imacdonald/phishing-email-detection-distilbert-executorch

Phishing email detection DistilBERT (ExecuTorch .pte, fp32), et-server model directory

An ExecuTorch export of cybersectony/phishing-email-detection-distilbert_v2.4.1 (Apache-2.0) for et-server POST /classify, and for lemonade's executorch recipe (/v1/classify) on the fork branch release/prpl-demo. Same model as lemonade-sdk/phishing-email-detection-distilbert-ONNX, so the three runtimes (ONNX Runtime, LiteRT, ExecuTorch) can be compared on one model.

• model.pte: 268147600 bytes, sha256 f4ed97a49b34c3016bfaa27da29542f8416918ce7dbc0a32695b8395bba0ae08. ExecuTorch 1.5.1, XNNPACK delegate, fp32. Four methods seq_64, seq_128, seq_256, seq_512, the same graph at four fixed lengths; inputs input_ids and attention_mask int64 [1, L], output float32 [1, 4].
• tokenizer.json, vocab.txt, tokenizer_config.json, special_tokens_map.json, config.json: from the source model.
• manifest.json: et-server's contract (task, id2label, max_length 512).
• validation.json: produced by et-server's tools/export_pte.py; every method matches the PyTorch source within 1e-6 (max score delta 2.9e-10), with tool versions and XNNPACK delegation coverage.

Labels are LABEL_0..LABEL_3, identical to the source model and the ONNX export. The source model's card maps them as: 0 legitimate_email, 1 phishing_url, 2 legitimate_url, 3 phishing_url_alt.

Memory with et-server v0.1.0 (musl, 2 threads, XNNPACK weights packed once and shared across methods): about 230 MiB anonymous with all four methods loaded; loading only --seq-lens 64,512 saves under 1 MiB.

Unofficial repackaging; the weights are unchanged from the source model.