Restaurants Data: Inspection Records, Global Footprints and Review Corpora · Head-to-head
Yelp structured datasets vs Hugging Face Datasets - Restaurant Search
Which restaurants data: inspection records, global footprints and review corpora data fits your job: Yelp structured datasets, or Hugging Face Datasets - Restaurant Search. API, files, or your warehouse. Daily, weekly, or hourly.
Yelp structured datasets
Hugging Face Datasets - Restaurant Search
Where the fields line up
No shared field names. These two answer different questions.
| Field | Yelp structured datasets | Hugging Face Datasets - Restaurant Search |
|---|---|---|
Business record | documented | not in this set |
Review record | documented | not in this set |
User record | documented | not in this set |
Check-in record | documented | not in this set |
Tip record | documented | not in this set |
Photo record | documented | not in this set |
Structured business attributes | documented | not in this set |
Category aliases | documented | not in this set |
Opening hours by weekday | documented | not in this set |
Star rating | documented | not in this set |
text | not in this set | documented |
label | not in this set | documented |
Coverage, side by side
| Yelp structured datasets | Hugging Face Datasets - Restaurant Search | |
|---|---|---|
| Geographic | 11 U.S. metropolitan areas; the named metro list ships with your sample report | Global and multilingual: English dominant, with Dutch, Czech, Vietnamese and other language corpora present |
| Temporal | One coherent multi-year snapshot; timestamps span several years inside the cut, vintage stated in your sample report | Static research corpora of varying vintages; each repository versions itself with its own last-modified stamp |
| Granularity | The establishment - one object per business, review, user, check-in, tip and photo | The example - one labelled sentence, token-tag sequence, dialogue act or tabular row per repository |
What each contains
Pick by fit, not by loyalty.
| Yelp structured datasets | Hugging Face Datasets - Restaurant Search | |
|---|---|---|
| Publisher | Yelp | Hugging Face - community-uploaded datasets on the Hugging Face Hub |
| Subject lens | One coherent educational cut of real Yelp activity: businesses, reviews, users, check-ins, tips and photos, restaurants dominating | A search over 975 'restaurant'-matching repositories: review sentiment, aspect-based sentiment, menu NER, task-oriented dialogue, recommendation |
| Unit of analysis | The establishment - one object per business, review, user, check-in, tip and photo | The example - one labelled sentence, token-tag sequence, dialogue act or tabular row per repository |
| Record scale | 6,990,280 reviews; 150,346 businesses; 200,100 photos | 975 matching repositories; size bands from under 1K rows (831 of them) up to 1M-10M rows |
| Geographic coverage | 11 U.S. metropolitan areas; the named metro list ships with your sample report | Global and multilingual: English dominant, with Dutch, Czech, Vietnamese and other language corpora present |
| Temporal coverage | One coherent multi-year snapshot; timestamps span several years inside the cut, vintage stated in your sample report | Static research corpora of varying vintages; each repository versions itself with its own last-modified stamp |
| Field dictionary | Documented at record-family level: business, review, user, check-in, tip, photo, plus attribute grid, categories, hours, stars | Seven documented payload fields verified during research: text, label, tokens, tags, dialog_act, target, gem_id |
| Best for | Volume-plus-context work: venue segmentation, recommender systems, reviewer behavior, foot-traffic and media analysis | Task-ready training: sentiment classification, ABSA benchmarks, menu NER, dialogue generation, cross-language transfer |
What each does better
Yelp structured datasets
Volume with joins intact. 6,990,280 reviews pointing back into 150,346 business records through stable identifiers, so sentiment by cuisine, reviewer lifetime value and venue-level visit patterns fall out of group-bys rather than reconstruction. The HF shelf has scale in exactly one place - the 1M-10M-row labelled Yelp subset - and no cross-repository join keys anywhere else.
The venue context labels never see. Structured attributes (ambience, parking, alcohol service), hours broken out per weekday, category aliases placing every establishment in the taxonomy. A sentence-level corpus cannot answer whether the brunch spot people praise has outdoor seating - this record can, on the same row.
Foot traffic under the ratings. Check-ins timestamp arrivals at venues, and 200,100 photos attach to businesses for listing-media work. Neither signal exists anywhere in the HF collection, which is why recommender-systems prototyping keeps coming back to the corpus even when labels are available elsewhere.
One vintage, stated once. Every row belongs to the same coherent cut, so replication studies hold together without auditing nine hundred separate last-modified stamps.
Hugging Face Datasets - Restaurant Search
Labels ready to train against today. Supervised work needs annotated examples, and the shelf carries them across four substantive families: 8 text-classification sets for review and aspect sentiment, 7 text-generation sets led by GEM/cs_restaurants, 6 question-answering corpora and 3 token-classification sets including tner/mit_restaurant. Starting from 6.99M unlabeled reviews means annotating from scratch; starting here means fine-tuning by Friday.
Benchmark lineage you can cite. SemEval-2014 aspect categories and GEM splits were battle-tested long before they landed here, which makes results comparable across papers instead of idiosyncratic to your annotation guidelines.
