arc-agi2-program-grid-sft

ARC: verified programs and grids

1,915,399 published training episodes from 239,520 fresh base episodes. Build status: constructing_episodes. Target: 1,120,000 fresh episodes, each expanded into up to eight distinct rotations/reflections. These counts describe the files in this build, not the earlier 24-row format preview.

Read an episode

Open the image viewer · Interactive gallery (the gallery currently displays the earlier corpus).

Select a task family in the viewer's subset menu. Each family contains its own combined episode records; the selected family's row count is not the total corpus count above. Family configurations limit conversion size because Hugging Face may truncate non-Parquet previews beyond 5 GB per configuration. They do not split programs from grid targets.

Each row displays a full-resolution, lossless PNG, readable Python program, output-grid JSON, prompt, and combined completion. Authoritative PNG/JSON pairs are grouped into downloadable TAR shards. Compressed Arrow support indexes embed exactly the same PNG bytes and records for the viewer, avoiding millions of individual repository files and failing archive image paths. Episode training shards are PNG/JSON TARs with Arrow viewer indexes, not Parquet. Separate internal reusable pair pools may use Parquet; they are not episode viewer inputs or additional training episodes. Hugging Face may create its own display cache. Source and index row counts are identical and are counted only once.

The PNG shows the expected query answer for human inspection. Do not feed it to the model. Use the record's prompt as input and completion as the target. The completion contains a Python program followed by the grid JSON.

Verification and augmentation

All demonstration and query grids must match the program exactly. Every distinct D4 variant must pass before its base episode is accepted; failed variants are not silently dropped. Programs operate in the visible grid coordinates, without undo-transform / native-program / redo-transform wrappers.

The same rotation/reflection and colour permutation apply to every pair within an episode. Symmetric duplicate variants are removed and counted. Shortfalls and low-diversity families are reported rather than filled with duplicate episodes.

Build manifest · Per-family reports · Exact size histograms · Allocation

The size allocation targets the measured query-size and input/output-shape distribution of the 1,120 public tasks. Demonstration selection also prefers each family's original uniform/mixed input dimensions and whether the query size was seen among the demonstrations. Required semantic contexts take priority; unsupported size patterns and deviations are recorded per family. Published measurements, not the target distribution alone, show whether a completed build achieves that match.

Approved teaching corrections are labelled in each record. Canonical benchmark files remain unchanged. Public evaluation families included here become training data. Model training is not launched by this dataset builder.

Upstream ARC-GEN: Google LLC, Apache-2.0. Reviewed additions and dataset pipeline: Ritwika Kancharla. This is an independent research dataset.