Best Ecommerce Datasets and Databases for ML Models (Free & Paid)

Last updated: August 2026

Search “ecommerce datasets” and five of the seven top-ranking roundups lead with a dataset collected between December 2010 and December 2011. None of them tell you that. None of them state a license either. Both omissions cost you more than they look like they should: a model trained on fifteen-year-old baskets learns fifteen-year-old buying behaviour, and a dataset you cannot legally ship is worth nothing to a production system no matter how clean the rows are.

So this article prints both, for every entry. Twenty ecommerce datasets and databases, free and paid, organized by seven ML use cases, each with six fields stated plainly: collection window, size, license, hosting and access, best-for task, and honest limitations. By the end you will be able to pick the right dataset for your model task in one read, and know exactly when a static dataset stops being enough.

How We Evaluated These Datasets

Every entry below is scored against the same rubric and prints the same Six-Field Spec Card: collection window, size and records, license, hosting and access, best-for ML task, and limitations. The card is the whole methodology, visible on every dataset.

Recency and license lead for a blunt reason: bad data is a P&L line item. In a 2025 IBM Institute for Business Value report, over a quarter of organizations estimated they lose more than $5 million annually to poor data quality, and 7% reported losses of $25 million or more. The stakes are compounding too: per Precedence Research (updated January 8, 2026), the AI in e-commerce market is valued at $9.01 billion in 2025 and $11.21 billion in 2026, a 23.59% CAGR over 2026–2035. Sourcing is now a first-order engineering decision, worth the same rigor you would apply through a data quality framework for external sources.

One observation shaped the rubric more than any stat: in our review of the seven top-ranking roundups for this keyword, none states the license or collection window for any dataset it recommends. Those two fields block production use most often, so they lead every card. All links click-verified August 4, 2026; no vendor paid for placement.

Ecommerce Datasets at a Glance

With the rubric set, here is the full field, details after. Four hubs supply most of the free entries: Kaggle, the UCI Machine Learning Repository, Hugging Face, and Google BigQuery public datasets, with Google Dataset Search as the discovery layer. Jump to the category matching your model task; every dataset gets a full spec card below. One caveat up front: Kaggle mirrors carry heterogeneous or unstated licenses, addressed per entry. And one teaser for why the window column matters: the most-recommended dataset on page 1, UCI Online Retail, is 541,909 rows collected December 2010 to December 2011. Roster current as of August 2026.

Best Ecommerce Datasets for Recommendation Systems

First of the seven categories on the roster. Recommendation models live on session and event data with timestamps: implicit signals (views, carts, orders) at scale beat small explicit-rating sets for nearly every production task. Note the distribution shift here too: Hugging Face load_dataset with Parquet-native files is the modern access pattern for large corpora; the Kaggle CSV download is the legacy one.

1. Amazon Reviews 2023 (McAuley Lab)

What’s in it. Reviews with ratings, text, fine-grained timestamps, and rich item metadata: 571.54M reviews spanning May 1996 through September 2023, 245.2% larger than the 2018 version (233.10M reviews).

Best for. Pretraining sequence recommenders and fine-tuning LLMs on commerce text. Overkill for a small explicit-feedback classifier.

Watch out. Route yourself off the 2014 and 2018 versions entirely; their duplicate and leakage issues were fixed only in the 2023 refresh, yet they still headline tutorials. Access is one line:

from datasets import load_dataset

dataset = load_dataset("McAuley-Lab/Amazon-Reviews-2023",
                       "raw_review_All_Beauty",
                       trust_remote_code=True)

2. OTTO Recommender Dataset

What’s in it. Click, cart, and order events per session, exactly the implicit-feedback shape production recommenders consume. At 12M sessions and 220M events under CC BY 4.0, it is the largest openly licensed real session dataset available.

Best for. Teams that need real sessions and a clean commercial-use answer. Smaller than Amazon Reviews 2023, but the license question has an actual answer, which matters more for a shipping model than another order of magnitude of rows.

