metadata
license: mit
library_name: transformers
tags:
- arc-agi
- test-time-training
- muon
- causal-lm
ARC Tiny Transformer checkpoints
Hugging Face checkpoints for N8python/arc-tiny-transformer.
| Folder | Parameters | Optimized tokens | Public evaluation, identity greedy |
|---|---|---|---|
7m-3.4b |
7,094,784 | 3,399,843,840 | 4.375% |
50m-100m |
50,372,096 | 100,073,472 | 0.750% |
50m-500m |
50,372,096 | 500,072,448 | 2.625% |
50m-1.13b |
50,372,096 | 1,133,150,208 | 5.125% |
50m-3.0b |
50,372,096 | 2,999,844,864 | 9.000% |
50m-3.4b |
50,372,096 | 3,399,843,840 | 8.500% |
440m-0.8b |
440,506,368 | 799,801,344 | 9.250% |
The headline test-time-training experiments use 50m-3.0b. One full-model TTT replica plus 128 greedy augmented candidates reaches 47.75% top-2 task-macro accuracy on the 400-task ARC-AGI-1 public evaluation; pooling three independently adapted replicas reaches 51.50%.
verifier-50m-epoch4 is the auxiliary binary classifier trained on correct, perturbed, and on-policy sequences. It is included for reproduction but did not improve the headline vote aggregation.
All causal-LM folders are standard Transformers/Qwen3-format checkpoints with the custom 19-token tokenizer. See the GitHub repository for exact tokenizer semantics, training code, model hashes, and evaluation commands.