datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
arc-agi-3-schema-traces
ARC-AGI-3 Schema Gameplay Trajectories
This release contains 50 ARC-AGI-3 gameplay trajectories and a dependency-free
scoring utility. The trajectories are split evenly across two collections:
gpt_5_6_sol/: 25 GPT-5.6 Sol trajectories.
claude_fable_opus/: 25 trajectories from Claude Opus 4.8 and Claude Fable 5.
Each trajectory directory includes run.json, a streamed events.jsonl event
log, sanitized session data, snapshots, and the shareable text/image files
produced during… See the full description on the dataset page: https://huggingface.co/datasets/schema-harness/arc-agi-3-schema-traces.arc-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.arc_agi_3_public_demo_human_testing
Dataset Card for ARC-AGI 3 Public Demo Human Testing
Dataset Summary
This dataset contains human gameplay logs and trajectories from the ARC-AGI 3 public demo. It is a fully open-source dataset created by the ARC Prize.
The primary purpose of publishing this dataset on Hugging Face is to make it easily accessible and convenient for participants in the Kaggle ARC Prize 2026 Competition.
The implementation and source code used to process and upload this dataset to… See the full description on the dataset page: https://huggingface.co/datasets/magic-sword/arc_agi_3_public_demo_human_testing.arc-agi-3-schema-traces
ARC-AGI-3 Schema Gameplay Trajectories
This release contains 50 ARC-AGI-3 gameplay trajectories and a dependency-free
scoring utility. The trajectories are split evenly across two collections:
gpt_5_6_sol/: 25 GPT-5.6 Sol trajectories.
claude_fable_opus/: 25 trajectories from Claude Opus 4.8 and Claude Fable 5.
Each trajectory directory includes run.json, a streamed events.jsonl event
log, sanitized session data, snapshots, and the shareable text/image files
produced during… See the full description on the dataset page: https://huggingface.co/datasets/JBrightmanAI/arc-agi-3-schema-traces.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.arc-agi-3-wm-traces
ARC-AGI-3 World Model Traces
This dataset contains ARC-AGI-3 transition traces in the same parquet schema used by HHazard/arc-agi-3.
Each row is one environment transition:
state, game_id, level_id, action_id, action_args, next_state,
level_done, frame_idx, origin, transformation, player
state and next_state are 64x64 ARC grids stored as nested integer arrays. action_id is the ARC-AGI-3 action kind; click actions use action_args.x and action_args.y.
Splits… See the full description on the dataset page: https://huggingface.co/datasets/fredericowieser/arc-agi-3-wm-traces.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-impabs-dpolr1e-7-beta0.01-classifiersft5e-7arc-agi-3-teaching-suite
ARC-AGI-3 teaching suite
100 grid-reasoning tasks across 16 puzzle families, with two splits:
split
rows
answers in this dataset
why
taught
80
yes
16 families x 5 worked examples, each documented in INSTRUCTIONS.md and reasoned through in CHAIN_OF_THOUGHT.md
heldout
20
no
16 unseen instances of the same families, plus 4 tasks that are provably not solvable by reasoning from their examples
The teaching material is what makes this a teaching suite rather than a… See the full description on the dataset page: https://huggingface.co/datasets/dnhkng/arc-agi-3-teaching-suite.arc-agi3-phase2-ar25
ARC-AGI-3 ar25 — ARA snapshots truncated by absolute inference-cost budget
Phase-2 dataset. 6 agents (harness×model) each played ar25, continuously crystallizing a structured world model (Agent-Native Research Artifact). Here each agent's ARA is sliced at fixed absolute cumulative inference-cost budgets — $13/$27/$51/$76 — so you can ask "for a spend of $B, what world model has each agent built?"
Structure
cost_05usd/ cost_15usd/ cost_40usd/ cost_75usd/
<agent>/… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-phase2-ar25.arc-agi-3arc-agi-synthetic
ARC-AGI Synthetic
A dataset of 4,940 self-defining ARC-AGI-style reasoning tasks, authored
as compressed Hy programs over a frozen DSL (library.hy) and materialized into
raw input/output grids. Each task is a single transformation rule expressed as a
constructive program for the inputs plus a structural to-output that derives
every output from its input — so the rule is guaranteed self-consistent across
all of the task's pairs.
This corpus was generated by an autonomous LLM… See the full description on the dataset page: https://huggingface.co/datasets/Fraser/arc-agi-synthetic.arc_agi_2_human_testing
ARC-AGI-2 Human testing data
This file contains data from human testing sessions on ARC-AGI tasks.
Each row represents a single test attempt by a human participant on a specific task-test pair in the "Public Train" or "Public Eval" ARC-AGI-2 datasets. Not all tasks in the released "Public Train"
sets were tested, so these results are not comprehensive. This data does not include tasks from "Semi Private Evaluation" or "Private Evaluation"
Column Descriptions… See the full description on the dataset page: https://huggingface.co/datasets/arcprize/arc_agi_2_human_testing.arc-agi-3-schema-traces-gpt56
ARC-AGI-3 Schema Gameplay Trajectories — GPT-5.6 Sol
This release contains every gpt-5.6-sol gameplay trajectory produced on our
cluster with the world_model_v5 agent harness — 100 runs across the 25 public
ARC-AGI-3 games — plus a dependency-free scoring utility.