Watch out. A frozen 2022–23 window from one retailer: benchmark and pretraining material, not current behavior.

3. Retailrocket Ecommerce Dataset

What’s in it. The classic view-cart-purchase funnel with a catalog attached, small enough to iterate on a laptop.

Best for. Prototyping event-join and negative-sampling logic before you scale it, ideally alongside the operational playbook for AI training data.

Watch out. The CC BY-NC-SA license makes this a prototyping dataset, not a shipping one. The pattern across all three entries: scale, freshness, and license clarity almost never arrive together in free recsys data.

Best Ecommerce Datasets for Demand Forecasting and Inventory

Second on the roster: demand forecasting. Forecasting models want long horizons, hierarchy (item, store, region), calendar effects, and per-SKU granularity. This category also has the sharpest teaching-data-versus-production-data divide, so each window below is printed next to the name, every time.

4. M5 Walmart Sales Forecasting

What’s in it. Daily unit sales with the hierarchy and calendar features real demand models need, which is why M5 remains the reference benchmark for retail forecasting methods.

Best for. Benchmarking. If an architecture cannot beat simple baselines on M5, it will not survive your production data.

Watch out. Benchmark, not behavior: one retailer, with history ending in June 2016.

5. Instacart Market Basket 2017

What’s in it. Order sequences with timing features you rarely get elsewhere: over 3 million orders from more than 200,000 users, released May 2017 under an as-is, non-commercial license.

Best for. Learning and benchmarking reorder models; the sequence and timing fields are the draw.

Watch out. The license is the hard stop. Non-commercial means the model it trains cannot ship in a commercial product. Prototype on it, then retrain on data you have rights to.

6. UCI Online Retail

What’s in it. Invoice-level transactions: 541,909 rows over exactly one year of one retailer’s trade, under CC BY 4.0.

Best for. The classroom. RFM and market-basket exercises work fine here, and the permissive license travels well in courses.

7. Maven Analytics Toy Store

What’s in it. A tidy relational schema that behaves like a small production database.

Best for. Rehearsing the SQL layer under a forecasting pipeline. The distinction this section turns on: teaching data builds skills, production data builds models, and no roundup tells you which is which.

Best Ecommerce Datasets for Product Search and Relevance

Third on the roster: search relevance. Labeled query-product pairs are the scarcest asset in ecommerce ML: they require real search traffic plus the kind of human judgment data labeling companies supply, and almost nobody publishes them. One dataset carries this category, and it is one of the few here you can legally build a commercial model on.

8. Amazon ESCI Shopping Queries Dataset

What’s in it. Each query-product pair is labeled Exact / Substitute / Complement / Irrelevant (the ESCI taxonomy) across three languages: 2.6M judgements over 130,652 queries. Fields include example_id, query, product_id, esci_label, loaded Parquet-native:

import pandas as pd
df_examples = pd.read_parquet('shopping_queries_dataset_examples.parquet')
df_products = pd.read_parquet('shopping_queries_dataset_products.parquet')

Best for. Relevance models you intend to ship. Said plainly: Apache-2.0 makes ESCI one of the only labeled search corpora that permits commercial model training. For multimodal work, the SQID (Shopping Queries Image Dataset) extension adds imagery on top.

Watch out. Frozen 2022, but relevance labels age far more slowly than prices or behavior, so it remains the canonical benchmark as of 2026. The honest caveat is distributional: these are Amazon’s queries and catalog; measure transfer to your own traffic rather than assuming it. It appears in none of the ranking roundups, which is your head start.

Best Ecommerce Datasets for Review NLP and Sentiment Models

Fourth on the roster: review NLP. Review NLP spans task scales: small labeled sets for classical classifiers, mid-size corpora for category models, and massive token counts for LLM fine-tuning and RAG over catalogs. Pick by task scale, not by whichever dataset a tutorial happened to use.

