TRIAD Multimodal Retrieval Benchmark Data
简体中文 · Dataset · Code · Leaderboard · Paper
TRIAD: Multimodal Retrieval across Abstraction, Constraints, and State
Introduction
Finding a visually similar image does not mean a model understands the retrieval request. TRIAD evaluates multimodal embedding models across multiple abstraction levels, multiple constraints, and multiple dialogue turns: finding broader concepts that an image belongs to, satisfying positive and negative requirements together, and updating retrieval requirements through dialogue.
This Hugging Face dataset repository provides queries, candidate records, all positives, diagnostic annotations, and the corresponding images together. The companion code repository provides a unified evaluation launcher, model adapters, and AP/mAP scoring.
Dataset overview
| Part | Retrieval task | Subsets | Queries |
|---|---|---|---|
| Abstraction | Instance-to-Concept I2C and Concept-to-Instance C2I | I2C, C2I | 1,000 each |
| Constraints | Satisfying multiple retrieval requirements together | Fashion, Cars | 1,000 each |
| State | Retrieving according to the final active requirements in the full dialogue | Fashion, Cars | 1,000 each |
| Self diagnostic | Matching instances and concepts at the same semantic level | I2C, C2I | 1,000 each |
| Same diagnostic | Same-item retrieval | Fashion, Cars | 500 each |
There are 6,000 main-task queries and 3,000 diagnostic queries. Every query has 1,000 candidates and may have multiple positives. mAP is the primary metric; report tasks separately and present Self / Same separately from the main tasks.
The Fashion and Cars global product pools contain 12,946 and 8,007 records, respectively, shared by Constraints / State / Same. Each query still uses its own 1,000 candidates.
Guide to TRIAD data
- Download JSON and images from the same Hugging Face dataset repository at the same dataset version.
- Annotations retain their task-native JSON structures and are read directly by the companion evaluator.
--project-rootpoints to the directory containingdata/.- See
SCHEMA.mdin the dataset repository for fields, and the companion code guides for environments, options, and CoT usage.
Download
Download sisisisis/TRIAD-data using the Hugging Face CLI. The command below selects the frozen v1.0 dataset tag, so JSON, images, and metadata come from the same version:
pip install -U huggingface_hub
export TRIAD_DATASET_REPO="sisisisis/TRIAD-data"
export TRIAD_DATA_REVISION="v1.0"
export TRIAD_DATA_ROOT="/path/to/benchmark-root"
hf download "$TRIAD_DATASET_REPO" \
--repo-type dataset \
--revision "$TRIAD_DATA_REVISION" \
--local-dir "$TRIAD_DATA_ROOT"
Prepare images
JSON retains its evaluation paths under data/. Images are provided in three archives under images/ in the same repository. After downloading, verify and extract them from the data root:
cd "$TRIAD_DATA_ROOT"
sha256sum -c SHA256SUMS
tar -xzf images/abstraction-images.tar.gz
tar -xzf images/fashion-images.tar.gz
tar -xzf images/cars-images.tar.gz
sha256sum -c MEDIA_SHA256SUMS
Archive members already contain the data/.../images/ paths shown below, so no extra moving or renaming is needed. MEDIA_MANIFEST.json records image counts and extraction locations; MEDIA_SHA256SUMS verifies the extracted original image bytes.
Evaluation-ready structure
Arrange the downloaded JSON and images as follows. Annotations comprise twelve JSON files, and images reside in the three images/ directories.
benchmark-root/
└── data/
├── abstraction/
│ ├── main/
│ │ ├── i2c_wordnet.json
│ │ └── c2i_wordnet.json
│ ├── self/
│ │ ├── i2c_wordnet.json
│ │ └── c2i_wordnet.json
│ └── images/
└── constraint_state/
├── fashion/
│ ├── pool.json
│ ├── constraints.json
│ ├── state.json
│ ├── same.json
│ └── images/
└── cars/
├── pool.json
├── constraints.json
├── state.json
├── same.json
└── images/
The companion evaluator reads these files directly, without format conversion. Abstraction / Self candidates are embedded in each query. Constraints / State / Same use candidate_pool to reference records in the domain's pool.json. Sharing a product pool does not imply identical per-query candidate sets.
Field descriptions and provenance files reside at the dataset repository root: SCHEMA.md, LICENSES.md, sources.json, provenance.csv, manifest.json, and checksum lists. See the code repository's data format guide for input construction and diagnostic-task usage.
Evaluation
See the companion code repository's README and evaluation guide for evaluation commands and model configurations. Set --project-root in those commands to the downloaded TRIAD_DATA_ROOT, the directory containing data/.
Leaderboard
See the TRIAD leaderboard for model results and submission instructions. Record the dataset version, evaluation code version, and model configuration with results, and report main tasks separately from Self / Same.
Related resources
Data downloads and task descriptions are provided by this Hugging Face repository; model adapters, dependencies, and evaluation tools are provided by the TRIAD code repository. Its guides cover:
| Guide | Contents |
|---|---|
| README | Project introduction, overview figure, results, and quick start |
| Models and dependencies | Major dependency versions, model sources, and external implementations |
| Evaluation guide | Task selection, model options, CoT, caches, and completion checks |
| Data format | Data layout, input fields, and CoT format |
Versioning and integrity
The dataset version is v1.0, with schema version 1 in original-evaluator-json format. manifest.json records preparation and metadata-update dates, file counts, checksums, and the compatible public evaluation code at commit 55e0645. Use the dataset tag and record the exact code commit when reporting results.
SHA256SUMS verifies the repository's JSON, image archives, and documentation; MEDIA_SHA256SUMS verifies extracted images. After downloading, run from the data root:
cd "$TRIAD_DATA_ROOT"
sha256sum -c SHA256SUMS
On macOS, replace sha256sum with shasum -a 256. JSON and images are frozen at the same dataset version; subsequent updates receive a new version and a change description in the repository history.
Citation
If you use TRIAD, please cite our paper. Please also cite the upstream datasets used in your experiments.
@misc{wu2026triad,
title={TRIAD: Multimodal Retrieval across Abstraction, Constraints, and State},
author={Xiangyang Wu and Chen Ju and Xu Chen and Likang Peng and Jinsong Lan and Haihong Tang and Yang Hu},
year={2026},
eprint={2611.12364},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2611.12364}
}
Sources and license
Abstraction / Self use ImageNet-21K images and WordNet concepts and definitions. Fashion and Cars are based on Fashion Product Images and Stanford Cars, respectively. The licensing scopes of TRIAD's original annotations and third-party images/text are described separately in LICENSES.md and sources.json; the code license does not automatically apply to these data.
TRIAD's original task annotations are licensed under CC BY 4.0. This license does not cover third-party images or text, which retain their respective terms. The repository uses license: other and links to the specific licensing document. Rights holders may submit item IDs, source details, and removal requests via wuxiangyangbuaa@gmail.com or the repository Community tab.
- Downloads last month
- 9