omrisap commited on
Commit
17e7836
·
verified ·
1 Parent(s): 54d8c90

Document epoch 25 checkpoint reconstruction

Browse files
Files changed (1) hide show
  1. README.md +11 -5
README.md CHANGED
@@ -3,12 +3,18 @@ license: apache-2.0
3
  ---
4
  # RuleLoopViT GL + NVARC bootstrap
5
 
6
- Phase 1E checkpoint after epoch 18 of the mixed ARC-AGI-1 RE-ARC and NVARC bootstrap run.
7
 
8
- - Epoch: 18
9
- - Optimizer step: 34,668
 
10
  - Effective batch size: 264
11
- - Checkpoint SHA-256: 679238e8053912a240a1319db5d7833670fbc635a180f5aa61cbb7768c3f583f
12
  - W&B run ID: gl-arc-nvarc-bootstrap-seed42-v1
13
 
14
- The checkpoint includes model, optimizer, scheduler, support sampler, negative sampler, and resume metadata. Training continued after this snapshot.
 
 
 
 
 
 
3
  ---
4
  # RuleLoopViT GL + NVARC bootstrap
5
 
6
+ Phase 1E checkpoints from the mixed ARC-AGI-1 RE-ARC and NVARC bootstrap run.
7
 
8
+ ## Latest snapshot: epoch 25
9
+
10
+ - Optimizer step: 48,150
11
  - Effective batch size: 264
12
+ - SHA-256: 81faae2407fd44a00affedf25f978632d40359fec88170b7ea0663aa8cd0110f
13
  - W&B run ID: gl-arc-nvarc-bootstrap-seed42-v1
14
 
15
+ The epoch-25 checkpoint is losslessly split into four files because the training host repeatedly dropped long uploads. Reconstruct it after downloading the repository:
16
+
17
+ cat epoch25_parts/checkpoint_epoch_0025.pt.part-* > checkpoint_epoch_0025.pt
18
+ sha256sum checkpoint_epoch_0025.pt
19
+
20
+ The reconstructed checkpoint includes model, optimizer, scheduler, support sampler, negative sampler, and resume metadata. Training continued after this snapshot.