PersonaGPT follows a GPT-2 based transformer architecture. It is built from the pretrained DialoGPT, which adapts the GPT-2 model to open-domain conversational responses using Reddit conversations as training corpus. PersonaGPT is fine-tuned on the Persona-Chat dataset, with added special tokens to better distinguish between conversational history and personality traits for dyadic conversations.
The goal of PersonaGPT is to create a create an open-source database that:
- captures the failure points of personalized conversational models, and
- provides training data to detect fake, personalized conversational agents.
Most conversational datasets capture near-perfect human-human dialog as a source of supervision. However, human-chatbot conversations can exhibit failure modes that are rare in normal human-human interactions: e.g., lack of consistent personality,
In an ongoing work, we are also examining cases where conversational agents do indeed successfully "fool" human evaluators. In the current era of "deep fakes", we are trying to identify more robust linguistic markers (which add an additional level of difficulty) and authentication algorithms to distinguish fake conversational agents from real ones.
Currently, we are collecting human-chatbot conversational data to create an active-learning style approach to improve open-domain conversational agents and to construct authentication algorithms. Our data-collection experiments are conducted here. If you are interested in participating, please read over the IRB consent information and sign-up on the website. You can do multiple experiments, track your data, and delete your account if you'd like to opt out at any point.
For reproducibility, this repository provides the following instruments to reconstruct the PersonaGPT from "scratch" (i.e., from pretrained DialoGPT or GPT-2, either of which are feasible as starting pre-trained models).
- Python 3.6+
- Pytorch (GPU preferred)
- transformers package
- dotenv
- tqdm
- (optional) apex for fp16 training
It is highly recommended that the
pytorchandtransformerspackages are installed under a virtual environment.
After cloning this repository, follow the directions below to set up the training environment.
- Go to the
.envfile and set thesave_pathto our local repository to store model checkpoints. Pointdata_pathto the/datafolder of the cloned repository. The.envfile also contains the hyperparameter configurations:
epochs = 3
learn_rate = 5e-5
gradient_accumulation_steps = 8
batch_size = 1
weight_decay = 0.0
logging_steps = 10
save_steps = 250
Replace epochs, batch_size, gradient_accumulation_steps and learn_rate with the desired hyperparameters of choice. If you have less than 12Gb GPU memory, consider using batch size = 1, with gradient accumulation steps between [4-8].
-
Run
preprocess_dataset.pyto preprocess/data/train_both_original_no_cands.txtand/data/valid_both_original_no_cands.txt. The original.txtfiles are obtained from the ConvAI2 Challenge, which may no longer be available since the ConvAI3 challenge has taken place. The ConvAI2 challenge data uses the Persona-Chat dataset which is what is provided under the/datafolder. -
Run
train.pyto train the PersonaGPT model. Results (i.e., loss per<logging_steps>) will be saved undersave_path/samples/. Model checkpoints are saved undersave_path/checkpoint/model.
-
Radford, Alec, et al. "Language models are unsupervised multitask learners." OpenAI Blog 1.8 (2019): 9.
-
Zhang, Yizhe, et al. "Dialogpt: Large-scale generative pre-training for conversational response generation." arXiv preprint arXiv:1911.00536 (2019).
-
Zhang, Saizheng, et al. "Personalizing dialogue agents: I have a dog, do you have pets too?." arXiv preprint arXiv:1801.07243 (2018).