AceGPT
AceGPT is a fully fine-tuned generative text model collection based on LlaMA2, particularly in the
Arabic language domain. This is the repository for the 13B-chat pre-trained model.
Model Developers
We are from the School of Data Science, the Chinese University of Hong Kong, Shenzhen (CUHKSZ), the Shenzhen Research Institute of Big Data (SRIBD), and the King Abdullah University of Science and Technology (KAUST).
Variations
AceGPT families come in a range of parameter sizes —— 7B and 13B, each size of model has a base category and a -chat category.
Input
Models input text only.
Output
Models output text only.
Model Evaluation Results
Experiments on Arabic Vicuna-80, Arabic AlpacaEval. Numbers are the average performance ratio of ChatGPT over three runs. We do not report the results of raw Llama-2 models since they cannot properly generate Arabic texts.
| Arabic Vicuna-80 | Arabic AlpacaEval | |
|---|---|---|
| Phoenix Chen et al. (2023a) | 71.92% ± 0.2% | 65.62% ± 0.3% |
| Phoenix–multiple-langs Chen et al. (2023b) | 71.67% ± 0.7% | 65.36% ± 0.1% |
| Jais-13B-chat Sengupta et al. (2023) | 75.40% ± 1.6% | 74.95% ± 0.2% |
| AceGPT-7B-chat | 94.82% ± 0.2% | 93.81% ± 0.1% |
| AceGPT-13B-chat | 100.88% ± 0.4% | 97.95% ± 0.1% |
You can get more detail at https://github.com/FreedomIntelligence/AceGPT/tree/main
- Downloads last month
- 12