README.md

---
license: cc-by-4.0
size_categories:
- 1M<n<10M
task_categories:
- other
pretty_name: ARC Intermediate Solving Steps
tags:
- arc
- arc-agi
- abstract-reasoning
- grid-puzzles
- intermediate-steps
- chain-of-thought
- synthetic
- procedural-generation
configs:
- config_name: arcgen_v1
  data_files:
  - split: train
    path: data/arcgen_v1/*.jsonl
- config_name: arcgen_v2
  data_files:
  - split: train
    path: data/arcgen_v2/*.jsonl
- config_name: rearc
  data_files:
  - split: train
    path: data/rearc/*.jsonl
---

# ARC Intermediate Solving Steps (`arc-steps`)

This dataset accompanies the paper [TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning](https://huggingface.co/papers/2607.29586).

**1,286,952 procedurally generated ARC-style records — 1,062,561 of them
with intermediate *solving steps*.** Each record is an `{input, steps, output}` triple:
`steps` is a sequence of intermediate grids tracing a semantically meaningful solution
path from the input to the output, captured at human-annotated checkpoints of the
program that produces each example.

**Code:** [github.com/LiuBinnan/TraceViT](https://github.com/LiuBinnan/TraceViT)

![Three records: input, then steps, the last step being the output](assets/filmstrip-hero.png)

*One record from each config — top: `arcgen_v1` task `007bbfb7` (fractal composition);
middle: `arcgen_v2` task `e5790162` (path tracing); bottom: `rearc` task `1e0a9b12`
(gravity, resolved per column). Leftmost frame = `input`; the frames after it are
`steps`, and the last one equals `output`.*

| config | tasks | records | records with steps | implementation |
|---|---:|---:|---:|---|
| `arcgen_v1` | 400 | 399,990 | 305,616 | ARC-GEN generators, ARC-AGI-1 training tasks |
| `arcgen_v2` | 500 | 487,091 | 486,091 | ARC-GEN generators, ARC-AGI-2 training tasks |
| `rearc` | 400 | 399,871 | 270,854 | RE-ARC generators + step-instrumented solver programs, ARC-AGI-1 training tasks |

Together the three configs cover **900 distinct ARC tasks**. `rearc` and `arcgen_v1`
share the same 400 ARC-AGI-1 task ids — two independent implementations of the same
puzzles, usable as a cross-implementation contrast set.

**No official ARC grids are included.** All records are synthetic; the official
train/test examples were removed from every config. No validation/test split is
provided — evaluate on the official ARC-AGI benchmarks.

## Schema

One JSON object per line (plain `*.jsonl`, one file per config):

| field | type | description |
|---|---|---|
| `task_id` | str | 8-hex ARC task id |
| `source` | str | `"arcgen_v1"` / `"arcgen_v2"` / `"rearc"` (matches the config) |
| `category` | str | `"selectable"` (task has annotated steps) or `"skip"` (no meaningful intermediates) |
| `input` | int[][] | input grid, values 0-9, at most 30×30 |
| `output` | int[][] | output grid, same constraints |
| `steps` | int[][][] | 0..N intermediate grids; **when non-empty, the last step equals `output`**; intermediate frames may differ in size from the output |
| `seed` | int \| null | generation seed (arcgen only) |
| `gen_kwargs` | str \| null | generator parameters as a JSON string (arcgen only; keys vary per task) |
| `palette` | int[10] \| null | global palette permutation applied at synthesis, `palette[0] == 0` (rare; arcgen only) |
| `background` | int \| null | background recolor layer; `0` means untouched (arcgen only) |
| `foreground_map` | str \| null | foreground color-permutation as a JSON string; `"{}"` means untouched (arcgen only) |
| `difficulty_rng` | float \| null | RE-ARC native difficulty knob (rearc only) |
| `difficulty_pso` | float \| null | RE-ARC native difficulty knob (rearc only) |
| `rule` | str \| null | model-written natural-language description of the task's transformation (`arcgen_v2` only — see "Task rule texts" below) |

Identity check for arcgen records: `background == 0 and foreground_map == "{}"` means
the record kept the generator's own colors. For tasks whose colors carry rule
semantics, one or both recolor layers were disabled entirely during human review;
per-task `bg_layer` / `fg_layer` flags are recorded in `stats/*.per_task.json`.

## Task rule texts (`arcgen_v2`)

Every `arcgen_v2` record carries a `rule` string: a natural-language description of
the task's transformation, written by a language model during the program-
instrumentation campaign (identical for all records of a task). Two caveats, stated
plainly:

- The texts are **model-generated and not individually human-verified** — treat them
  as auxiliary annotations, not ground truth.
- They describe the task in its **canonical (official) colors**. Records whose
  recolor layers are non-identity (`background != 0` or `foreground_map != "{}"`)
  will not match the rule's color words, though the underlying transformation is the
  same up to the recorded color permutation.

`arcgen_v1` and `rearc` records have `rule: null`.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("lbn32/arc-steps", "arcgen_v2", split="train")

# only records that carry intermediate steps
with_steps = ds.filter(lambda r: len(r["steps"]) > 0)

# ±steps ablation: same records, just ignore the steps column
```

Per-task record counts, per-task with-steps counts, and the drop ledger are in
`stats/*.per_task.json`.

## Limitations

- Per-task record counts vary: sampling runs under per-attempt and per-task time
  budgets, and a small number of hard-to-sample tasks fall short of the per-task
  target (exact counts in `stats/`).
- Records are procedurally generated and validated against structural invariants,
  not against a per-instance solvability proof.
- `rearc` excludes 129 generated records (all from task `e26a3af2`)
  whose grids exceeded the 30×30 ARC limit.
- Steps are annotated at the task level (checkpoint positions in the program), not
  per-instance.

## License and attribution

The dataset is released under **CC-BY-4.0** (see `LICENSE`).

It is generated with, and gratefully builds on, the following projects (no upstream
code and no official ARC grids are redistributed here; license copies in `licenses/`):

| project | role | license |
|---|---|---|
| [ARC-AGI](https://github.com/fchollet/ARC-AGI) / [ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2) | task definitions the generators mimic | Apache-2.0 |
| [ARC-GEN](https://github.com/google/ARC-GEN) | procedural generators (`arcgen_v1`, `arcgen_v2`) | Apache-2.0 |
| [RE-ARC](https://github.com/michaelhodel/re-arc) | procedural generators + solver programs (`rearc`) | MIT |