PR #1513

open

Non-record JEPA submission: VRS (Void Rescue System)

by ikermoelView on GitHub
val_bpb
1.8658
Architecture
Transformer
Optimizer
Artifact Size
15,980,840 bytes

Training Techniques

Architecture
JEPA-style regression transformer
Causal transformer trained to predict next-token embeddings with an MSE objective instead of direct token classification.
parameters: null
MLP
Small auxiliary rescuer decoder that maps raw regression latents v_void to corrected latents v_rescued before final decoding.
parameters: {"parameters":524288}
Regularization
weight decay
parameters: null
Compression
zlib
level: null

Novel Contributions

  • JEPA-style regression language model for the Parameter Golf challenge
  • Auxiliary rescuer decoder that corrects regression latents before token decoding
  • Regression-only training with MSE against target token embeddings
  • Three-seed stability evidence under the 10-minute 8xH100 budget
  • Comparison against standalone regression-only baselines