9. Women’s E-Commerce Clothing Reviews

What’s in it. A clean labeled classification set: 23,486 rows with rating, recommended flag, and product-hierarchy features.

Best for. Sentiment baselines where real commercial text matters more than scale.

Watch out. This dataset is itself a live example of the undated-Kaggle-upload problem: no collection window is disclosed anywhere. Useful default when you meet an undated upload: assume the data predates its upload year by years.

10. Zenodo Ecommerce Text Classification

What’s in it. 50,425 product-text rows in four categories: solid headroom for supervised category models without GPU ceremony.

Best for. Multi-class text classification beyond the clothing-reviews scale.

Watch out. Taxonomy-scale categorization needs bigger label spaces than four classes.

Cross-reference, not a duplicate entry: for LLM-scale review text, go back to Amazon Reviews 2023: 30.14B review tokens plus 30.78B metadata tokens with second-level timestamps. That token-count framing signals dataset builders now target LLM fine-tuning and RAG consumers, not just recsys teams. For the licensing and provenance side of such corpora, see public vs private data for AI training.

Best Ecommerce Datasets for Product Image and Categorization Models

Fifth on the roster: product imagery. Catalog vision covers categorization, attribute extraction, image search, and increasingly 3D and AR. The two entries here are a deliberate pairing no roundup makes: the toy benchmark everyone knows and the real catalog corpus most teams have never opened, both permissively licensed.

11. Fashion-MNIST

What’s in it. Ten classes of 28×28 grayscale garment images, as easy to load as MNIST and meaningfully harder to solve. Author Han Xiao of Zalando Research framed the motivation: “MNIST is too easy. Convolutional nets can achieve 99.7% on MNIST. Classic machine learning algorithms can also achieve 97% easily.” He adds that “MNIST can not represent modern CV tasks.”

Best for. Sanity-checking a vision architecture in minutes.

12. Amazon Berkeley Objects (ABO)

What’s in it. Real listings at scale: 398,212 catalog images across 147,702 listings, plus 3D models and 360° spins, under CC BY 4.0.

Best for. The production-grade counterpart to Fashion-MNIST. Benchmark data proves the architecture; catalog data proves the product.

Watch out. Storage and preprocessing are the real costs; budget pipeline time before training time.

Best Ecommerce Datasets for Customer Behavior, CLV and Churn

Sixth on the roster: customer behavior, the last use-case section before the paid marketplaces. Behavior and CLV modeling need event-level user journeys with timestamps, and real ones are scarce because they are full of PII and competitive signal. What exists splits three ways: real-but-old, obfuscated, and synthetic. Pick by whether you need signal or schema.

13. REES46 Multi-Category Store

What’s in it. The largest genuine public behavior corpus here: 285M real events across seven months, October 2019 through April 2020.

Best for. Anything needing real event sequences at scale, from propensity models to churn features.

Watch out. Print the window and the caveat together, always: the final months sit inside the COVID demand shock, and March–April 2020 behavior generalizes poorly to anything since. Catching that kind of silent skew is what data observability for third-party datasets is for.

14. GA4 Merchandise Store Sample (BigQuery)

What’s in it. A real GA4 export shape: nested event parameters, ecommerce items, the schema your production pipeline will meet.

Best for. Building GA4-to-warehouse pipeline patterns before pointing them at production exports.

Watch out. Google’s own documentation flags the limited internal consistency, an honest caveat straight from the source. Schema practice, not behavioral ground truth.

15. theLook eCommerce (BigQuery)

What’s in it. A full relational ecommerce schema queryable in place, free, with no download step: the operational contrast with everything Kaggle-hosted here. A CLV starter you can run now (table and column names check against the public dataset’s documented schema; run it in your own BigQuery project):

SELECT
  u.id AS user_id,
  COUNT(DISTINCT o.order_id) AS orders,
  SUM(oi.sale_price)         AS lifetime_revenue
FROM `bigquery-public-data.thelook_ecommerce.users` u
JOIN `bigquery-public-data.thelook_ecommerce.orders` o  ON o.user_id = u.id
JOIN `bigquery-public-data.thelook_ecommerce.order_items` oi ON oi.order_id = o.order_id
GROUP BY user_id
ORDER BY lifetime_revenue DESC
LIMIT 100;

Best for. Query-in-place development of the relational layer under CLV and churn models.

