File size: 1,734 Bytes
0dc4ea9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | ---
license: mit
task_categories:
- text-generation
tags:
- arc-agi
- re-arc
- test-time-training
- synthetic-data
pretty_name: ARC Tiny Transformer corpus and evaluation candidates
---
# ARC Tiny Transformer data
Data and vote-level artifacts for [N8python/arc-tiny-transformer](https://github.com/N8python/arc-tiny-transformer).
## Corpus
`corpus/documents_3k_per_task.jsonl.zst` is the exact pretraining corpus:
- 1,200,000 JSONL documents;
- 3,000 documents for each of 400 RE-ARC generator families;
- 3,401,127,124 ARC tokens;
- 10,413,648,837 uncompressed bytes;
- uncompressed SHA-256 `5d1336f8d6f45358a377af2e3fd05d43c4bfe91621b0c6a44c53dad0e34ca430`.
Each record contains token IDs and metadata for task ID, sampled difficulty band, demonstration count, rejection statistics, and generation seed. The fixed vocabulary and sampler are in the GitHub repository. `manifests/` contains the complete task-level corpus provenance.
## Evaluation artifacts
`artifacts/` contains the full 128-candidate records for:
- the frozen 50M model;
- TTT replicas seeded 57, 58, and 59;
- the 7M TTT run.
These files are sufficient to recompute rank-1, raw-frequency top-2, historical hierarchical top-2, oracle accuracy, candidate-scaling curves, and multi-replica ensembles without rerunning model inference.
`viewer/data/` contains the compact per-query data backing the browser-based 384-candidate explorer.
## Licensing note
Release code and metadata are MIT licensed. The corpus is procedurally generated with the pinned [RE-ARC](https://github.com/michaelhodel/re-arc) generators/verifiers; ARC-AGI task files are not duplicated here. Users should also review the upstream RE-ARC and ARC-AGI licenses for their use case.
|