OmniRetriever-Bench

The first 12-direction audio-video-text retrieval benchmark.
3,782 held-out triples scored on a shared gallery across all 6 single-modal
and 6 dual-modal directions — with human-corrected captions.

📄 Paper · 🏠 Project page · 💻 Code & evaluator · 🤖 Model

---

Why this benchmark exists

Standard audio–text and video–text benchmarks evaluate only single-modal directions and never pool both video and audio in the same gallery. That leaves the central capability of a unified audio-video-text encoder — the joint (T, V, A) embedding it produces on every forward pass — untested at evaluation time.

OmniRetriever-Bench closes that gap. It scores all 12 single- and dual-modal retrieval directions on one shared 3,782-clip gallery:

            ┌──────────────── target ────────────────┐
            │  T       V       A      AV     TV    AT │
   T   ──── │  ─    t→v    t→a    t→av   ─     ─    │
   V   ──── │ v→t    ─     v→a    ─      ─     ─    │
query A ─── │ a→t   a→v    ─      ─     a→tv   ─    │
   AV  ──── │ av→t   ─     ─      ─      ─     ─    │
   TV  ──── │  ─     ─    tv→a    ─      ─     ─    │
   AT  ──── │  ─    at→v   ─      ─      ─     ─    │
            └─────────────────────────────────────────┘
            6 single-modal pairs + 6 dual-modal pairs = 12 directions

	T	V	A	AV	TV	AT
T  query	—	✓	✓	✓	—	—
V  query	✓	—	✓	—	—	—
A  query	✓	✓	—	—	✓	—
AV  query	✓	—	—	—	—	—
TV  query	—	—	✓	—	—	—
AT  query	—	✓	—	—	—	—

---

Sample clips

All captions are reviewed and corrected by trained human annotators starting from a Gemini 3.0 Pro draft — no machine-only captions, no unfiltered scrapes.

---

What the trained retriever learns

OmniRetriever-7B, the released companion model, is trained with fusion-as-teacher distillation (L_D) and Tuple-InfoNCE (L_T) on top of pairwise InfoNCE (L_A):

This benchmark is the stress test that exposes whether a unified AVT encoder actually uses its joint z_TVA embedding — or just treats it as a passive average of three pair geometries.

---

Schema

OmniRetriever-Bench.csv is a 4-column UTF-8 CSV with a header row and 3,782 data records:

Column	Type	Meaning
url	str	TikTok player URL of the source video
start_sec	float	clip start time within the source (seconds)
end_sec	float	clip end time within the source (seconds)
caption	str	human-validated dense caption for the (video, audio) clip

Quick peek

from datasets import load_dataset
ds = load_dataset("YunzeLiu/OmniRetriever-Bench", split="test")
print(ds)
print(ds[0])

Dataset({features: ['url', 'start_sec', 'end_sec', 'caption'], num_rows: 3782})

{'url': 'https://www.tiktok.com/player/v1/7515451532880530710',
 'start_sec': 30.0, 'end_sec': 43.0,
 'caption': 'A woman with long blonde hair stands indoors, holding up a birthday card ...'}

---

How to use

The benchmark does not redistribute media. Reproduce the clips with two steps:

1. Download the source at url (use your preferred tool: yt-dlp, TikTok APIs, …).
2. Cut the window [start_sec, end_sec] with ffmpeg.

import csv, subprocess, pathlib

with open("OmniRetriever-Bench.csv", newline="", encoding="utf-8") as f:
    rows = list(csv.DictReader(f))

def materialise(row, out_dir="clips"):
    pathlib.Path(out_dir).mkdir(parents=True, exist_ok=True)
    src = f"{out_dir}/source.mp4"
    # use yt-dlp or your downloader of choice on row['url'] -> src
    out = f"{out_dir}/clip.mp4"
    subprocess.check_call([
        "ffmpeg", "-y", "-ss", str(row["start_sec"]),
        "-to", str(row["end_sec"]), "-i", src,
        "-c", "copy", out,
    ])
    return out, row["caption"]

This release follows the standard video-benchmark protocol of distributing source identifiers + captions only (see VATEX, ColPali).

---

Scoring

The benchmark uses diagonal relevance — one positive target per query, identical across all 12 retrieval directions. Concretely: for any q2g direction, the gallery embedding at index i is the positive match for the query embedding at index i.

The reference evaluator (Recall@1/5/10, NDCG@10, MRR, median rank) lives in the Omni-Retriever GitHub repo.

After extracting one embedding per modality per record:

import numpy as np
blob = np.load("output/embeddings/omniretriever_bench.npz")

ids = sorted({k.split("__")[0] for k in blob.files})
def stack(mod): return np.stack([blob[f"{i}__{mod}"] for i in ids])

# Example: text -> audio+video direction.
queries = stack("text")
gallery = stack("av")
sim     = queries @ gallery.T          # embeddings are L2-normalised

omniretriever.evaluation.metrics.all_metrics(sim) then yields Recall@1/5/10, NDCG@10, MRR, and median rank.

---

Headline numbers from the paper

R@1 across the 12 directions (averaged):

Model	AVG-single	AVG-dual	AVG-all	Δ vs. open SOTA
WAVE-7B (frozen)	19.27	31.37	25.32	—
Omni-Embed-Nemotron (open)	21.79	31.84	26.81	+1.49
Gemini Embedding 2 (closed)	25.44	40.80	33.12	—
OmniRetriever-7B (ours)	28.63	41.05	34.84	+8.03 open / +1.72 vs. closed Gemini

Where OmniRetriever-7B wins (R@1 on the audio-anchored routes):

Direction	OmniRetriever-7B	Gemini Embedding 2	Δ
a→t	11.92	1.48	+10.44
a→tv	23.45	6.00	+17.45
v→a	25.46	13.80	+11.66
tv→a	26.07	16.79	+9.28

---

Stats

Property	Value
Held-out triples	3,782
Number of retrieval directions	12
Caption source	Gemini 3.0 Pro draft + trained human annotator review
Caption language	English
Median clip duration	2.16 s
p99 clip duration	16.17 s
Author audit κ (10 % sample)	0.78

---

Intended use & limitations

Intended use. Benchmarking cross-modal retrieval models; research on unified AVT representation learning; building retrieval evaluation harnesses for multimodal RAG systems.

Out of scope / prohibited. Biometric identification, recognition, profiling, or surveillance of natural persons in any deployed system — explicitly disallowed by the license clause below.

Limitations.

• English-only captions. Clip captions are not translated.
• Source-URL dependence. Underlying source videos are third-party UGC; we do not redistribute them. If a source URL becomes unavailable (uploader deletion, platform takedown), the corresponding row is no longer reproducible. Snapshot at evaluation time if long-term reproducibility matters.
• Diagonal relevance only. One positive target per query — no relevance-graded ranking signal.
• Short clips. Median 2 s, p99 16 s — not a test of long-form retrieval.

---

License

Released under Apache-2.0 with an additional clause prohibiting deployment for biometric identification of natural persons. By using this dataset you agree not to deploy any system derived from it that identifies, recognises, profiles, or surveils natural persons by their face, voice, body, gait, or other biometric attribute.

The underlying source videos referenced by url remain the property of their original uploaders and are governed by the source platform's terms of service.

---

Citation

@article{liu2026omniretriever,
  title         = {OmniRetriever: Any-to-Any Audio-Video-Text Retrieval via Fusion-as-Teacher Distillation},
  author        = {Liu, Yunze and Wu, Chi-Hao and Zhou, Enmin and Shen, Junxiao},
  year          = {2026},
  eprint        = {2605.26641},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2605.26641}
}