Watch out. Synthetic data can teach a pipeline; it cannot validate a behavioral hypothesis. If your claim is about how humans buy, synthetic tables will happily confirm whatever your features imply.

16. Olist Brazilian E-Commerce

What’s in it. Real orders with payments, reviews, and geolocation joined across tables: the most production-shaped free relational set here.

Best for. End-to-end practice on real multi-table commerce data.

Watch out. The NC license is confirmed on the Kaggle page, so commercial training is off the table. Prototype here, then retrain on data you have rights to.

Best Paid Ecommerce Datasets and Data Marketplaces

That completes the use-case sections of the roster. What paid buys you is specific: recency tiers, scale, support, and a defined schema. What it does not buy, automatically, is freshness forever. Published marketplace rate cards and a 2026 provider roundup put the economics in a narrow band: subscriptions observed at $59–$1,000 per month, and per-record marketplace pricing around $250 per 100,000 records, all observed 2026-08-04.

A structural note before the entries. The outline gave pricing data its own category, on one condition: that at least one free source still published on a rolling refresh. Verification broke the condition (details in the first entry below), so the two pricing entries sit here with the marketplaces, where refresh is a product tier you pay for rather than a promise in a README.

17. octaprice ecommerce-product-dataset

What’s in it. CSV/JSON drops of around 1,000 live product records each, published through 2025.

Best for. Practicing on price observations newer than anything else in the free roster, as long as you date-stamp them honestly.

Watch out. This entry is why the pricing category collapsed into this section. We planned to present it as the one free source with a rolling weekly refresh, which is what its README still promises. The commit history says otherwise: the last drop landed December 9, 2025, eight months before this audit. Free freshness did not just have a low ceiling. On verification, it had none.

18. Electronic Products and Pricing Data (Datafiniti)

What’s in it. Structured price fields across more than 15,000 electronics products; a reasonable schema to practice feature extraction on.

Best for. Learning the shape of multi-source price data before paying for the real thing.

19. Bright Data Dataset Marketplace

What’s in it. Pre-collected product, pricing, and review snapshots at marketplace scale, priced around $250 per 100,000 records (roughly $0.0025 per record) on published rate cards as of August 2026.

Best for. Major-platform coverage at a scale free data cannot reach, without running extraction yourself.

Watch out. Run the per-record math at your refresh cadence: at ~$0.0025 per record, a 10M-record snapshot refreshed monthly is a recurring line item, not a one-time buy. And the schema is the vendor’s; if your model needs a field the dataset does not carry, no tier fixes that.

20. Similarweb Shopper Intelligence

What’s in it. Amazon sales, search, and product metrics delivered as CSV/JSON through a Batch API, with up to 36 months of history depending on plan. That plan-gating is worth noticing: even in paid products, history depth is a tier you buy.

Best for. Feature pipelines that join Amazon market context (search volume, category share) onto your own data.

Datarade (where to scope the paid landscape)

Datarade is not a dataset; it is a marketplace aggregator for comparing many vendors’ ecommerce datasets in one place, useful for scoping what exists and what it costs, alongside a broader view of alternative data vendors. One misconception to park here and settle after the comparison table: paid does not mean always-fresh.

How Do All 20 Ecommerce Datasets Compare?

One screen, side by side: the table that exists nowhere else on page 1. Scan the window column first, then license, then size; that order disqualifies faster than it qualifies, which is the point. (Datarade is an aggregator, not a dataset, so it takes no row.)

When Does a Static Dataset Stop Being Enough?

