Todd is a library designed to provide anomaly scorer and the derived filters for text generation. It supports three levels of anomaly detection: input level, generated sequence level and token level. The input level anomaly detection is used for OOD detection. Whereas the generated sequence level aims at selecting the best candidates in a beam and the token level aims at detecting potentially wrong tokens in the generated sequence.

It revolves around two concepts: scorers and filters. The scorers are used to score the input, token or ouput and filters leverage these scores to produce masks to decide what to keep or reject.

For benchmarking purpose we tend to only use scorers to dump the scores and compte statistics. In practice though, we would want to use filters to select the best candidates.

@software{Darrin_Todd_A_tool_2023,
author = {Darrin, Maxime and Faysse, Manuel and Staerman, Guillaume and Picot, Marine and Dadalto Camara Gomez, Eduardo and Colombo, Pierre},
month = {2},
title = {{Todd: A tool for text OOD detection.}},
url = {https://github.com/icannos/Todd},
version = {0.0.1},
year = {2023}
}

git clone [email protected]:icannos/Todd.git
cd Todd
pip install -e .

pip install todd

Different examples are provided in the examples folder.

# Extract features from the reference set
ref_embeddings, _ = extract_embeddings(model, tokenizer, in_val_loader, layers=[6])

# Initialize the Mahalanobis detector
maha_detector = MahalanobisScorer(layers=[6])
# Fit the detector with the reference set
maha_detector.fit(ref_embeddings)

with torch.no_grad():
    for batch in loader:
        inputs = tokenizer(
            batch["source"], padding=True, truncation=True, return_tensors="pt"
        )
        output = model.generate(
            **inputs,
            return_dict_in_generate=True,
            output_hidden_states=True,
            output_scores=True,
        )

        print(maha_detector(output)) 
        # Output the scores for each input in the batch
        # True means the input is In-Distribution (ie to be kept) and False means OOD

• Add reorderers to reorder the beam search candidates
• Split filters and anomaly scorers (WIP, maxime)
• Add power means and cosines ood scores