ARC-like data

  

TL;DR: generate datasets that require reasoning via the composition of visually intuitive "programs" such as translation of blocks of pixels, or identifying block endpoints, or denoising.

Examples

Each plot has 10 input-output pairs lined up together.

Details

The neurallambda project (also on X) aims to develop generic architectures that support reasoning. This project requires datasets that demonstrate reasoning, and small toy problems help prove and iterate on the research quickly.

The original ARC Prize dataset contains visual puzzles, but as a benchmark only, therefore not enough data to train on, and some puzzles are prohibitively large.

The arc-like repo generates 1D puzzles constructed via composing simple combinator functions, eg:

puzzles = [
    # translate a block by 4 pixels
    ('translate', compose([gen_some_blocks(colors), translate(4)])),

    # identify the endpoints of blocks of pixels
    ('endpoints', compose([gen_some_blocks(colors), endpoints])),

    # compose `translate` and `endpoints`
    ('translate + endpoints', compose([gen_some_blocks(colors),  translate(4), endpoints]))
]

A "combinator" is very simply a function Sequence -> Sequence, which you can see chain together nicely:

@dataclass
class Sequence:
    ''' An input-output pairing, and metadata that might be used by downstream combinators. '''
    inputs: List[Any]
    outputs: List[Any]
    metadata: Any

Combinator = Callable[[Sequence], Sequence]

def translate(n: int) -> Combinator:
    """ Translate the sequence by n positions. """
    def f(seq: Sequence) -> Sequence:
        outputs = seq.outputs
        new_outputs = outputs[-n:] + outputs[:-n]
        return Sequence(seq.inputs, new_outputs, seq.metadata)
    return f

Everything's currently in puzzles.py, no dependencies!:

git clone https://github.com/neurallambda/arc-like
cd arc-like

# Run the demo (demo only depends on `torch` and `matplotlib`)
python arc_like/puzzles.py

# Use in your code
cd your_code
pip install -e path-to/arc-like
import arc_like.puzzles as puzzles

Please add a puzzle as a PR, or, just tell me what you want on X!

Gratitude

• arcticbio: Thank you for adding a ton of new puzzles, and cleaning up some logic!
• rybla: Thank you for adding composable combinators, this makes for a much cleaner, more extensible approach!