Every commercial page answers this in one line: “free data is outdated, buy data.” That is an assertion, not an argument, so here is the argument in five rungs. The Escalation Ladder decides when to climb, and each rung is priced by facts already printed above.

1. Free static datasets. Prototyping and benchmarking: UCI, Retailrocket, Fashion-MNIST prove code and architecture cheaply.
2. Modern academic corpora. Amazon Reviews 2023, ESCI, OTTO, ABO: pretraining and fine-tuning within license terms.
3. Paid static datasets. Snapshots at roughly $0.0025 per record: scale and a defined schema, still frozen between refreshes.
4. Subscription refreshed feeds. Recency at the cadence you pay for, history depth plan-gated (Similarweb’s 36-month gate).
5. Managed custom extraction. Your websites, your schema, your cadence, quality-assured continuously; for production models whose inputs cannot be someone else’s snapshot.

Climb a rung when one of four walls hits: drift in production metrics, a coverage gap on the websites or SKUs you need, a schema mismatch between vendor fields and your features, or a license wall blocking commercial use. Drift is the one with a number attached: in the first systematic study of AI aging, Vela et al. tested 4 standard ML models against 32 datasets across four industries, 128 (model, dataset) pairs, and “observed temporal model degradation in 91% of cases” (Scientific Reports, 2022). In production, drift is a when, not an if. That is why extracting ecommerce data at scale on a controlled cadence, or having a partner such as Forage AI run that extraction as a managed service, sits at the ladder’s top rather than as a vendor afterthought; the build vs buy decision up there deserves its own analysis.

Beyond Static Datasets: Managed Ecommerce Data Extraction

Rung 5 made concrete. Managed Data Extraction means a team runs the pipeline for you: the websites you specify, a schema built for your features (price history, seller-level granularity, event timestamps), refreshed on your cadence, quality-assured on every run, and owned outright, with no redistribution clauses or non-commercial ambiguity attached.

That last property dissolves the license wall: half the free datasets above are blocked from commercial use or undated into uselessness, while extraction on your own behalf produces data whose terms you never re-audit. Forage AI operates in exactly this space: Web Data Extraction across 500M+ websites, delivered as a managed service in which Forage AI handles selector drift, anti-bot evolution, and schema changes as part of the service, with a 3x QA team on every delivery and onboarding in 1-2 weeks.

Honest fit boundaries, because the ladder cuts both ways: managed extraction is overkill for prototyping and benchmarking, where static datasets remain the right tool. It is the right call when a production model needs fresh, governed inputs and one of the four walls has already cost an incident or a quarter of engineering time.

The field’s broader shift backs this ordering. Andrew Ng, launching the data-centric AI movement in 2021, put it in four words: “Data is food for AI,” arguing that practitioner focus belongs on training-data quality rather than model architecture, and observing at the time that 80% of an AI developer’s time goes to data preparation (a 2021 figure, cited as context, not a current survey number).

Frequently Asked Questions

Where can I find free ecommerce datasets?

Four hubs cover nearly everything: Kaggle, the UCI Machine Learning Repository, Hugging Face, and Google BigQuery public datasets, with Google Dataset Search for discovery and GitHub hosting the newer releases (ESCI, OTTO). The roster table earlier in this article lists 18 free entries.

Can I use free ecommerce datasets commercially?

It depends entirely on the license, and the honest answer is usually no. ESCI (Apache-2.0), OTTO and ABO (CC BY 4.0), and UCI Online Retail (CC BY 4.0) permit commercial use. Instacart is explicitly non-commercial; Olist, Retailrocket, and Datafiniti’s pricing sample all carry CC BY-NC-SA 4.0 on their Kaggle pages; and plenty of uploads elsewhere state no license at all: a shipping risk, not a bargain. License terms also change; re-verify the current terms on the hosting page, and get a legal read, before shipping a commercial model.

Which ecommerce dataset is best for recommendation systems?

