This repo contains code and info of parspec assignment for DataScientist Role
Checkout the APP HERE -> https://parspec-bert-finetuned-lighting.streamlit.app/
There are 2 ways to run the code / make predictions
1 - To ease the process, I have already deployed the fine-tuned model [ RECOMMENDED ] You can click this link -> https://parspec-bert-finetuned-lighting.streamlit.app/ and make preds
2 - You can checkout the notebook which I have created. Load the finetuned model and make preds.
PS < to make it more accessible, I have pushed the finetuned model to huggingface hub : Check here -> https://huggingface.co/luci007/LightingData-Bert-Finetuned/tree/main >
- Clone the repo
- Install the requirements.txt file
- Move into code folder ->
cd /parspec_assignment/code - Run the
tutorial.ipynbfile to make preds
PS -> You need to have proper modules installed with versions to do so
There are 5 Folders inside the repo
1 - src/ : Folder containing all the main code for Streamlit APP for deployment + Frontend
- main.py : Main Streamlit code
- model.py : Loading model and making preds
- utils.py : Code for Downloading PDF from URL and Extracting text from pdf and cleaning text from extracted text
2 - code/ : This is the main experiment code folder.
- BERT_BASED_UNCASED_FINETUNED_CLEAN_DATA.ipynb : The Main experiment notebook which has code for loading the data, partitioning, DataLoader, Dataset Class and FineTuning the BERT BASED UNCASED MODEL on Cutom Data and Benchmarking it
- tutorial.ipynb : A tutorial notebook which you can run manually to check the entire flow and make preds.
- rough_code : dir : This has my rough code
- tuts_file : experimenting files which were stored for tutorial purpose. You can use them if you like
- push_to_hf_hub.ipynb : NB to push my finetuned model to HF hub
3 - clean_data/ : This folder has 2 .csv file
- training and test csv : These CSV's are created after extracting texts from pdf and cleaning them
4 - holdout_score/ : Folder which has the benchmark score [results.csv] on the holdout data provided. Accuracy - 91.25 %
5 - images : Folder having readme.md images
- : Download the pdf file from the URL and save them in directory [ ~ 1hrs ] - time taken to download, not for code [ start date - 30/11/23 ]
- : Extract text from the pdf files and save them in a .txt file [ ~ 30 mins ] - time taken to extract text, not for code [ start date - 30/11/23 ]
- : Convert the extracted text into a csv file [ for both train and test ]. This csv will be used for finetuning model
- : Cleaning and pre-processing the texts
- : Finetuned BERT base model and BERT large model, had issues in bert large model. Satisfied with bert base
- : Try out Debertav3 model from MSFT for text classification [ couldn't do it, since model is large and no compute power]
- : Benchmarking the scores on test data [ Update : scored 91.25 % Accuracy ]
- : Add a tutorial NB for the team to run the code and check preds
- : Uploaded to huggingface hub [link -> https://huggingface.co/luci007/LightingData-Bert-Finetuned]
- : Host on Streamlit Cloud [ final link -> https://parspec-bert-finetuned-lighting.streamlit.app/]
- : Try out MultiModality Model [ Text + Images ] to be used for context [ time constraint, couldn't do]