Code for paper "Do Language Models Have Beliefs? Methods for Detecting, Updating, and Visualizing Model Beliefs"

This is the codebase for the paper: Do Language Models Have Beliefs? Methods for Detecting, Updating, and Visualizing Model Beliefs

Directory Structure

data/ --> data folder including splits we use for FEVER, zsRE, Wikidata5m, and LeapOfThought
training_reports/ --> folder to be populated with individual training run reports produced by main.py
result_sheets/ --> folder to be populated with .csv's of results from experiments produced by main.py
aggregated_results/ --> contains combined experiment results produced by run_jobs.py
outputs/ --> folder to be populated with analysis results, including belief graphs and bootstrap outputs
models/ --> contains model wrappers for Huggingface models and the learned optimizer code
data_utils/ --> contains scripts for making all datasets used in paper
main.py --> main script for all individual experiments in the paper
metrics.py --> functions for calculing metrics reported in the paper
utils.py --> data loading and miscellaneous utilities
run_jobs.py --> script for running groups of experiments
statistical_analysis.py --> script for running bootstraps with the experimental results
data_analysis.Rmd --> R markdown file that makes plots using .csv's in result_sheets
requirements.txt --> contains required packages

Requirements

The code is compatible with Python 3.6+. data_analysis.Rmd is an R markdown file that makes all the plots in the paper.

The required packages can be installed by running:

pip install -r requirements.txt

If you wish to visualize belief graphs, you should also install a few packages as so:

sudo apt install python-pydot python-pydot-ng graphviz

Making Data

We include the data splits from the paper in data/ (though the train split for Wikidata5m is divided into two files that need to be locally combined.) To construct the datasets from scratch, you can follow a few steps:

1. Set the DATA_DIR environment variable to where you'd like the data to be stored. Set the CODE_DIR to point to the directory where this code is.
2. Run the following blocks of code

Make FEVER and ZSRE

cd $DATA_DIR
git clone https://github.com/facebookresearch/KILT.git
cd KILT
mkdir data
python scripts/download_all_kilt_data.py
mv data/* ./
cd $CODE_DIR
python data_utils/shuffle_fever_splits.py
python data_utils/shuffle_zsre_splits.py

Make Leap-Of-Thought

cd $DATA_DIR
git clone https://github.com/alontalmor/LeapOfThought.git
cd LeapOfThought
python -m LeapOfThought.run -c Hypernyms --artiset_module soft_reasoning -o build_artificial_dataset -v training_mix -out taxonomic_reasonings.jsonl.gz
gunzip taxonomic_reasonings_training_mix_train.jsonl.gz taxonomic_reasonings_training_mix_dev.jsonl.gz taxonomic_reasonings_training_mix_test.jsonl.gz taxonomic_reasonings_training_mix_meta.jsonl.gz
cd $CODE_DIR
python data_utils/shuffle_leapofthought_splits.py

Make Wikidata5m

cd $DATA_DIR
mkdir Wikidata5m
cd Wikidata5m
wget https://www.dropbox.com/s/6sbhm0rwo4l73jq/wikidata5m_transductive.tar.gz
wget https://www.dropbox.com/s/lnbhc8yuhit4wm5/wikidata5m_alias.tar.gz
tar -xvzf wikidata5m_transductive.tar.gz
tar -xvzf wikidata5m_alias.tar.gz
cd $CODE_DIR
python data_utils/filter_wikidata.py

Experiment Replication

Experiment commands require a few arguments: --data_dir points to where the data is. --save_dir points to where models should be saved. --cache_dir points to where pretrained models will be stored. --gpu indicates the GPU device number. --seeds indicates how many seeds per condition to run. We give commands below for the experiments in the paper, saving everything in $DATA_DIR.

