wtf is cdfm?

CDFM, short for Causal Discovery Foundation Model, is a pretrained machine learning model that tries to figure out cause and effect relationships hidden in a dataset. Given a table of numbers with no labels telling you what causes what, the model looks at the raw data and predicts a causal graph, meaning a diagram of which variables in the data appear to influence which other variables, in a single pass rather than through a lengthy statistical search. The README explains that CDFM was trained on a large and varied collection of synthetic datasets built from artificial cause and effect structures, and that this training lets it recognize the kinds of statistical patterns that show up when one variable causes another. Because of that training, it works in what the paper calls a zero-shot setting: the same pretrained model can be applied to a new dataset, of a different size and number of variables, without retraining it first. The README states it outperforms other methods it was compared against across 15 different types of underlying data generating processes and on real-world benchmark datasets. Using the model is meant to be simple. It is installed with a single pip command, and Python code loads it from Hugging Face Hub with one line and produces a prediction with another, calling model.predict on your data to get back a causal graph as a grid of zeros and ones. The library also includes a separate feature for filling in missing values in a dataset, which the README shows performing noticeably better than a simple average based fill-in method on an example dataset. The project is research code released alongside an academic paper on arXiv, and it targets people doing causal inference or structural learning research in Python, such as academics or data scientists who need to estimate causal structure from observational data without hand designing a model for each new dataset. It is released under the Apache 2.0 license.