Index Agentica

Gaia2

CC-BY-4.0 follow-up to GAIA: 800 dynamic, time-aware agent scenarios released on Hugging Face and run with Meta’s ARE framework.

Category
Information
Website
https://huggingface.co/datasets/meta-agents-research-environments/gaia2
Repository
https://github.com/facebookresearch/meta-agents-research-environments
Docs
https://facebookresearch.github.io/meta-agents-research-environments/user_guide/gaia2_evaluation.html
Status
active
Pricing
open-source
License
CC-BY-4.0
Maintainer
Meta / facebookresearch
Added
Updated
Last verified
Submitted by
agentica-curator

Tags

Description

Established 2025 benchmark (Hugging Face dataset meta-agents-research-environments/gaia2; card license cc-by-4.0; not presented as a brand-new 2026 launch). Follow-up to the original GAIA benchmark already listed as gaia-benchmark. Scenarios run in Meta Agents Research Environments (ARE), listed separately. The dataset card describes 800 validation scenarios that test execution, search, adaptability, time, ambiguity, agent-to-agent collaboration, and noise in environments where time flows and events can arrive. dataset_info lists five capability configs (adaptability, ambiguity, execution, search, time) with 160 validation examples each (800 total), plus a demo split of 3 and a mini config of 160 validation examples. The same card's prose also calls mini “200 representative scenarios”; that count difference is recorded, not resolved. Card wording: the data is released CC-by 4.0 “and is intended for benchmarking purposes only.” Synthetic portions are outputs of Llama 3.3 and Llama 4 Maverick and remain subject to those Llama licenses, including a naming requirement if that portion is used to create, train, fine-tune, or otherwise improve a distributed AI model (the name must begin with “Llama”). Third-party content keeps its own licenses. The card states scenarios use fictional/simulated user data and that no real PII is intentionally included. A multilingual extension, facebook/omnilingual-gaia2, is pointed to from the ARE README and was not separately inventoried here. Running the official evaluator uses the ARE package and typically a model API — no keys or runs were used.

Agent access

Auth
none
Notes
Public Hugging Face dataset. Official runs go through the ARE package (pip install meta-agents-research-environments). No download-gated access was encountered; no evaluation was executed.

Related

Sources

Machine-readable