DetLM: Detoxified Language model via PPLM

Author: Taehee Jung

The code is originally from (https://github.com/uber-research/PPLM) and modified by Taehee Jung for the purpose of final project in STAT2651, Spring 2021 at University of Pittsburgh.

Paper: Plug and Play Language Models: a Simple Approach to Controlled Text Generation

Authors: Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu

Paper link: https://arxiv.org/abs/1912.02164 Blog link: https://eng.uber.com/pplm

Requirement

We use python 3.8. Please run pip install -r requirement.txt to install python dependencies.

Train hate-offensive classifier with GPT-2

To download HateOffensive dataset, please visit https://github.com/dykang/xslue

python pplm_discrim_train.py \
    --dataset HateOffensive \
    --dataset_path PATH\TO\YOUR\DATASET \
    --epochs 20 \
    --save_model \
    --cached

Inference with hate-offensive classifier

For our experiments, we use test_selected_prompt_out_2.txt under dataset folder.

python pplm_discrim_eval.py \
    --discriminator_path PATH\TO\YOUR\CLASSIFIER\SAVED
    --discriminator HateOffensive
    --sentences TEST\SET\TXT\

Train and Generate text from unmodified / PPLM models

Our default set up for class_label is 2, which generates PPLM outputs with 'neither offensive nor hate-speech text'. Also, the other hyperparameter are equal to the original code. Here, cond_text should be txt file with a prompt of each sentence in one line. We use /output/test_selected_prompt.txt

python pplm.py -D HateOffensive \
    --discriminator_path PATH\TO\YOUR\CLASSIFIER\SAVED \
    --class_label 2 \
    --cond_text TEST\PROMPT\TXT \
    --length 30 \
    --gamma 1.0 \
    --num_iterations 5 \
    --num_samples 1 \
    --stepsize 0.04 \
    --kl_scale 0.01 \
    --gm_scale 0.95 \
    --sample

buptpriswang / DetLM