To train the task and prepare the necessary data for training learned optimizers, run:

python run_jobs.py -e task_model --seeds 5 --dataset all --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e write_LeapOfThought_preds --seeds 5 --dataset LeapOfThought --do_train false --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

To get the main experiments in a single-update setting, run:

python run_jobs.py -e learned_opt_main --seeds 5 --dataset all --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

For results in a sequential-update setting (with r=10) run:

python run_jobs.py -e learned_opt_r_main --seeds 5 --dataset all --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

To get the corresponding off-the-shelf optimizer baselines for these experiments, run

python run_jobs.py -e base_optimizers --seeds 5 --do_train false  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e base_optimizers_r_main --seeds 5 --do_train false  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

To get ablations across values of r for the learned optimizer and baselines, run

python run_jobs.py -e base_optimizers_r_ablation --seeds 1 --do_train false  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

Next we give commands for for ablations across k, the choice of training labels, the choice of evaluation labels, training objective terms, and a comparison to the objective from de Cao (in order):

python run_jobs.py -e learned_opt_k_ablation --seeds 1 --dataset ZSRE  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e learned_opt_label_ablation --seeds 1 --dataset ZSRE --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e learned_opt_eval_ablation --seeds 1 --dataset ZSRE  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e learned_opt_objective_ablation --seeds 1 --dataset all  --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR
python run_jobs.py -e learned_opt_de_cao --seeds 5 --dataset all --data_dir $DATA_DIR --save_dir $DATA_DIR --cache_dir $DATA_DIR

Analysis

Statistical Tests

After running an experiment from above, you can compute confidence intervals and hypothesis tests using statistical_analysis.py.

To get confidence intervals for the main single-update learned optimizer experiments, run

python statistical_analysis -e learned_opt_main -n 10000

To run hypothesis tests between statistics for the learned opt experiment and its baselines, run

python statistical_analysis -e learned_opt_main -n 10000 --hypothesis_tests true

You can substitute the experiment name for results for other conditions.

Belief Graphs

Add --save_dir, --cache_dir, and --data_dir arguments to the commands below per the instructions above.

Write preds from FEVER model:
python main.py --dataset FEVER --probing_style model --probe linear --model roberta-base --seed 0 --do_train false --do_eval true --write_preds_to_file true

Write graph to file:
python main.py --dataset FEVER --probing_style model --probe linear --model roberta-base --seed 0 --do_train false --do_eval true --test_batch_size 64 --update_eval_truthfully false --fit_to_alt_labels true --update_beliefs true --optimizer adamw --lr 1e-6 --update_steps 100 --update_all_points true --write_graph_to_file true --use_dev_not_test false --num_random_other 10444

Analyze graph:
python main.py --dataset FEVER --probing_style model --probe linear --model roberta-base --seed 0 --test_batch_size 64 --update_eval_truthfully false --fit_to_alt_labels true --update_beliefs true --use_dev_not_test false --optimizer adamw --lr 1e-6 --update_steps 100 --do_train false --do_eval false --pre_eval false --do_graph_analysis true

Combine LeapOfThought Main Inputs and Entailed Data:
python data_utils/combine_leapofthought_data.py

Write LeapOfThought preds to file:
python main.py --dataset LeapOfThought --probing_style model --probe linear --model roberta-base --seed 0 --do_train false --do_eval true --write_preds_to_file true --leapofthought_main main

Write graph for LeapOfThought:
python main.py --dataset LeapOfThought --leapofthought_main main --probing_style model --probe linear --model roberta-base --seed 0 --do_train false --do_eval true --test_batch_size 64 --update_eval_truthfully false --fit_to_alt_labels true --update_beliefs true --optimizer sgd --update_steps 100 --lr 1e-2 --update_all_points true --write_graph_to_file true --use_dev_not_test false --num_random_other 8642

Analyze graph (add --num_eval_points 2000 to compute update-transitivity):
python main.py --dataset LeapOfThought --leapofthought_main main --probing_style model --probe linear --model roberta-base --seed 0 --do_train false --do_eval true --test_batch_size 64 --update_eval_truthfully false --fit_to_alt_labels true --update_beliefs true --optimizer sgd --update_steps 100 --lr 1e-2 --do_train false --do_eval false --pre_eval false --do_graph_analysis true

Plots

The data_analysis.Rmd R markdown file contains code for plots in the paper. It reads data from aggregated_results and saves plots in a ./figures directory.