Amazon Reviews 2023 for scale and LLM-era text, OTTO for cleanly licensed real sessions, Retailrocket for a compact starter. Use the 2023 Amazon version specifically; the 2014/2018 versions are superseded.

Is the UCI Online Retail dataset good for demand forecasting?

For teaching RFM and market-basket analysis, yes. For production forecasting, no: it is a single UK gift retailer’s transactions from December 2010 to December 2011, collected fifteen years ago.

How much do paid ecommerce datasets cost?

Observed pricing (August 2026): marketplace rate cards around $250 per 100,000 records with refresh tiers from daily to one-time, and provider subscriptions from $59 to $1,000 per month per a 2026 provider roundup. Total cost scales with refresh cadence, so price the cadence, not the file.

Is the Amazon reviews dataset still available?

Yes. The current version is Amazon Reviews 2023 from the McAuley Lab, hosted on Hugging Face via load_dataset. The 2014 and 2018 versions still circulate but are superseded, with duplicate and leakage issues fixed only in the 2023 refresh.

How much data do you need to train an ecommerce ML model?

Task-dependent, so beware any single number. Classical classifiers work on tens of thousands of labeled rows; sequence recommenders want millions of events; LLM fine-tuning consumes billions of tokens. Size the dataset to the task, then validate against a held-out time window, not just a held-out sample.

Make Dataset Sourcing a Cadence, Not a One-Time Pick

You now hold the two artifacts this article promised: a Six-Field Spec Card for every dataset worth considering, and an Escalation Ladder for the day the cards stop being enough. The uncomfortable part is the sustainability turn: datasets freeze and models age, with temporal degradation observed in 91% of tested model-dataset pairs. Sourcing is a cadence, not a one-time pick, so put a re-audit on the calendar: re-check the windows, licenses, and links your models train on, on a schedule, the way you patch dependencies.

Apply the six fields to the next dataset you download, whoever recommends it. And when you hit one of the four walls, that is the point where the ladder runs out of files and the data needs extracting fresh. That is the work Forage AI’s ecommerce data service exists for. Talk to our expert.

Sources

• Precedence Research (2026): Artificial Intelligence in E-commerce Market, updated January 8, 2026 – precedenceresearch.com
• IBM Institute for Business Value (2025): The True Cost of Poor Data Quality – ibm.com
• Vela, D., Sharp, A., Zhang, R., Nguyen, T., Hoang, A., Pianykh, O.S. (2022): Temporal quality degradation in AI models, Scientific Reports 12:11654 – nature.com
• Hou et al. (2024): Bridging Language and Items for Retrieval and Recommendation (Amazon Reviews 2023 paper) – arXiv:2403.03952
• Reddy et al. (2022): Shopping Queries Dataset (ESCI paper) – arXiv:2206.06588
• Xiao, H., Rasul, K., Vollgraf, R. (2017): Fashion-MNIST – arXiv:1708.07747
• Gil Press, Forbes (2021): Andrew Ng Launches A Campaign For Data-Centric AI – forbes.com
• Jeremy Stanley, Instacart (2017): 3 Million Instacart Orders, Open Sourced – tech.instacart.com
• Makridakis, S., Spiliotis, E., Assimakopoulos, V. (2022): The M5 competition: Background, organization, and implementation, International Journal of Forecasting – sciencedirect.com

Related Articles

• AI Training Data: The Operational Playbook – What to do with a dataset after sourcing: validation, labeling QA, and drift management.
• Ecommerce Data Scraping at Scale – Pipeline architecture for extracting fresh ecommerce data across 1M+ SKUs.
• Top Ecommerce Data Providers: How to Evaluate – The provider-side companion when no static dataset fits.
• Data Quality Framework & Quality Checklist for External Sources – How to vet any third-party dataset before training on it.
• AI Training Data Providers: How to Evaluate the Top Vendors – Cross-vertical vendor evaluation for teams outgrowing free datasets.