setup.cfg

[metadata]
name = crfm-helm
version = 0.4.0
author = Stanford CRFM
author_email = contact-crfm@stanford.edu
description = Benchmark for language models
long_description = file: README.md, docs/tutorial.md
long_description_content_type = text/markdown
keywords = language models benchmarking
license = Apache License 2.0
classifiers =
    Programming Language :: Python :: 3 :: Only
    Programming Language :: Python :: 3.8
    License :: OSI Approved :: Apache Software License
url = https://github.com/stanford-crfm/helm

[options]
python_requires = >=3.8,<3.11
package_dir =
    =src
packages = find:
zip_safe = False
include_package_data = True

install_requires=
    # Common
    cattrs~=22.2
    dacite~=1.6
    importlib-resources~=5.10
    Mako~=1.2
    numpy~=1.23
    pyhocon~=0.3.59
    retrying~=1.3
    spacy~=3.5
    tqdm~=4.64
    zstandard~=0.18.0
    # sqlitedict==2.0.0 is slow! https://github.com/RaRe-Technologies/sqlitedict/issues/152
    # Keep sqlitedict version at 1.7.0.
    sqlitedict~=1.7
    bottle~=0.12.23

    # Basic Scenarios
    datasets~=2.15
    pyarrow>=11.0.0,  # Pinned transitive dependency for datasets; workaround for #1026
    pyarrow-hotfix~=0.6  # Hotfix for CVE-2023-47248

    # Basic metrics
    nltk~=3.7
    pyext~=0.7
    rouge-score~=0.1.2
    scipy~=1.10
    uncertainty-calibration~=0.1.4
    scikit-learn~=1.1

    # Models and Metrics Extras
    transformers~=4.37  # For anthropic_client, vision_language.huggingface_vlm_client, huggingface_client, huggingface_tokenizer, test_openai_token_cost_estimator, model_summac (via summarization_metrics)
    # TODO: Upgrade torch - we need > 2.0.0 for newer versions of transformers
    torch>=1.13.1,<3.0.0  # For huggingface_client, yalm_tokenizer, model_summac (via summarization_metrics)
    torchvision>=0.14.1,<3.0.0  # For huggingface_client, yalm_tokenizer, model_summac (via summarization_metrics)

    # Metrics Extras
    google-api-python-client~=2.64  # For perspective_api_client via toxicity_metrics

[options.extras_require]
proxy-server =
    gunicorn~=20.1.0

human-evaluation =
    scaleapi~=2.13.0
    surge-api~=1.1.0

scenarios =
    gdown~=4.4.0  # For disinformation_scenario, med_mcqa_scenario, med_qa_scenario: used by ensure_file_downloaded()
    sympy~=1.11.1  # For numeracy_scenario
    xlrd~=2.0.1  # For ice_scenario: used by pandas.read_excel()

metrics =
    numba~=0.56.4  # For copyright_metrics
    pytrec_eval==0.5  # For ranking_metrics
    sacrebleu~=2.2.1  # For disinformation_metrics, machine_translation_metrics

summarization =
    summ-eval~=0.892  # For summarization_metrics

plots =
    colorcet~=3.0.1
    matplotlib~=3.6.0
    seaborn~=0.11.0

decodingtrust =
    fairlearn~=0.9.0

slurm =
    simple-slurm~=0.2.6

cleva =
    unidecode==1.3.6
    pypinyin==0.49.0
    jieba==0.42.1
    opencc==1.1.6
    langdetect==1.0.9

images =
    accelerate~=0.25.0  # For the newer versions of Transformers
    pillow~=10.2

mongo =
    pymongo~=4.2.0

# Model extras
aleph-alpha =
    aleph-alpha-client~=2.14.0
    tokenizers>=0.13.3

amazon = 
    boto3~=1.28.57
    awscli~=1.29.57
    botocore~=1.31.57

anthropic =
    anthropic~=0.2.5
    websocket-client~=1.3.2  # For legacy stanford-online-all-v4-s3

mistral =
    mistralai~=0.0.11

openai =
    openai~=1.0
    tiktoken~=0.3.3

google =
    google-cloud-aiplatform~=1.38.1

tsinghua =
    icetk~=0.0.4

yandex =
    sentencepiece~=0.1.97

models =
    crfm-helm[aleph-alpha]
    crfm-helm[amazon]
    crfm-helm[anthropic]
    crfm-helm[google]
    crfm-helm[mistral]
    crfm-helm[openai]
    crfm-helm[tsinghua]
    crfm-helm[yandex]

