Skip to main content
CARLA HQ
Legal & Public Safety Public Services & Risk use casesNACE O84.23Binary ClassificationTarget: two_year_recidivismOpenML #42192 ↗

Compas Two Years

Canonical benchmark containing 5,278 criminal defendant profiles from Broward County used to evaluate algorithmic fairness and predict two-year recidivism risk.

✨ Try in CarlaInstall Chrome Extension ↗100% In-Browser WebGPU • Zero Cloud Upload
5,278Records (Rows)
13Predictive Features
14 / 0Numeric / Categorical
0.0%Missing Value Ratio
0.27011 - AUC Error
5.81sWebGPU Mean Split
BUSINESS CONTEXT & OBJECTIVEPublic Safety, Judicial Analytics & Algorithmic Auditing (O84.23)

Business Objective: Compas Two Years

01Business Context

Pre-trial and sentencing risk assessment algorithms evaluate criminal defendants to guide bail, parole, and supervision determinations within municipal and state justice systems.

02Analytical Objective

Predict whether a defendant will be rearrested for a new offense within a two-year observation window following their assessment.

03Economic & Decision Impact

Accurate assessment balances public safety with individual civil liberties. High false positive rates lead to unwarranted detention, severe societal disruption, and disproportionate demographic harm, while false negatives expose communities to preventable public safety risks.

ML Benchmark Narrative

Standard supervised classifiers (XGBoost, LightGBM, and Random Forest) typically reach an AUROC between 0.72 and 0.74 on this benchmark, reflecting intrinsic label noise and systemic reporting bias. In-context tabular foundation models such as TabICL on Carla match tree-ensemble discriminative power zero-shot without manual hyperparameter tuning, simplifying algorithmic fairness audits directly in tabular workflows.

Source Origin:ProPublica Analysis of Broward County Sheriff's Office Records (Hosted on OpenML)
Creator:Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner (ProPublica) (2016)
License:Open Data Commons Attribution License (ODC-By)
FEATURE SPECIFICATION & DATA DICTIONARY

Target & Feature Column Definitions

Exhaustive business definitions, measurement units, roles, and target variables across all 14 columns.

ColVariable / Feature NameRoleTypeUnit / ScaleDescription & Business Meaning
Atwo_year_recidivismTwo-Year Recidivism StatusTARGETNUM
Binary target variable indicating if the individual was rearrested within two years (1 = Recidivated, 0 = No Recidivism).
Bsex_indicatorDefendant SexFeatureNUM
Binary encoding of defendant biological sex (1 for Male, 0 for Female).
Cage_yearsDefendant AgeFeatureNUMyears
Recorded age of the defendant in years at the time of the COMPAS screening.
Djuvenile_felony_countJuvenile Felony CountFeatureNUMcount
Total count of prior juvenile felony offenses recorded on the defendant's record.
Ejuvenile_misdemeanor_countJuvenile Misdemeanor CountFeatureNUMcount
Total count of prior juvenile misdemeanor charges.
Fjuvenile_other_offense_countJuvenile Other Offense CountFeatureNUMcount
Total count of prior non-felony, non-misdemeanor juvenile infractions or charges.
Gprior_adult_convictions_countPrior Charges CountFeatureNUMcount
Total count of prior adult criminal charges and convictions.
His_age_25_to_45Age Bracket: 25 to 45FeatureNUM
Binary one-hot encoded flag indicating if the defendant's age falls between 25 and 45 inclusive (1 = True, 0 = False).
Iis_age_greater_than_45Age Bracket: Over 45FeatureNUM
Binary one-hot encoded flag indicating if the defendant is older than 45 years (1 = True, 0 = False).
Jis_age_under_25Age Bracket: Under 25FeatureNUM
Binary one-hot encoded flag indicating if the defendant is younger than 25 years (1 = True, 0 = False).
Kis_race_african_americanRace: African AmericanFeatureNUM
Binary one-hot encoded flag indicating defendant identification as African American (1 = True, 0 = False).
Lis_race_caucasianRace: CaucasianFeatureNUM
Binary one-hot encoded flag indicating defendant identification as Caucasian (1 = True, 0 = False).
Mis_felony_chargeCurrent Charge: FelonyFeatureNUM
Binary one-hot flag indicating the current arrest charge is classified as a Felony (1 = True, 0 = False).
Nis_misdemeanor_chargeCurrent Charge: MisdemeanorFeatureNUM
Binary one-hot flag indicating the current arrest charge is classified as a Misdemeanor (1 = True, 0 = False).
DATASET PREVIEW

Interactive Data Table

Explore rows, feature values, and target labels for Compas Two Years.

Loading dataset...
LOCAL EXECUTION TELEMETRY

TabICLv2 WebGPU Benchmark Results

Complete performance metrics on persisted IID or non-IID evaluation splits, executed 100% locally in the browser sandbox.

