By embedding robust human-intuitive analogies into ARC-style tasks, GIFARC guides AI agents to evaluate the task analogically before engaging in brute-force pattern search, thus efficiently reducing problem complexity and build a more concise and human-understandable solution.
- 1,614 ARC style puzzles made from GIF with analogy.
- Pair‑wise ground‑truth mappings + rich textual rationales for supervised or in‑context use.
- Easy Play generation pipeline - extend or remix new analogy families with gif in a few minutes.
- Review artifacts and generated examples are distributed through the anonymous supplementary material during double-blind review.
We highly command to using docker. To setting with docker check SETUP.md.
git clone <GIT_url>
cd gifarc
pip install -r requirements.txt
pip install -r requirements-dev.txt During double-blind review, use the dataset bundle from the anonymous supplementary material. Place downloaded JSONL files under a local artifact directory and keep large raw outputs out of Git. This anonymous review artifact supports inspection of the processed GIFARC dataset and running the generation pipeline on user-provided GIF files. It does not include the original GIF files, raw GIPHY crawl metadata, or source GIF URLs used in the internal data collection stage. Therefore, it is not intended to exactly reproduce the original GIPHY crawling stage during double-blind review.
Place your own GIF files under data/GIF/, then open description_executor.ipynb and run the pipeline cells.
| Split | #Tasks | #Unique GIFs | Size |
|---|---|---|---|
| Train | 1,614 | 1,614 | < 100 MB |
Every task packages looks as follows:
{
"source": "<source code>", # python code string
"examples": [
[<input_grid_1>,<output_grid_1>], # pair 1
[<input_grid_2>,<output_grid_2>], # pair 2
...
],
"seeds": [
"<file_name_1>",
"<file_name_2>",
...,
"<file_name_N>",
"<Concept_and_description>"
],
"url": "<minified_url>"
}
See the anonymous supplementary material for licensing, intended use, and data statements during review.
- Modular & Easy generation – Place review-safe GIF inputs under
data/GIF/, then rundescription_executor.ipynbto generate local outputs. - Stable environment setting enable easy set up with docker and devcontainer.
- All intermediate artifacts are cached for reproducibility.
Detailed instructions live in GENERATION.md.
## Project Structure
./GIFARC
├── data
│ └── GIF
├── description_executor.ipynb # use this to execute
├── docker-compose.yml
├── docs
│ ├── ANONYMIZATION.md
│ ├── GENERATION.md
│ ├── SETUP.md
│ └── THIRD_PARTY_NOTICES.md
├── README.md
├── requirements-dev.txt
├── requirements.txt
└── src
├── execution.py
├── experiments.py
├── generate_descriptions.py
├── generate_problems.py
├── generate_visualization_html.py
├── GIFARC_data_batch
├── GIFARC_utils
├── misc
├── parse_batch_description_samples.py
├── prompts
├── seeds
├── utility
└── visualize_problems.py
Generated outputs such as results/, experiment_results/, loggings/, and caches are intentionally ignored.
@misc{anonymous2026gifarc,
title={GIFARC: A GIF-Derived Resource for Analogy-Guided Task Framing},
author={Anonymous Authors},
note={Under review},
year={2026}
}- GIPHY for powering the GIF search API.
- BARC – our generation pipeline stands on the shoulders of this excellent project.
- Some seed-program filenames preserve upstream identifiers for reproducibility; see Third-Party Notices.
- GIFARC wouldn’t be possible without the open‑source community and our amazing reviewers.
Distributed under the MIT License. See LICENSE.