README.md

metadata

language: vi
tags:
  - vi
  - vietnamese
  - gpt2
  - text-generation
  - lm
  - nlp
datasets:
  - oscar
widget:
  - text: Việt Nam là quốc gia có

GPT-2

Pretrained model on Vietnamese language using a causal language modeling (CLM) objective. It was introduced in this paper and first released at this page.

How to use the model

import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel

tokenizer = GPT2Tokenizer.from_pretrained('NlpHUST/gpt2-vietnamese')
model = GPT2LMHeadModel.from_pretrained('NlpHUST/gpt2-vietnamese')

text = "Albert Einstein là nhà vật lý học tạo ra thuyết lượng tử"
input_ids = tokenizer.encode(text, return_tensors='pt')
max_length = 100

sample_outputs = model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,
                                   do_sample=True,
                                   max_length=max_length,
                                   min_length=max_length,
                                   top_k=40,
                                   num_beams=5,
                                   early_stopping=True,
                                   no_repeat_ngram_size=2,
                                   num_return_sequences=3)

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

Model architecture

A 12-layer, 768-hidden-size transformer-based language model.

Training

The model was trained on Vietnamese Oscar dataset (32 GB) to optimize a traditional language modelling objective on v3-8 TPU for around 6 days. It reaches around 13.4 perplexity on a chosen validation set from Oscar.

GPT-2 Finetuning

The following example fine-tunes GPT-2 on WikiText-2. We're using the raw WikiText-2 (no tokens were replaced before the tokenization). The loss here is that of causal language modeling.

The script here .

python run_clm.py \
    --model_name_or_path NlpHUST/gpt2-vietnamese \
    --dataset_name wikitext \
    --dataset_config_name wikitext-2-raw-v1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --do_train \
    --do_eval \
    --output_dir /tmp/test-clm