Hugging Face Datasets — Beer Search (83 datasets)

It is the Hugging Face Datasets Hub's entire 'beer' result set, catalogued as one product: 83 dataset repositories, dominated by consumer review corpora and NLP benchmarks, each publishing typed column schemas through its card's live data viewer.

The set splits into three tiers. The flagship tier carries the volume: SWilliams20/Kaggle_Beer_Reviews at 1,586,614 rows (~180 MB) with thirteen fields per review, and tlam25/beeradvocate_full1m5 at about 1.5 million rows - between them the deepest public consumer-opinion base the beer industry has. The benchmark tier repurposes or complements that material for machine-learning work: qi2peng2/beerqa for question answering, deniseiras/ABSA_beer for aspect-based sentiment, arize-ai/beer_reviews_label_drift_neg and _neutral for model-monitoring drills, matchbench/Beer and Beer-SM for entity-matching. The long tail holds smaller topical sets - television advertising, youth-drinking studies, Dutch-language question pairs.

One honest caveat belongs up front: a name-only hub search mixes genuine beer-domain corpora with incidental matches - robotics repositories whose titles happen to contain 'beer'. The true count of industry-relevant datasets sits below 83, which is why Datadory treats relevance filtering as part of curation rather than leaving it to the consumer. Get a sample of this dataset cut to the genuinely relevant corpora, or read the rows below first.

Two consecutive reviews of the same brewery, exactly as captured during the August 2026 research pass:

brewery_name : Vecchio Birraio     beer_name   : Sausa Weizen
beer_style   : Hefeweizen          beer_abv    : 5.0        brewery_id: 10325
review_overall: 1.5   review_aroma: 2.0   review_appearance: 2.5
review_palate : 1.5   review_taste: 1.5   review_time: 1234817823
review_profilename: stcules             beer_beerid: 47986

brewery_name : Vecchio Birraio     beer_name   : Red Moon
beer_style   : English Strong Ale  beer_abv    : 6.2        brewery_id: 10325
review_overall: 3.0   review_aroma: 2.5   review_appearance: 3.0
review_palate : 3.0   review_taste: 3.0   review_time: 1235915097
review_profilename: stcules             beer_beerid: 48213

Read the pair and the corpus's design shows itself. One reviewer, one brewery, two styles, and a full point of separation between the verdicts - the Hefeweizen lands at 1.5 overall while the English Strong Ale takes 3.0, with every axis moving in step. That structure is what five separate rating columns buy: you can ask whether this reviewer punishes appearance more than taste (here, no - both rise together), and repeat the question across 1.59 million rows. Half-point granularity throughout, Unix-timestamp timing on both rows, and integer IDs ready to serve as join keys. Your sample arrives pre-cut to the corpora and fields you name.

Thirteen fields define each review in the flagship corpus, dividing into identity, product description and judgment. The identity block - brewery_id, brewery_name, beer_beerid, beer_name, review_profilename - gives every row three resolvable entities: who made it, what was drunk, who said it. The product block - beer_style and beer_abv - carries the categorical and numeric attributes most analyses segment on.

The judgment block is the differentiator: review_overall, review_aroma, review_appearance, review_palate and review_taste decompose one opinion into five independent axes at half-point resolution. A star rating collapses all five; these columns let a model learn that a beer can smell wonderful and drink disappointing. review_time closes the schema as a Unix timestamp, which turns the whole table into a longitudinal panel rather than a snapshot.

Handle two columns deliberately. beer_abv has real gaps - treat zeros and blanks as missing, not as near-beer products. And review_profilename is a free-text username, useful for within-reviewer consistency panels but requiring normalization before any deduplication claim.

