requirements.txt

torch>=2.0
ai2-tango[torch,transformers,fairscale,beaker,wandb,gs]>=1.3

# Sometimes you want to insist on the latest tango.
# Uncomment the following if you do.
# ai2-tango[torch,transformers,fairscale,beaker,wandb,gs] @ git+https://github.com/allenai/tango.git@main

# Sometimes you want to insist on the latest beaker-py.
# Uncomment the following if you do.
# beaker-py @ git+https://github.com/allenai/beaker-py.git@main

torchmetrics==0.11.1
more_itertools
spacy>=3.0.0
wget
datasets>=2.14.6
accelerate
bettermap
tiktoken  # For Salesforce/xgen-* models

# For the P3 datasets, which we get from huggingface datasets
protobuf<=3.20

# For lm-eval
bert_score             # For summarization tasks
sacrebleu>=1.5.0
scikit-learn>=0.24.1   # Eleuther uses this for metrics. Can we replace it with torchmetrics?
iso639>=0.1.4
rouge-score>=0.0.4  # Can we replace this with torchmetrics?
# The Eleuther test harness depends on these even at runtime, but does not declare them.
mypy_extensions
pytest

# For promptsource
jinja2
pyyaml>=5