KoGPT2 chatbot

Simple Chit-Chat based on KoGPT2

Purpose

• Simple chatbot experiment using publicly available Hangul chatbot data and pre-trained KoGPT2

• We explore the potential for various uses of KoGPT2 and qualitatively evaluate its performance.

Architecture

Hello! is suitable for training data. Applying GPT-2, we designed the following architecture

• Learn a model that can maximize P(|, ) by mapping the Q field of the data to utterance, the A field to utterance, and the emotion label to . Emotion labels follow the definitions here (everyday occurrence 0, breakup (negative) 1, love (positive) 2).

Install

 git clone --recurse-submodules https://github.com/haven-jeon/KoGPT2-chatbot.gitcd KoGPT2-chatbot
pip3 install -r requirements.txt

How to Train

 CUDA_VISIBLE_DEVICES=0 python train_torch.py ​​--gpus 1 --train --max_epochs 2

How to Chat!

 # CUDA_VISIBLE_DEVICES=0 python train.py --chatCUDA_VISIBLE_DEVICES=0 python train_torch.py ​​--gpus 1 --chat
user > jackpot
 Simsimi > It will be fine.

user > What is your name?Simsimi > I am Wirobot.

user > What do we do on Parents' Day? Simsimi > Play with me.

user > It's raining
Simsimi > Take an umbrella.

user > It's snowing
Simsimi > Build a snowman.

user > Thunder strikes
Simsimi > I miss you too.

user > Spring has come.
Simsimi > I'm always excited about spring.

user > I have an upset stomach.
Simsimi > Go home and rest quickly.

user > Who made you?Simsimi > You are the person who made you.

Colab tutorials

• PyTorch

•