README.md

---
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](https://github.com/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.