Isda SMS classification (spam / ham / smishing)

Binary Keras LSTM models trained for Philippine SMS: spam vs ham, then smishing vs spam. Preprocessing and thresholds are defined in preprocessing_config.json.

Files

lstm_spam_ham_detection_model.h5 — first-stage spam gate
lstm_smish_ham_detection_model.h5 — second-stage smishing vs spam
preprocessing_config.json — vocab_size, sentence_len, decision thresholds, Tagalog stopwords

Labels returned by the reference API: 0 ham, 1 spam (non-smish), 2 smish.

Usage

Install: pip install tensorflow scikit-learn nltk huggingface_hub

Download the repo locally:

huggingface-cli download Anecra/isda-sms-lstm --local-dir ./isda-hub

Load config and models from ./isda-hub, then run the same pipeline as in the Isda TrainingFiles/preprocess.py module (one_hot + pad_sequences with values from preprocessing_config.json).

Citation

If you use these weights, cite or link the upstream Isda project repository you obtained them from.

Downloads last month: -