System

Models

0 LightGBM · 0 contrastive
LightGBM

None active

Personal embedding

No personal embedding trained yet

No model history yet. Performance ribbon will appear after the first train.

Training

Run these in order whenever your cohort has changed materially (new approveds, new rejected_seed imports, large vetting batch committed). Each step feeds the next.

1

Contrastive (personal embedding)

Learns a 256-dim space where approveds cluster apart from rejecteds from your swipe history. LightGBM consumes this as a feature, so it must be retrained first.

2

LightGBM (ranker)

Trains the approve/reject classifier on the updated personal embedding + foundation features. New row appears below with overall accuracy, rejected recall, AUC, and gate.

3

Re-evaluate vetting

Re-scores every candidate currently in the vetting queue against the freshly-trained models. Runs in the background — vetting queue refreshes as decisions update.

4

Re-evaluate spring clean

Re-scores every flagged profile in the existing pool against the freshly-trained models. Independent of step 3 — they can run in either order or in parallel.

LightGBM history

No LightGBM models trained yet.

Contrastive adapter history

No contrastive models trained yet.

Couldn't reach the pipeline. Check health.