Datasets:

omrisap
/

arc-agi-1-zloopvit-traces

license: apache-2.0
task_categories:
  - image-to-image
language:
  - en
tags:
  - arc-agi
  - abstract-reasoning
  - intermediate-reasoning
  - visual-reasoning
pretty_name: ARC-AGI-1 ZLoopViT Traces
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl

ARC-AGI-1 ZLoopViT Traces

Canonical intermediate grid trajectories for all 400 training tasks in ARC-AGI-1. Each row corresponds to one original official train or test pair; generated augmentations are not included.

Dataset contents

400 tasks
1,718 trajectories: 1,302 demonstration/train pairs and 416 test pairs
5,059 visible intermediate transitions
Exact final-output validation on every official pair

Important columns:

task_id: official ARC task identifier
train_or_test: whether the source pair is an official demonstration or test pair
input, output: original ARC grids
steps: ordered intermediate grid checkpoints
events: structured descriptions aligned one-to-one with steps
trace_id, tracer_version, tracer_sha256: trace provenance
Z_steps_k_4, Z_steps_k_8, Z_steps_k_16, Z_steps_k_32: latent transition-token sequences aligned one-to-one with steps

The file summary.json contains aggregate export statistics and effective code counts.

For every row and every available K, len(Z_steps_k_K) == len(steps). The K=8 and K=16 models use 7 and 12 effective codes respectively; their unused nominal token IDs do not appear in the corresponding columns.

Construction and validation

Every task has one deterministic traced program. Each published trajectory was regenerated from its official ARC-AGI-1 input and required to end exactly at the official expected output. The traces were also reviewed through representative HTML filmstrips for transformation correctness, checkpoint granularity, and event label fidelity.

Generated RE-ARC or ARC-GEN examples were used only as stress-test evidence and are not rows in this dataset.

Two tasks have explicit generated-source exceptions while remaining exact on all official pairs:

a64e4611: 937/1,000 sampled ARC-GEN examples matched; clipped-edge labels can depend on hidden generator geometry.
e5062a87: 624/1,000 sampled ARC-GEN examples matched; some sprite labels depend on hidden placement metadata not uniquely recoverable from the rendered grid.

These exceptions do not alter the official trajectories in this dataset.

Intended use

The dataset is intended for research on ARC intermediate supervision, learned refinement operators, latent step discovery, planner/refiner models, and trace alignment. Intermediate checkpoints are authored computational traces, not claims about a unique human reasoning process.

License and source

The underlying ARC-AGI data is distributed under Apache License 2.0. This traced derivative is released under the same license.