rerobbins / w266_final_project

Final project repository for UCB MIDS W266. Group: Bronte Baer, Jean-Luc Jackson, Richard Robbins

w266_final_project

Final project repository for UCB MIDS W266. Group: Bronte Baer, Jean-Luc Jackson, Richard Robbins

Project File Organization

We utilize our Google Drive project folder to avoid GitHub’s large file restrictions. All datasets, models, and predictions can be found in our Google Drive.

Repo File Organization

code

Contains working folders for each team member. These are unstructured.

data

Points you to our Google Drive project folder — see note above and section below.

Google Drive File Organization

Data

Contains a folder for each dataset, formatted as <dataset-name>.hf. Each folder contains the post-processed train and validation data, named <model>_train_pairs.csv and <model>_valid_pairs.csv respectively. Here, <model> indicates how the data was formatted — as input into T5, BART, etc.

• quac.hf
  • … (miscellaneous files/folders)
  • ...train_pairs.csv
  • ...valid_pairs.csv
• squad.hf
  • …
  • ...train_pairs.csv
  • ...valid_pairs.csv
• trivia_qa_rc.hf
  • …
  • ...train_pairs.csv
  • ...valid_pairs.csv
• …

Models

Contains a folder for each model, formatted as <checkpoint>_<framework>_<training-dataset(s)>. Here, <checkpoint> refers to the Hugging Face model checkpoint we used (with underscores in place of hyphens or spaces). <framework> refers to the machine learning coding framework, either PyTorch (pt) or TensorFlow/Keras (tf). <training-dataset(s)> refers to the datasets used to fine-tune the model, listed alphabetically and separated by underscores.

• T5_base_pt_squad
  • checkpoints
  • … (other model-related files/folders)
• T5_base_tf_squad
  • checkpoints
  • … (other model-related files/folders)
• …

Notes

Contains miscellaneous notes related to our project.

Predictions

Contains a folder for each model with folder names matching the formatting of the Models folder above. Each model folder here contains a predictions.csv with two columns: target | prediction. The target is the question we are aiming to generate and the prediction is the question actually generated by the model.

This folder also contains a notebook generate_predictions.ipynb that generates the described CSV and evaluates the predictions saved in each subfolder. These evaluations are summarized in our project report.

Reference Material

Contains related literature to our project, including inspiration and generally related works.