Restaurant review data at Yelp scale: the corpus, delivered as rows

Every conversation about restaurant review data starts with the same benchmark: 6,990,280 reviews of 150,346 businesses across 11 U.S. metropolitan areas, restaurants dominating the business mix throughout, each establishment carrying hours by weekday, aggregated star ratings and structured attributes - ambience, parking availability, good-for-kids, alcohol service, outdoor seating. The corpus itself is superb. What trips people up is treating it as a single undifferentiated thing, when it actually answers four different questions:

The first row hands over complete, dated review prose at scale. The second buys currency and consumer-display rights. The third supplies labels for supervised work. The fourth trades prose depth for global breadth and observed footfall. Datadory delivers all four as products on stable schemas, so the choice becomes a routing decision instead of a research project.

Only the first row gives you full essay-length reviews - which is why most stacks start there and bolt the others on.

Two things separate it from ordinary review piles. First, coherence: one cut, one schema, every review joined to both its author and its venue, so the whole reads as a connected graph rather than loose text. Second, the attribute grid beside the prose - "family-friendly patios serving beer in walkable neighborhoods" is a filter pass over typed columns, not a reading exercise.

The grain is six record families - one object per business, review, user, check-in and tip, plus 200,100 photos - all joined on stable identifiers. That grain is the payoff: sentiment by cuisine, reviewer lifetime value and venue-level visit patterns fall out of group-bys once the families arrive pre-joined, which is exactly how they ship through Datadory. Business records carry category aliases, hours broken out per weekday, star ratings and attribute flags cast as booleans.

Within Datadory's catalog of 1,744 datasets averaging 7.81 on quality, this record scores 8/10 - credited as the largest review corpus in the slice, docked because a snapshot cannot answer questions that need this week's signal. Its scale has no rival on the shelf; its vintage is stated in your sample report so nobody trains on an assumption. Workflows that do need this week's signal belong with the daily-refreshed records below.

Hugging Face Datasets - Restaurant Search indexes 975 community corpora matching "restaurant," and several are purpose-built for supervised work:

1. mrcaelumn/yelp_restaurant_review_labelled - review text paired with sentiment labels across 1M-10M rows in parquet shards, the collection's heavyweight.
2. yqzheng/semeval2014_restaurants - the classic aspect-based sentiment benchmark.
3. tomaarsen/setfit-absa-semeval-restaurants - few-shot ABSA on the SetFit recipe.
4. tner/mit_restaurant - tokens/tags arrays for menu-and-query named-entity recognition.
5. GEM/cs_restaurants - Czech dialogue data-to-text generation with dialog_act and delexicalized fields.
6. cmotions/NL_restaurant_reviews - Dutch-language review coverage.

The honest arithmetic behind the "975" headline: 831 of those sets hold fewer than 1,000 rows, tag aggregation puts 774 in tabular-classification (mostly small synthetic tables) against just 8 in text-classification, and the whole collection is static material with per-repo last-modified dates. Substantive NLP corpora are a minority - reasons the record scores 7 rather than higher in our catalog. As fixtures for model development they are exactly right; as a maintained feed they are not that at all.

The trade-offs sit record-by-record in Yelp-scale review corpus vs Hugging Face Datasets - Restaurant Search, and the data-scientists lens on restaurants ranks the slice for ML work.

Foursquare Places API is the commercially strongest answer when the question is "is this place decent, right now, anywhere on earth." Its place graph aggregates records from more than 100,000 sources into 100 million-plus POIs across 200+ countries, organized through a 1,100-plus category taxonomy whose Dining and Drinking branch holds the restaurant tree.

Each place record can carry rating, stats.total_ratings, total_photos and total_tips, a visit-derived popularity score, veracity_rating, price tier, amenity flags, and three lifecycle stamps: date_created, date_refreshed and date_closed. That closure field alone outclasses any frozen dump - a chain's store-closure analysis needs to know which locations shut, not just which ones score well.

