Howzer v2.5.4-DE: a new model generation — sharpened on real member feedback
Three core models retrained against blind spots measured on real ADAC reviews: severity now catches signals buried in long texts, risk no longer inflates pure praise, sentiment recognizes mixed and angry feedback far more reliably (mixed recall 40% → 91%). Every model version is now revision-pinned, and the LLM response path is production-validated on GPU.
By Howzer Team, Product · product ·
What's in v2.5.4-DE
v2.5.4-DE is a model-generation release. Instead of tuning thresholds around model weaknesses, we measured three core models against real member feedback, retrained them on what we found, and removed the workarounds the old models needed. The result: analysis quality moves up across the board, and the system carries fewer crutches.
Severity: no more long-text blind spot
Long reviews used to dilute critical signals — a safety-relevant sentence buried in a long text could be scored as routine. The retrained severity model hits every signal that is visible in its analysis window, keeps false alarms on signal-free long reviews at zero, and improves short-text accuracy at the same time.
Risk: praise is not a risk
The previous risk model over-reacted to enthusiastic praise — pure compliments could score as critical risk, and the pipeline compensated with thresholds and a sentiment gate. The new model scores praise low on its own, so both workarounds are gone, while real escalations (legal threats, churn announcements, safety incidents) keep their priority.
Sentiment: hearing anger and ambivalence
Real feedback is messy: furious all-caps complaints, polite cancellations, praise that ends in a service failure. The previous model misread a measurable share of these as positive. The retrained model fixes every documented case of that class and lifts mixed-sentiment recall from 40% to 91% — the single biggest quality jump in this release.
Model updates are now deliberate
Every ML model version is now pinned to an exact revision in the deployed configuration. A model publish changes production only through an explicit, reviewed version bump — behavior can no longer drift silently underneath a running installation.
LLM responses: validated on GPU
The hybrid response engine was validated end to end on GPU with the production model: real cases, zero template fallbacks, zero hallucinated names, median latency around 2.5 seconds. Installations without a GPU keep the honest deterministic fallback.