DataEval

DataEval analyzes datasets and models to give users the ability to train and test performant, unbiased, and reliable AI models and monitor data for impactful shifts to deployed models.

The dataeval package provides a rigorous and reliable set of tools for developing and analyzing computer vision datasets and the resulting impact on models.

To view our extensive collection of tutorials, how-to's, explanation guides, and reference material, please visit our documentation on Read the Docs

Why DataEval?

DataEval addresses the critical need underlying every AI model -- the data. The difference between a great dataset and a poor dataset can have drastic consequences on AI model performance. Data collected in the wild is noisy, often imbalanced, and doesn't always cover the entire spectrum of conditions need for deployment. DataEval provides AI practitioners with a library of rigorous, algorithm-backed metrics for performance estimation, bias analysis, dataset cleaning and assessment, and data distribution shifts. Throughout all stages of the machine learning lifecycle -- from initial data collection through operational monitoring -- DataEval identifies data problems before they become model failures.

DataEval is easy to install, supports a wide range of Python versions, and is compatible with many of the most popular packages in the scientific and T&E communities.

Target Audience

DataEval is intended to help data scientists, developers, and T&E engineers who want to evaluate and enhance their datasets for optimum performance. For users of the JATI product suite, DataEval has native interoperability when using MAITE-compliant datasets and models.

Getting Started

Python versions: 3.10 - 3.14

Choose your preferred method of installation below or follow our installation guide.

• Installing with pip
• Installing with uv
• Installing with conda/mamba
• Installing from GitHub

Installing with pip

You can install DataEval directly from pypi.org using the following command.

pip install dataeval

By default, PyTorch is installed from PyPI, which bundles CUDA support on Linux and is a much larger download than the CPU build. To choose a specific PyTorch variant, install torch from that variant's wheel index first, then install DataEval — it accepts the build already present in the environment:

# 1. Pick your PyTorch build (cpu / cu118 / cu128)
pip install torch --index-url https://download.pytorch.org/whl/cu128

# 2. Install DataEval
pip install dataeval

Use --index-url, not --extra-index-url, to pick a CUDA build. --extra-index-url adds an index instead of replacing PyPI, and pip then takes the highest version across both. The CUDA indexes lag the latest PyTorch release, so PyPI usually wins and you silently get the default CUDA-bundled build — the install succeeds with no warning. --index-url replaces the index outright, so it is reliable. (For CPU only, pip install dataeval --extra-index-url https://download.pytorch.org/whl/cpu does work, because the CPU index tracks the latest release.)

The cpu / cu118 / cu128 extras do not select a PyTorch variant under pip. All three declare the same requirements (torch, torchvision); what distinguishes them is [tool.uv.sources], which routes those packages to the right wheel index. That is project metadata applied by uv when resolving from source — it is not part of the published wheel. Select the variant with --index-url under pip, --torch-backend under uv pip, and use the extras only for source installs.

torchvision (optional)

torchvision is not a DataEval dependency, and nothing imports it until you reach for TorchvisionTransform — the escape hatch for running a torchvision v2 transform across a dataset view. If you want that class, install torchvision yourself, from the same index as your torch build:

pip install torchvision --index-url https://download.pytorch.org/whl/cu128

Do not mix indexes. A torchvision from PyPI alongside a torch installed from a wheel index resolves and installs cleanly, then fails on import torchvision with RuntimeError: operator torchvision::nms does not exist. torchvision's compiled ops are built against one specific torch build, so both packages must come from the same index — which is also why pip install dataeval[cpu] is the wrong way to obtain it.

Installing with uv

uv pip install dataeval --torch-backend cpu   # or cu118 / cu128 / auto

Installing with conda

DataEval can be installed from conda-forge:

conda install -c conda-forge dataeval

Alternatively, create an environment from the provided environment.yml at the repository root. PyTorch is installed into that environment from PyPI via pip, so this path gives you the CPU/CUDA-bundled PyPI build rather than a specific variant.

micromamba create -f environment.yml

Installing from GitHub

To install DataEval from source locally on Ubuntu, pull the source down and change to the DataEval project directory.

git clone https://github.com/aria-ml/dataeval.git
cd dataeval

Contributing rather than just installing? Use uvx --with nox-uv nox -s dev to build a full development environment — tests, linting, type checking and docs tooling included. See Development Setup for the options it takes.