Mean Split Latency?Mean wall-clock execution time per persisted evaluation split running 100% locally via WebGPU inside Chrome.Explore Latency Guide →
5.81s9 splits • 8 ensembles/split
Total Evaluation Runtime?Total wall-clock duration across all selected splits, including in-context encoding, WebGPU shader execution, and result aggregation.Explore Split Protocol →
52.28sIncludes warmup & sync
Evaluation Protocol?Precomputed row indices ensure the browser and Python runners evaluate identical out-of-sample observations without leakage.Explore Split Protocol →
3×3 SplitsRepeated IID • 9 total
TabICLv2 Score?Primary task performance score (1 - AUC Error) achieved by TabICLv2 across out-of-sample evaluation splits.Explore Metric Formula →
0.27011 - AUC Error
Evaluation SplitTrain RowsTest Rows
Split Duration?Wall-clock inference time taken for this evaluation split running 100% locally via WebGPU.Learn more →
Accuracy?Proportion of correct test predictions across all classes.Google ML Guide & Details →
ROC-AUC?Area under the ROC curve evaluating ranking capability across all thresholds.Google ML Guide & Details →
F1 Score?Harmonic mean of Precision and Recall, robust against class imbalance.Learn more →
Precision?Proportion of predicted positives that were actually positive.Google ML Guide & Details →
Recall?Proportion of actual positive cases successfully captured.Google ML Guide & Details →
R1 / F13,5181,7605.90s67.44%0.72460.63100.67590.5918
R1 / F23,5191,7595.78s69.64%0.75320.65770.70080.6196
R1 / F33,5191,7595.80s67.71%0.73280.61730.69710.5538
R2 / F13,5181,7605.80s67.50%0.73830.62610.68230.5785
R2 / F23,5191,7595.82s67.65%0.73300.63460.67760.5966
R2 / F33,5191,7595.79s69.24%0.74140.65520.69270.6215
R3 / F13,5181,7605.78s70.68%0.75950.67710.70260.6534
R3 / F23,5191,7595.80s63.39%0.69140.47730.72770.3551
R3 / F33,5191,7595.79s67.88%0.74380.64620.67010.6239
Mean ± Std5.81s67.90% ± 1.93%0.735 ± 0.0180.625 ± 0.0550.6920.577
UNIFIED BENCHMARK LEADERBOARD

Foundation Models vs Traditional ML Baselines

Side-by-side evaluation of our in-browser Carla engine (WebGPU), open-source tabular foundation models, and OpenML baselines.

RankAlgorithm / ModelModel FamilyRuntime
Predictive Accuracy?Proportion of correct test predictions across cross-validation splits.Google ML Guide & Details →
ROC-AUC?Area under the ROC curve evaluating ranking capability.Learn more →
F-Measure?Harmonic mean of precision and recall.Learn more →
Reference / Repo
#1Streaming nanotabiclTabular Foundation ModelPyTorch/Python68.42%0.73870.6446soda-inria/nanotabicl ↗
#2TabPFN v3Tabular Foundation ModelPyTorch/Python68.35%0.73850.6432PriorLabs/tabpfn ↗
#3EXAONE TabularTabular Foundation ModelPyTorch/Python68.35%0.73880.6418LGAI-Research/EXAONE-Tabular ↗
#4Google TabFM v1.0Tabular Foundation ModelPyTorch/Python68.20%0.73810.6422google-research/tabfm ↗
#5nanotabicl VanillaTabular Foundation ModelPyTorch/Python68.17%0.73910.6395soda-inria/nanotabicl ↗
#6Carla Engine (TabICLv2 WebGPU)Tabular Foundation ModelBrowser67.90%0.73530.6247100% In-Browser
#7Full TabICLv2 (PyTorch Reference)Tabular Foundation ModelPyTorch/Python67.15%0.72990.6371soda-inria/tabicl ↗
STATISTICAL DISTRIBUTIONS

Variable Schema & Summary Distributions

Observed numerical ranges, category cardinalities, missing rates, and sample values across 5,278 rows.

ColVariable NameTypeMissingDistinctSummary Stats / DistributionSample Values
Atwo_year_recidivismTARGETNUM02[0, 1] μ=0.5 σ=0.5111
Bsex_indicatorNUM02[0, 1] μ=0.8 σ=0.4111
Cage_yearsNUM062[18, 80] μ=34.4 σ=11.7342441
Djuvenile_felony_countNUM09[0, 10] μ=0.1 σ=0.4000
Ejuvenile_misdemeanor_countNUM010[0, 13] μ=0.1 σ=0.5000
Fjuvenile_other_offense_countNUM08[0, 7] μ=0.1 σ=0.5010
Gprior_adult_convictions_countNUM036[0, 38] μ=3.5 σ=4.90414
His_age_25_to_45NUM02[0, 1] μ=0.6 σ=0.5101
Iis_age_greater_than_45NUM02[0, 1] μ=0.2 σ=0.4000
Jis_age_under_25NUM02[0, 1] μ=0.2 σ=0.4010
Kis_race_african_americanNUM02[0, 1] μ=0.6 σ=0.5110
Lis_race_caucasianNUM02[0, 1] μ=0.4 σ=0.5001
Mis_felony_chargeNUM02[0, 1] μ=0.7 σ=0.5111
Nis_misdemeanor_chargeNUM02[0, 1] μ=0.3 σ=0.5000
FREQUENTLY ASKED QUESTIONS

Frequently Asked Questions: Compas Two Years

Common questions regarding the Compas Two Years dataset, machine learning task formulations, and in-browser tabular inference.

What is the COMPAS Two Years dataset used for?

The COMPAS Two Years dataset is a premier benchmark for evaluating binary classification models predicting criminal recidivism and researching algorithmic fairness across demographic subgroups.

Can I run zero-server predictions on this dataset in Google Sheets?

Yes, Carla HQ enables zero-server in-context tabular prediction directly in spreadsheets using TabICL foundation models.

What machine learning models perform best on the COMPAS Two Years dataset?

Gradient boosting frameworks like LightGBM and XGBoost, alongside in-context tabular models like TabICL, achieve top validation ROC-AUC scores (~0.72-0.74) on this benchmark.

LIVE EVALUATION IN GOOGLE SHEETS

Test TabICLv2 on Compas Two Years Yourself

Open the pre-loaded Google Sheet and let Carla configure the target and task for local, zero-cloud tabular machine learning.

Step 1

Launch Carla

Click once to open the spreadsheet and Carla side panel together.

Step 2

Review the Setup

Carla selects the dataset target and task from this page automatically.

Step 3

Evaluate & Predict

Run predictions and compute metrics with zero server uploads.