Prediction Modes

Magika’s deep learning model returns each prediction with a confidence score (from 0.0 to 1.0). A common challenge with classification models is determining the minimum score required to trust a result.

Instead of a single, global threshold, Magika uses per-content-type thresholds. The rationale is that the model is naturally more confident about some types than others. For example, our experiments show that most valid PDFs are detected with over 99% confidence, so a prediction with an 80% score might be questionable. In contrast, an 80% score for a JavaScript file is often a very reliable prediction.

Magika manages these confidence levels in two ways:

• Pre-tuned Thresholds: Each model ships with carefully tuned, per-content-type thresholds derived from evaluating the model on our large validation dataset.
• Prediction Modes: Because the impact of a misidentification varies by use case, Magika allows you to select a prediction mode. This lets you balance precision (accuracy of predictions) and recall (number of identified files). The available modes are high-confidence, medium-confidence, and best-guess.

The high-confidence mode offers higher precision at the cost of lower recall. In contrast, best-guess provides the highest recall—potentially with lower precision—as it returns the model’s prediction regardless of its confidence score. This can be selected via a command-line flag or as an option in the language bindings.