lightonai / ArabicWeb24-ablation-model-v1

Model summary

This model is trained on the ArabicWeb dataset V1. It was trained on 25B tokens using the AraGPT-2 tokenizer. The model has 900 million parameters with a context length of 1024 tokens and uses the Mamba2 architecture.

• License: odc-by
• Languages: Arabic

Model Description

The ArabicWeb Ablation Model V1 is trained on a diverse corpus of Arabic text, including news articles, art and entertainment, and encyclopedia entries. This makes it suitable for a variety of Arabic text generation tasks. For more details, you can read the blog post.

• Model Type : Language Model
• Architecture : Mamba
• Training Data : ArabicWeb dataset
• Training Objective : Text generation

Usage

This model was primarily trained to assess the quality of the ArabicWeb dataset and is designed for text generation in Arabic. Please note that this is an ablation model that was not instruction-tuned. The primary intended use case is to compare its performance with other models trained under the same configuration but with different versions of datasets.

Training

Model

• Architecture: Mamba2 model
• Pretraining tokens: 25B
• Scheduler: Cosine
• d_model: 2304
• d_intermediate: 0
• n_layer: 18

Hardware

• Platform: HPE Cray node
• Hardware: 8 NVIDIA H100 GPUs
• Cloud Provider: Orange Cloud Avenue