arabic-nano-gpt-v1

This model is a fine-tuned version of openai-community/gpt2 on the arabic wikimedia/wikipedia dataset.

Repository on GitHub: e-hossam96/arabic-nano-gpt

The model achieves the following results on the held-out test set:

• Loss: 3.02885

How to Use

import torch
from transformers import pipeline

model_ckpt = "e-hossam96/arabic-nano-gpt-v1"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


lm = pipeline(task="text-generation", model=model_ckpt, device=device)

prompt = """المحرك النفاث هو محرك ينفث الموائع (الماء أو الهواء) بسرعة فائقة \
لينتج قوة دافعة اعتمادا على مبدأ قانون نيوتن الثالث للحركة. \
هذا التعريف الواسع للمحركات النفاثة يتضمن أيضا"""

output = lm(prompt, max_new_tokens=128)

print(output[0]["generated_text"])

Model description

• Embedding Size: 384
• Attention Heads: 4
• Attention Layers: 4

Training and evaluation data

The entire wikipedia dataset was split into three splits based on the 90-5-5 ratios.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 512
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.01
• num_epochs: 24

Training Loss

Validation Loss

Framework versions

• Transformers 4.45.2
• Pytorch 2.5.0
• Datasets 3.0.1
• Tokenizers 0.20.1