Languages beyond the eleven metros. Dutch reviews (cmotions/NL_restaurant_reviews), Czech dialogue, Vietnamese and other language corpora sit in the same result set - cross-language dining vocabularies the single-country Yelp cut cannot reach.
Problem types no venue schema anticipates. Task-oriented dialogue, few-shot ABSA via tomaarsen/setfit-absa-semeval-restaurants, recommendation pairs (ManikaSaini/zomato-restaurant-recommendation) - coverage of how machines learn to read restaurants, not just what restaurants are.
Where they're equivalent
More than the corpus-versus-shelf framing suggests. Both feed the same downstream jobs - review mining, recommender systems, NLP coursework, market research on dining language - and both are static research material rather than monitoring feeds: the Yelp cut is one coherent snapshot whose vintage is stated in your sample report, and the HF repositories version independently with their own last-modified stamps.
Both score above or near the catalog mean (8/10 and 7/10 against 7.81), and both carry an honesty clause in the dictionary itself. Yelp documents its corpus at the record-family level rather than promising column-level detail it does not publish; the Hub collection admits that 831 of its 975 repositories sit under 1K rows, mostly small synthetic tables inflating the count. Neither asks you to trust a headline number uncritically - including the number nine-seventy-five.
And both are complements, not substitutes, because their weaknesses are symmetric: the corpus has no labels, the shelves have almost no context.
The verdict
Verdict: sample both - the pairing is a pipeline, and neither half completes it alone.
Take Yelp structured datasets if your question starts with a place or a person. Venue segmentation by attributes and cuisine, reviewer lifetime value, foot-traffic patterns from check-ins, listing-media analysis over 200,100 photos, city-level market structure across the metros. Anything answered by describing establishments and their reviewers thoroughly.
Take Hugging Face Datasets - Restaurant Search if your question starts with a model. Sentiment classifiers, aspect-based sentiment, menu-entity recognition, dialogue-state generation, recommendation ranking - anything answered by training on annotated examples with benchmark lineage behind them.
Data scientists & ML engineers usually need both halves and know it; market researchers can often stop at the corpus; developers & builders prototyping a language feature start on the shelf and grow into the volume once the prototype works.
Sample both, pick by fit. See Yelp structured datasets · See Hugging Face Datasets - Restaurant Search
Or take both in one feed
Yes - and the largest single corpus on the HF shelf proves the pattern, because it is already derived from Yelp reviews. The defensible workflow runs volume first, labels second: prototype and validate the approach on a labelled subset from the shelf, then retrain against the full unlabeled corpus once the method is worth scaling - weak supervision seeded from label conventions, aspect lexicons grown from SemEval categories, NER patterns transferred from tner/mit_restaurant onto tip text.
Three seams decide whether the merge holds. First, lineage: the labelled set inherits its platform's terms through derivation, so pin the chain of custody before any production use - the sample report traces it per repository. Second, distribution shift: benchmarks carry older vintages than the corpus cut, so evaluate the transferred model on held-out rows from the corpus itself, not only on the benchmark test split. Third, schema drift: label indices and tag inventories differ per repository, so normalize them into one mapping before pooling.
Browse the rest of the shelf at the restaurants data hub, see how the slice ranks in the best restaurants datasets list, or read the publishers' profiles - Yelp as a source and Hugging Face as a source.
Datadory ships either record alone or both merged onto one calendar, delivered daily, weekly, or hourly - your call. Or take both in one feed.
API, files, or your warehouse. Daily, weekly, or hourly.
Fair questions
Is Yelp structured datasets better than Hugging Face Datasets - Restaurant Search?
Better at different jobs. The corpus wins volume and context: 6,990,280 reviews joined to 150,346 venues with attributes, hours, check-ins and photos across 11 U.S. metros. The Hub shelf wins labels: sentiment classes, aspect categories, entity tags and dialogue acts across four substantive task families. Sample both.
Do Yelp structured datasets and Hugging Face Datasets - Restaurant Search contain the same kind of data?
They share review text about restaurants and little else structurally. One is a corpus of six joined record families keyed by establishment; the other is a search over 975 community repositories whose schemas change per dataset. The bridge between them is real - the shelf's largest corpus derives from Yelp reviews - but the units of analysis never coincide.
Which one should I train a restaurant sentiment model on?
Start on Hugging Face Datasets - Restaurant Search: yelp_restaurant_review_labelled ships pre-labelled reviews in the 1M-10M row band, with SemEval-2014 aspect benchmarks beside it. Move to the Yelp structured datasets when the method deserves full volume - 6,990,280 unlabeled reviews to retrain against beat any annotating-from-scratch plan.
Which dataset covers more geography?
Depends on the axis. The Yelp structured datasets goes deep in one country: 11 U.S. metropolitan areas with the named metro list reported with your sample. The Hub collection goes wide by language instead - Dutch, Czech, Vietnamese and other corpora alongside English-dominant sets - reaching diners no U.S. metro frame contains.
Can Datadory deliver both Yelp structured datasets and Hugging Face Datasets - Restaurant Search together?
Yes - sample both and pick by fit, or take both in one feed. Each arrives normalized to its documented dictionary, with sample rows and the vintage statement attached before anything ships, delivered daily, weekly, or hourly - your call. Or take both in one feed.