---
pretty_name: MIRAGE
license: cc-by-sa-4.0
task_categories:
- image-text-to-text
- visual-question-answering
size_categories:
- 10K<n<100K
source_datasets:
- AskExtension
configs:
- config_name: MMMT_Direct
  default: true
  data_files:
  - split: train
    path: MMMT_Direct/train-0000-of-0001.parquet
  - split: dev
    path: MMMT_Direct/dev-0000-of-0001.parquet
  - split: test
    path: MMMT_Direct/test-0000-of-0001.parquet
- config_name: MMMT_Decomp
  data_files:
  - split: train
    path: MMMT_Decomp/train-0000-of-0001.parquet
  - split: dev
    path: MMMT_Decomp/dev-0000-of-0001.parquet
  - split: test
    path: MMMT_Decomp/test-0000-of-0001.parquet
- config_name: MMST_Standard
  data_files:
  - split: train
    path: MMST_Standard/train-0000-of-0001.parquet
  - split: test
    path: MMST_Standard/test-0000-of-0001.parquet
- config_name: MMST_Contextual
  data_files:
  - split: test
    path: MMST_Contextual/test-0000-of-0001.parquet
dataset_info:
- config_name: MMMT_Direct
  features: &features
  - name: images
    list:
    - name: bytes
      dtype: 'null'
    - name: path
      dtype: string
  - name: id
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      list:
      - name: type
        dtype: string
      - name: text
        dtype: string
  - name: raw_metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 3734905
    num_examples: 3876
  - name: dev
    num_bytes: 844393
    num_examples: 878
  - name: test
    num_bytes: 844960
    num_examples: 861
  download_size: 1325714410
  dataset_size: 1325714410
- config_name: MMMT_Decomp
  features: *features
  splits:
  - name: train
    num_bytes: 4407679
    num_examples: 3876
  - name: dev
    num_bytes: 995362
    num_examples: 878
  - name: test
    num_bytes: 992906
    num_examples: 861
  download_size: 1326686099
  dataset_size: 1326686099
- config_name: MMST_Standard
  features: *features
  splits:
  - name: train
    num_bytes: 17348634
    num_examples: 17532
  - name: test
    num_bytes: 7962774
    num_examples: 8184
  download_size: 6948107240
  dataset_size: 6948107240
- config_name: MMST_Contextual
  features: *features
  splits:
  - name: test
    num_bytes: 6599964
    num_examples: 3934
  download_size: 1341825286
  dataset_size: 1341825286
---

# MIRAGE

![MIRAGE sample images](static/MIRAGE_sample.webp)

MIRAGE is a benchmark built from real conversations between farmers or gardeners and agricultural experts on the AskExtension platform. The authors collected about 285,000 consultations from 2012 to 2025 and kept 37,512 high-quality ones, covering more than 7,600 plants, pests and diseases. Many questions leave out key details, so a good model has to notice what is missing and ask, not just answer.

The authors tested 22 vision-language models. Even the best, GPT-4.1, reached only 43.9% identification accuracy.

- Paper: [MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations](https://arxiv.org/abs/2506.20100) (NeurIPS 2025 Datasets and Benchmarks)
- Project page: [mirage-benchmark.github.io](https://mirage-benchmark.github.io/)
- Original dataset: [MIRAGE-Benchmark/MIRAGE](https://huggingface.co/datasets/MIRAGE-Benchmark/MIRAGE)

## Configs

| Config | What it is | Train | Dev | Test |
|---|---|---|---|---|
| `MMST_Standard` | Single-turn questions that can be answered from the text and images | 17,532 | | 8,184 |
| `MMST_Contextual` | Single-turn questions with missing context (place, time or background) the model should notice | | | 3,934 |
| `MMMT_Direct` | Multi-turn dialogues where the model decides to ask a clarifying question or answer, then writes it | 3,876 | 878 | 861 |
| `MMMT_Decomp` | The same dialogues, with the known and missing goals spelled out | 3,876 | 878 | 861 |

## Structure

```
MMST_Standard/
  train-0000-of-0001.parquet
  test-0000-of-0001.parquet
  images/
    image-mmst_standard-000-of-003.zip ... image-mmst_standard-002-of-003.zip
    path_to_shard.parquet
MMST_Contextual/
  test-0000-of-0001.parquet
  images/
MMMT_Direct/
  train-0000-of-0001.parquet
  dev-0000-of-0001.parquet
  test-0000-of-0001.parquet
  images/
MMMT_Decomp/
  (same as MMMT_Direct)
Source_Data_For_MMMT.json
```

Each config has its own `images/` folder. Unique images per config: `MMST_Standard` 46,874, `MMST_Contextual` 8,032, `MMMT_Direct` and `MMMT_Decomp` 7,995 each.

