Skip to content
D-CSIL

Model Registry · FieldTransformer family

FieldTransformer v1.4

MultiScaleFieldLayer (k=3 + k=7 Laplacian) + causal attention · flux / Schwarzian / latent-diffusion / entropy regularizers

FIELDTRANSFORMER V1.4 — TRAINING LOSS (DEMO DATA)172,800 STEPS · MIN 0.01
0.00.51.11.62.10k43k86k130k173k

Benchmarks

MetricValueNotes
v1.2 final loss0.0109PPL 1.01 — memorization at this scale, labeled as such
v1.3cKILLEDstopped at 3,500 steps

Known Limitations

  • Tiny-corpus results only; PPL 1.01 on Shakespeare demonstrates fit, not generalization.
  • Custom IP terms apply (see repository notice).