lvwerra/gpt2-imdb finetuned using Multi-Objective Reinforcement Learning (MORL).
Using positive and negative movie review classifier (sentiment-analysis) we try to optimize the reward:
0.1 * positive score + 0.9 * negative score.
Other scalar combinations were tried as well. Results are below:
image/png
Usage
Use the model.generate() function to generate reviews that, in this case, should reflect a pretty negative sentiment.
