README.md

---
extra_gated_prompt: >-
  This is a re-packaging of the GAIA validation split, redistributed under
  upstream's requirement that it live only in a gated or private repository. By
  requesting access you agree to the same terms as
  https://huggingface.co/datasets/gaia-benchmark/GAIA — in particular, do not
  reshare this data in a crawlable format.
extra_gated_fields:
  I agree to not reshare this dataset outside of a gated or private repository: checkbox
task_categories:
- question-answering
language:
- en
tags:
- agents
- agentic-benchmark
- evaluation
- gaia
- tool-use
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*.parquet
  - split: no_multimodal
    path: data/no_multimodal-*.parquet
  - split: smoke
    path: data/smoke-*.parquet
---

# GAIA validation split (nearai-bench packaging)

A **flat, self-contained repackaging** of the
[GAIA](https://arxiv.org/abs/2311.12983) validation split — 165 questions
**with ground-truth answers**, the only GAIA split whose answers are public
(the test split's are withheld for the upstream leaderboard).

Questions, answers, levels and attachments are **unmodified** upstream content.

## 🔒 Gated — and it has to stay that way

Upstream's terms say: *"you agree to not reshare this dataset outside of a gated
or private repository on the HF hub"* and *"do not reshare the validation or
test set in a crawlable format."* This mirror is gated for exactly that reason.
**Do not make it public.** If you need wider access, point people at
[gaia-benchmark/GAIA](https://huggingface.co/datasets/gaia-benchmark/GAIA) and
let them accept the terms themselves.

Gating also protects the benchmark: GAIA is a live leaderboard, and crawlable
validation answers are how it stops measuring anything.

## Why this exists

Upstream ships per-split `metadata.parquet` alongside a flat directory of
attachment files, so a harness has to resolve `file_name` against a second
download. Here each question is **one row**, with its attachment inline as a
deterministic base64 `tar.gz`, plus a **ready-to-send `prompt`** — GAIA's own
reference system prompt already prepended, matching
[nearai-bench](https://github.com/nearai/benchmarks)'s adapter byte-for-byte.

```python
from datasets import load_dataset
ds = load_dataset("NEAR-AI/gaia", split="validation")
```

## Splits

| Split | Rows | What it is |
|---|---|---|
| `validation` | 165 | The full public-answer split |
| `no_multimodal` | 152 | Text-only subset — drops questions needing image/audio/video understanding |
| `smoke` | 8 | Tiny subset for pipeline checks |

`no_multimodal` and `smoke` are strict subsets of `validation`, and match
`suites/gaia-no-multimodal.toml` / `suites/gaia-smoke.toml` upstream in
nearai-bench.

## Columns

| Column | Type | Notes |
|---|---|---|
| `task_id` | string | GAIA task UUID |
| `question` | string | Verbatim upstream `Question` |
| `final_answer` | string | **Ground truth** (upstream `Final answer`) |
| `level` | int64 | `1` \| `2` \| `3` — GAIA difficulty tier |
| `file_name` | string | Attachment filename, `""` when the question has none |
| `file_sha256` | string | Checksum of the attachment bytes |
| `assets_tar` | string | base64(tar.gz) of the attachment; `""` when none. 38 of 165 rows carry one |
| `prompt` | string | **Ready to send**: `system_prompt` + question (+ a note about the attachment when present) |
| `system_prompt` | string | GAIA's reference system prompt from the paper |
| `annotator_steps` | string | Human annotator's solution walkthrough |
| `annotator_num_steps` | string | Step count |
| `annotator_tools` | string | Tools the annotator needed |
| `annotator_num_tools` | string | Tool count |
| `annotator_time` | string | Wall-clock the annotator took |

## Running a question

```python
import base64, io, tarfile, pathlib, tempfile

row = ds[0]
ws = pathlib.Path(tempfile.mkdtemp())
if row["assets_tar"]:
    blob = base64.b64decode(row["assets_tar"])
    with tarfile.open(fileobj=io.BytesIO(blob), mode="r:gz") as t:
        t.extractall(ws)          # lands at ws/<file_name>

answer = my_agent(row["prompt"], cwd=ws)   # prompt already carries the system preamble
```

## Scoring

Extract the `FINAL ANSWER:` line from the response (fall back to the whole
response if the model ignored the template), then apply upstream's
`question_scorer`: numeric equality for numbers, element-wise compare for
comma/semicolon-separated lists, normalized string equality otherwise.

Keep it **bug-for-bug identical to upstream** or your numbers stop being
comparable to published GAIA results. The known quirk worth preserving: a
ground truth like `"3,676"` is parsed as a **two-element list**, not the number
3676. A reference port lives in `src/scoring.rs::gaia_match` /
`gaia_extract_final_answer` in
[`nearai/benchmarks`](https://github.com/nearai/benchmarks).

GAIA is search-heavy — ~76% of questions require web browsing per the paper,
and ~30% need multi-modality — so scores mostly reflect whether your agent has a
working search/fetch path. Compare across agents only with that held constant.

## ⚠️ Contamination warning

This split has **public ground-truth answers**, and `annotator_steps` contains
full human solution walkthroughs. Do not train on it, and do not let it into a
crawlable location. Treat it as an eval holdout.

## Provenance & license

- **Upstream**: [gaia-benchmark/GAIA](https://huggingface.co/datasets/gaia-benchmark/GAIA)
  — Mialon et al., *"GAIA: a benchmark for General AI Assistants"*
  ([arXiv:2311.12983](https://arxiv.org/abs/2311.12983)). Built from
  `2023/validation/metadata.parquet` plus that split's attachment files.
- **This repackaging**: same terms as upstream, gated. Content unmodified; the
  additions are the assembled `prompt` column, `file_sha256`, and the base64
  container format.
- **Packaged by**: [NEAR AI](https://near.ai) for
  [nearai-bench](https://github.com/nearai/benchmarks) (adapter: PR #317).

```bibtex
@misc{mialon2023gaia,
  title={GAIA: a benchmark for General AI Assistants},
  author={Gr{\'e}goire Mialon and Cl{\'e}mentine Fourrier and Craig Swift and
          Thomas Wolf and Yann LeCun and Thomas Scialom},
  year={2023}, eprint={2311.12983}, archivePrefix={arXiv}
}
```