---
title: "Domain Adaptation Datasets: Benchmarks and Your Own Data"
url: https://www.brightinfo.com/blog/domain-adaptation-datasets/
site: https://www.brightinfo.com/
description: "Which domain adaptation datasets researchers use, what source and target domains they cover, and why real company data is the target domain AI buyers pay for."
last_updated: 2026-11-20
---

# Domain Adaptation Datasets: Benchmarks and Your Own Data

Published 2026-11-20 · Bright Info

Domain adaptation datasets are collections split into a source domain and a target domain, so researchers can test whether a model trained on one kind of data still works on another. The public benchmarks (Office-Home, PACS, WILDS, LibriAdapt and others) are where methods get compared. The target domain that production AI systems actually need is real operating data: your emails, tickets, CRM records and documents. That is the data AI buyers compete for, and a company that offers it to several buyers at once, rather than one, sets the price through competition.

This guide covers both sides. First, what a domain adaptation dataset is and which ones are standard in vision, text, speech and graphs. Then what the benchmark landscape tells a business owner about the value of the data they already hold.

## TL;DR

- Domain adaptation means a model trained on a source domain must perform on a target domain with the same task but a different input distribution.
- Standard benchmarks include Office-Home and PACS for images, GTA and Cityscapes for driving scenes, IMDB and Amazon reviews for text, and LibriAdapt for speech.
- Every benchmark is a stand-in for a real target domain, such as a particular company’s support tickets, sales emails or call recordings.
- If you hold that kind of data, the way to find its value is to let multiple AI buyers bid on it. Bright Info runs that process at no cost to you unless a deal closes.

## Table of Contents

- [What a domain adaptation dataset actually is](#what-a-domain-adaptation-dataset-actually-is)
- [The kinds of shift benchmarks cover](#the-kinds-of-shift-benchmarks-cover)
- [Established computer-vision datasets](#established-computer-vision-datasets)
- [Text, speech and graph datasets](#text-speech-and-graph-datasets)
- [How to choose a benchmark, and what no benchmark can tell you](#how-to-choose-a-benchmark-and-what-no-benchmark-can-tell-you)
- [Where your data fits: the target domain nobody else has](#where-your-data-fits-the-target-domain-nobody-else-has)
- [Why one offer is the wrong way to price it](#why-one-offer-is-the-wrong-way-to-price-it)
- [A worked example](#a-worked-example)
- [Questions to ask about your own data before you start](#questions-to-ask-about-your-own-data-before-you-start)
- [Get a valuation of your company data](#get-a-valuation-of-your-company-data)
- [FAQ](#faq)
- [Sources](#sources)

## What a domain adaptation dataset actually is

A machine-learning model learns from training data and is then judged on test data. When both come from the same distribution, performance is predictable. When they do not, accuracy drops. A model trained on studio product photos meets blurry phone photos. A sentiment classifier trained on movie reviews meets insurance complaints.

IEEE’s Technology Navigator describes [domain adaptation as a specific form of transfer learning](https://technav.ieee.org/topic/transfer-learning/) where the source and target tasks are the same but the input distributions differ. The MIT Foundations of Computer Vision textbook [frames it as adapting the training domain to look like the test domain, or the reverse](https://visionbook.mit.edu/transfer_learning.html), and notes the common unsupervised case: labels exist only for the source, while plenty of unlabeled inputs exist for the target.

So a domain adaptation dataset has a few defining parts:

- **Two or more domains** with a documented shift between them (style, sensor, geography, time, vocabulary).
- **The same task across domains**, such as the same set of object classes.
- **A protocol** stating which domains are source, which are target, and whether target labels may be used.

That last part matters more than newcomers expect. The 2016 Journal of Big Data survey by Weiss, Khoshgoftaar and Wang [documents that terminology is inconsistent across papers](https://link.springer.com/article/10.1186/s40537-016-0043-6): the same labeled-source, limited-labeled-target setup is called supervised by some authors and semi-supervised by others. Always read the protocol, not just the dataset name.

## The kinds of shift benchmarks cover

The WILDS benchmark from Stanford, presented at ICML 2021, is a useful map of shift types. [It organizes real-world distribution shift](https://cs.stanford.edu/people/jure/pubs/wilds-icml21.pdf) into three groups:

| Shift type | What changes | WILDS example |
| --- | --- | --- |
| Domain generalization | Test domains are disjoint from training domains | Train on some hospitals, test on a different hospital (CAMELYON17-WILDS) |
| Subpopulation shift | Same domains, different proportions | Demographic groups in CIVILCOMMENTS-WILDS |
| Hybrid | Both at once | Satellite images split by year and region (FMOW-WILDS) |

Notice what these have in common: each shift is about where the data came from. A hospital, a camera trap, a region. In business terms, the “domain” is the environment that produced the data. For your company, that environment is your customers, your product, your jargon and your workflows.

