---
license: apache-2.0
base_model: cybersectony/phishing-email-detection-distilbert_v2.4.1
library_name: onnx
pipeline_tag: text-classification
tags:
  - onnx
  - lemonade
  - text-classification
---
# phishing-email-detection-distilbert-ONNX

ONNX export of [`cybersectony/phishing-email-detection-distilbert_v2.4.1`](https://huggingface.co/cybersectony/phishing-email-detection-distilbert_v2.4.1), packaged for the [Lemonade](https://github.com/lemonade-sdk/lemonade) router classifier backend ([`ort-server`](https://github.com/lemonade-sdk/ort-server)).

- **Base model:** [`cybersectony/phishing-email-detection-distilbert_v2.4.1`](https://huggingface.co/cybersectony/phishing-email-detection-distilbert_v2.4.1)
- **Task:** text-classification
- **Labels:** `LABEL_0`, `LABEL_1`, `LABEL_2`, `LABEL_3`

## Files

| file | purpose |
|------|---------|
| `model.onnx` | the exported model (`input_ids`/`attention_mask` → logits) |
| `tokenizer.json` | the original HuggingFace tokenizer |
| `manifest.json` | task / labels / normalization for ort-server |
| `export.py` | the exact script used to produce & validate these files |

## Validation after export

Exported from source with 🤗 Optimum and **validated against the original PyTorch model** on fixtures (ONNX Runtime CPU vs HF): max softmax delta **0.0** (0 = identical).

CPU-EP latency (ONNX Runtime, single input): **~4.01 ms** p50 @ 12 tokens.

## Reproduce

```bash
pip install "optimum[onnxruntime]" transformers torch onnxruntime sentencepiece
python export.py cybersectony/phishing-email-detection-distilbert_v2.4.1 ./out --task text-classification
```

See `validation.json` for the recorded parity result.

## License

Follows the base model [`cybersectony/phishing-email-detection-distilbert_v2.4.1`](https://huggingface.co/cybersectony/phishing-email-detection-distilbert_v2.4.1); refer to it for terms.