Motivation :
The goal of the project was to adapt large language models for the Arabic language and create a new state-of-the-art Arabic LLM. Due to the scarcity of Arabic instruction fine-tuning data, not many LLMs have been trained specifically in Arabic, which is surprising given the large number of Arabic speakers.
Our final model was trained on a high-quality instruction fine-tuning (IFT) dataset, generated synthetically and then evaluated using the Hugging Face Arabic leaderboard.
Training :
This model is the 9B version. It was trained for a week on 4 A100 GPUs using LoRA with a rank of 128, a learning rate of 1e-4, and a cosine learning rate schedule.
Evaluation :
Metric | Slim205/Barka-9b-it |
---|---|
Average | 61.71 |
ACVA | 73.68 |
AlGhafa | 54.42 |
MMLU | 52.52 |
EXAMS | 52.51 |
ARC Challenge | 59.14 |
ARC Easy | 59.69 |
BOOLQ | 86.41 |
COPA | 58.89 |
HELLAWSWAG | 38.04 |
OPENBOOK QA | 56.16 |
PIQA | 72.01 |
RACE | 48.71 |
SCIQ | 66.43 |
TOXIGEN | 85.35 |
Please refer to https://github.com/Slim205/Arabicllm/ for more details.
- Downloads last month
- 1,789