argpt2-goodreads

This model is a fine-tuned version of gpt2-medium on an goodreads LABR dataset. It achieves the following results on the evaluation set:

Loss: 1.4389

Model description

Generate sentences either positive/negative examples based on goodreads corpus in arabic language.

Intended uses & limitations

the model fine-tuned on arabic language only with aspect to generate sentences such as reviews in order todo the same for other languages you need to fine-tune it in your own. any harmful content generated by GPT2 should not be used in anywhere.

Training and evaluation data

training and validation done on goodreads dataset LABR 80% for trainng and 20% for testing

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mofawzy/argpt2-goodreads")

model = AutoModelForCausalLM.from_pretrained("mofawzy/argpt2-goodreads")

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 16
eval_batch_size: 16
seed: 42
distributed_type: tpu
num_devices: 8
total_train_batch_size: 128
total_eval_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 20.0

Training results

train_loss = 1.474

Evaluation results

eval_loss = 1.4389

train metrics

epoch = 20.0
train_loss = 1.474
train_runtime = 2:18:14.51
train_samples = 108110
train_samples_per_second = 260.678
train_steps_per_second = 2.037

eval metrics

epoch = 20.0
eval_loss = 1.4389
eval_runtime = 0:04:37.01
eval_samples = 27329
eval_samples_per_second = 98.655
eval_steps_per_second = 0.773
perplexity = 4.2162

Framework versions

Transformers 4.13.0.dev0
Pytorch 1.10.0+cu102
Datasets 1.16.1
Tokenizers 0.10.3