DeividasM's picture
Update README.md
32931d6
|
raw
history blame
2.5 kB
metadata
datasets:
  - wikipedia
language:
  - lt
license: apache-2.0
tags:
  - text-generation
widget:
  - text: 'Lietuva yra viena '

Model description

Dalle-mini

GPT-2 model from Lithuania using Wikipedia corpus dataset based on GPT-2 small model.

This is only the first version of the model, over time model will be improved using a bigger dataset and better data preparation.

Training data

This model was pre-trained with 180MB of Lithuanian Wikipedia. The texts are tokenized using a byte-level version of Byte Pair Encoding (BPE).

Training

The model was trained on wiki-corpus for 40 hours using NVIDIA Tesla P100 GPU.

How to use

Load model

from transformers import AutoTokenizer, TFAutoModelWithLMHead
import tensorflow as tf

tokenizer = AutoTokenizer.from_pretrained("DeividasM/gpt2_lithuanian_small")
model = TFAutoModelWithLMHead.from_pretrained("DeividasM/gpt2_lithuanian_small")

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

model.eval()

Generate text

text = "tekstas "
inputs = tokenizer.encode(text, return_tensors="tf")


outputs = model.generate(inputs, eos_token_id=50256, pad_token_id=50256, 
                         do_sample=True,
                         max_length=40,
                         top_k=40)
                         
print(tokenizer.decode(outputs[0]))

Limitations and bias

The training data used for this model come from Lithuanian Wikipedia. We know it contains a lot of unfiltered content from the internet, which is far from neutral. As the openAI team themselves point out in their model card:

"Because large-scale language models like GPT-2 do not distinguish fact from fiction, we don’t support use-cases that require the generated text to be true. Additionally, language models like GPT-2 reflect the biases inherent to the systems they were trained on, so we do not recommend that they be deployed into systems that interact with humans > unless the deployers first carry out a study of biases relevant to the intended use-case. We found no statistically significant difference in gender, race, and religious bias probes between 774M and 1.5B, implying all versions of GPT-2 should be approached with similar levels of caution around use cases that are sensitive to biases around human attributes."

Author

Lithuanian GPT-2 small was trained and evaluated by Deividas Mataciunas (https://www.linkedin.com/in/deividasmataciunas/)