataset Formats

To accommodate various modeling architectures and evaluation protocols, the data is provided in two distinct representations:

- **`native`**: The complete, benchmark-ready full-spectrum representation with original m/z and intensity arrays.
- **`canonical`**: A processed representation in a standardized 234-dimensional space, specifically designed for uniform benchmark evaluation across heterogeneous model architectures.

## Repository Contents

The dataset is distributed across three primary data groups, each offered in both `native` and `canonical` formats:

- **`massive_kb`**: Benchmark-ready data tables derived from MassIVE-KB (train/validation/test splits).
- **`prospect`**: Benchmark-ready data tables derived from the PROSPECT dataset (train/validation/test splits).
- **`ood`**: Cross-species out-of-distribution (OOD) evaluation tables (7 species-specific subsets).

## Quick Start / Sample Data

The full dataset release exceeds 4 GB. For rapid preliminary inspection of the schema, post-translational modification (PTM) normalization, and overall data formatting, a lightweight sample package is provided in the `sample_data` config.