What is Jade AI?
Jade's a Contextual chatbot, based on modern NLP techniques. The primary driver of the most recent version of Jade is based on OpenAI's GPT-2 (see this post). GPT-2 uses a tranformer model to effectively and efficiently generate tokenized sequences.
Demo & Stats
You can add the bot to your server, and run JDT [message] (currently under a different prefix than normal due to testing) or JD [message] (for her previous, v5, version based on K-means & beam search)
Pics or it didn't happen.
Link to Code
Note: the codebase is currently incomplete!!! The trained model and training script will be uploaded soon.
A Bit of History
I started the Jade project in my freshman year in high school, knowing absolutely nothing about NLP (natural language processing) and having done very little research on chatbots or artificial intelligence algorithms.
The very first version was really simple: it was a text file with a ton of questions and answers, and I used python to try and match the input sentence to the sentence with the most similar words in the text file, then returned the next line.
Over the years, I've explored LSTM, Seq2Seq, and K-means (with beam search) algorithms to improve the accuracy of the Jade chatbot.
Today
Over the quarrantine period, Jade v6 was completed, acheiving a suprising 88% accuracy on new data (taken from discord). It is able to run on low-end hardware, (1 vCPU, 1.25gb of memory), and is able to maintain context/learning between discord servers without mixing up conversations.
The program runs on Python, uses the discord.py (1.3.3) library for discord interactions, pytorch for machine learning, and GPT-2 as the framework for the neural network.
Trained on my own open-source dataset, Jade is out there and ready to converse over a lot of interesting things...
Top comments (0)