DevInterp is a python library for conducting research on developmental interpretability, a novel AI safety research agenda rooted in Singular Learning Theory (SLT). DevInterp proposes tools for detecting, locating, and ultimately controlling the development of structure over training.
Read more about developmental interpretability.
⚠️ This library is still in early development. Don't expect things to work on a first attempt. We are actively working on improving the library and adding new features.
To install devinterp
, simply run pip install devinterp
.
from devinterp.slt import sample, LLCEstimator
from devinterp.optim import SGLD
from devinterp.utils import optimal_temperature
# Assuming you have a PyTorch Module and DataLoader
llc_estimator = LLCEstimator(..., temperature=optimal_temperature(trainloader))
sample(model, trainloader, ..., callbacks = [llc_estimator])
llc_mean = llc_estimator.sample()["llc/mean"]
To see DevInterp in action, check out our example notebooks:
For more advanced usage, see the Diagnostics notebook and for a quick guide on picking hyperparameters, see the Calibration notebook.
. Documentation can be found here.
For papers that either inspired or used the DevInterp package, click here.
- The current implementation does not work with transformers out-of-the-box. This can be fixed by adding a wrapper to your model, for example passing Unpack(model) to sample() where unpack is defined by:
class Unpack(nn.Module):
def __init__(model: nn.Module):
self.model = model
def forward(data: Tuple[torch.Tensor, torch.Tensor]):
return self.model(*data)
- LLC Estimation is currently more of an art than a science. It will take some time and pain to get it work reliably.
If you run into issues not mentioned here, please first check the github issues, then ask in the DevInterp Discord, and only then make a new github issue.
See CONTRIBUTING.md for guidelines on how to contribute.
This package was created by Timaeus. The main contributors to this package are Stan van Wingerden, Jesse Hoogland, and George Wang. Zach Furman, Matthew Farrugia-Roberts and Edmund Lau also made valuable contributions or provided useful advice.
If this package was useful in your work, please cite it as:
@misc{devinterp2024,
title = {DevInterp},
author = {Stan van Wingerden, Jesse Hoogland, and George Wang},
year = {2024},
howpublished = {\url{https://github.com/timaeus-research/devinterp}},
}