# ARC Task Gen — Synthetic ARC-AGI Task Generator

> Generate original ARC-AGI-1-style visual reasoning tasks that are distribution-matched to the official evaluation set, useful for benchmarking and training.

## Install

Save in your project root:

# ARC Task Gen — Synthetic ARC-AGI Task Generator

## Quick Use
```bash
pip install arc-task-gen
arc-task-gen generate --count 100 --output tasks.json
```

## Introduction
ARC Task Gen is a Python tool by Pathway that generates original ARC-AGI-1-style visual reasoning tasks. Each generated task is distribution-matched to the public evaluation set, giving researchers a scalable way to create training and benchmarking data for abstract reasoning systems without manually designing puzzles.

## What ARC Task Gen Does
- Generates novel grid-based visual reasoning tasks matching the ARC-AGI-1 format
- Produces input-output example pairs with verifiable transformation rules
- Ensures statistical alignment with the official ARC evaluation distribution
- Outputs tasks in the standard ARC JSON schema for direct use with evaluation harnesses
- Supports configurable difficulty and transformation-type filters

## Architecture Overview
The generator uses a pipeline of composable transformation primitives — rotations, reflections, color mappings, object manipulations — that are sampled and chained according to a learned distribution model. A validation layer checks each generated task against solvability constraints and distribution metrics before emitting it. The system runs entirely on CPU and produces hundreds of tasks per minute.

## Self-Hosting & Configuration
- Install from PyPI or clone the repository and install with pip
- Configure generation parameters via CLI flags or a YAML config file
- Set a random seed for reproducible task batches
- Output format supports JSON and the ARC-AGI evaluation harness directory layout
- No external API keys or GPU required

## Key Features
- Distribution-matched output ensures generated tasks reflect real ARC difficulty curves
- Deterministic seeding for reproducible research experiments
- Extensible primitive library for adding custom transformation types
- Built-in deduplication against the official ARC public training and evaluation sets
- Lightweight with zero heavy dependencies

## Comparison with Similar Tools
- **Manual ARC authoring** — slow and labor-intensive; ARC Task Gen automates the process
- **LLM-based task generation** — can hallucinate invalid puzzles; ARC Task Gen validates every output
- **RE-ARC** — another procedural generator; ARC Task Gen focuses on tighter distribution matching
- **ARC-AGI official set** — fixed 800 tasks; ARC Task Gen provides unlimited new ones

## FAQ
**Q: Are the generated tasks guaranteed to be solvable?**
A: Yes. Each task passes a validation step that confirms at least one consistent transformation rule.

**Q: Can I use generated tasks for model training?**
A: Yes. The output follows the standard ARC JSON format and is ready for training pipelines.

**Q: How closely do generated tasks match the official distribution?**
A: The generator is calibrated against the public eval set and reports distribution metrics per batch.

**Q: Does it require a GPU?**
A: No. Task generation is CPU-only and runs on any Python 3.10+ environment.

## Sources
- https://github.com/pathwaycom/arc-task-gen
- https://arcprize.org

---
Source: https://tokrepo.com/en/workflows/asset-06eeec94
Author: AI Open Source