## Established computer-vision datasets

Vision is where domain adaptation benchmarks are most mature.

### Office-Home

The [Hugging Face card for Office-Home](https://huggingface.co/datasets/flwrlabs/office-home) describes four domains (Artistic, Clip Art, Product, Real-World) with 65 object categories and 15,588 training rows, each with image, domain and label fields. The card also flags a quirk worth knowing: a “Clock” subfolder is referenced in a CSV but absent, while “Alarm-Clock” is present, so counting the phantom class would give 66 categories. Small details like this are why teams check a dataset’s documentation before trusting a leaderboard number.

### PACS, Digits-5 and Office+Caltech-10

IEEE DataPort’s [PCMDA package](https://ieee-dataport.org/documents/datasets-pcmda) bundles three benchmarks:

- **Digits-5:** 32x32 digit images from MNIST, MNIST-M, USPS, SynthDigits and SVHN.
- **PACS:** 9,944 images at 224x224 across Photo (1,792), Art Painting (2,048), Cartoon (2,344) and Sketch (2,760), with seven labels.
- **Office+Caltech-10:** 9,000 images at 252x252 across Amazon (2,533), Webcam (795), DSLR (498) and Caltech (4,174), with ten office-object labels.

The package is a 1.19 GB zip, and access requires a login.

### Driving scenes: synthetic to real

Synthetic-to-real is the classic setup in autonomous driving. The [da-dataset repository](https://github.com/JiahongChen/da-dataset) documents downloads for GTA (14,966 synthetic pixel-annotated images), SYNTHIA_RAND_CITYSCAPES (9,400 synthetic images) and Cityscapes (5,000 annotated real traffic images, which requires account registration). Synthetic data is cheap to label; real data is what the deployed system sees. Adaptation bridges the gap.

![Image counts across synthetic and real benchmarks](https://www.brightinfo.com/assets/blog/domain-adaptation-datasets/image-1.jpg)

## Text, speech and graph datasets

### Text

Most NLP adaptation work leans on classification corpora from Zhang et al.'s 2015 character-level CNN paper. The fastText repository’s [classification-results.sh script](https://github.com/facebookresearch/fastText/blob/7842495a4d64c7a3bb4339d45d6e64321d002ed8/classification-results.sh) downloads eight of them: AG News, Sogou News, DBPedia, Yelp review polarity and full, Yahoo Answers, and Amazon review full and polarity. Label structures differ by set. AG News has 4 classes, DBPedia 14, and Yahoo Answers 10, according to [one maintained set of dataset notes](https://github.com/Renovamen/Text-Classification/blob/master/notes/datasets.md). Amazon reviews offer five star classes or a polarity version that groups stars into negative and positive.

Researchers create domain shifts by training on one corpus (say, product reviews) and testing on another (say, restaurant reviews). The vocabulary changes, the task does not.

### Speech

The [LibriAdapt repository](https://github.com/akhilmathurs/libriadapt) is a speech recognition adaptation set with three shift axes: six microphones, three accents (US, Indian and British English, the latter two generated with text-to-speech on LibriSpeech transcripts) and four noise conditions (clean, rain, wind, laughter). Partitions download in parts of up to 12 GB.

A separate [2022 study in Entropy](https://mdpi-res.com/d_attachment/entropy/entropy-24-01250/article_deploy/entropy-24-01250-v3.pdf?version=1663750316) adapts speech emotion recognition across three corpora: EmoDB (535 acted German samples), eNTERFACE (42 speakers, 14 nationalities) and CASIA (1,200 Chinese samples). Pairing them yields six directed source-to-target tasks.

### Graphs

GDABench, [published on arXiv in July 2024](https://arxiv.org/pdf/2407.11052), benchmarks unsupervised graph domain adaptation on five datasets, Airport, Blog, ArnetMiner, Twitch and MAG, spanning 74 source-target pairs. Its authors chose them for the complexity of the shift, the scale, and the potential for downstream applications.

## How to choose a benchmark, and what no benchmark can tell you

When a research team picks a dataset, a practical checklist helps:

1. **Match the shift to your problem.** Style shift (photos to sketches) is not sensor shift (one microphone to another) or label shift.
2. **Read the protocol.** Which domains are source and target, and are target labels allowed?
3. **Check access terms.** Some sets need registration or login, as with Cityscapes and the PCMDA package.
4. **Check the documentation for quirks,** like the Office-Home class count.
5. **Watch for negative transfer.** IEEE’s overview warns that [adaptation can hurt when domains are too dissimilar](https://technav.ieee.org/topic/transfer-learning/).

