omrisap/ruleloopvit-gl-sft019200-all400-epoch33

RuleLoopViT GL — all 400 ARC-AGI-1 tasks, epoch 33

This is the preserved epoch-33 checkpoint from the second grounded-language RuleLoopViT training stage. Training uses all 400 ARC-AGI-1 training tasks, RE-ARC examples, grounded demonstration memory (G), and rules generated by the 19,200-example SFT language model.

Contents

• checkpoint_epoch_0033.pt: complete PyTorch training checkpoint.
• run_manifest.json: architecture, data, rendering, and parameter metadata.
• granite_full_rule_tokens.pt: tokenized 400-task generated-rule cache.
• arc1_all400_sft019200_rules.json: all-400 training partition manifest.
• arc1_generator_principal_300x64_seed42.json: diagnostic 300/50/50 split.
• eval_epoch33_partitioned.json: valid evaluation separated by that split.

Epoch-33 valid evaluation

Partition	Tasks	Evaluation pairs	CE	Token accuracy	Exact match
SFT train	300	306	0.125608	0.952438	0.326797
Generator dev	50	55	0.252097	0.922574	0.272727
Joint dev	50	55	0.222909	0.919155	0.163636
All tasks	400	416	0.155740	0.943891	0.300481

The checkpoint uses the custom RuleLoopViT/LoopViT implementation and is not a Transformers AutoModel checkpoint.