Galileo
Searchโ€ฆ
๐Ÿ”
Glossary

A per sample holistic data quality score to identify samples in the dataset contributing to low or high model performance i.e. โ€˜pullingโ€™ the model up or down respectively. In other words, the DEP score measures the potential for "misfit" of an observation to the given model. Read Moreโ€‹

The subset of data samples with highest DEP scores, thereby "hard" for the model to learn from during training or "hard" for the model to make predictions on at test time. These samples can be hard due to one or more of the following reasons: boundary samples, noisy / corrupt samples, mislabelled samples, misclassified samples, out-of-distribution samples etc. Read Moreโ€‹

The subset of data samples with lowest DEP scores, thereby "easy" for the model to learn from during training, or "easy" for the model to make predictions on at test time. Typically these "easy" samples are clean, noise free data samples that the model had no issues training/predicting on. Read More

The ground truth label of a sample as specified by the user. In other words, the target label used by the model for a sample.

The predicted label of a sample as specified by the model. In other words, the label with highest prediction probability by the model.

Use interchangeably with Gold, and is the ground truth of a sample as specified by the user. In other words, the target used by the model for a sample.

In a text_multi_label run, the task corresponds to the task your model is making a prediction for. For example:
input: I can't believe today is the day!
outputs: task happiness: true, task nervousness: true, task anger: false
In this case, your tasks are happiness, nervousness, and anger

The particular modeling exercise your model is working towards. Some task types in Galileo:
  • text_classification
  • text_multi_label
  • text_ner
โ€‹
Copy link
On this page
DEP Score
Hard for the model
Easy for the model
Gold
Pred
Label
Task
Task Type