HELM VHELM - Fair Face: leaderboard
The "Fair Face" subtask of HELM VHELM. Stanford HELM's VHELM project, the holistic evaluation of vision-language models: image-grounded question answering, OCR, chart and diagram reading, spatial reasoning, robustness, bias, fairness and toxicity scenarios scored under one ha...
Metric: PEM. Source: crfm.stanford.edu. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 69.9 | #344 |
| 2 | Gemini 2.0 Flash (Preview) | 69.87 | #366 |
| 3 | Gemini 2.5 Pro (Preview 03-25) | 69.65 | #124 |
| 4 | Gemini 2.0 Pro (Preview 02-05) | 68.48 | #305 |
| 5 | Gemini 2.0 Flash Lite | 68.16 | #453 |
| 6 | Gemini 1.5 Pro (002) | 66.98 | #350 |
| 7 | GPT-4 Turbo | 66.95 | #406 |
| 8 | Gemini 1.5 Flash (002) | 65.32 | #585 |
| 9 | Llama 4 Scout Instruct | 64.1 | #542 |
| 10 | Llama 4 Maverick Instruct FP8 | 62.36 | #412 |
| 11 | GPT-4.1 Mini | 58.8 | #351 |
| 12 | Claude 3 Haiku (20240307) | 58.43 | #777 |
| 13 | Claude 3 Sonnet (20240229) | 55.82 | #681 |
| 14 | GPT-4.1 (2025-04-14) | 53.95 | #250 |
| 15 | Claude 3 Opus (20240229) | 52.55 | #413 |
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-fair-face · How It Works · Data refreshed daily, snapshot 2026-10-23.