phishing-email-detection-distilbert-v2.4.1-CoreAI — Apple Core AI

A Core AI (.aimodel) build of cybersectony/phishing-email-detection-distilbert_v2.4.1 for on-device inference on Apple Silicon. Upstream ships PyTorch weights and no Core AI build.

What the model is. A DistilBERT sequence classifier for phishing detection in emails and URLs (4 classes). Upstream claims 99.58% accuracy on its own dataset; that figure is upstream's, not measured here.

Outputs

main takes input_ids and attention_mask (int32, fixed length 512). The runtime names outputs after the producing op, not after the wrapper's keys, so select by shape:

Class names are from the upstream README; config.json itself only says LABEL_0..3:

0 legitimate_email
1 phishing_url
2 legitimate_url
3 phishing_url_alt

Probes and results

Probes are in cases-phishing.txt; the predicted index is the argmax of the logits (ref-output.json), and was re-checked from the Hub download.

Measured equivalence

Exported graph vs the original PyTorch model (float32), same inputs, raw outputs diffed.

• Worst absolute difference over all outputs: 3.815e-06
• Worst residual relative to output magnitude: 3.8e-07 (float32 rounding)
• Minimum cosine between embeddings: 1.0000000000

This is arithmetic equivalence only. There is no quality benchmark here (no F1, no accuracy, no comparison with upstream's reported numbers), and no speed or memory measurement.

Built by

ahmedandaloes — Apple Silicon builds (Core AI, MLX, GGUF) of security and language models that ship none: ahmedandaloes.com

Need a model in a format that is not here — another runtime, another quantisation, your own checkpoint? That is the same job. Ask.

Use

import asyncio, numpy as np
from coreai.runtime import AIModel, NDArray, SpecializationOptions
from transformers import AutoTokenizer

async def run(text):
    tok = AutoTokenizer.from_pretrained(".")
    model = await AIModel.load("phishing-email-detection-distilbert-v2.4.1.aimodel", SpecializationOptions.default())
    fn = model.load_function("main")
    enc = tok([text], padding="max_length", truncation=True, max_length=512, return_tensors="np")
    out = await fn({"input_ids": NDArray(enc["input_ids"].astype("int32")),
                    "attention_mask": NDArray(enc["attention_mask"].astype("int32"))})
    arrs = [np.asarray(v.numpy()) for v in out.values()]
    logits = next(a for a in arrs if a.shape == (1, 4))
    return int(logits.argmax())

Files

phishing-email-detection-distilbert-v2.4.1.aimodel/ (graph), manifest.json, validation.json, ref-output.json, cases-phishing.txt, arabic_encoder_coreai.py (exporter and measurer), requirements.txt, tokenizer/config files.

Limits

• Fixed at 512 tokens; longer inputs need chunking.
• 8 English probes measured (listed above); nothing else was.
• Softmax and thresholding are the caller's steps.
• The bare-URL probes disagree with the README's class names. All three URL probes — one phishing-looking and two plainly legitimate (wikipedia.org, github.com) — got index 3 (phishing_url_alt), in the original PyTorch model as well as here. So index 3 does not mean "phishing" on bare URLs, or the README names for 2/3 are off. This conversion reproduces upstream exactly; it does not repair it. Verify the class semantics on your own data before use.

Licence and attribution

apache-2.0, as declared in the upstream Hub card. A derivative of cybersectony/phishing-email-detection-distilbert_v2.4.1; credit for the model is theirs.