Skip to content

Latest commit

 

History

History
21 lines (18 loc) · 1.72 KB

README.md

File metadata and controls

21 lines (18 loc) · 1.72 KB

Got Headlines: Abstractive Title generation for News articles

Usage

Run the experiments by doing:
cd src first and then running one of:
python3 main.py [experiment name] to run the training and evaluation for the generalization experiments.
python3 practical_year_model.py --train-year [year number] to run the training and evaluation experiments for the year-by-year experiments. Requires the model from the previous year to be trained first for any given year.
python3 evaluate_model.py [experiment name] to run the evaluation experiments for the generalization experiments on all the out of distribution articles using the model trained for general articles.\

Bash Script Usage

If on an HPC, the above can be run by submitting one of the following jobs:
deepspeed_startup.sh to run the generalization experiments using DeepSpeed (requires the data to be tokenized first without deepspeed)
startup.sh to run the generalization experiments without parallelizattion
fewshot_startup.sh to run the out of distribution evaluations for generalization experiments
year_train.sh to run the year-by-year training experiments \

Experiments names possible

generic_allyears: Extract articles only from the publications CNN, Reuters and The New York Times, for all years and train/evaluate on T5
peg_ally: Extract articles only from the publications CNN, Reuters and The New York Times, for all years and train/evaluate using Pegasus
pegx_ally: Extract articles only from the publications CNN, Reuters and The New York Times, for all years and train/evaluate using Pegasus-X
pegx_ally2019: Extract articles only from the publications CNN, Reuters and The New York Times, for all years and train/evaluate using Pegasus-X