The main objective of this module is to fine-tune and evaluate a model (pre-trained on a large-scale dataset) on domain-specific data. Finetuning will improve the performance of the model on domain specific tasks. The pre-trained models can be finetuned on a number of downstream tasks based on their architecture. Here, I have taken an example of finetuning sequence-to-sequence models such as T5, BART, Pegasus on an abstractive summarization task using the Trainer API from Hugging Face.

• A number of pre-trained models can be finetuned such as:
  • T5 (small, base, large, 3B, 11B)
  • BART (base, large-cnn, large-mnli)
  • Longformer Encoder Decoder (allenai/led-base-16384, allenai/led-large-16384)
  • Pegasus (large, xsum, multi_news)

Checkout pre-trained models to see the checkpoints available for each of them.

Finetuning with custom dataset placed at data/:

python run.py \
    --model_name_or_path facebook/bart-base \
    --train_file data/news_summary_train_small.csv \
    --validation_file data/news_summary_valid_small.csv \
    --text_column Text \
    --summary_column Summary \
    --output_dir output/ \
    --overwrite_output_dir \
    --do_train \
    --do_eval \
    --num_beams=3 \
    --min_summ_length=100 \     
    --max_summ_length=250 \   
    --length_penalty=1.0 \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --predict_with_generate 

To see all the possible command line options, run:

python run.py --help

If you are using Google Colab, Open colab/finetuning.ipynb in Colab, save a copy in Drive and follow the instructions.