Using Poetry

Install DataEval.

poetry install

Enable Poetry's virtual environment.

poetry env activate

Using uv

Install DataEval with dependencies for development.

uv sync

Enable uv's virtual environment.

source .venv/bin/activate

Working with data

DataEval has two input paths depending on which part of the library you are using.

dataeval.core provides stateless functions that operate directly on NumPy arrays — embeddings, labels, image hashes, and statistics. No dataset object is required. Call these functions with arrays and get results back directly. Examples include compute_stats, label_errors, divergence_mst, and ber_knn.

dataeval.quality, dataeval.bias, dataeval.shift, and dataeval.performance provide stateful evaluator classes (Duplicates, Outliers, Prioritize, Balance, drift detectors, and so on). These accept either NumPy arrays or Modular AI Trustworthy Engineering (MAITE)-compliant datasets depending on the evaluator.

If your data is not yet in MAITE format, the sections below show what is required and how to wrap a common format, for both image classification and object detection tasks.

Image classification dataset

A MAITE-compliant image classification dataset implements __len__ and __getitem__, where each item is a tuple of (image, label, metadata). Images must be NumPy arrays of shape (H, W, C). Labels must be one-hot encoded arrays of shape (num_classes,). Metadata must be a DatumMetadata object with at minimum an id field.

import maite.protocols as mp
import maite.protocols.image_classification as ic
import numpy as np


class MyImageClassificationDataset(ic.Dataset):
    metadata: mp.DatasetMetadata

    def __init__(self, images: list[np.ndarray], labels: list[int], num_classes: int) -> None:
        # images: list of np.ndarray, each shape (H, W, C)
        # labels: list of int (class indices)
        self._images = images
        self._labels = labels
        self._num_classes = num_classes

        self.metadata = mp.DatasetMetadata(
            id="my_image_classification_dataset",
            index2label={i: f"class_{i}" for i in np.unique(labels)},  # example mapping
        )

    def __len__(self) -> int:
        return len(self._images)

    def __getitem__(self, idx: int) -> tuple[ic.InputType, ic.TargetType, ic.DatumMetadataType]:
        return (
            self._images[idx],  # np.ndarray (H, W, C)
            np.eye(self._num_classes, dtype=np.float32)[self._labels[idx]],  # np.ndarray (num_classes,)
            ic.DatumMetadataType(id=idx),
        )

Object detection dataset

A MAITE-compliant object detection dataset follows the same three-tuple structure, but the label element is replaced by a detection target object carrying per-box labels, bounding boxes, and scores. Bounding boxes use (x0, y0, x1, y1) format. Labels and scores are per-box, not per-image.

import maite.protocols as mp
import maite.protocols.object_detection as od
import numpy as np


class DetectionTarget(od.TargetType):
    """Holds per-box labels, boxes, and one-hot scores for one image."""

    def __init__(self, labels: list[int], boxes: list[list[float]], num_classes: int):
        # labels: list of int, one per box
        # boxes:  list of [x0, y0, x1, y1], one per box
        self._labels = labels
        self._boxes = boxes
        self._scores = np.eye(num_classes)[labels]

    @property
    def labels(self) -> mp.ArrayLike:
        return self._labels

    @property
    def boxes(self) -> mp.ArrayLike:
        return self._boxes

    @property
    def scores(self) -> mp.ArrayLike:
        return self._scores


class MyObjectDetectionDataset(od.Dataset):
    def __init__(
        self, images: list[np.ndarray], labels: list[list[int]], boxes: list[list[list[float]]], num_classes: int
    ) -> None:
        # images: list of np.ndarray, each shape (H, W, C)
        # labels: list of list[int] — per-box class indices, one list per image
        # boxes:  list of list[[x0,y0,x1,y1]] — one list per image
        self._images = images
        self._labels = labels
        self._boxes = boxes
        self._num_classes = num_classes

        self.metadata = mp.DatasetMetadata(
            id="my_object_detection_dataset",
            index2label={i: f"class_{i}" for i in np.unique(labels)},  # example mapping
        )

    def __len__(self) -> int:
        return len(self._images)

    def __getitem__(self, idx: int) -> tuple[od.InputType, od.TargetType, od.DatumMetadataType]:
        return (
            self._images[idx],  # np.ndarray (H, W, C)
            DetectionTarget(self._labels[idx], self._boxes[idx], self._num_classes),
            od.DatumMetadataType(id=idx),
        )

