Kaggle Automobile Dataset (UCI Mirror)

A compact tabular benchmark where every row is one 1985 US-market import, described well enough to predict its price from its engineering. The file holds 205 car model variants across 22 manufacturers - alfa-romero at the front of the alphabet, volvo at the back, with audi, bmw, jaguar, mercedes-benz, porsche, saab, toyota and volkswagen in between. Content traces to the 1985 Ward's Automotive Yearbook, the same compilation behind the classic UCI Automobile entry; this mirror repackages it as one directly loadable CSV with a header row, which is precisely what the original lacked.

Why automobile manufacturer teams still reach for it in 2026: it is the rare public record that publishes price and an insurance-risk rating alongside the specifications that plausibly explain them. Registration counts tell you how many vehicles exist; production indexes tell you how many were built. Neither tells you how a hatchback's displacement, curb weight and drive configuration translated into a sticker price and an actuarial risk grade - this does, row by row, for every import on the 1985 market.

Scale is deliberately tiny - the whole thing reads into memory in one line. What the file lacks in volume it returns in tooling: the community has published 229 notebooks demonstrating manufacturer-based price and risk analysis on exactly these columns, so a new analyst inherits worked examples rather than a blank notebook.

Get a sample of this dataset - real rows come back before you commit to anything.

The first row of the file, transposed for readability - one line per field:

make              : alfa-romero
fuel-type         : gas       aspiration   : std
num-of-doors      : two       body-style   : convertible
drive-wheels      : rwd       engine-location : front
wheel-base        : 88.6      length       : 168.8     width : 64.1
curb-weight       : 2548      engine-type  : dohc
num-of-cylinders  : four      engine-size  : 130
fuel-system       : mpfi      bore         : 3.47      stroke : 2.68
compression-ratio : 9.0       horsepower   : 111       peak-rpm : 5000
city-mpg          : 21        highway-mpg  : 27
price             : 13495     symboling    : 3
normalized-losses : (missing)

That single row demonstrates why the benchmark teaches well: a rear-drive turbo-free convertible with a double-overhead-cam four, priced at $13,495 and graded symboling 3 - risky - while its normalized-losses cell shows the '?' placeholder the dataset is famous for. Specifications, economy, price and risk sit side by side in one flat line, no joins required.

Roughly 26 attributes per row, grouped into four families. Identity and configuration: make, fuel-type (gas or diesel), aspiration (std or turbo), num-of-doors, body-style (convertible, hatchback, sedan, wagon, hardtop), drive-wheels (4wd, fwd, rwd) and engine-location. Dimensions: wheel-base, length, width, height and curb-weight in pounds. Engine: engine-type (dohc, ohcv, ohc among others), num-of-cylinders spelled out, engine-size in cubic inches (61-326), bore, stroke, compression-ratio (7-23), horsepower (48-288), peak-rpm (4150-6600) and fuel-system. Economics and risk: city-mpg (13-49), highway-mpg (16-54), price ($5,118-$45,400), symboling (-3 safe to +3 risky) and normalized-losses (65-256). The table below carries the fields with verified definitions; the remainder fold into the footnote.

Geography - one market: imported automobiles sold in the United States for the 1985 model year. Domestic nameplates are not in scope, so treat any non-import conclusion as out of sample.

Temporal - a static snapshot, not a series. Every observation dates to 1985 via the Ward's Automotive Yearbook, and nothing about the content has changed since the mirror was posted in May 2017. There is no longitudinal axis here; pair it with a multi-year feed when trend questions appear.

Granularity - one row per car model variant: 205 rows total, each carrying the full attribute set. No monthly cuts, no regional splits, no hierarchy above the variant itself - which is exactly what makes regression and classification exercises reproducible.

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

• Price-model teaching and prototyping - regress price against engine-size, horsepower, curb-weight and body-style on a table small enough to diagnose every residual by hand; the 229 community notebooks give instructors and new analysts worked starting points rather than a cold start.
• Risk-analysis exercises - the symboling column grades each variant from -3 (safe) to +3 (risky), letting a team explore which configurations insurers penalized and test whether normalized-losses tracks the grade.
• Classification drills - body-style and drive-wheels give two clean categorical targets with balanced-enough class counts, useful for demonstrating encoding choices and evaluation without big-compute overhead.
• Missing-data practice - six attributes carry '?' placeholders whose handling visibly moves results, making this the standard sandbox for imputation comparisons.
• Dashboard and pipeline demos - 4,991 bytes loads instantly, so builders prototype joins, charts and validation logic on real automotive semantics before pointing the pipeline at production volumes.

Data scientists and ML engineers get the cleanest kind of teaching artifact: every variable inspectable in one screen, targets that behave sensibly, and a quarter-million-strong body of prior work to compare approach against. Journalists, academics and students get a citable 1985 market cross-section with formal provenance back to the Ward's Yearbook - ideal for classroom exercises and methods demonstrations. Developers and data-product builders get realistic automotive semantics at demo scale, handy for prototyping before wiring up larger feeds.

Persona-specific breakdowns: data scientists working in automobile manufacturers, journalists and academics and developers and builders.

Who it serves poorly, honestly: competitive-intel, sales, investor and e-commerce teams all score zero relevance here - a static 1985 table says nothing about current competitor products, demand signals or inventory. Those teams belong on the automobile manufacturers data hub picking live feeds instead.

Three things worth knowing upfront. First, the temporal ceiling is hard: nothing newer than 1985 exists in this file, so any current pricing, registration or production question needs a different record - the EPA Fuel Economy Dataset covers 43 model years of US configurations, and OICA carries modern world production. Second, plan for the '?' placeholders: normalized-losses, bore, stroke, horsepower, peak-rpm and price all have gaps, and imputation choices materially affect price and horsepower results, so settle the convention before modeling rather than after. Third, header spelling varies between dotted and underscored renderings of the UCI names depending on copy - verify against your delivered schema mapping during sample review, since downstream code keys on exact column names.