Trident-Eval

Official evaluation data for Chaos in the Text: Revealing the Modality Preference in Mixed-Modality Retrievers

Overview

Real-world corpora mix text-only, image-only, and fused text-image documents. Trident-Eval provides, for every dataset, two versions of the retrieval corpus:

Setting	File	Description
Origin	origin_corpus.jsonl	The dataset's original single-modality corpus.
Mix	mix_corpus.jsonl	The mixed-modality version of the corpus, constructed as described in Section 2.1 of the paper.

Evaluating the same model on both settings reveals how much retrieval quality degrades when the corpus modality composition changes ("Chaos in the Text").

Two benchmarks are included:

Benchmark	Paper	Datasets
vdoc — Visual document retrieval	Table 1	ChartQA, DocVQA, InfoVQA, SlideVQA, ViDoSeek, Dude
natural — Natural-image retrieval	Table 4 (Appendix F)	Google_WIT, MSCOCO, VisualNews, OVEN (Mix corpus follows MixBench)

Directory Structure

Trident-Eval/
├── README.md
├── ChartQA/
│   ├── queries.jsonl
│   ├── origin_corpus.jsonl
│   ├── mix_corpus.jsonl
│   ├── qrels.jsonl
│   └── images.tar.gz        # extracts to ./images/
├── DocVQA/      ...
├── InfoVQA/     ...
├── SlideVQA/    ...
├── ViDoSeek/    ...
├── Dude/        ...
├── Google_WIT/  ...
├── MSCOCO/      ...
├── VisualNews/  ...
└── OVEN/        ...

After downloading and extracting, the layout matches what the Trident code expects under data/:

data/
├── ChartQA/
│   ├── queries.jsonl
│   ├── origin_corpus.jsonl
│   ├── mix_corpus.jsonl
│   ├── qrels.jsonl
│   └── images/
│       └── ...
└── ...

Download & Extract

Option 1: one-line script (recommended)

The Trident repository ships a helper script that downloads and extracts the data in the expected layout:

git clone https://github.com/OpenBMB/Trident.git && cd Trident

bash data/download_data.sh eval                         # all ten datasets
bash data/download_data.sh eval --benchmark vdoc        # Table 1
bash data/download_data.sh eval --benchmark natural     # Table 4

Option 2: manual

# 1. Download (everything, or pass --include "ChartQA/*" for a single dataset)
hf download OpenBMB/Trident-Eval \
    --repo-type dataset --local-dir data

# 2. Extract the image archives in place
for d in data/*/; do
    [ -f "${d}images.tar.gz" ] && tar -xzf "${d}images.tar.gz" -C "$d"
done

# 3. (optional) remove archives to save disk space
# rm data/*/images.tar.gz

Or download from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="OpenBMB/Trident-Eval",
    repo_type="dataset",
    local_dir="data",
    allow_patterns=["ChartQA/*"],   # omit to download everything
)

Then run the evaluation from the code repository, e.g. with the released Trident-Qwen3VL-2B (its LoRA is already merged, so do not pass --base_model):

hf download OpenBMB/Trident-Qwen3VL-2B --local-dir checkpoints/Trident-Qwen3VL-2B

bash scripts/eval.sh trident_qwen3vl checkpoints/Trident-Qwen3VL-2B vdoc

Data Format

All .jsonl files contain one JSON object per line. Image paths are relative to the dataset directory (data/<DATASET>/) and use forward slashes.

queries.jsonl

{"id": "q1", "text": "How many stores did Saint Laurent operate in Western Europe in 2020?"}

origin_corpus.jsonl / mix_corpus.jsonl

Each document has an id and a text field, an image field, or both:

{"id": "d1", "image": "images/d1.png"}
{"id": "d2", "text": "A bar chart showing ..."}
{"id": "d3", "text": "Short summary ...", "image": "images/d3.png"}

Document type	Fields present
Image-only	id, image
Text-only	id, text
Fused (text + image)	id, text, image

In origin_corpus.jsonl every document has the dataset's original modality. In mix_corpus.jsonl documents are a mixture of the three types above. Document ids are shared between the two corpora, so the same qrels.jsonl applies to both.

qrels.jsonl

{"query_id": "q1", "doc_id": "d1", "score": 1}

All ids (id, query_id, doc_id) are strings.

Dataset Statistics

Dataset	Benchmark	#Queries	#Docs
ChartQA	vdoc	1250	1509
DocVQA	vdoc	591	741
InfoVQA	vdoc	718	459
SlideVQA	vdoc	556	1284
ViDoSeek	vdoc	1142	5385
Dude	vdoc	496	27955
Google_WIT	natural	1000	4421
MSCOCO	natural	984	984
VisualNews	natural	981	981
OVEN	natural	1000	1000

Mix Setting

The Mix corpus of each dataset is built following Section 2.1 of the paper (every document is kept in exactly one of the text-only / image-only / fused forms; text versions are generated by Qwen3-VL-235B-A22B-Instruct).

During evaluation, the Mix corpora of all datasets in a benchmark are merged into a single retrieval pool. This merge is performed on the fly by scripts/run_all_datasets.sh; the per-dataset files in this repository are not pre-merged.

License

The annotations and generated text released here are under the MIT License. The images and the original queries, corpora, and relevance labels originate from third-party datasets and remain subject to their original licenses; please make sure your use complies with them.

Citation

If you find this work useful, please cite:

@article{trident2026chaos,
  title   = {Chaos in the Text: Revealing the Modality Preference in Mixed-Modality Retrievers},
  author  = {Sun, Yubo and Peng, Chunyi and Yan, Yukun and Liu, Zhenghao and Xu, Zhipeng and Mei, Sen and Xin, Linlin and Zeng, Zheni and Sun, Maosong},
  journal = {arXiv preprint arXiv:2610.11816},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.11816}
}