---
license: mit
pretty_name: MixBench
task_categories:
  - text-ranking
task_ids:
  - document-retrieval
language:
  - en
multilinguality: monolingual
annotations_creators:
  - machine-generated
dataset_creator: Binxu Li et al.
dataset_info:
  features:
    - name: id
      dtype: string
    - name: text
      dtype: string
    - name: image
      dtype: image
configs:
  - config_name: MSCOCO
    data_files:
      - split: queries
        path: MSCOCO/queries.parquet
      - split: corpus
        path: MSCOCO/corpus.parquet
      - split: mixed_corpus
        path: MSCOCO/mixed_corpus.parquet
  - config_name: Google_WIT
    data_files:
      - split: queries
        path: Google_WIT/queries.parquet
      - split: corpus
        path: Google_WIT/corpus.parquet
      - split: mixed_corpus
        path: Google_WIT/mixed_corpus.parquet
  - config_name: VisualNews
    data_files:
      - split: queries
        path: VisualNews/queries.parquet
      - split: corpus
        path: VisualNews/corpus.parquet
      - split: mixed_corpus
        path: VisualNews/mixed_corpus.parquet
  - config_name: OVEN
    data_files:
      - split: queries
        path: OVEN/queries.parquet
      - split: corpus
        path: OVEN/corpus.parquet
      - split: mixed_corpus
        path: OVEN/mixed_corpus.parquet
tags:
  - retrieval
  - image
  - text
  - multimodal
  - benchmark
---

# MixBench: A Benchmark for Mixed Modality Retrieval

**MixBench** is a benchmark for evaluating retrieval across text, images, and multimodal documents. It is designed to test how well retrieval models handle queries and documents that span different modalities, such as pure text, pure images, and combined image+text inputs.

MixBench includes **four subsets**, each curated from a different data source:

- **MSCOCO**
- **Google_WIT**
- **VisualNews**
- **OVEN**

Each subset contains:

- `queries.jsonl`: each entry contains a `query_id`, `text`, and/or `image`  
- `mixed_corpus.jsonl`: each entry contains a `corpus_id`, a `text` or an `image` or a multimodal document (`text` and `image`) 
- `qrels.tsv`: a tab-separated list of relevant query-document pairs (`query_id`, `corpus_id`, `score=1`)  
- `corpus.jsonl`: the original corpus

This benchmark supports diverse retrieval settings including unimodal-to-multimodal and cross-modal search.

---

## 🔄 Load Example

You can load a specific subset of MixBench using the `name` argument:

```python
from datasets import load_dataset

# Load the MSCOCO subset
ds_query = load_dataset("andy0207/mixbench", name="MSCOCO", split='query')
ds_corpus = load_dataset("andy0207/mixbench", name="MSCOCO", split='mixed_corpus')
ds_query = load_dataset("andy0207/mixbench", name="MSCOCO", split='qrel')
# Load other subsets (corpus)
ds_gwit = load_dataset("andy0207/mixbench", name="Google_WIT", split='mixed_corpus')
ds_news = load_dataset("andy0207/mixbench", name="VisualNews",split='mixed_corpus')
ds_oven = load_dataset("andy0207/mixbench", name="OVEN", split='mixed_corpus')