vlm =
    # VLM models
    crfm-helm[openai]
    torch~=2.1.2    # For IDEFICS

    # VLM scenarios
    crfm-helm[images]
    crfm-helm[image2structure]

image2structure =
    crfm-helm[images]
    # Latex
    # You will need to install LaTeX separately.
    # You can run `sudo apt-get install texlive-full` on Ubuntu.
    latex~=0.7.0
    pdf2image~=1.16.3
    opencv-python~=4.7.0.68
    lpips~=0.1.4
    selenium~=4.17.2

heim =
    # HEIM scenarios
    gdown~=4.4.0

    # HEIM models
    diffusers~=0.24.0
    jax~=0.4.13
    jaxlib~=0.4.13
    crfm-helm[openai]

    # For model, kakaobrain/mindall-e
    einops~=0.6.0
    omegaconf~=2.3.0
    pytorch-lightning~=2.0.5

    # For model, craiyon/dalle-mini and craiyon/dalle-mega
    flax~=0.6.11
    ftfy~=6.1.1
    Unidecode~=1.3.6
    wandb~=0.13.11

    # HEIM perturbations
    google-cloud-translate~=3.11.2

    # HEIM metrics
    autokeras~=1.0.20
    clip-anytorch~=2.5.0
    google-cloud-storage~=2.9.0
    lpips~=0.1.4
    multilingual-clip~=1.0.10
    NudeNet~=2.0.9
    opencv-python~=4.7.0.68
    pytorch-fid~=0.3.0
    tensorflow~=2.11.1
    timm~=0.6.12
    torch-fidelity~=0.3.0
    torchmetrics~=0.11.1

    # Shared image dependencies
    crfm-helm[images]

# Install everything
all =
    crfm-helm[proxy-server]
    crfm-helm[human-evaluation]
    crfm-helm[scenarios]
    crfm-helm[metrics]
    crfm-helm[plots]
    crfm-helm[decodingtrust]
    crfm-helm[slurm]
    crfm-helm[cleva]
    crfm-helm[images]
    crfm-helm[models]
    crfm-helm[mongo]
    crfm-helm[heim]
    crfm-helm[vlm]
    # crfm-helm[dev] is excluded because end-users don't need it.
    # crfm-helm[summarize] is excluded because it requires torch<2.0
    # TODO(#2280): Add crfm-helm[summarize] back.

# Development only
# Do not include in all
dev =
    pytest~=7.2.0
    pre-commit~=2.20.0
    # Errors produced by type checkers and linters are very version-specific
    # so they are pinned to an exact version.
    black==22.10.0
    mypy==1.5.1
    flake8==5.0.4

[options.entry_points]
console_scripts = 
    helm-run = helm.benchmark.run:main
    helm-summarize = helm.benchmark.presentation.summarize:main
    helm-server = helm.benchmark.server:main
    helm-create-plots = helm.benchmark.presentation.create_plots:main
    crfm-proxy-server = helm.proxy.server:main
    crfm-proxy-cli = helm.proxy.cli:main

[options.packages.find]
where = src
exclude =
    tests*

# Settings for Flake8: Tool For Style Guide Enforcement
[flake8]
max-line-length = 120
exclude =
    venv/*
    src/helm/proxy/clients/image_generation/dalle_mini/*
    src/helm/proxy/clients/image_generation/mindalle/*

# Ignore completely:
# E203 - White space before ':', (conflicts with black)
# E231 - Missing whitespace after ',', ';', or ':'
# E731 - do not assign a lambda expression, use a def
# W503 - line break before binary operator, (conflicts with black)
# W605 - invalid escape sequence '\', (causes failures)
ignore = E203,E231,E731,W503,W605

# Settings for Mypy: static type checker for Python 3
[mypy]
ignore_missing_imports = True
check_untyped_defs = True
# TODO: Remove disable_error_code
disable_error_code = annotation-unchecked
# TODO: Change disallow_untyped_defs to True
disallow_untyped_defs = False
exclude = dalle_mini|mindalle

[tool:pytest]
addopts =
    # By default, we don't test models because doing so will
    # make real requests and spend real money
    -m 'not models'
markers =
    # Marker for tests that make real model requests
    models