Fugi1128

/

ScamScan

Like 0

Replaces the scam classifier with the version retrained on public call datasets and generated Indian-context calls, from ScamScan code commit 37af8ac (branch rithvik/manual-calls-and-classifier).

• scamscan_classifier_htp.onnx: retrained MiniLM, same static [1,128] int32 graph as before. On held-out calls it misses 2.0% of scam windows (previously 46.6%) and flags 1.4% of benign windows (previously 2.8%). With the app's smoothing, 313 of 314 held-out scam calls reach Critical and 2.7% of benign calls reach a false Critical (previously 7.7%).
• scam_keywords_in.json: matches the code, with 8 more digital-arrest phrases. Keywords only choose which advice text to show.
• runtime-manifest.json: new sizes and SHA-256 values for both files; code_commit set to 37af8ac.
• README: restore steps through scripts/setup_model_assets.py, a manual download command that skips the files Git tracks, gated-access notes, the classifier's training data and licenses, and the new results.

After merging, set RUNTIME_REVISION in scripts/setup_model_assets.py to the new main revision so new checkouts restore this classifier.

• models/minilm_classifier/: the trained PyTorch MiniLM that the ONNX classifier was exported from, with its config and tokenizer. The app does not download it.