Benchmark
Bridging the Gap Belebele Xhosa (xh) human-translated
Multilingual benchmarks compare model performance across languages. Scores are percentages. Higher is better.
- Models and systems measured
- 46
- Scores recorded
- 52
- Sources
- 1
- IQ weight
- 2.25
Score distribution
- 42.2%
- Mean
- 42%
- Median
- 8.6%
- Std dev
- 36.8%
- 10th percentile
- 45.7%
- 90th percentile
- 29.1%
- Min
This benchmark has a weight of 2.25 for chat models, measured over 6 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 1 | GPT 4otable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 82.8% |
| 2 | GPT-4 Turbotable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 76.3% |
| 3 | Llama 3 70B Instructtable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| 4 | Llama 3 8B Instructtable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 33% |
| 5 | GPT-3.5 Turbotable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 31.8% |
| 6 | Phi-3 Mini 4K Instructtable: 2, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 29.1% |
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| Llama 3 70B Instructtable: A.26, row: base, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B Instructtable: A.27, row: base, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B Instructtable: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: High, quantity: 0%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B Instructtable: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: Low, quantity: 0%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B Instructtable: A.31, dataset: Winogrande, trainingLanguage: xh, quality: High, quantity: 0%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B Instructtable: A.31, dataset: Winogrande, trainingLanguage: xh, quality: Low, quantity: 0%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 36.8% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: aftable: A.26, training: Winogrande train (small), trainingLanguage: af, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 42.4% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: amtable: A.26, training: Winogrande train (small), trainingLanguage: am, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 43.2% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: bmtable: A.26, training: Winogrande train (small), trainingLanguage: bm, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 30.3% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: entable: A.26, training: Winogrande train (small), trainingLanguage: en, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 37.4% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: igtable: A.26, training: Winogrande train (small), trainingLanguage: ig, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 41.1% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: nsotable: A.26, training: Winogrande train (small), trainingLanguage: nso, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 44.2% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: sntable: A.26, training: Winogrande train (small), trainingLanguage: sn, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 43.3% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: sttable: A.26, training: Winogrande train (small), trainingLanguage: st, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 42.4% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: tntable: A.26, training: Winogrande train (small), trainingLanguage: tn, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 44.4% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: tstable: A.26, training: Winogrande train (small), trainingLanguage: ts, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 37.2% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: xhtable: A.26, training: Winogrande train (small), trainingLanguage: xh, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 42% |
| Llama 3 70B ITLoRA, dataset: Winogrande train (small), language: zutable: A.26, training: Winogrande train (small), trainingLanguage: zu, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 41.6% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: aftable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: af, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 38.8% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: amtable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: am, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 42.5% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: bmtable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: bm, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 37.7% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: entable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: en, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 36.5% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: igtable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: ig, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 44% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: nsotable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: nso, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 41.6% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: sntable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: sn, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 43.8% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: sttable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: st, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 44.4% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: tntable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: tn, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 41.4% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: tstable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: ts, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 43.3% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: xhtable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: xh, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 45.7% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine dev + test + val, language: zutable: A.27, training: MMLU College Medicine dev + test + val, trainingLanguage: zu, task: Belebele, language: xh, shots: 0· Bridging the Gap (arXiv:2412.12417v1) data | 43.9% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: High, quantity: 100%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: High, quantity: 100%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 47.3% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: High, quantity: 25%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: High, quantity: 25%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 41.1% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: High, quantity: 50%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: High, quantity: 50%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 44.2% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: High, quantity: 75%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: High, quantity: 75%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 45.6% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: Low, quantity: 100%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: Low, quantity: 100%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 46.1% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: Low, quantity: 25%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: Low, quantity: 25%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 40.5% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: Low, quantity: 50%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: Low, quantity: 50%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 39.3% |
| Llama 3 70B ITLoRA, dataset: MMLU College Medicine, language: xh, quality: Low, quantity: 75%table: A.31, dataset: MMLU College Medicine, trainingLanguage: xh, quality: Low, quantity: 75%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 42.3% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: High, quantity: 100%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: High, quantity: 100%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 45.6% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: High, quantity: 25%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: High, quantity: 25%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 43% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: High, quantity: 50%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: High, quantity: 50%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 46.5% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: High, quantity: 75%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: High, quantity: 75%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 43.5% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: Low, quantity: 100%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: Low, quantity: 100%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 41% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: Low, quantity: 25%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: Low, quantity: 25%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 40.5% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: Low, quantity: 50%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: Low, quantity: 50%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 44.5% |
| Llama 3 70B ITLoRA, dataset: Winogrande, language: xh, quality: Low, quantity: 75%table: A.31, dataset: Winogrande, trainingLanguage: xh, quality: Low, quantity: 75%, task: Belebele· Bridging the Gap (arXiv:2412.12417v1) data | 37.1% |
Data credits: Tuka Alhanai, Adam Kasumovic, Mohammad Ghassemi, Aven Zitzelberger, Jessica Lundin and Guillaume Chabot-Couture. Bridging the Gap: Enhancing LLM Performance for Low-Resource African Languages with New Benchmarks, Fine-Tuning, and Cultural Adjustments, arXiv:2412.12417v1. CC BY 4.0 article numeric/protocol extract, restructured with values unchanged. No task or model rights asserted. (licence)