GAIA — license & access-terms note
===================================

This directory does NOT vendor any real GAIA data. GAIA is an ACCESS-GATED
dataset and its terms/license are accepted on HuggingFace at ingest time by the
person who fetches it; agenttic does not redistribute it.

Upstream
--------
- Dataset:  gaia-benchmark/GAIA
- URL:      https://huggingface.co/datasets/gaia-benchmark/GAIA
- Authors:  Mialon, Fourrier, Swift, Wolf, LeCun, Scialom (Meta AI / HuggingFace).
- Paper:    "GAIA: A Benchmark for General AI Assistants", arXiv:2311.12983 (2023).
- License:  The dataset is released by its authors under their stated terms on
            the HuggingFace dataset page (CC-BY-4.0 per the dataset card at the
            time of writing). It is, however, **access-gated**: you must be
            logged in to HuggingFace and **accept the dataset's terms/conditions**
            before the files can be downloaded. The held-out TEST split answers
            are NOT public; only the VALIDATION split ships with final answers.

What is in this directory
-------------------------
- LICENSE.txt              — this note.
- ATTRIBUTION.md           — source, citation, and what is / is not vendored.
- gaia_validation.sample.jsonl — a tiny, CLEARLY-LABELED, NON-OPERATIONAL
                            PLACEHOLDER sample. It is fabricated stand-in content
                            that mirrors only the GAIA validation record SCHEMA.
                            It is NOT real GAIA data.

agenttic implements GAIA's *methodology* (normalized exact-match answer accuracy
on the validation split). To score the real dataset you must bring your own
access: accept the terms upstream and authenticate (see ATTRIBUTION.md).