Wrapping a PyTorch dataset

If your data is in a PyTorch Dataset, wrap it to conform to the MAITE protocol. Note that torchvision tensors are (C, H, W) — permute to (H, W, C) before passing to DataEval.

import maite.protocols as mp
import maite.protocols.image_classification as ic
import numpy as np
import torch
from torchvision import transforms
from torchvision.datasets import CIFAR10

tv_cifar10 = CIFAR10(root="./data", train=True, download=True, transform=transforms.ToTensor())


class MyCIFAR10Wrapper(ic.Dataset):
    def __init__(self, source: CIFAR10) -> None:
        self._source = source
        self.metadata = mp.DatasetMetadata(
            id="tv_cifar10",
            index2label={
                0: "airplane",
                1: "automobile",
                2: "bird",
                3: "cat",
                4: "deer",
                5: "dog",
                6: "frog",
                7: "horse",
                8: "ship",
                9: "truck",
            },
        )

    def __len__(self) -> int:
        return len(tv_cifar10)

    def __getitem__(self, idx: int) -> tuple[ic.InputType, ic.TargetType, ic.DatumMetadataType]:
        tv_datum: tuple[torch.Tensor, int] = tv_cifar10[idx]
        image = tv_datum[0].permute(1, 2, 0).numpy()  # Permute image from (C, H, W) to (H, W, C)
        label = np.eye(10, dtype=np.float32)[tv_datum[1]]  # Convert label to one-hot encoding
        return image, label, mp.DatumMetadata(id=idx)


dataset: ic.Dataset = MyCIFAR10Wrapper(tv_cifar10)

Run your first evaluation

The example below uses Duplicates from dataeval.quality to detect near-duplicate images by finding groups of embeddings that are similar in embedding space. Duplicates inflate benchmark scores and cause models to overfit to repeated collection events rather than generalizing to new conditions.

from torch.nn import Flatten

from dataeval.extractors import TorchExtractor
from dataeval.flags import ImageStats
from dataeval.quality import Duplicates

# Configure a feature extractor using a pre-trained PyTorch model.
# Here we use a simple Flatten layer for demonstration, but in practice
# you would use a more powerful model like a pre-trained ResNet or ViT.
extractor = TorchExtractor(Flatten())

# Find near-duplicates using only embedding-based clustering.
# An aggressive cluster_threshold of 1.5 should produce detections
# of near duplicates even with a simple Flatten extractor.
evaluator = Duplicates(
    flags=ImageStats.NONE,
    cluster_algorithm="hdbscan",
    cluster_threshold=1.5,
    extractor=extractor,
    batch_size=64,
)
result = evaluator.evaluate(dataset)

# Near duplicates are grouped into sets of indices that are within
# the specified cluster_threshold in embedding space.
print(result)

shape: (3, 5)
┌──────────┬───────┬──────────┬────────────────┬─────────────┐
│ group_id ┆ level ┆ dup_type ┆ item_indices   ┆ methods     │
│ ---      ┆ ---   ┆ ---      ┆ ---            ┆ ---         │
│ i64      ┆ str   ┆ str      ┆ list[i64]      ┆ list[str]   │
╞══════════╪═══════╪══════════╪════════════════╪═════════════╡
│ 0        ┆ item  ┆ near     ┆ [18586, 39942] ┆ ["cluster"] │
│ 1        ┆ item  ┆ near     ┆ [23157, 31426] ┆ ["cluster"] │
│ 2        ┆ item  ┆ near     ┆ [32024, 49135] ┆ ["cluster"] │
└──────────┴───────┴──────────┴────────────────┴─────────────┘

A result with many large groups is a signal that your dataset contains repeated collection events. Before training, remove all but one sample from each group. See the deduplication how-to guide for a complete walkthrough, including how to choose which sample to keep.

Where to go next

Not sure what to evaluate first? Use the Which tool should I use? guide to find the right evaluator for your situation.

Know which tool to use, then check out What data does each tool need? for a quick-reference table of every algorithm's inputs, and the Functional Overview for task applicability.

Want to just explore the documentation? The Where to go next page allows you to jump around between the different areas of the documentation with small summaries of what each page covers.

Contact Us

If you have any questions, feel free to reach out to us!

Acknowledgement

CDAO Funding Acknowledgement

This material is based upon work supported by the Chief Digital and Artificial Intelligence Office under Contract No. W519TC-23-9-2033. The views and conclusions contained herein are those of the author(s) and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government.