README.md

---
language: en
license: apache-2.0
tags:
  - setfit
  - sentence-transformers
  - dlp
  - data-loss-prevention
  - pii
  - security
  - spidercob
base_model: sentence-transformers/all-MiniLM-L6-v2
metrics:
  - accuracy
model-index:
  - name: dlp-intent-classifier
    results:
      - task:
          type: text-classification
        metrics:
          - type: accuracy
            value: 0.9425
---

# DLP Intent Classifier

Fine-tuned [SetFit](https://github.com/huggingface/setfit) model for Data Loss Prevention intent classification. Part of the [Spidercob](https://spidercob.com) DLP platform.

## Model Description

This model classifies text snippets that triggered DLP regex patterns into one of four categories to reduce false positives:

| Label | Description |
|---|---|
| `REAL_DATA` | Genuine sensitive data (PII, credentials, secrets) |
| `TEST_DATA` | Synthetic/mock data used in tests or examples |
| `DOCUMENTATION` | Documentation examples, tutorials, placeholder values |
| `NOISE` | Random strings, hashes, or non-sensitive matches |

**Base model:** `sentence-transformers/all-MiniLM-L6-v2`  
**Framework:** SetFit (few-shot fine-tuning)  
**Test accuracy:** 94.25%

## Usage

```python
from setfit import SetFitModel

model = SetFitModel.from_pretrained("Sumeetgpt/dlp-intent-classifier")
predictions = model.predict(["sk-proj-abc123XYZ", "test_api_key_placeholder"])
print(predictions)  # ['REAL_DATA', 'TEST_DATA']
```

## Intended Use

Used inside the Spidercob DLP engine to verify whether regex-matched findings are genuinely sensitive before blocking or alerting. Blocks on `REAL_DATA` confidence > 0.7; allows through on `TEST_DATA` confidence > 0.8.

## Training Data

Trained on curated examples spanning:
- Real PII samples (anonymized): emails, SSNs, credit cards, API keys, passwords
- Synthetic test data from `faker`, `factory_boy`, pytest fixtures
- Documentation examples from popular libraries
- Noise patterns (random strings, base64, UUIDs)

## Limitations

- Optimized for English text
- Best performance on short text snippets (< 512 tokens)
- Not a replacement for regex-based pattern matching — meant as a downstream filter