Range estimation sets the numerical bounds used to derive a quantizer’s scale and zero-point from observed activations. If calibration data misses production inputs, activations can exceed those bounds, causing clipping and degraded answers. section 4.0
Named range estimation algorithms include:
- Absolute max section 4.0
- Percentile section 4.0
- MSE section 4.0
- KLD section 4.0
Mismatched calibration data
Calibrating a support chatbot on Wikipedia text can miss order IDs and code in customer tickets. Those inputs may push activations beyond the calibrated range and get clipped. section 4.0
