datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Nemotron-SFT-ARC-AGI-v1
Dataset Description:
Nemotron-SFT-ARC-AGI-v1 is a supervised fine-tuning (SFT) dataset of multi-turn agentic reasoning traces produced by open-weight large language models attempting to solve ARC-AGI visual-reasoning puzzles. Each ARC puzzle (a set of (input grid, output grid) demonstration pairs plus one or more test inputs, where grids are 2D integer arrays representing colors) is formatted as a text prompt and given to an agent powered by one of nine open-weight reasoning… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-SFT-ARC-AGI-v1.arc-agi-2arc-agi3-kimi-k2.7-ar25
ARC-AGI-3 ar25 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game ar25, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-ar25.arc-agi3-agy-gemini3.1pro-tr87
ARC-AGI-3 tr87 — Agent Trajectories (agy-gemini3.1pro)
Gameplay trajectories from the harness×model pair agy-gemini3.1pro playing the
ARC-AGI-3 game tr87, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-agy-gemini3.1pro-tr87.arc-agi3-agy-gemini3.1pro-g50t
ARC-AGI-3 g50t — Agent Trajectories (agy-gemini3.1pro)
Gameplay trajectories from the harness×model pair agy-gemini3.1pro playing the
ARC-AGI-3 game g50t, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-agy-gemini3.1pro-g50t.Nemotron-RL-ARC-AGI-v1
Dataset Description:
Nemotron-RL-ARC-AGI-v1 is a reinforcement-learning (RL) gym environment dataset of single-step ARC-AGI puzzle prompts intended for RL post-training of large language models. Each row corresponds to one ARC puzzle (a set of (input grid, output grid) demonstration pairs plus a single test input grid) rendered as a text prompt; reward is binary (1.0 / 0.0) determined by exact-match comparison against the ground-truth output grid. No LLM judge is used, no… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-ARC-AGI-v1.arc-agi3-agy-gemini3.1pro-su15
ARC-AGI-3 su15 — Agent Trajectories (agy-gemini3.1pro)
Gameplay trajectories from the harness×model pair agy-gemini3.1pro playing the
ARC-AGI-3 game su15, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-agy-gemini3.1pro-su15.kepler-arc-agi-3-traces
Kepler 1.0 ARC-AGI-3 trace corpus
Run artifacts from Kepler 1.0, an open-source agent harness for the 25 public
ARC-AGI-3 games. A stock CLI coding agent
encodes its theory of each game as an executable world_model.py, certifies it
against the full recorded interaction history, plans inside the certified
model, and acts through a guarded channel that voids the plan on the first
misprediction.
Project page ·
Code ·
Paper ·
Integrity record
The canonical release contains two… See the full description on the dataset page: https://huggingface.co/datasets/cveinnt/kepler-arc-agi-3-traces.arc-agi-labelledARC-AGI-v1arc-agi-2-grids
ARC-AGI-2 Grids — training + analysis corpus (NVARC-compatible)
Companion dataset for the Kaggle ARC Prize 2026 (ARC-AGI-2) solver built on
sorokin/qwen3_4b_grids15_sft139 + per-task rank-256 LoRA (NVARC lineage).
Everything here is generated from public canonical data only (1,000
training / 120 evaluation tasks); no hidden competition data is included.
Contents
Path
Rows
Description
train/train_tasks.jsonl
1,000
canonical training tasks (full I/O)… See the full description on the dataset page: https://huggingface.co/datasets/Nabidnur/arc-agi-2-grids.arc_agi_1_augmentedRPS-ARC-AGI-1-and-2This is the DPO dataset used in the RPS paper ( https://github.com/iamjasonfeng/RPS-Paper )
This dataset is based on the following dataset from Trelis: https://huggingface.co/datasets/Trelis/arc-agi-2-reasoning-5
r1-trajectories-arcagi-barcarc-agi-1-zloopvit-traces
ARC-AGI-1 ZLoopViT Traces
Canonical intermediate grid trajectories for all 400 training tasks in
ARC-AGI-1. Each row corresponds to one original official train or test pair;
generated augmentations are not included.
Dataset contents
400 tasks
1,718 trajectories: 1,302 demonstration/train pairs and 416 test pairs
5,059 visible intermediate transitions
Exact final-output validation on every official pair
Important columns:
task_id: official ARC task identifier… See the full description on the dataset page: https://huggingface.co/datasets/omrisap/arc-agi-1-zloopvit-traces.arc-agi-2-cot-sft
ARC-AGI-2 CoT-Solving SFT Dataset
Companion to Nabidnur/arc-agi-2-grids.
Grid-based chain-of-thought transcripts for the ARC-AGI-2 Kaggle competition,
NVARC format compatible (Qwen chat template, {" "}-separated digit rows).
Tiers
program-verified — a transformation rule fitted ONLY on the demonstrations
reproduces all of them exactly (verified=true). Hypothesis + per-pair
verification + application.
program-true (synthetic) — rules from generator ancestry… See the full description on the dataset page: https://huggingface.co/datasets/Nabidnur/arc-agi-2-cot-sft.arc-agi-2-synthetic-v1
ARC-AGI-2 Synthetic Curriculum v1
Program-generated ARC-like tasks with full ancestry records and
independent verification (every task's rule must be re-fittable from its
own demonstrations; 44 degenerate tasks rejected).
Families (7): d4_transform, color_remap, tile_k, crop_bbox, symmetry_fill,
hole_recolor, path_propagation.
Counts
family
verified
d4_transform
450
color_remap
450
tile_k
450
crop_bbox
450
symmetry_fill
439
hole_recolor
417… See the full description on the dataset page: https://huggingface.co/datasets/Nabidnur/arc-agi-2-synthetic-v1.regular-arcagi-3k-r1-020925arc-agi-1-ruleloopvit-rules
ARC-AGI-1 RuleLoopViT Rules
This dataset contains one canonical, task-specific English rule for each of the
400 official ARC-AGI-1 training tasks. Rules were inferred only from official
demonstration input/output pairs. Official test inputs, test outputs, and test
traces were excluded from rule authoring.
Each row includes:
a concise standalone core_rule_text;
a five-section full_rule_text;
the corresponding structured sections;
augmentation-aware references for colors… See the full description on the dataset page: https://huggingface.co/datasets/omrisap/arc-agi-1-ruleloopvit-rules.ARC-AGI-v1-5050arc_agi_2_augmentedr1-masked-arcagi-v0r1-masked-arcagi-v1