Datasetpublic

Nabidnur/arc-agi-3-github-collection

ARC-AGI-3 GitHub Collection — every useful repo, mirrored & organized Mirrored 2026-09-29 · 45 repositories covering the full public ARC-AGI-3 landscape, discovered via GitHub API search (6 query buckets, 152 unique hits → 45 selected for direct competition relevance). Companion Space: Nabidnur/arc-agi-3-command-center · Models index: Nabidnur/arc-agi-3-models-and-analysis · Datasets: Nabidnur/arc-agi-3-datasets-combined Layout repos/<owner>__<name>/ — full… See the full description on the dataset page: https://huggingface.co/datasets/Nabidnur/arc-agi-3-github-collection.

sourceHugging Faceupdated 9d agoView on Hugging Face

0likes181downloads

Dataset Card

ARC-AGI-3 GitHub Collection — every useful repo, mirrored & organized

Mirrored 2026-09-29 · 45 repositories covering the full public ARC-AGI-3 landscape, discovered via GitHub API search (6 query buckets, 152 unique hits → 45 selected for direct competition relevance).

Companion Space: [Nabidnur/arc-agi-3-command-center](https://huggingface.co/spaces/Nabidnur/arc-agi-3-command-center) · Models index: Nabidnur/arc-agi-3-models-and-analysis · Datasets: Nabidnur/arc-agi-3-datasets-combined

Layout

—repos/<owner>__<name>/ — full source of each repo (.git stripped, exact snapshot 2026-09-29)
—repos/GITHUB-INDEX.md — annotated index: what each repo is, why it matters, star count
—provenance/github_discovery_ranked.json — raw ranked discovery dump (152 repos with stars/topics/queries)

The crown jewels

Path	Why
`repos/Tufalabs__duck-harness`	Milestone-1 winner (our base harness) + full example-run trace corpus (events.jsonl + transcripts + prompts for all 25 games × levels)
`repos/ryanbbrown__Retrodict`	183/183, 99.86% RHAE, $654 — log-as-context + plan-queue verification + playbook memory
`repos/NIMI-research__Tycho`	183/183, 100.00 RHAE — deterministic Moore-machine rendering
`repos/synchopate__arc-agi-crystalline`	180/183 — source-reading + ACT-R cognitive memory
`repos/Alexyskoutnev__TWIN-ARC-AGI-3`	179/183 — executable game twin, plan-in-twin, ExecuteChecked
`repos/alexisfox7__PRO-LONG`	+18pp, 4.2–5.8× token reduction — append-only searchable log
`repos/studio-dots-ai__TELL`	43.9% single-conversation experiential learning
`repos/NVIDIA__dream-team`	NVIDIA's multi-agent ARC-AGI-3 solver (AVO lineage)
`repos/feng-rrRay__Continual-Harness-ARC-AGI-3`	Continual skill library across games
`repos/arcprize__*`	Official: ARC-AGI-3-Agents, Kaggle-Starter, benchmarking, ARC-AGI-1/2 data, community leaderboard
`repos/Hisernberg__checkmpobbyacckraggle`	Complete Kaggle research archive: 237-submission history, leaderboard snapshots, kernel index

Licenses

Each mirrored repo keeps its original LICENSE file. All content © original authors — mirrored for research access with attribution; see per-repo provenance in repos/GITHUB-INDEX.md.

Milestone-2 Update (2026-10-02)

ARC-AGI-3 Milestone 2 = leaderboard snapshot at Sept 30, 2026 23:59 UTC (open-source notebooks). Podium: dfranzen 27.89 · lordhansolo 23.84 · sirikilohit/rellik13 22.53 · richardcsaky 20.00. All four run the same lineage: Tufa Duck/TAAF harness + Qwen3.8-Flash-Next (Intel W4A16 AutoRound or mixed NVFP4/FP8) + MTP speculative decoding + FP8 E4M3 KV cache (1.0–1.4M-token pools) + 45–57k rolling history with blockwise hysteresis trims + score-aware scheduling.

The one ablation that matters: FP8 KV + longer history alone took rellik13 14.49 → 22.53 with zero prompt changes; KV pool was 91–97% full — memory, not decode TPS, was the bottleneck.

Full deconstruction incl. the 4→27 thought process, per-notebook serving configs, LB scoring factors (level = min(115, (baseline/your_actions)²×100)), top-notebook gap analysis, and the 30–40+ fusion plan: see docs/27-milestone2-analysis.md in arc-agi-3-models-and-analysis and this repo's milestone2/ folder.