How MEDIUM_CONFIDENCE Prediction Mode Works in Magika: Threshold Logic and Fallback Behavior

In Magika's MEDIUM_CONFIDENCE mode, a file is assigned its deep-learning predicted content type only when the model's confidence score meets or exceeds the global medium_confidence_threshold (default 0.5); otherwise, the prediction falls back to generic TXT or UNKNOWN labels.

Google's Magika library provides three prediction modes—BEST_GUESS, MEDIUM_CONFIDENCE, and HIGH_CONFIDENCE—that determine how strictly the model interprets its own confidence scores. When you configure Magika to use MEDIUM_CONFIDENCE mode, it applies a single global threshold to decide whether to trust the neural network's output or return a conservative generic type, making it ideal for applications requiring balanced accuracy without per-content-type customization.

The Confidence Threshold Logic

The core validation logic resides in python/src/magika/magika.py within the _get_output_label_from_dl_label_and_score method. For every file analyzed, the deep-learning model returns a content-type label (dl_label) and a floating-point confidence score (score). The MEDIUM_CONFIDENCE mode evaluates this score against exactly one threshold:

# Conceptual flow for MEDIUM_CONFIDENCE

if score >= self._model_config.medium_confidence_threshold:
    # Trust the model output

    output_label = dl_label
else:
    # Fall back to generic label

    output_label = ContentTypeLabel.TXT or ContentTypeLabel.UNKNOWN

According to the source code analysis, MEDIUM_CONFIDENCE uses a single global threshold defined in the model configuration file (assets/models/standard_v3_3/config.min.json). Unlike HIGH_CONFIDENCE mode—which checks per-type thresholds first—MEDIUM_CONFIDENCE applies the same threshold universally across all content types.

Configuration and Default Values

The threshold value is stored in the model's JSON configuration:

• File: assets/models/standard_v3_3/config.min.json
• Key: medium_confidence_threshold
• Default value: 0.5

You can inspect this configuration to verify the threshold used by your specific model version. If a prediction's score falls below this value, Magika triggers the fallback mechanism regardless of the specific content type being predicted.

Comparison with HIGH_CONFIDENCE Mode

Understanding MEDIUM_CONFIDENCE requires contrasting it with the stricter HIGH_CONFIDENCE alternative:

Mode	Threshold Strategy	Behavior When Condition Fails
BEST_GUESS	None (always trust)	Never falls back
MEDIUM_CONFIDENCE	Global threshold only	Falls back to TXT/UNKNOWN
HIGH_CONFIDENCE	Per-type threshold first, then global fallback	Falls back to TXT/UNKNOWN

In HIGH_CONFIDENCE mode, Magika first checks if the score meets a content-type-specific threshold (e.g., 0.9 for certificate files) before falling back to the global medium threshold. MEDIUM_CONFIDENCE skips the per-type check entirely, using only the global medium_confidence_threshold for all file types.

Fallback Behavior When Confidence Is Low

When operating in MEDIUM_CONFIDENCE mode, scores below the threshold trigger the overwrite logic found in magika.py:

overwrite_reason = OverwriteReason.LOW_CONFIDENCE
if self._get_ct_info(output_label).is_text:
    output_label = ContentTypeLabel.TXT
else:
    output_label = ContentTypeLabel.UNKNOWN

This means:

• Text-like files receive the TXT label
• Binary files receive the UNKNOWN label
• The overwrite_reason field is set to LOW_CONFIDENCE in the prediction result

Code Examples

Initializing Magika with MEDIUM_CONFIDENCE Mode

from magika import Magika
from magika.types import PredictionMode

# Explicitly select medium confidence mode

magika = Magika(prediction_mode=PredictionMode.MEDIUM_CONFIDENCE)

result = magika.identify_path("document.pdf")
print(f"Label: {result.prediction.output.label}")
print(f"Score: {result.prediction.score}")
print(f"Overwrite reason: {result.prediction.overwrite_reason}")

Observing the Threshold Boundary

from magika import Magika
from magika.types import PredictionMode, OverwriteReason

magika = Magika(prediction_mode=PredictionMode.MEDIUM_CONFIDENCE)

# Process multiple files

paths = ["image.png", "script.sh", "tiny_fragment.bin"]
results = magika.identify_paths(paths)

for r in results:
    if r.prediction.overwrite_reason == OverwriteReason.LOW_CONFIDENCE:
        print(f"{r.path}: Below threshold (score={r.prediction.score:.2f}) → {r.prediction.output.label}")
    else:
        print(f"{r.path}: Accepted (score={r.prediction.score:.2f}) → {r.prediction.output.label}")

Programmatic Threshold Check Logic

While you cannot change the threshold at runtime without modifying the model config, you can verify the current threshold programmatically:

magika = Magika(prediction_mode=PredictionMode.MEDIUM_CONFIDENCE)
threshold = magika._model_config.medium_confidence_threshold
print(f"Current medium confidence threshold: {threshold}")  # Typically 0.5

Summary

• MEDIUM_CONFIDENCE applies a single global threshold (medium_confidence_threshold, default 0.5) to all content types equally.
• When scores fall below this threshold, Magika returns TXT for text-like content and UNKNOWN for binary content, setting overwrite_reason to LOW_CONFIDENCE.
• This mode is less strict than HIGH_CONFIDENCE (which uses per-type thresholds) but more conservative than BEST_GUESS (which never falls back).
• The implementation resides in python/src/magika/magika.py, specifically within the _get_output_label_from_dl_label_and_score method.
• Configuration values are loaded from config.min.json in the model assets directory.

Frequently Asked Questions

What is the difference between MEDIUM_CONFIDENCE and HIGH_CONFIDENCE in Magika?

MEDIUM_CONFIDENCE uses only the global medium_confidence_threshold (typically 0.5) for every file type, while HIGH_CONFIDENCE first checks per-content-type thresholds defined in the model configuration before falling back to the global threshold. This makes HIGH_CONFIDENCE stricter for specific file types that require higher certainty, whereas MEDIUM_CONFIDENCE applies uniform validation across all formats.

How does Magika handle predictions that fail the MEDIUM_CONFIDENCE threshold?

When a prediction's confidence score is below the medium_confidence_threshold, Magika overwrites the model output with a generic label. Text-like files are labeled as TXT, while binary files are labeled as UNKNOWN. The result object includes overwrite_reason: LOW_CONFIDENCE to indicate that the deep-learning prediction was rejected due to insufficient confidence.

Can I customize the medium confidence threshold value?

The threshold is defined in the model configuration file (assets/models/standard_v3_3/config.min.json) and loaded at initialization. You cannot override this value via the Python API without modifying the underlying configuration JSON or loading a custom model. The default value of 0.5 is optimized for balanced precision-recall trade-offs across the standard model.

When should I use MEDIUM_CONFIDENCE instead of HIGH_CONFIDENCE?

Use MEDIUM_CONFIDENCE when you need consistent behavior across all file types without requiring higher certainty for specific formats, or when you want to avoid the complexity of per-type thresholds. It is appropriate for batch processing pipelines where uniform confidence standards are preferable to type-specific scrutiny. Choose HIGH_CONFIDENCE when certain content types (like executables or certificates) require stricter validation before classification.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}