This repository is used for storing the implementation part of my master's thesis.
The goal of the thesis is to utilize the dataset from "I Never Said That: A dataset, taxonomy and baselines on response clarity classification" paper and LLMs to improve the metrics of the paper and explore different DL techniques.
Progress so far:
-
Created a cleaner version of train and test set: Train data have been taken from https://huggingface.co/datasets/ailsntua/QEvasion, while test data exist on data folder. The proccess to clean the data is on data_cleaning.py file. The results of that proccess are 4 files in preproccessed_data folder, two for training set and two for test set. For every set there is a full version incliding all the columns of initial datasets and a standard one including only the part needed for training.
-
Experiments on how to perform LoRA finetuning to LLMs: The experiments can be found on the lora_llm.ipynb notebook.