---
pretty_name: OneSearch-MI-Bench
language:
- en
task_categories:
- visual-question-answering
tags:
- benchmark
- visual-search
- multi-image
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: test.jsonl
---

# OneSearch-MI

OneSearch-MI contains 301 questions and 910 unique PNG images. Each question
uses multiple images in a fixed order. The images retain their original bytes,
resolution and aspect ratio; no resizing or recompression is applied.

## Files

```text
README.md
prepare.py
test.jsonl
media.jsonl
media/
  0000.png
  ...
```

Media is distributed in `mi-00001.zip`. Extract all media archives into this
directory, preserving their paths. Each archive is at most 10 GB
(10,000,000,000 bytes). All images needed by this benchmark are included;
the video benchmark is not required.

`test.jsonl` contains one record per question:

| Field | Meaning |
| --- | --- |
| `id` | Stable sample ID, represented as a string |
| `question` | Question given to the model |
| `answer` | Reference answer for evaluation; not an actor input |
| `images` | Ordered image paths relative to this MI dataset directory |
| `n_images` | Number of images in this sample |
| `structure` | Question structure |
| `difficulty` | Difficulty label |
| `num_hops` | Annotated reasoning hop count |
| `knowledge_domain` | Knowledge category |

**Preserve the order of `images`.** The first list element becomes `img_1`,
the second becomes `img_2`, and so on. Numeric filenames are media identifiers,
not the input sequence; sorting the filenames can change the sample.
Multiple questions may share an image. `media.jsonl` has one entry per unique
PNG, including its relative path, byte size, width and height.

## Prepare inference inputs

The loader requires Python 3.9 or newer and uses only the standard library.
It makes no model or judge calls and does not modify the images.

Check image byte sizes, PNG dimensions and sample references:

```bash
python prepare.py --verify-only
```

Expose this MI dataset directory through an HTTP(S) media server reachable by
the inference backend. Set `MEDIA_BASE_URL` to that directory's URL, then run:

```bash
python prepare.py --media-base-url "$MEDIA_BASE_URL"
```

This produces:

- `runtime/mi.csv`: `index,question,answer` for inference and evaluation.
- `runtime/mi_manifest.jsonl`: ordered image URLs for the multi-image pipeline.

Each manifest record has `case_id` equal to the sample `id`, `ok: true`,
`modality: "multi_image"`, `n_images`, and ordered `image_urls`. It contains
no reference answer, source-video metadata or annotations. The question is
loaded from the CSV; the reference answer is reserved for evaluation.

From the OneSearch-VL code repository, with `DATA_ROOT` pointing to this MI
dataset directory:

```bash
python onesearch_vl/run_infer_mi.py \
  --model 8b --checkpoint /models/your-checkpoint \
  --csv "$DATA_ROOT/runtime/mi.csv" \
  --manifest "$DATA_ROOT/runtime/mi_manifest.jsonl" \
  --output-dir outputs/onesearch_mi

bash onesearch_vl/run_eval.sh image \
  --csv "$DATA_ROOT/runtime/mi.csv" \
  --traj_dir outputs/onesearch_mi
```

Configure inference, search and evaluation credentials in the code repository
before running those commands. No evaluation results are included in this package.