Granite Arabic Card - Question Answering - Belebele (Dial): leaderboard
IBM Research / MBZUAI Granite 4.0 H 3B Arabic public checkpoint card; lm-evaluation-harness five-shot evaluation. Strict generative exact-match question answering across English, Modern Standard Arabic and dialects, with the card's published aggregate weights. Source abbreviations g4.0-h-micro_hf and g3.3-2b-instruct remain source identities without an unverified checkpoint merge. Exact Match (%); higher is better.
Metric: Exact Match (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 4B 2507 Instruct | 70.67 | #762 |
| 2 | granite-4.0-h-3B-ar | 68.76 | #857 |
| 3 | Nile-Chat-4B | 66.38 | #1165 |
| 4 | Gemma 3 4B (IT) | 63.67 | #992 |
| 5 | Falcon-H1-3B-Instruct | 63.62 | #1221 |
| 6 | Atlas-Chat-2B | 56.29 | #1284 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=granite-arabic-card-question-answering-belebele-dial · How It Works · Data refreshed daily, snapshot 2026-10-20.