Family
LightGBM binary (ONNX) + RuleClassifierONNX shim fallback
in-developmentMITtabularinternal
Scribe Signal Classifier.
LightGBM binary classifier gating which conversation turns trigger Scribe LLM extraction. PR 2 ships with RuleClassifierONNX shim filling the LightGBM slot (Option E); PR 7 trains the real model.
Foundation · LightGBM binary (ONNX) + RuleClassifierONNX shim fallback (MIT)
What it does
The honest description.
LightGBM binary classifier gating which conversation turns trigger Scribe LLM extraction. PR 2 ships with RuleClassifierONNX shim filling the LightGBM slot (Option E); PR 7 trains the real model.
Architecture
How it works.
Turn-level features extracted by ScribeFeatureBuilder: message length, token count, question flag, numeric value flag, named entity count, IndicBERT embedding norms (8-dim), prior-turn acceptance rate. Binary signal score ∈ [0, 1]. Scores ≥ 0.6 trigger LLM extraction; scores in [0.3, 0.6) enter abstention zone (shim overrides); scores < 0.3 bypass extraction entirely.
Training data & drift
What it learned from, and how we watch it.
Source
Inputs: Turn-level features extracted by ScribeFeatureBuilder: message length, token count, question flag, numeric value flag, named entity count, IndicBERT embedding norms (8-dim), prior-turn acceptance rate.
Drift baseline · current quarter
Drift baseline not yet documented for this model.
What it's NOT
- PR 7 trains a real LightGBM on Scribe corpus; abstention zone becomes a real divergence point.
- LightGBM is MIT-licensed.
- Task 2.5 (same PR) registers binary_imbalanced drift strategy.
Who calls this model
Across the house and the API.
Internal routes
No public routes documented yet.
Featured in
/features/your-memoryVersion history · immutable once published
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.