This is a multilabel classification prototype on assigning correct labels to Stackoverflow questions using question title and body.

Project info

• Date: 31/05/2023
• Done by: Baran Nama
• Requested By: Coresystems
• Dataset: https://www.kaggle.com/datasets/stackoverflow/stacksample?datasetId=265

Folder structure

• utils: Python module for some small common utility functions among notebooks.
• notebooks: Jupyter notebooks which have been sorted by subtask order.
• models: Any trained models will be saved here.
• datasets: Input and processed datasets will be stored here.

Notebook structure

• 01_data_processing: Data cleaning operations has been made such as filtering by question score and question label frequency, basic visualizations and text cleaning.
• 02_baseline_model: Several shallow classifiers such as logistic regression, tree based and gradient boosting based classifiers have been chained with TF-IDF vectorization.
• 03_lm_model: A small Transformer based LM has been evaluated on zero shot setup as well as fine-tuned.
• 04_llm_model: A LLM model has been fine-tuned using LoRA and State-of-the-art Parameter-Efficient Fine-Tuning (PEFT).

Steps to reproduce the results and running notebooks

1. Download the dataset from the above URL, create a folder called datasets and unpack the downloaded dataset. There should be 3 CSV files.
2. Setup Conda environment using environment.yml file provided.
3. Register the environment to Jupyter and select the environment while running notebooks.

 python -m ipykernel install --user --name stackoverflow