FairFace
The balanced face-attribute dataset and models built to measure and reduce racial bias in age, gender, and race estimation.
About
FairFace was created to fix a measurement problem: popular face-attribute datasets were overwhelmingly white and male, so models trained on them performed worst on everyone else. The authors collected 108,501 Flickr images balanced across seven race groups, genders, and age bands, and trained attribute models (race_4, race_7, plus age and gender) with a ResNet backbone.
It became a standard fairness benchmark — if you claim your attribute model is unbiased, reviewers expect FairFace numbers. The dataset and pretrained models are freely downloadable, though distribution is via a Google Drive folder rather than a polished release, and the repo does not prominently state a license, so treat it as research use until confirmed.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
The 2019 paper reports the dataset's balance statistics and baseline attribute accuracies as provider-reported figures; the value of FairFace is as a benchmark other models are measured against, not as a leaderboard entry itself. No independent benchmark of the bundled models found.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics, metrics-v1). Evidence: peer-reviewed arXiv paper (1908.04913) by named UCLA researchers; the dataset is widely used as the fairness benchmark in attribute-estimation literature. Strengths: filled a real gap, genuine standard for bias measurement. Docked for: (1) informal distribution (Google Drive folder) and no prominently stated license; (2) models are baselines rather than competitive state-of-the-art. License: no license file prominently visible on the release — treat as research use until confirmed.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Kärkkäinen & Joo (UCLA), arXiv 1908.04913, 2019; 108,501 images balanced across 7 race groups, genders, age bands.
- · Standard fairness benchmark for attribute estimation; bundled race_4/race_7, age, gender models are baselines.
- · Repo is low-activity; no independent benchmark of the bundled models found.