RuleLoopViT LoopViT-Small epoch-10 fork

This repository contains the immutable epoch-10 checkpoint used to fork the matched RuleLoopViT Phase 1 experiments:

  • Branch A: vision-only continuation, epochs 11โ€“30;
  • Branch B: frozen Granite-30M rule conditioning, epochs 11โ€“30.

The checkpoint contains the model, AdamW optimizer, cosine scheduler, AMP GradScaler, epoch, best validation accuracy, and training configuration states. It is therefore suitable for an exact training continuation rather than only inference.

Architecture and training point

  • LoopViT-Small reconstruction (loop_vit1)
  • embedding width: 512
  • attention heads: 8
  • shared loop-core depth: 1
  • MLP width: 1024
  • recurrent iterations: 24
  • image canvas: 64 ร— 64
  • patch size: 2
  • completed epoch: 10
  • next epoch: 11

Integrity

The expected SHA-256 digest of checkpoint_epoch_0010.pt is recorded in manifest.json. Verify the digest before resuming training.

2d4857df682cade2a89f746a53e9f0a88e3168f3551dad5d465527d8905e4eb6

The checkpoint is tied to the RuleLoopViT code and the pinned LoopViT source used by the project; it is not a standalone Transformers checkpoint.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support