---
license: mit
pretty_name: ARC bad predictions
task_categories:
- image-to-image
size_categories:
- 1K<n<10K
language:
- en
---

# ARC Bad Predictions

Visualization of bad predictions: [ARC-AGI training](https://neoneye.github.io/simon-arc-lab-web/model/2024-oct-17-1318/arcagi_training/), [ARC-AGI evaluation](https://neoneye.github.io/simon-arc-lab-web/model/2024-oct-17-1318/arcagi_evaluation/).

My goal during the ARC-AGI contests has been to make a `stepwise refinement` algorithm, that can improve on earlier predictions.

This repo is intended for `stepwise refinement` algorithms. This dataset contains incorrect predictions that are somewhat close to the target.
I have manually inspected these predictions and removed the worst predictions. However there may still be more bad predictions.

A bad prediction, can be a hint towards solving the puzzle.

A bad prediction, can also be misleading.

The color values 0-9 are the same as the ARC-AGI colors. Color values outside this range sometimes occur, when a solver failed to predict that pixel.

The `dataset` and `task` correspond to a file found in the [arc-dataset-collection](https://github.com/neoneye/arc-dataset-collection) repo.
A few puzzles have multiple versions, these are suffixed with `_v2` and `_v3`.