---
title: Dark_humor_generator
app_file: app.py
sdk: gradio
sdk_version: 5.6.0
---
DISCLAIMER!!! THIS PROJECT IS DEDICATED ONLY FOR EDUCATIONAL AND ENTERTAINMENT PURPOSE AND NOT TO OFFEND OR HUMILIATE SOMEONE!

This project is dedicated for the creation of dark humor generating bot(english language). It is not only using the concept of RAG but 
is also dedicated to prompt hacking exploration. The helping context for model prompt is generated by files with offensive speech, to make LLM less constricted, twitter and reddit jokes. All data files have ending _preprocessed.json. 
Data preprocessing can be seen in data_playground.py. 
app.py is responsible for using LLM using gradio.
Retriever.py is used for getting 'n' docs by similiraties using BM25 and/or semantic from embeddings using 'sentence-transformers/all-distilroberta-v1'. 
Semantic embeddings are stored partically to avoid RAM overusing. However even this can be not enough for process to avoid state where he is being killed due to RAM shortage so maybe you will need to use swap memory. Also the logic of getting doc embeddings should be reworked in retriever.py with the use of device.cuda. I have created embeddings in kaggle kernels with separate from project create-embeddings-for-semantic-score.ipynb. 
For tokenization I used lowercasing + NLTK library(It's usage is in tokenizing.py file). 
In LLM_usage.py it is described the logic of retrieving responces from 'groq/llama3-8b-8192'. 
I have considered to not use reranker because of lack of time mostly :( I will probably implement it later 
This project is not working perfectly, I consider that finetuning the model with my dataset using LORA or other methods could show
better results

Using instructions
python 3.10.12 or higher
install requirements.txt to your python env (If something is missing, install it as well)
Maybe, you will need to change 'base_path' in app.py, LLM_usage.py and retriever.py
Then to launch you just need to write in terminal gradio app.py and follow the link