The fifth point leads to the limit of every public benchmark. The domains in them are stand-ins: product photos, review sites, read-aloud audiobooks. A company that builds an AI agent for insurance claims, field-service dispatch or B2B sales does not have a benchmark domain. It has a real target domain, and the only way to represent it is with real examples from that kind of business.

That is why AI developers look for company data: Slack threads, email history, CRM records, support tickets, codebases, call recordings and financial exports. A model adapted on public text sees how people write reviews. A model adapted on real support tickets sees how customers describe their problems and how your team resolves them.

## Where your data fits: the target domain nobody else has

If you run a company with an established data system, you already hold something no benchmark offers: a long, consistent record of one real operating environment. Businesses with 20 to 500 staff and more than three years of history are the typical profile for the sellers we work with at Bright Info.

Several types of company data map directly onto adaptation use cases:

- **Customer support tickets and chat logs:** domain-specific language and resolution patterns for text models and agents.
- **CRM records and sales email:** how deals progress in a given industry. Our page on [CRM data monetization for AI training](https://www.brightinfo.com/blog/crm-data-monetization/) covers this in detail.
- **Call recordings:** speech in real acoustic conditions, the real-world version of what LibriAdapt simulates with microphones and noise.
- **Financial exports:** structured records that buyers assess for type and quality, as explained in our guide to [financial datasets for AI](https://www.brightinfo.com/blog/financial-dataset-ai/).
- **Sensor and maintenance logs:** see our overview of [predictive maintenance datasets](https://www.brightinfo.com/blog/predictive-maintenance-datasets/).
- **Human feedback and preference data:** covered in our explainer on [RLHF datasets](https://www.brightinfo.com/blog/rlhf-datasets/).

If you are weighing whether your records count, the question is not whether they look like a clean benchmark. It is whether they represent a real environment that a model would otherwise never see.

## Why one offer is the wrong way to price it

Most companies that consider selling data receive, or go looking for, a single offer. A single offer has no reference point. You cannot know whether it is half or double what another buyer would pay, because the buyers value the same data differently depending on what they are training.

We built Bright Info around that problem. Our description of the market is that company data typically ranges from $100K to $4M in value, and the spread is why a competitive process matters. Our [AI data buyers directory](https://www.brightinfo.com/buyers/) profiles who buys company data, and our [November 2026 market report](https://www.brightinfo.com/blog/ai-data-licensing-market-report/) tracks the licensing market.

Here is how the process works on our side:

- **Intake takes about two minutes.** There are no upfront fees.
- **We act as a sell-side broker for you,** the data owner, not for buyers. We invite every qualified buyer to bid competitively in sealed bids.
- **Privacy is built into the process.** De-identification is done by an independent partner that you approve, and no company data moves until a deal is signed.
- **License terms are controlled and reviewed.** Legal review and transparent terms are part of the process, and our [seller checklist for AI data licensing agreements](https://www.brightinfo.com/blog/ai-data-licensing-agreement/) lays out what to look for.
- **You pay nothing.** The winning buyer pays our commission, and only if a deal closes. Because we are paid the same way regardless of who wins, we have no commission bias toward any particular buyer.

## A worked example

Take a hypothetical logistics software company with 120 employees and six years of history. It holds roughly 200,000 support tickets, a CRM with a full record of renewals and churn, and years of internal Slack discussion about how customer issues were diagnosed. This example is illustrative, not a client result.

A buyer building a customer-support agent for B2B software values the tickets and Slack threads most. A buyer building sales-forecasting tools values the CRM history. A buyer training coding assistants may care mainly about the codebase. Each values a different slice, and each will bid up to a different level. If the company talked to only one, it would learn one number, and that number would reflect only that buyer’s use case.

Putting the whole dataset in front of all qualified buyers lets those different needs surface as competing bids. That is the mechanism behind the range from $100K to $4M, and it is why the starting point is a valuation, not a guess. Our [company data valuation page](https://www.brightinfo.com/blog/company-data-valuation/) explains what goes into it.

## Questions to ask about your own data before you start

You can borrow the benchmark mindset to assess your data’s readiness:

1. **What is the domain?** Name the environment: an industry, a product line, a customer type.
2. **How long is the record?** Longer, consistent histories show change over time, which is a kind of shift in its own right.
3. **Is it labeled or labelable?** Ticket categories, deal stages and resolution codes work like benchmark labels.
4. **Who owns it, and what do contracts say?** Customer agreements, employee communications and third-party data each have terms. This is where legal review earns its place.
5. **What would need removing?** Names, emails, account numbers and other identifiers. De-identification is the step that makes sharing viable.

