Sqwish Labssqwish-d1 training
LIVEWaiting for data

sqwish-d1 is getting ready

A model for classification, scoring and estimation. Ask multiple questions about one context, with probabilities trained to reflect how often the answers are right.

LIVEWaiting for the first snapshot Refreshes every minute
Connecting to the live feed
SQWISH LABS/ sqwish-d1

About sqwish-d1

The model

Multiple questions, one shared context. sqwish-d1 returns probabilities for yes/no questions, distributions over choices, and scores against ordered criteria. Training rewards both accuracy and calibration.

Training data

Dataset details appear with the live feed.

3 independent seeds, 2 passes through the data, with a fresh shuffle each time.

Evaluation

Held-out questions measure accuracy, probability quality and calibration against Live Jev and the untrained model. JevBench is excluded from training and model selection.