Benchmark
CharXiv Descriptive Questions - CNTG
Vision benchmarks measure how well models understand images. Scores are percentages. Higher is better.
- Models measured
- 19
- Scores recorded
- 19
- Sources
- 1
- IQ weight
- 1.34
Score distribution
- 72.1%
- Mean
- 72.3%
- Median
- 11.8%
- Std dev
- 56.2%
- 10th percentile
- 86.3%
- 90th percentile
- 51.2%
- Min
This benchmark has a weight of 1.34 for chat models, measured over 19 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | o1· CharXiv Leaderboard data docs | 93.1% |
| 2 | Claude 3.5 Sonnet· CharXiv Leaderboard data docs | 90.1% |
| 3 | Qwen2-VL-72B· CharXiv Leaderboard data docs | 86.3% |
| 4 | GPT 4o· CharXiv Leaderboard data docs | 85.5% |
| 5 | GPT-4o mini· CharXiv Leaderboard data docs | 79.1% |
| 6 | Molmo 72B· CharXiv Leaderboard data docs | 78.1% |
| 7 | Pixtral 12B· CharXiv Leaderboard data docs | 78.1% |
| 8 | Gemini 1.5 Pro· CharXiv Leaderboard data docs | 76.3% |
| 9 | Reka Flash· CharXiv Leaderboard data docs | 72.5% |
| 10 | Claude 3 Sonnet· CharXiv Leaderboard data docs | 72.3% |
| 11 | Reka Core· CharXiv Leaderboard data docs | 71.3% |
| 12 | Claude 3 Opus· CharXiv Leaderboard data docs | 70.5% |
| 13 | Phi-3.5 Vision Instruct· CharXiv Leaderboard data docs | 67.7% |
| 14 | Molmo 7B-D· CharXiv Leaderboard data docs | 64.1% |
| 15 | gemini-1.0-pro· CharXiv Leaderboard data docs | 62.6% |
| 16 | Claude 3 Haiku· CharXiv Leaderboard data docs | 61.8% |
| 17 | Qwen-VL Plus· CharXiv Leaderboard data docs | 56.2% |
| 18 | Reka Edge· CharXiv Leaderboard data docs | 52.2% |
| 19 | Qwen-VL Max· CharXiv Leaderboard data docs | 51.2% |
Data credits: CharXiv (Wang et al., NeurIPS 2024), Princeton NLP. https://charxiv.github.io