Espritchatbot V1.2

This is a chatbot designed to answer questions about Esprit University. It was created by a group of Esprit's Students called INNOVISION using the Rasa framework.

The unindexed intents are generated using the RWKV-4-Raven Model

Prerequirments

Python 3.4+ Virtualenv pip

Installation

To open the chatbot, clone the repository to your local machine:

git clone https://github.com/kimou6055/Espritchatbot.git

All of the following will be built into a virtualenv

open the cmd in the root folder do :

cd ../

Then do the follow:

python -m venv myenv

Then activate the environnement

windows :

myenv\Scripts\activate

linux :

source myenv/bin/activate

Make sure you have Python 3.x installed on your machine. You can install the required Python libraries by running:

pip install rasa
pip install langdetect
pip install translate
pip install pynvml
pip install rwkv
pip install ninja 

You need to install torch with CUDA

pip3 install torch(version)+cu(CUDA version) torchvision(version)+cu(CUDA version) torchaudio --index-url https://download.pytorch.org/whl/cu117

Updgrade the RWKV-4 package :

pip install rwkv --upgrade

open the RWKV folder and drop the RWKV-4-Raven-7B-V10 model

click on it to download

Here we used the V11X 99% English 1% other 8192 ctx (V11+LaMini instructions)

you can download and use any other version found HERE

Make sure to install CUDA : Linux Windows

Usage

To use the chatbot, navigate to the Espritchatbot directory and run these following commands:

1- if you are using the chatbot for the first time you will need to train it :

rasa train

2- Run the chatbot :

rasa shell

3- You will need to run the actions

Open another cmd access the root then run :

rasa run actions --port 5055

PS : choose the proper strategy for your own hardware configuration while loading the RWKV model in the actions.py file

Here we tested it with a custom cuda fp16 *12 -> cuda fp16i8 *1 -> cpu fp32

you can see other options HERE

RWKV python implementation can be found HERE

Hyperparameters of RWKV can be found in the actions.py file

4- Run the chatbot with nlu accuracy:

rasa shell nlu

Credits

Credits for the RWKV-4-Raven model goes to BlinkDL

Credits for Rasa project structure and implementation goes to INNOVISION

INNOVISION Team Members:

• Med Karim Akkari
• Nadia Bedhiafi
• Sarra Gharsallah
• Karim Aloulou
• Yosr Abbassi
• Med Hedi Souissi
• Aziz Jebabli