Text Classification

Text · HuggingFace · Fine-tuned

The Text Classification detector runs any HuggingFace text-classification pipeline model on asset content. You control the model checkpoint, confidence threshold, and a severity map that translates predicted labels into finding severity levels.

When to use

Spam and phishing detection — run a dedicated spam classifier on email content or support tickets.
Toxicity and content moderation — flag harmful language in user-generated text before it reaches moderation queues.
Sentiment analysis — score feedback, reviews, or support interactions to identify negative patterns.
Custom topic labelling — deploy a fine-tuned model trained on your own labelled data to detect domain-specific categories.
Confidence-gated findings — use confidence_threshold to surface only high-confidence predictions and reduce noise.

How it works

At scan time, Classifyre loads the specified HuggingFace model (either from the Hub or a local path) and runs the text-classification pipeline on the asset content. The top-k predicted labels above the confidence_threshold become findings. The severity_map translates specific label names to severity levels; unmatched labels fall back to the severity default.

For long documents, set chunk_size to split the text before classification. Each chunk is scored independently; findings reference the originating chunk position.

Configuration

Parameter	Type	Required	Description	Default	Constraints
model	string	Yes	HuggingFace hub ID (e.g. 'mrm8488/bert-tiny-finetuned-sms-spam-detection') or absolute local directory path.	—	—
model_revision	string	No	Git branch, tag, or commit hash when fetching from the HuggingFace hub.	null	—
device	string	No	Inference device: 'cpu' (default), 'cuda', 'mps', or a CUDA device string like 'cuda:0'.	cpu	—
top_k	integer	No	Maximum number of top predictions to return. When null all labels above confidence_threshold are returned.	null	—
function_to_apply	string	No	Score normalization: 'softmax' for single-label, 'sigmoid' for multi-label, 'none' for raw logits.	null	—
confidence_threshold	number	No	Minimum prediction confidence to report a label as a finding (0-1).	0.7	min 0, max 1
severity	string	No	Default severity when no severity_map rule matches.	info	—
severity_map	array	No	Ordered rules mapping predicted labels to severity levels. First matching rule wins.	null	—
max_length	integer	No	Override the tokenizer's maximum sequence length.	null	—
chunk_size	integer	No	Split text into chunks of this many characters before classification.	null	—
chunk_overlap	integer	No	Character overlap between consecutive chunks.	0	—

Examples

SMS spam detection

{
  "type": "TEXT_CLASSIFICATION",
  "model": "mrm8488/bert-tiny-finetuned-sms-spam-detection",
  "confidence_threshold": 0.85,
  "severity_map": [
    { "label": "spam", "severity": "high" }
  ],
  "severity": "info"
}

Toxicity classifier with chunking for long documents

{
  "type": "TEXT_CLASSIFICATION",
  "model": "unitary/toxic-bert",
  "device": "cpu",
  "confidence_threshold": 0.7,
  "chunk_size": 512,
  "chunk_overlap": 64,
  "severity_map": [
    { "label": "toxic", "severity": "high" },
    { "label": "obscene", "severity": "medium" }
  ]
}