README.md

metadata

language:
  - en
license: other
pretty_name: P-ARC (PotARCin Test Set) tabular export
configs:
  - config_name: default
    data_files:
      - split: test
        path: p_arc_dataset.csv
size_categories:
  - n<1K
tags:
  - arc
  - arc-agi
  - potarcin
  - program-synthesis
task_categories:
  - other

P-ARC CSV export (PotARCin Test2)

One CSV file in UTF-8. Each row is one of the fifty P-ARC tasks from PotARCin (t1.json through t50.json). Besides the usual train/test grids, each row includes the fifty-sample bundle from t<n>_samples_50.json as compact JSON (same structure as the file, without the extra whitespace from pretty-printing), plus the generator.py and verifier.py sources from the matching task folder.

Files

File	Description
p_arc_dataset.csv	One row per task (t1–t50); column details below and in SCHEMA.json
README.md	Dataset card (what you are reading now)
SCHEMA.json	Same column layout in JSON for scripts

Columns

Column	Description
task_id	t1 … t50
train_demonstrations_json	JSON array of training pairs {input, output}; grids are nested lists of integers
test_input_json	JSON grid for test[0].input
test_output_json	JSON grid for test[0].output
stable_instances_50_json	Compact JSON for the object in t<n>_samples_50.json
generator_py	Full generator.py source
verifier_py	Full verifier.py source

Grids follow the usual ARC convention: each row is a list of cell integers.

Size and reading the CSV

The file is fairly large because big JSON blobs and full Python files sit inside cells. In Python, the standard library csv module caps field length by default; bump it before you read:

csv.field_size_limit(sys.maxsize)

Stdlib example

import csv, json, sys

csv.field_size_limit(sys.maxsize)

with open("p_arc_dataset.csv", encoding="utf-8", newline="") as f:
    for row in csv.DictReader(f):
        train = json.loads(row["train_demonstrations_json"])
        test_in = json.loads(row["test_input_json"])
        test_out = json.loads(row["test_output_json"])
        samples = json.loads(row["stable_instances_50_json"])
        # row["generator_py"], row["verifier_py"]

pandas

import sys, pandas as pd
import csv as _csv

_csv.field_size_limit(sys.maxsize)
df = pd.read_csv("p_arc_dataset.csv", encoding="utf-8")