Each row has:

- `images`: list of `{"bytes": null, "path": <sha256>}`, up to 3 per row, in the source's `image_1`, `image_2`, `image_3` order. The path is the SHA256 of the image bytes and points into a zip shard through that config's `images/path_to_shard.parquet`.
- `id`: `mirage_<config>_<split>_<counter>`, for example `mirage_mmst_standard_test_0`.
- `messages`: one user turn (images, then text) and one assistant turn.
    - In the MMST configs, the text is the question and the assistant turn is the expert's answer.
    - In the MMMT configs, the text is the dialogue so far (`dialog_context`, with `user:` and `expert:` lines), and the assistant turn is the next expert utterance.
- `raw_metadata`: JSON string.
    - MMST: `original_id`, `subset`, `answer_type`, `category`, `entity_type`, `entity_name`, `entity_scientific_name`, `entity_common_names`, `state`, `county` and `asked_time`. `MMST_Contextual` also has `location_related`, `time_related`, `location_related_analysis` and `time_related_analysis`.
    - MMMT: `original_id`, `subset`, `answer_type`, `decision` (such as `<Clarify>`), `dialog_turns`, `known_goal` and `missing_goal`.

Example (texts shortened):

```json
{
  "images": [{"bytes": null, "path": "cbff90a5ba89a76488f0533e5a98dc2a33095785a299e90eb7c09664a058e773"}],
  "id": "mirage_mmst_standard_test_0",
  "messages": [
    {"role": "user", "content": [
      {"type": "image", "text": null},
      {"type": "text", "text": "My arborvitae's have little gold, looking seeds on the ends and are also turning brown. Part of it m..."}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "The seeds are just that. Just like some get pine cones, this is the arborvitae's..."}]}
  ],
  "raw_metadata": "{\"original_id\": \"#885475\", \"subset\": \"MMST_Standard\", \"answer_type\": \"oeqa\", \"category\": \"Plant Care and Gardening Guidance\", \"entity_type\": \"plant\", \"entity_name\": \"arborvitae\", \"entity_scientific_name\": \"\", \"entity_common_names\": [], \"state\": \"Minnesota\", \"county\": \"Stearns County\", \"asked_time\": \"2024-09-16 16:35:00\"}"
}
```

### Changes from the original data

- Each source config becomes a config here, with its own source splits. `dev` keeps its source name.
- `image_1`, `image_2` and `image_3` become the `images` list. Empty slots are left out.
- The MMMT dialogue history stays as one block of text in the user turn, exactly as in `dialog_context`. It is not split into separate turns.
- Some source fields are renamed in `raw_metadata`: `id` becomes `original_id`, and `meta_data_state`, `meta_data_county` and `meta_data_asked_time` become `state`, `county` and `asked_time`.
- `raw_metadata` also has two fields that are not in the source:
    - `subset`: the config name.
    - `answer_type`: added during conversion. It is `mcqa` when the question lists at least two lettered or numbered options, or the answer is a single letter or digit, and `oeqa` otherwise. MMMT rows are always `oeqa`.
- `MMMT_Direct` rows have `known_goal` and `missing_goal` set to null, since only `MMMT_Decomp` has them in the source.
- Images are stored once per config, deduplicated by SHA256. `MMMT_Direct` and `MMMT_Decomp` share the same images, so each holds its own copy.
- `Source_Data_For_MMMT.json` from the original dataset is included unchanged.

## Usage

```python
from agml import loadImageTextToTextDataset

dataset = loadImageTextToTextDataset("Project-AgML/MIRAGE", config="MMST_Standard")

sample = dataset["test"][0]
sample["images"]    # list of PIL images
sample["messages"]
```

## Citation

```bibtex
@article{dongre2025mirage,
  title={MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations},
  author={Dongre, Vardhan and Gui, Chi and Garg, Shubham and Nayyeri, Hooshang and Tur, Gokhan and Hakkani-T{\"{u}}r, Dilek and Adve, Vikram S},
  journal={arXiv preprint arXiv:2506.20100},
  year={2025}
}
```

This dataset is indexed and structured on https://project-agml.github.io/ as part of Project AgML @ UC Davis. This dataset was reformatted from its original format to match HuggingFace's Imagefolder standards but requires an external module (agml) that processes and returns a HF Dataset object faster than HF module functions.