Arko007/zenyx-arc-synth

Zenyx ARC-Synth

597,000+ Programmatically Synthesized ARC-Style Spatial & Inductive Reasoning Puzzles

Correct by Construction • 37 Rule Families • Zero Benchmark Leakage • Apache-2.0

---

Overview

Zenyx ARC-Synth is a large-scale, high-fidelity synthetic dataset designed to foster abstract reasoning, spatial transformation understanding, and few-shot program induction in language models.

Inspired by François Chollet's Abstraction and Reasoning Corpus (ARC / ARC-AGI), this dataset contains 597,098 training tasks and 2,000 validation tasks (~480 Million tokens) spanning 37 diverse transformation families.

Core Design Philosophy: "Correct by Construction"

Unlike synthetic datasets generated by prompting commercial LLMs (which frequently suffer from hallucinated labels, off-by-one spatial errors, and broken logical consistency), every task in Zenyx ARC-Synth is synthesized programmatically via deterministic algorithmic generators. The input grids are programmatically constructed, and the target output grids are computed via exact transformation kernels. As a result, every single label is mathematically guaranteed to be 100% correct.

---

Dataset Summary

Attribute	Details
Total Examples	597,098 (Train) + 2,000 (Validation)
Total Estimated Tokens	~480.2 Million tokens
Average Tokens per Row	804.3 tokens (P90: 1,486 tokens)
Synthesis Method	Programmatic (Python / NumPy generator; no LLM involved)
Transformation Families	37 distinct mathematical and spatial categories
Chained Transformations	~25% of tasks compose two sequential rules
License	Apache 2.0
Decontamination	Rigorously scrubbed against all 800 public ARC-AGI-1 tasks

---

Dataset Structure

Each record represents a complete self-contained few-shot inductive reasoning problem:

• instruction (string): The full prompt containing 2 to 5 solved few-shot demonstration pairs (Input: ... Output: ...) followed by the unlabelled Test Input: grid.
• reasoning (string): A concise, formal one-sentence description of the underlying transformation rule (e.g., "Rule: Keep only the largest object, cut out as a rectangle"). About 47.2% of the dataset includes this rule, while ~52.8% leaves this field empty to train direct-answer latent inference.
• answer (string): The ground-truth 2D target grid corresponding to the Test Input.
• family (string): The primary rule family (one of 37 categories).
• n_train (int64): Number of demonstration pairs provided in the context (ranges from 2 to 5).
• tokens_est (int64): Estimated token count under the Zenyx Byte-Pair Encoding (BPE) tokenizer.

Grid Representation

Grids are represented as 2D numeric matrices using ASCII digit strings:

• Digit 0 represents the neutral background.
• Digits 1 through 9 represent distinct foreground object colors (randomly permuted per task to prevent color bias).

---

The 37 Transformation Families

The tasks span fundamental priors of human intelligence and visual cognition:

1. Topological & Object Primitives:

   • crop_largest: Identify connected components and crop the bounding box of the largest object.
   • crop_bbox & crop_color: Extract objects by bounding box or specific color filters.
   • count_objects: Abstract count of discrete topological clusters.
   • odd_one: Identify and filter outlier components differing in size, shape, or color.
   • denoise: Distinguish structural patterns from background noise artifacts.

2. Physics & Cellular Dynamics:

   • gravity: Simulated vertical falling of non-zero pixels down each column.
   • move & roll: Directional spatial translation and cyclic wraparound.
   • connect: Draw lines connecting isolated anchor points of matching colors.
   • rays: Project orthogonal/diagonal light rays from emitter points.
   • floodfill & fill_holes: Enclosed region detection and flood filling.

3. Geometry, Symmetry & Morphology:

   • symmetry: Reflectional and rotational symmetry completion.
   • mirror_concat: Concatenate reflections across horizontal/vertical axes.
   • upscale & downscale: Nearest-neighbor block expansion or pooling.
   • tile: Periodic lattice replication and tessellation.
   • border & interior: Outer perimeter extraction vs internal cavity isolation.
   • checker & diag: Checkerboard interleaving and diagonal matrix traversals.
   • cross & stamp: Motif stamping and orthogonal cross projections.

4. Algebraic, Set & Color Mappings:

   • bool: Pixel-wise boolean logic (AND, OR, XOR) across sub-grids.
   • colormap & recolor_all: Permutational color remapping and canonical recoloring.
   • recolor_size: Color objects as a function of their component pixel area.
   • swap: Swap color pairs or sub-matrix blocks.
   • histogram & sort_bars: Bar chart generation and sorting by object frequency.
   • most_freq & parity: Statistical color dominance and parity-based decision rules.
   • grid_summary: Dimensionality reduction into abstract categorical summaries.

---

Rigorous Quality & Decontamination Audit

1. Correctness by Construction: All 37 transform algorithms pass programmatic self-testing (scripts/gen_arc_syn.py --selftest). Every answer grid is verified against its input transformation.
2. Strict ARC-AGI Decontamination: The generator audited every synthesized task against all 800 public ARC-AGI-1 tasks (both training and evaluation sets). Any task that shared an input-output pair with an official ARC task (under any color permutation) was discarded (902 overlapping candidates dropped). As a result, the dataset teaches generalizable abstract spatial priors without leaking benchmark evaluation tasks.
3. Data Integrity: Verified zero null entries, zero corrupted lines, and 100% syntactically valid integer grids.

---

Intended Use

• Pre-training & Continued Pre-training: Ingesting spatial grid structures teaches models to allocate attention heads to 2D coordinates and coordinate place values.
• Supervised Fine-Tuning (SFT): Teaching models to generate explicit reasoning scratchpads (<think> Rule: ... </think>) before outputting spatial structures.
• Inductive Logic Research: Serving as a high-throughput training bed for ARC-AGI research, neural program synthesis, and test-time compute scaling.

---

How to Use

Loading via Hugging Face datasets

from datasets import load_dataset

# Load train and validation splits
dataset = load_dataset("Arko007/zenyx-arc-synth")

print(dataset["train"][0])

Example Record

sample = dataset["validation"][0]

print("=== FAMILY ===")
print(sample["family"])

print("=== INSTRUCTION ===")
print(sample["instruction"])

print("=== REASONING ===")
print(sample["reasoning"])

print("=== ANSWER ===")
print(sample["answer"])

---

Citation & License

This dataset is released under the Apache 2.0 License.

@dataset{zenyx_arc_synth_2026,
  author       = {Arko and Zenyx AI Contributors},
  title        = {Zenyx ARC-Synth: Programmatically Synthesized Spatial & Abstract Reasoning Grids},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {https://huggingface.co/datasets/Arko007/zenyx-arc-synth}
}