BOLD: leaderboard
Bias in Open-Ended Language Generation dataset with 23,679 prompts across profession, gender, race, religion, and political ideology, scored with toxicity and bias metrics.
Metric: BOLD score. Source: github.com. Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Opus | 0.76 | #571 |
| 2 | Mistral-7B-v0.3 | 0.74 | #1355 |
| 3 | Gemini 1.5 Flash (001) | 0.74 | #577 |
| 4 | GPT-4 Preview (1106) | 0.74 | #418 |
| 5 | Gemma 2 9B | 0.74 | #958 |
| 6 | Mixtral 8x7B Instruct (v0.1) | 0.73 | #1073 |
| 7 | GPT-3.5 Turbo (0125) | 0.73 | #984 |
| 8 | Bielik-11B-v2.3-Instruct | 0.73 | #693 |
| 9 | Llama 2 70B Chat (HF) | 0.73 | #1264 |
| 10 | Qwen 1.5 72B Chat | 0.72 | #776 |
| 11 | Llama 2 13B Chat Base | 0.72 | #1422 |
| 12 | Mistral 7B Instruct (v0.2) | 0.72 | #1333 |
| 13 | Mistral 7B Instruct (v0.3) | 0.71 | #1274 |
| 14 | Yi Chat 34B | 0.68 | #976 |
| 15 | Llama 2 7B Chat | 0.68 | #1499 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bold · How It Works · Data refreshed daily, snapshot 2026-10-20.