The substrate is what separates it from directories assembled by hand: one billion-plus photos, tips and reviews and sixteen billion-plus historical check-ins stand behind the signals. Tips themselves are short-form reactions, not essay-length reviews - treat them as signal density rather than prose corpora, and pair them with the snapshot above when the writing matters.

For surrounding geography, Overture Maps Places contributes about 9.98 million Food and Drink features among roughly 74 million POIs per monthly release, each carrying an existence-confidence score between 0 and 1 that doubles as the primary cleaning lever. Joining the two gives a fully cleared base layer under whichever review route answers the question.

If stars were only ever a proxy for "is this place decent and safe," two regulators publish better evidence continuously, and neither comes with reuse anxiety.

NYC DOHMH Restaurant Inspection Results posts about 295,000 violation-level rows across 31 columns: letter grades, critical flags, numeric scores, cuisine descriptions, geocodes down to census tract. Inspections run through August 18, 2026 in the August 2026 pull, with a record date of August 20 - the canonical benchmark dataset behind NYC's letter-grading site, scoring 9/10 in our catalog.

UK Food Hygiene Rating Scheme (FHRS) Open Data rates food businesses 0-5 across roughly 380 local authority feeds - England, Wales and Northern Ireland under FHRS, Scotland under FHIS - with hygiene, structural and confidence-in-management sub-scores, rating dates per business, and a live central-London query returning 9,777 establishments within two kilometres at research time. Camden alone publishes 4,260 records. It scores a perfect 10/10.

Between them sits Data.gov Restaurant Catalog Search, discovery across a 552,271-dataset federal catalog whose restaurant thread reaches City of Chicago Food Inspections at roughly 314,000 rows running January 1, 2010 to present.

One establishment and one review, read straight off the shape that lands in your warehouse:

record        : Yelp-scale review corpus (snapshot cut)
business_key  : <stable establishment identifier>
city          : <one of the 11 metro areas>    state : <two-letter code>
categories    : <category alias list - restaurants dominate>
stars         : 4.0        review_count : <integer>
attributes    : ambience=casual, outdoor_seating=true, alcohol=beer_and_wine
hours         : Mon-Thu 11:00-22:00, Fri-Sat 11:00-23:00, Sun 10:00-21:00

review_key    : <stable review identifier>       business_key : <joins to the row above>
stars         : 5                                date : 2019-03-14
text          : "Best patio in the neighborhood - staff remembered our order..."
useful_votes  : 12                               cool_votes : 4

The join keys are the point. Identity, geography, classification, commerce signals and the attribute grid arrive as separate typed columns in one row - which is what lets a Philadelphia cut join cleanly against a Phoenix cut without remapping anything, and lets the labelled corpora attach to real venue attributes instead of floating text.

Every shipment includes validation rows, a coverage profile mapped to the metro areas and categories you name, and the vintage statement for the cut. Whatever you prototype on the sample survives delivery intact.

Data scientists train sentiment and recommendation models on the snapshot's 6.99M joined reviews, then validate against the labelled corpora before committing to production weights. Market researchers and consultants read cuisine-level sentiment and attribute grids for positioning work - who says what about patios, parking and kid-friendliness, by metro. Competitive intel and product teams track chains through the Foursquare place graph, using date_closed to separate expansion stories from shrinkage ones. PropTech and site-selection analysts pair review volume with inspection grades to score trade areas on reputation and safety together. Hospitality operators benchmark their own ratings and tip density against the competitive set in their metros.

Adjacent workflows on this blog: the restaurants data guide maps all twelve pooled records for the industry, and the best restaurants datasets ranking sorts the slice by quality score.

Name the metros, categories, layers and rhythm when you request a sample and the extract arrives already cut to that scope, field dictionary attached.

This cluster sits inside a twelve-record pool spanning snapshot corpora, community NLP collections, commercial place graphs and regulator transparency feeds. The restaurants data hub lists every product with coverage and delivery options in one table, and the restaurants data guide maps how review corpora fit beside inspection feeds and the Overture place layer.

Start from the question, not the format: the Yelp-scale review corpus product page shows sample rows, the full field dictionary and coverage chips, and the sample ships first either way.