notesum.ai
Published at December 5Arabic Stable LM: Adapting Stable LM 2 1.6B to Arabic
cs.CL
Released Date: December 5, 2024
Authors: Zaid Alyafeai1, Michael Pieler, Hannah Teufel, Jonathan Tow, Marco Bellagente, Duy Phung, Nikhil Pinnaparaju, Reshinth Adithyan, Paulo Rocha, Maksym Zhuravinskyi, Carlos Riquelme
Aff.: 1Stability AI Language Team

| Model | Params | STEM | Social Science | Humanities | Language | Other | Average |
|---|---|---|---|---|---|---|---|
| AraGPT2-base | 135M | 28.3 | 32.0 | 32.7 | 33.0 | 33.1 | 31.7 |
| AraT5v2-base-1024 | 220M | 26.3 | 29.4 | 28.1 | 27.7 | 30.3 | 28.3 |
| AraGPT2-medium | 370M | 28.5 | 32.0 | 33.8 | 32.1 | 35.1 | 32.2 |
| jais-family-590m | 590M | 31.8 | 35.6 | 37.2 | 37.7 | 37.3 | 35.7 |
| jais-family-590m-chat | 590M | 29.5 | 30.7 | 32.2 | 30.9 | 33.9 | 31.4 |
| AraGPT2-large | 792M | 28.5 | 33.0 | 34.1 | 33.2 | 35.0 | 32.6 |
| jais-family-1p3b | 1.3B | 32.9 | 37.1 | 39.6 | 39.5 | 39.4 | 37.5 |
| jais-family-1p3b-chat | 1.3B | 32.0 | 34.0 | 34.9 | 31.7 | 40.1 | 34.6 |
| AraGPT2-mega | 1.46B | 29.8 | 32.9 | 35.7 | 34.6 | 34.3 | 33.3 |
| Qwen2-1.5B | 1.5B | 28.4 | 30.9 | 29.5 | 33.3 | 32.1 | 30.5 |
| Qwen2-1.5B-Instruct | 1.5B | 29.0 | 30.9 | 29.8 | 33.7 | 32.5 | 30.8 |
| bloom-1b7 | 1.72B | 29.2 | 33.7 | 33.4 | 34.9 | 34.7 | 33.0 |
| bloomz-1b7 | 1.72B | 29.9 | 33.8 | 33.9 | 35.7 | 35.7 | 33.5 |
| jais-family-2p7b | 2.7B | 35.2 | 38.1 | 42.7 | 39.9 | 39.0 | 39.0 |
| jais-family-2p7b-chat | 2.7B | 32.1 | 35.1 | 36.4 | 37.0 | 41.5 | 36.1 |
| jais-family-6p7b | 6.7B | 35.8 | 39.2 | 45.0 | 41.5 | 42.3 | 40.7 |
| jais-family-6p7b-chat | 6.7B | 35.1 | 37.3 | 40.9 | 37.3 | 42.8 | 38.7 |
| AceGPT-7B | 7B | 35.0 | 41.3 | 44.3 | 42.1 | 42.1 | 40.9 |
| AceGPT-7B-chat | 7B | 34.4 | 39.0 | 39.0 | 40.0 | 39.5 | 38.2 |
| SILMA-9B-Instruct-v1.0 | 9B | 28.4 | 29.7 | 30.0 | 33.3 | 35.2 | 30.8 |
| AceGPT-13B | 13B | 37.1 | 41.4 | 42.6 | 41.4 | 40.9 | 40.7 |
| AceGPT-13B-chat | 13B | 36.2 | 42.0 | 42.7 | 41.5 | 43.5 | 41.1 |
| AceGPT-v1.5-13B | 13B | 35.5 | 40.1 | 42.3 | 41.1 | 43.0 | 40.3 |
| AceGPT-v1.5-13B-chat | 13B | 35.8 | 40.5 | 42.6 | 41.5 | 45.0 | 40.9 |
| jais-13b | 13B | 35.3 | 39.9 | 44.2 | 40.8 | 44.5 | 40.9 |
| jais-13b-chat | 13B | 33.8 | 39.6 | 44.1 | 42.0 | 45.3 | 40.7 |
| jais-family-13b | 13B | 37.1 | 39.8 | 46.1 | 43.5 | 44.2 | 41.9 |
| jais-family-13b-chat | 13B | 34.9 | 38.5 | 44.1 | 38.5 | 43.4 | 39.9 |
| ar-stablelm-2-base | 1.64B | 36.1 | 39.6 | 44.9 | 44.1 | 41.8 | 41.1 |
| ar-stablelm-2-chat | 1.64B | 37.8 | 40.9 | 50.2 | 45.4 | 55.1 | 45.5 |