Overview

This research repository maintains the code and the results for the research paper: SETLEXSEM CHALLENGE: Using Set Operations to Evaluate the Lexical and Semantic Robustness of Language Models.

"Set theory has become the standard foundation for mathematics, as every mathematical object can be viewed as a set." -Stanford Encyclopedia of Philosophy

TL;DR Install Package from PyPI and Use

To install the package, please run: pip install setlexsem You can generate the dataset by:

from setlexsem.generate.generate_sets import make_sets

data_config = {
    "set_types": ["words"],
    "n": [1000],
    "m": [2, 4, 8, 16],
    "item_len": [1, 2, 3],
    "decile_group": [5],
    "swap_status": None,
    "overlap_fraction": [0.5],
}

make_sets(
    set_types=["numbers"],
    n=[10],
    m=[1,2],
    item_len=[1],
    decile_group=None,
    swap_status=None,
    overlap_fraction=[None],
    seed_value=292,
    number_of_data_points= 3
)
# --- or, this can be config based:
config = {
    "set_types": ["numbers"],
    "n": [10],
    "m": [10],
    "item_len": [2],
    "decile_group": None,
    "swap_status": None,
    "overlap_fraction": [0.5],
}
out_data = make_sets(
    config=config,
    number_of_data_points= 5,
    seed_value=292
)

You can generate the prompts by:

from setlexsem.generate.generate_prompts import create_prompts

data_config = {
    "set_types": ["words"],
    "n": [1000],
    "m": [2, 16],
    "item_len": [2],
    "decile_group": [5],
    "swap_status": None,
    "overlap_fraction": [0.5],
}

prompt_config = {
    "op_list": ["union", "intersection"],
    "k_shot": [0, 1, 3],
    "prompt_type": ["formal_language"],
    "prompt_approach": ["baseline", "chain_of_thought"],
    "is_fix_shot": [True]
}

prompt_and_ground_truth = create_prompts(
    # data config
    data_config=data_config,
    number_of_data_points=5,
    random_seed_value=292,
    # prompt config
    prompt_config=prompt_config,
    add_roles=False)

Development

When installing, it's important to upgrade to the most recent pip. This ensures that setup.py runs correctly. An outdated version of pip can fail to run the InstallNltkWordnetAfterPackages command class in setup.py and cause subsequent errors.

/usr/bin/python3 -mvenv venv
. venv/bin/activate
python3 -m pip install --upgrade pip
pip install -e .
pip install -e ."[dev, test]"

NLTK words

If you get errors from nltk about package words not being installed while executing the code in this repository, run:

import nltk
nltk.download("words")

Note that words should be automatically installed by pip when you follow the installation instructions for this package.

Project layout

• configs/
  • configs/experimetns contains configuration files which specify hyperparamter settings for running experiments.
  • configs/generation_data contains configuration files for dataset generation
  • configs/generation_prompt contains configuration files for prompt generation based on the data previously stored
  • configs/post_analysis contains a configuration file which can be used for analysis of cost, latency, and performance metrics for one set of hyperparameters for a particular study. This config is used in the script scripts/anaylsis_for_one_study.py
  • configs/post_hypothesis contains a configuration file which specifies filtering criterias for generating figures for various hypotheses.
• notebooks/ has a Jupyter notebook for generating figures that are used in the research paper
• scripts/ contains Python scripts for running experiments, post-processing the results, and analysis of results
• setlexsem/ is the module which has all the important functions and utils for analysis, experimentation, generation of data, samplers.
  • analyze contains code for error_analysis of post-processed results, visualizaiton code and utils neeeded for generating figures for hypothesis.
  • experiment contains code for invoking LLMs and running experiments for a particular hypothesis/study.
  • generate contains code for generating data, sample synthetic sets, prompts and utils needed for data generation.
  • prepare contains helper functions for partitioning words according to their frequencies.

Generate the dataset

Create the sets

Sample sets of numbers or words

To make the CSV file containing sets of words and numbers, run:

python setlexsem/generate/generate_sets.py --config-path "configs/generation_sets/numbers.yaml" --seed-value 292 --save-data

python setlexsem/generate/generate_sets.py --config-path "configs/generation_sets/words.yaml" --seed-value 292 --save-data

Sample sets based on training-set frequency

To sample sets based on their training-set frequency, we use an approximation based on rank frequency in the Google Books Ngrams corpus.

This requires wget (brew install wget or apt install wget). After installing wget, you need to create deciles.json. The following command downloads the English unigram term frequencies of the Google Books Ngram corpus, filters them by the vocabulary of the nltk.words English vocabulary, and stores the vocabulary, separated by deciles of rank frequency, in data/deciles.json.

scripts/make-deciles.sh

This will take ~10 minutes or more, depending on your bandwidth and the speed of your computer.

To make the CSV file containing sets of words sampled by the approximated training-set frequency, run:

python setlexsem/generate/generate_sets.py --config-path "configs/generation_sets/deciles.yaml" --seed-value 292 --save-data

Sample "deceptive" sets

To sample semantically "deceptive" sets (see paper for details), create hyponyms.json by running the following command:

python scripts/make_hyponyms.py --output-path data/hyponyms.json

To make the CSV file containing deceptive sets:

python setlexsem/generate/generate_sets.py --config-path "configs/generation_sets/deceptive.yaml" --seed-value 292 --save-data

Create the prompts

Once you've sampled the sets, create the prompts. The prompts are written as CSV files in the prompts directory.

Example: Prompts based on the config file

To make the CSV file containing prompts sets of words and numbers, run:

python setlexsem/generate/generate_prompts.py --config-path "configs/generation_prompt/sample_config.yaml" --save-data

Run the evalution

1. Create a config file like configs/experiments/anthr_sonnet.yaml
2. Run the prompts:

python setlexsem/experiment/run_experiments.py --account-number ACCOUNT_NUMBER --save-file --load-previous-run --config-file configs/experiments/anthr_sonnet.yaml

Note: Currently, our experiments are dependent on AWS Bedrock and need an AWS account number to be provided. However, you have the capability to run experiments using OPENAI_KEY. We will add more instructions soon.

3. Post-process the results. (Check whether your study_name is present in the STUDY2MODEL dict in setlexsem/constants.py)

python scripts/save_processed_results_for_study_list.py

4, Analysis of cost, latency, and performance metrics for one set of hyperparameters for a particular study - enter hyperparameter values in the configs/post_analysis/study_config.json

python scripts/analysis_for_one_study.py

• Generate figures using notebooks/Hypothesis Testing - Manuscript.ipynb. Validate the filtering criterias in configs/post_hypothesis/hypothesis.json

Test

To test the full-suite of tests, you need to provide the Account Number.

pytest -s .

You will be prompted to provide your Account Number after that.

Coverage report

pip install pytest-cov
pytest --cov=setlexsem --cov-report=term-missing

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.