> Researchers have spent a decade building benchmarks because the real target domains are hard to get. A model can look excellent on Office-Home or Amazon reviews and still stumble the day it meets a real company’s tickets, emails or calls.
>
> We see that gap as the reason company data has value. A long, consistent record of how one real business operates is exactly what adaptation methods are designed to learn from, and it cannot be recreated by scraping the public web.
>
> Our position is simple. The owner of that data should decide who sees it, on what terms, and for how much. That is why we work for the seller, invite every qualified buyer to bid, keep data in place until a deal is signed, and charge sellers nothing. A single offer tells you one buyer’s opinion. Sealed competing bids tell you what the data is worth.
>
> *Erik*

## Get a valuation of your company data

If your company has three or more years of operating data in email, CRM, support systems, call recordings or code, the next step is to find out what AI buyers would bid for it. We offer a free valuation with a two-minute intake and no upfront fees. You pay nothing unless a deal closes, and then the buyer, not you, pays our commission.

Start at [Bright Info](https://www.brightinfo.com/) and tell us what systems you run. We take it from there, with de-identification and legal review built in.

## FAQ

### What is the difference between domain adaptation and transfer learning?

Transfer learning is the broad idea of reusing knowledge from one task or dataset on another related one. Domain adaptation is the narrower case where the task stays the same but the input distribution changes, as [IEEE’s overview of transfer learning](https://technav.ieee.org/topic/transfer-learning/) explains. Classifying the same objects in photos versus sketches is domain adaptation.

### Which domain adaptation dataset should a beginner start with?

Office-Home is a common starting point for images because it has four clearly different domains and 65 categories, and it is hosted in a convenient format on [Hugging Face](https://huggingface.co/datasets/flwrlabs/office-home). For text, Amazon and Yelp review sets let you train on one product area and test on another. Whichever you pick, read its protocol first.

### Can a company’s own data serve as a domain adaptation dataset?

Yes. Real support tickets, CRM histories, call recordings and documents represent a specific operating environment, which is the target domain many production AI systems must handle. Before any use, identifying details must be removed and the license terms must be clear. At Bright Info, de-identification is done by an independent partner you approve, and no data moves until a deal is signed.

### How much is company data worth to AI buyers?

We have found that company data typically ranges from $100K to $4M in value, depending on its type, volume and history, and on how many buyers compete for it. A free valuation from us is the quickest way to see where your data falls, and you pay nothing unless a deal closes.

## Sources

- [Transfer learning, IEEE Technology Navigator](https://technav.ieee.org/topic/transfer-learning/)
- [Transfer Learning and Adaptation, MIT Foundations of Computer Vision](https://visionbook.mit.edu/transfer_learning.html)
- [A survey of transfer learning, Journal of Big Data](https://link.springer.com/article/10.1186/s40537-016-0043-6)
- [WILDS: A Benchmark of in-the-Wild Distribution Shifts](https://cs.stanford.edu/people/jure/pubs/wilds-icml21.pdf)
- [flwrlabs/office-home, Hugging Face](https://huggingface.co/datasets/flwrlabs/office-home)
- [Datasets of PCMDA, IEEE DataPort](https://ieee-dataport.org/documents/datasets-pcmda)
- [JiahongChen/da-dataset, GitHub](https://github.com/JiahongChen/da-dataset)
- [fastText classification-results.sh, GitHub](https://github.com/facebookresearch/fastText/blob/7842495a4d64c7a3bb4339d45d6e64321d002ed8/classification-results.sh)
- [Text-Classification dataset notes, GitHub](https://github.com/Renovamen/Text-Classification/blob/master/notes/datasets.md)
- [LibriAdapt, GitHub](https://github.com/akhilmathurs/libriadapt)
- [Adapting Multiple Distributions for Bridging Emotions, Entropy 2022](https://mdpi-res.com/d_attachment/entropy/entropy-24-01250/article_deploy/entropy-24-01250-v3.pdf?version=1663750316)
- [Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation, arXiv](https://arxiv.org/pdf/2407.11052)

## Recommended

- [Predictive Maintenance Datasets](https://www.brightinfo.com/blog/predictive-maintenance-datasets/)
- [RLHF Datasets: What They Contain and How They Work](https://www.brightinfo.com/blog/rlhf-datasets/)
- [CRM Data Monetization for AI Training](https://www.brightinfo.com/blog/crm-data-monetization/)
- [Financial Dataset AI: Types, Quality Checks and Value](https://www.brightinfo.com/blog/financial-dataset-ai/)

---

More posts: https://www.brightinfo.com/blog/

<!-- topical-links:start -->
## Continue your research

- [AI data buyers: programs, intermediaries and terms](/buyers/)
<!-- topical-links:end -->
