Awesome

HC-Var (Human and ChatGPT texts with Variety)

This is a repository for training binary classifcation models to distinguish human texts and ChatGPT (GPT3.5-Turbo) generated texts. We collect a new dataset HC-Var (Human and ChatGPT texts with Variety) to fulfill our objective. This dataset includes the texts which are generated / human written to accomplish various language tasks with various approaches. The included language tasks and topics are summarized below. The HC-Var dataset is available now in hugging face: https://huggingface.co/datasets/hannxu/hc_var.

Dataset Summary

This dataset contains human and ChatGPT texts to fulfill 4 distinct language tasks, including news composing (News), review (Review), essay writing (Writing) and question answering (QA). Under each task, we collect the human and ChatGPT generated texts with one or multiple topics. For each language task, this dataset considers 3 different prompts to inquire ChatGPT outputs.

Environments

The code is primary runned and examined under python 3.10.12, torch 2.0.1. To install other required packages using the command:

pip install -r requirements.txt

To train the model

This repository currently supports training classification models under RoBERTa-base, RoBERTa-large and T5 (we test under T5-base). An example command to run the code to train a RoBERTa-base classification model and test the model on the domain "review".

python -m detector.train_roberta --domain review 

For details, the training process includes 3 major steps:

1. Load the training, validation and test dataloaders.

train_loader, valid_loader, test_loader = Loader(batch_size = 32, domain=domain, cache_dir = cache_dir)

2. Initilize the classification model and optimizer:

model_name = 'roberta-large' 
tokenizer = RobertaTokenizer.from_pretrained(model_name)
model = RobertaForSequenceClassification.from_pretrained(model_name).to(device)
optimizer = AdamW(model.parameters(), lr=learning_rate)

3. Train the model.

def train(model, tokenizer, optimizer, device, loader):
    model.train()
    for i, dat in enumerate(loader):
        texts, labels = dat
        texts = list(texts)
        result = tokenizer(texts, return_tensors="pt", padding = 'max_length', max_length = 256, truncation=True)
        texts, masks, labels = result['input_ids'].to(device), result['attention_mask'].to(device), labels.to(device)
        aa = model(texts, labels=labels, attention_mask = masks)
        loss = aa['loss']
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

To evaluate the model

We define 3 types of test data loaders to evaluate the models performance facing different varieties. For example, to evaluate a model's performance when test samples are divided in different tasks:

test_loader = Domain_loader(domain= "TaskName", cache_dir = cache_dir)  ## TaskName can be News, Review, Writing, QA

Or when test samples are divided in different topics in the same task, i.e., QA:

test_loader = Topic_loader(domain= 'QA', topic = "TopicName", cache_dir = cache_dir)  ## TopicName can be history, finance, medical, science

Or when test samples are divided in different prompts in the same task, i.e., QA:

test_loader = Prompt_loader(domain= 'QA', prompt = promptid, cache_dir = cache_dir)  ## promptid can be "P1", "P2", "P3"

To cite our dataset, code or paper:

@misc{xu2023generalization,
      title={On the Generalization of Training-based ChatGPT Detection Methods}, 
      author={Han Xu and Jie Ren and Pengfei He and Shenglai Zeng and Yingqian Cui and Amy Liu and Hui Liu and Jiliang Tang},
      year={2023},
      eprint={2310.01307},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}