zhangdw/GAIA2-mini

Download   ·   ARE ↗   ·   Paper ↗

GAIA2-mini packages the official GAIA2 mini / validation subset with its published filesystem resources for native ARE evaluation. 160 scenarios, 32 per capability: Execution · Search · Adaptability · Time · Ambiguity.

🗂️ What's inside

GAIA2-mini/
├── data/validation-00000-of-00001.parquet # 160 native ARE scenarios
├── filesystem/demo_filesystem/            # 294 original resource files
├── tasks.jsonl                            # Row index, IDs and categories
├── data-manifest.json                     # File sizes and SHA-256 hashes
├── provenance.json                        # Source revisions and validation
├── known-missing-files.json               # Missing upstream file references
├── upstream/                              # Original dataset cards
├── assets/                                # Dataset card artwork
├── README.md                              # Overview and usage
└── NOTICE.md                              # Attribution, licenses and citation

📦 Get the data

With uv installed, download the complete ~422 MB bundle:

uvx --from huggingface_hub hf download zhangdw/GAIA2-mini --type dataset --local-dir GAIA2-mini

The scenario table is in data/; the 294 resource files are in filesystem/demo_filesystem/. Keep the resource paths intact.

Python access & native ARE

from datasets import load_dataset

tasks = load_dataset("zhangdw/GAIA2-mini", "mini", split="validation")

Each row's data field is a native ARE scenario JSON string; tasks.jsonl indexes the original row order. To use the downloaded resources, set DEMO_FS_PATH to the absolute path of GAIA2-mini/filesystem/demo_filesystem before starting ARE. Install ARE and configure your agent and judge models.

---

Sources: Meta's GAIA2 and filesystem. Revisions & validation · Checksums · Licenses & citation.

All 160 scenarios passed ARE 1.1.0 loading and oracle preprocessing. The resource gap report lists four missing upstream file references. The upstream authors request that this data be used for evaluation.

Related project · Bara

Bara — Better ARE ReAct Agent collects ARE ReAct source code for studying recursive self-improvement (RSI) on GAIA2. Its five research reimplementations draw on public code and paper appendices from AutoSaddler, ActiveSaddler, ParetoSaddler (two variants) and HarnessMM. Each agent has a source directory you can install, modify and evaluate with the bara CLI.

Browse sources & papers → · Evaluation guide →