This repository provides an implementation of Rank-One Model Editing (ROME) on auto-regressive transformers (GPU-only). We currently support OpenAI's GPT-2 XL (1.5B) and EleutherAI's GPT-J (6B). The release of a 20B GPT-like model from EleutherAI is expected soon; we hope to support it ASAP.
Feel free to open an issue if you find any problems; we are actively developing this repository and will monitor tickets closely.
We recommend conda
for managing Python, CUDA, and PyTorch-related dependencies, and pip
for everything else. To get started, simply install conda
and run:
./scripts/setup_conda.sh
notebooks/causal_trace.ipynb
demonstrates Causal Tracing, which can be modified to apply tracing to the processing of any statement.
notebooks/rome.ipynb
demonstrates ROME. The API is simple; one simply has to specify a requested rewrite of the following form:
request = {
"prompt": "{} plays the sport of",
"subject": "LeBron James",
"target_new": {
"str": "football"
}
}
Several similar examples are included in the notebook.
Details coming soon!
See baselines/
for a description of the available baselines.
experiments/evaluate.py
can be used to evaluate any method in baselines/
.
To get started (e.g. using ROME on GPT-2 XL), run:
python3 -m experiments.evaluate \
--alg_name=ROME \
--model_name=gpt2-xl \
--hparams_fname=gpt2-xl.json
Results from each run are stored at results/<method_name>/run_<run_id>
in a specific format:
results/
|__ ROME/
|__ run_<run_id>/
|__ params.json
|__ case_0.json
|__ case_1.json
|__ ...
|__ case_10000.json
To summarize the results, you can use experiments/summarize.py
:
python3 -m experiments.summarize --dir_name=ROME --runs=run_<run_id>
Running python3 -m experiments.evaluate -h
or python3 -m experiments.summarize -h
provides details about command-line flags.
Say you have a new method X
and want to benchmark it on CounterFact. To integrate X
with our runner:
HyperParams
into XHyperParams
and specify all hyperparameter fields. See ROMEHyperParameters
for an example implementation.hparams/X/gpt2-xl.json
and specify some default values. See hparams/ROME/gpt2-xl.json
for an example.apply_X_to_model
which accepts several parameters and returns (i) the rewritten model and (ii) the original weight values for parameters that were edited (in the dictionary format {weight_name: original_weight_value}
). See rome/rome_main.py
for an example.X
to ALG_DICT
in experiments/evaluate.py
by inserting the line "X": (XHyperParams, apply_X_to_model)
.Finally, run the main scripts:
python3 -m experiments.evaluate \
--alg_name=X \
--model_name=gpt2-xl \
--hparams_fname=gpt2-xl.json
python3 -m experiments.summarize --dir_name=X --runs=run_<run_id>
We currently only support methods that edit autoregressive HuggingFace models using the PyTorch backend. We are working on a set of general-purpose methods (usable on e.g. TensorFlow and without HuggingFace) that will be released soon.
@article{meng2022locating,
title={Locating and Editing Factual Associations in {GPT}},
author={Kevin Meng and David Bau and Alex Andonian and Yonatan Belinkov},
journal={Advances in Neural Information Processing Systems},
volume={35},
year={2022}
}