Prefix-tuning.md

Prefix-tuning

You can use Prefix-tuning module (though only for BART, T5, and GPT-2) through setting efficient_methods and efficient_kwargs.

For example, you can use Prefix-tuning in BART following below commands:

python run_textbox.py \
    --model=BART \
    --model_path=facebook/bart-base \
    --dataset=samsum \
    --efficient_methods=\[\'prefix-tuning\'\] \
    --efficient_kwargs=\{\'prefix_length\':\ 100,\ \'prefix_dropout\':\ 0.1,\ \'prefix_mid_dim\':\ 512\}

or yaml equivalent:

efficient_methods: ['prefix-tuning']
efficient_kwargs: {'prefix_length': 100, 'prefix_dropout': 0.1, 'prefix_mid_dim': 512}