It is the GPT-5.6 Sol member of a family built by the same harness and the same
sanitizer, so trajectories can be compared game by game:
arc-agi-3-schema-traces-fable5 — Claude Fable 5, best per game (25)… See the full description on the dataset page: https://huggingface.co/datasets/guanning/arc-agi-3-schema-traces-gpt56.big-arc-agi-3arc-agi-3-schema-traces-opus48
ARC-AGI-3 Schema Gameplay Trajectories — Claude Opus 4.8
This release contains the best claude-opus-4-8 / max trajectory for each of
the 25 public ARC-AGI-3 games, plus a dependency-free scoring utility. It is the
Opus 4.8 counterpart of
arc-agi-3-schema-traces-fable5,
produced by the same agent harness (world_model_v5) and the same sanitizer, so
the two collections can be compared game by game.
Each trajectory directory includes run.json, a streamed events.jsonl event
log… See the full description on the dataset page: https://huggingface.co/datasets/guanning/arc-agi-3-schema-traces-opus48.arc-agi-mixed-max4096-impabs-dpo-lr1e-7-beta0.1-16samp-rerun-flat-curr-validNemotron-RL-ARC-AGI-v1-prompt-only
Nemotron-RL-ARC-AGI-v1-prompt-only
Prompt-only extraction from nvidia/Nemotron-RL-ARC-AGI-v1.
Files:
prompts.csv: one prompt extraction record per source row. Records include
prompt, separated system_prompt, and structured tools when the source row
defines available tools. Nested values are JSON-encoded inside CSV cells.
summary.md: source row counts, extracted row counts, count deltas, and failed prompt counts.
null_or_empty_rows.md: row indexes where prompt extraction… See the full description on the dataset page: https://huggingface.co/datasets/jamesdborin/Nemotron-RL-ARC-AGI-v1-prompt-only.arc-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.Nemotron-SFT-ARC-AGI-v1-prompt-only
Nemotron-SFT-ARC-AGI-v1-prompt-only
Prompt-only extraction from nvidia/Nemotron-SFT-ARC-AGI-v1.
Files:
prompts.csv: one prompt extraction record per source row. Records include
prompt, separated system_prompt, and structured tools when the source row
defines available tools. Nested values are JSON-encoded inside CSV cells.
summary.md: source row counts, extracted row counts, count deltas, and failed prompt counts.
null_or_empty_rows.md: row indexes where prompt extraction… See the full description on the dataset page: https://huggingface.co/datasets/jamesdborin/Nemotron-SFT-ARC-AGI-v1-prompt-only.arc-agi-3-schema-traces-gpt56-xhigh
ARC-AGI-3 Schema Gameplay Trajectories — GPT-5.6 Sol (xhigh)
The best gpt-5.6-sol trajectory at xhigh reasoning effort for each of the
25 public ARC-AGI-3 games, produced with the world_model_v5 agent harness.
This release exists to make the cross-model comparison single-effort on all
sides. Its siblings are each one model at one effort, but the
gpt-5.6-sol collection in
arc-agi-3-schema-gameplay
is a mix of xhigh and max (16 games + 9 games), so it is not directly
comparable to… See the full description on the dataset page: https://huggingface.co/datasets/guanning/arc-agi-3-schema-traces-gpt56-xhigh.arc-agi-mixed-max4096-qwenabsgen-all-flat-train-Qwen3-4B-Thinking-2507-samp1-abs-all-validplan-crl-arc-agi-2
PLAN-CRL ARC-AGI-2
This is a Hugging Face mirror/packaging of the public ARC-AGI-2 tasks used by the PLAN-CRL ARC-AGI-2 evaluation harness.
Source: https://github.com/arcprize/ARC-AGI-2
Splits
training: 1000 tasks
evaluation: 120 tasks
Columns
task_id: upstream ARC task id.
train: ARC training examples with input and output grids.
test_inputs: test input grids only.
test_outputs: reference output grids used by the evaluator.
prompt_json: compact JSON shown… See the full description on the dataset page: https://huggingface.co/datasets/1337xyz1337xyz/plan-crl-arc-agi-2.arc-agi-mixed-max4096-fulltrain-absrespgen-witho4correctsolsarc-agi-2-vision
ARC-AGI-2 Vision Dataset
Image-rendered ARC-AGI-2 grids for vision-based solvers (e.g. Qwen-VL class
models). Companion to Nabidnur/arc-agi-2-grids
(text/grid SFT) and Nabidnur/arc-agi-2-cot-sft
(reasoning transcripts).
Render spec
Standard ARC 10-color palette (see meta/manifest.json).
White 4px margin, 1px dark grid lines, cell scale = max(8, 448//max_dim) px.
image column: PNG bytes (HF Image-compatible {bytes, path} struct).
Contents… See the full description on the dataset page: https://huggingface.co/datasets/Nabidnur/arc-agi-2-vision.arc_agi_1_augmentedarc-agi-1-kimi-k2arc-agi-1-gpt-5-high