gpt2-small-turkish

Turkish GPT2 Model Finetuned

Türkçe GPT2 Modeli

Model description

This is a GPT2-Small English based model finetuned and additionaly trainied with Wikipedia Articles in Turkish as of 28-10-2020

Live demo based on this work at : https://www.metayazar.com/

Fine tuned writer on this model: https://huggingface.co/gorkemgoknar/gpt2-turkish-writer

Work has been done on Pierre Guillou tutorial as on this page. (https://github.com/piegu/fastai-projects/blob/master/finetuning-English-GPT2-any-language-Portuguese-HuggingFace-fastaiv2.ipynb)

Code is converted to work with Fastai 2.X .

Using Google Colab for training.

Additional tutorial and source will be in https://github.com/gorkemgoknar in later stage.

Current accuracy 33 % , Perplexity : 51.88

Models are available:

• [gpt2-small-tuned-tr] (https://huggingface.co/gorkemgoknar/gpt2-small-turkish)
• [gpt2-small-turkish-writer] (https://huggingface.co/gorkemgoknar/gpt2-turkish-writer)

Intended uses & limitations

How to use

Install

from transformers import AutoTokenizer, AutoModelWithLMHead
import torch

tokenizer = AutoTokenizer.from_pretrained("gorkemgoknar/gpt2-small-turkish")
model = AutoModelWithLMHead.from_pretrained("gorkemgoknar/gpt2-small-turkish")

# Get sequence length max of 1024
tokenizer.model_max_length=1024 

model.eval()  # disable dropout (or leave in train mode to finetune)

Generate 1 word

# input sequence
text = "Bu yazıyı bilgisayar yazdı."
inputs = tokenizer(text, return_tensors="pt")

# model output
outputs = model(**inputs, labels=inputs["input_ids"])
loss, logits = outputs[:2]
predicted_index = torch.argmax(logits[0, -1, :]).item()
predicted_text = tokenizer.decode([predicted_index])

# results
print('input text:', text)
print('predicted text:', predicted_text)

# input text: 
# predicted text:  

Generate Full Sequence

# input sequence
text = "Bu yazıyı bilgisayar yazdı."
inputs = tokenizer(text, return_tensors="pt")

# model output using Top-k sampling text generation method
sample_outputs = model.generate(inputs.input_ids,
                                pad_token_id=50256,
                                do_sample=True, 
                                max_length=50, # put the token number you want
                                top_k=40,
                                num_return_sequences=1)

# generated sequence
for i, sample_output in enumerate(sample_outputs):
    print(">> Generated text {}\\\\
\\\\
{}".format(i+1, tokenizer.decode(sample_output.tolist())))

# >> Generated text
#    

Limitations and bias

The training data used for this model come from Turkish Wikipedia. We know it contains a lot of unfiltered content from the internet, which is far from neutral.

Training data

Wikipedia Turkish article dump as of 28-10-2020

Training procedure

Eval results

#Epoch 0 on Tesla T4, others on V100