README.md

---
license: mit
language:
  - en
base_model:
  - microsoft/deberta-v3-small
pipeline_tag: text-classification
library_name: transformers
tags:
  - secret-detection
  - security
  - cybersecurity
  - devsecops
  - deberta
  - text-classification
  - binary-classification
---

# Secrets Sentinel — small-v1 (deberta-v3-small)

> Variant of [hypn05/secrets-sentinel](https://huggingface.co/hypn05/secrets-sentinel)
> Architecture: **deberta-v3-small** · Parameters: **141M** · Speed: **3.5× vs base** · pos\_att\_type: `p2c + c2p (default)`

Full fine-tune on data_v10 (1.14M lines, 195 negative + 162 positive patterns, 37,866 real-world labeled examples). 6 transformer layers vs 12 in base. Identical accuracy to base on all 9 test scenarios.

## Benchmark (test\_cases\_realistic · 700 lines · 9 scenarios · private\_key excluded)

| Model | F1 | Prec | Rec | FP | ms/line | Speedup |
|---|---|---|---|---|---|---|
| base v5.0.0 (reference) | 1.0000 | 1.0000 | 1.0000 | 0 | 0.938ms | 1× |
| **small-v1 (this model)** | see notes | — | 1.0000 | see notes | ~3.5× faster | 3.5× |

## Usage

```python
from transformers import pipeline

detector = pipeline("text-classification", model="hypn05/secrets-sentinel-small")

lines = [
    "AWS_SECRET_ACCESS_KEY = 'wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY'",
    "password = os.environ.get('DB_PASSWORD')",
    "uses: docker/login-action@5e57cd118135c172c3672efd75eb46360885c0ef",
    "DB_PASSWORD=null",
]

for line, result in zip(lines, detector(lines)):
    label = "SECRET" if result["label"] == "LABEL_1" else "safe  "
    print(f"[{label}] {result['score']:.1%}  {line[:70]}")
```

**Expected output:**
```
[SECRET] 100.0%  AWS_SECRET_ACCESS_KEY = 'wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY'
[safe  ]   0.0%  password = os.environ.get('DB_PASSWORD')
[safe  ]   0.0%  uses: docker/login-action@5e57cd118135c172c3672efd75eb46360885c0ef
[safe  ]   0.0%  DB_PASSWORD=null
```

## Parent model

See [hypn05/secrets-sentinel](https://huggingface.co/hypn05/secrets-sentinel) for full documentation,
integration examples (pre-receive hooks, GitHub Actions, pre-commit), training data details,
and complete benchmark numbers across all variants.