hari31416/search-anything-benchmark

Search Anything benchmark

Project repository: Hari31416/search-anything. The repository contains the search pipeline, dataset preparation scripts, and evaluation runner.

A compact, frozen multimodal retrieval benchmark with 1,150 candidate assets and 350 text queries, sampled from seven existing datasets.

We created these subsets so we could index and evaluate our search pipeline quickly with a low storage footprint on a local machine. They support repeated regression checks and ranking experiments without downloading or indexing the full source datasets. They are evaluation subsets, not training data or replacements for the full benchmarks.

Contents

Task	Original dataset	Candidate assets	Development queries	Held-out queries
text	NanoSciFact	500 documents	30	20
image	MMEB MSCOCO_t2i	200 images	30	20
audio	Clotho evaluation	100 environmental-sound clips	30	20
video	MSR-VTT test_1k	50 videos	30	20
speech	MTEB Spoken SQuAD	100 synthesized speech clips	30	20
image_hard	COLA	100 images	30	20
natural_speech	SLUE-SQA-5 verified test	100 read-speech passages	30	20
Total		1,150	210	140

The original collections remain under search-benchmark/, spoken-squad-benchmark/, and expanded-benchmark/. Each contains its frozen manifest and media. The expansion uses the corrected version 2 query split. Its archived version 1 is excluded.

The package includes selected media, query text, relevance judgments, original source IDs, source revisions, crop metadata where available, and SHA-256 checksums. It excludes full source Parquet files, downloaded archives, query embeddings, generated transcripts, indexing state, credentials, and machine-specific results.

Runtime and storage

Initial indexing measurements on a local M4 Pro on October 10, 2026 were approximately 20.4 minutes for the original 850 assets, 3.0 minutes for Spoken SQuAD, and 5.8 minutes for the expansion. These include pilots and retries or model warmup as recorded by each collection. They depend on the pipeline and hardware. Existing model caches and container images are excluded.

package-summary.json records the exact media payload size. The uploaded package is smaller than the original preparation directories because it omits their cached source files. Speech duration is 22.4 minutes for Spoken SQuAD and 49.9 minutes for SLUE.

Usage

from datasets import load_dataset
from huggingface_hub import snapshot_download

repo = "hari31416/search-anything-benchmark"
corpus = load_dataset(repo, "corpus", split="corpus")
development = load_dataset(repo, "queries", split="dev")
held_out = load_dataset(repo, "queries", split="test")

# Media paths in corpus rows are relative to this downloaded directory.
root = snapshot_download(repo_id=repo, repo_type="dataset")

corpus has one row per asset, including id, collection, task, source_id, path, content_type, bytes, sha256, source_url, and optional duration_seconds. Media paths are strings, not automatically decoded media objects. Original manifests contain additional source-specific metadata.

Queries contain id, collection, task, text, split, and relevant, a list of relevant corpus IDs. COLA queries also have pair_id and hard_negative. Candidate IDs are unique across the package. Preserve them when computing metrics.

Search the full candidate pool for the query's task. The same candidate pool is shared by development and held-out queries. Use development queries for tuning and held-out queries to check a selected change. The existing Search Anything runner accepts each downloaded collection directory through --root; it creates a new local index state.

Evaluate Recall@5, Recall@10, MRR@10 and binary nDCG@10 per task. These are local subset scores, not official MMEB, MTEB, COLA or SLUE leaderboard results. Combining all tasks into one search pool requires additional relevance judgments.

Selection and limitations

Selection is deterministic and retains all labeled positives for selected queries. The original text, image and environmental-audio tasks add lexical neighbors and random distractors. Video selection spans source categories, with one query per selected video. Relevant assets do not cross development and held-out query splits.

Spoken SQuAD contains synthesized speech about Super Bowl 50. Source paragraphs supplying queries stay within one split, and related sentences provide distractors.

COLA retains 50 contrasting image pairs. Captions from 25 pairs supply 50 queries; the remaining pairs add distractors. Pairs stay within one query split. Labels identify the positive and paired negative but may omit other relevant images in the pool. This tests object/attribute composition, not dedicated OCR.

SLUE contains natural human readings of Wikipedia. Each of its five question sources, Natural Questions, SQuAD, CuratedTREC, TriviaQA and WebQuestions, contributes six development and four held-out queries. Positive audio passages are cropped around all labeled answer occurrences to at most 30 seconds and converted to 16-bit PCM WAV. Crop offsets and answer times remain in the original manifest. Questions come from distinct articles and speakers; the full candidate pool covers 98 articles and 98 speakers. This tests bounded passage retrieval, not full-recording localization or noisy conversation.

Reference captions and answers belong to evaluation metadata. Ingest only the media files and original text documents into the search pipeline. Do not ingest query text or relevance metadata. Opaque media filenames avoid caption leakage. Caption relevance judgments can be incomplete, and small score changes need per-query inspection.

Source references and terms

NanoSciFact's mirror declares CC BY 4.0. Clotho captions are for experimental and non-commercial use, and publications using them must acknowledge the Audio Research Group at Tampere University. The full copyright notice is in source-notices/clotho-LICENSE.txt. source-notices/clotho-selected-attributions.jsonl maps every packaged environmental-audio clip to its creator, original source URL, individual license and upstream excerpt offsets.

This compilation does not assign a single new license to its third-party content. Each source retains its original terms. Source notices are included under source-notices/; original dataset links and pinned revisions are recorded below and in the manifests. Check the original sources for citations, attribution, and permitted uses.

Subset	Original dataset and references	Source terms and provenance
NanoSciFact	NanoSciFact, SciFact	Original scientific document text and claim relevance labels; retain upstream notices.
MSCOCO_t2i	COCO, COCO terms, MMEB-V3	COCO annotations and individual images have distinct terms; images retain their original rights. MMEB supplies retrieval metadata.
Clotho	Clotho v2.1, Clotho paper, Freesound	Caption license is the upstream Tampere University non-commercial attribution license. Audio licenses and creator credits vary by Freesound clip; see the included metadata and license.
MSR-VTT	MSR-VTT, source mirror	Original video and annotation rights remain with their respective owners; the mirror does not grant a new blanket media license.
Spoken SQuAD	Original Spoken SQuAD, MTEB retrieval mirror, SQuAD	The MTEB mirror declares Apache-2.0; underlying speech derives from SQuAD text, whose source terms also apply.
COLA	COLA author repository, self-contained mirror, Visual Genome	COLA annotations are MIT; underlying Visual Genome images retain their original source terms.
SLUE-SQA-5	SLUE Phase 2, SLUE repository, Spoken Wikipedia	Spoken Wikipedia audio is CC BY-SA 4.0. Question terms depend on source: TriviaQA is Apache-2.0; the other question sources are CC BY-SA 4.0. Cropping and PCM conversion are the modifications in this subset.

Pinned source revisions:

VLM2Vec/MMEB-V3: 4a5560b2b64384204b6fea8a82ea986eba51f5aa
zeta-alpha-ai/NanoSciFact: 309f1d1ae3ae2e092444a8a0c25bed59b82318bc
VLM2Vec/MSR-VTT: 947b139d0c591235a3f311e45a31a88461bad0d2
Clotho: Zenodo record 4783391, evaluation split
mteb/spoken-squad-t2a: b311394fa5ab5e003012304799d0d724d64303b3
array/cola: daee86b15a60308cfcc277ea11ab7f6be60fdfbc
asapp/slue-phase-2: e2989c55a53593a8e39b8f8ebdb47ccaccbe484a