README.md

---
pretty_name: AetherSearch Search-R1 Full Evaluation Set
language:
- en
license: other
task_categories:
- question-answering
size_categories:
- 10K<n<100K
---

# AetherSearch Search-R1 Full Evaluation Set

This repository contains the complete Search-R1 `test.parquet` used by the
current AetherSearch full-data evaluation recipe. It is an exact, unmodified
mirror of the test file published in
[`PeterJinGo/nq_hotpotqa_train`](https://huggingface.co/datasets/PeterJinGo/nq_hotpotqa_train).

## Contents

| File | Rows | SHA-256 |
|---|---:|---|
| `test.parquet` | 51,713 | `30aa887b6d47e06e8c0f6f5307c88fe4e13461ac25a20ec0a5433ad7a4fe25dc` |

Source distribution:

| Data source | Rows |
|---|---:|
| 2WikiMultiHopQA | 12,576 |
| Bamboogle | 125 |
| HotpotQA | 7,405 |
| MuSiQue | 2,417 |
| Natural Questions | 3,610 |
| PopQA | 14,267 |
| TriviaQA | 11,313 |

The AetherSearch evaluation manifest preserves the original parquet row order
and covers every row. No sampled subset is used by the current public recipe.

## Provenance And Licensing

The file was obtained from the official Search-R1 dataset release. Search-R1
constructs the merged test parquet from upstream QA benchmarks, primarily via
`RUC-NLPIR/FlashRAG_datasets`. This mirror does not claim ownership or grant a
new license over those source datasets. Users must follow the terms of
Search-R1 and each underlying benchmark. The upstream Hugging Face repository
does not currently declare a single dataset license, so this card uses
`license: other` rather than asserting an unsupported license.

## Use With AetherSearch

Set the following local path after downloading the dataset:

```bash
export AETHERSEARCH_VALIDATION_DATA=/path/to/test.parquet
```

The public recipe verifies the file SHA-256, total row count, per-source
counts, and full row-identity manifest before evaluation.