TAROT-PPO

Task-Oriented Authorship Obfuscation Using Policy Optimization Methods

Fine-tuned text rewriting model with proximal policy optimization for authorship obfuscation.

ArXiv paper: https://arxiv.org/abs/2407.21630v1

Model description

• Model type: Authorship obfuscation model using GPT2-based text rewriting
• Reward models: rrivera1849/LUAR-MUD & Alibaba-NLP/gte-large-en-v1.5
• Finetuned from model: philippelaban/keep_it_simple
• Dataset: Yelp/yelp_review_full
• Repository: https://github.com/hornetsecurity/tarot

Example use

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gabrielloiseau/TAROT-PPO")
model = AutoModelForCausalLM.from_pretrained("gabrielloiseau/TAROT-PPO")

paragraph = """I had dinner at Bella's Bistro last night, and it was a delightful experience. 
As soon as I walked in, I was greeted warmly by the hostess, and the cozy, rustic decor made me feel right at home. 
I started with the bruschetta, which was so fresh and flavorful—I could have eaten a whole meal of just that!"""

inputs = tokenizer([paragraph + "<|endoftext|>"], return_tensors="pt", padding=True)
outputs = model.generate(**inputs, do_sample=True, max_new_tokens=128)

outputs = outputs[:, inputs["input_ids"].shape[1]:]
tokenizer.batch_decode(outputs,skip_special_tokens=True)