The thirteen-column spine covers every review row; the extensions below fold into additional fields on request because they depend on the shape of the delivery:

1. The sibling corpora. Entity-matching pairs (matchbench/Beer, matchbench/Beer-SM), aspect-based sentiment annotations (ABSA_beer), the beerqa question-answer corpus, drift-monitoring derivatives, and the 1.5M-row BeerAdvocate corpus - each folded in when named rather than shipped by default.
2. Relevance-filtered cuts. The false positives a name-only search drags in - robotics repos and unrelated titles containing 'beer' - dropped at curation time so your pipeline never meets them.
3. Cross-catalog joins. Brewery-name-keyed links onto venue-level establishment tables elsewhere in the brewers catalog, putting review scores and physical footprints in one table.
4. Derived aggregates. Average ratings by style, ABV band or brewery; review-volume trends by period; reviewer-panel counts informing weighting choices.
5. Schema documentation. Per-column type and population-rate metadata from each card's viewer, joined onto the dictionary so coverage expectations travel with the columns.

Name the ones your models need when you request the sample; column naming locks against live records at that point.

• Geography: global and deliberately unspecified. Review corpora aggregate BeerAdvocate and Kaggle community bases - heavily North American and European in practice - without shipping an explicit country column, so geography enters through brewery names and joins rather than a dedicated field. Provenance gets checked corpus-by-corpus during curation.
• Temporal: roughly fifteen years of consumer opinion. Review timestamps in the flagship corpus reach back toward 2009 and forward to collection, making trend, seasonality and reputation-trajectory work possible without stitching annual snapshots. The surrounding hub listings refresh continuously as maintainers push changes.
• Granularity: one row per review, thirteen fields per row, nothing aggregated above the individual verdict. The sentiment and QA sets run per-document; the entity-matching benchmarks run per candidate pair. No rollups hide underneath, so any 'average rating for style X' number is computable rather than inherited.

API, files, or your warehouse. Daily, weekly, or hourly.

Name the corpora, fields and any filters when you request the sample and it lands shaped to that scope with the thirteen-field dictionary unchanged - rating axes kept separate, identifiers kept as integers, timestamps kept raw for your timezone layer to interpret. Most teams take the full 1.59-million-row corpus once and keep a refresh rotating at whatever cadence matches their model-retraining cycle; key structure stays identical either way, so each pull appends cleanly to the last.

• Preference and style analytics - five independent rating axes turn 1.59 million reviews into style-cohort benchmarks: which categories over-index on palate versus appearance, where ABV stops predicting overall liking; the workflow continues on our market sizing page.
• NLP training and evaluation - review text paired with ratings yields supervised sentiment and regression corpora, with ABSA, QA and drift sets supplying ready-made evaluation baselines.
• Entity matching - the matchbench pairs exist to test fuzzy linkage of brewery and beer names, the exact problem distributor and marketplace catalogs face; methods continue on our citation-grade research page.
• Competitive monitoring - track a named beer's five-axis trajectory over fifteen years against its style cohort, catching reputation shifts before sales reports do.
• Demand-signal features - review volume and rating momentum feed forecasting models as leading indicators alongside production and tax-ledger series; see demand forecasting.

Data scientists and ML engineers get a million-scale labeled corpus with typed schemas and benchmark derivatives already split off - the difference between starting at feature engineering and starting at collection consent forms. Market researchers and consultants gain category-level preference claims grounded in five-axis scores rather than survey anecdotes. Brand and category managers can benchmark their products against style cohorts over time and tell whether a reformulation moved taste perception or merely optics. Developers and data-product builders ship recommendation and review-analytics features on Parquet-shaped files whose columns document themselves. Journalists, academics and students cite review-level evidence traceable to individual rows across a decade and a half of craft-beer opinion.

Scope note - '83 datasets' counts the hub's search results at research time, not 83 vetted industry sources; after removing false positives and near-duplicates the genuinely relevant count is materially lower, and Datadory's curation reflects that rather than the raw figure.

Provenance note - the review corpora carry no explicit geography column; location enters via brewery names and joins. Treat any regional breakdown built directly off these rows as an approximation until joined against a venue-level table.

Where to go next - the rail below pairs this consumer-opinion corpus with the venue-level establishment table it joins onto, the regulator's production ledger, and the trade association's interpretation layer.