Data Artifacts

This directory contains all model predictions and evaluations. Results for all experiments and runs are also available in a Google Sheet here.

At the top level, this directory is divided into tasks TaskA and TaskB.

TaskA

TaskA-ValidationSet-SubmissionFormat.csv contains the shared task validation set in the submission format for easier evaluation with model outputs. This file can be re-generated with scripts/convert_to_submission_format.py.
predictions contains the model outputs for three runs
results contains the metrics after evaluating the outputs of each run

TaskB-ValidationSet-SubmissionFormat.csv contains the shared task validation set in the submission format for easier evaluation with model outputs. This file can be re-generated with scripts/convert_to_submission_format.py.
predictions contains the model outputs for three runs
results contains the metrics after evaluating the outputs of each run
predictions and results are further divided by approach, into fine-tuning and in-context-learning
in-context-learning is further divided according to the ablation into filtered and unfiltered and then random and similar, and finally note_only and dialogue_note
human_eval contains all the resources used in the human evaluation (see human_eval/README.md for more details)

Contains raw token length counts and histograms for the training and validation sets of all tasks. Further divided by tokenizer used ("gpt-4" in openai or "google/flan-t5-large" in huggingface). Can be re-generated with scripts/count_and_plot_tokens.py.