MIRAGE

MIRAGE is a benchmark for multimodal, expert-guided agricultural consultations, introduced in "MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations" (arXiv:2506.20100, NeurIPS 2025). It's built from real crop health and pest diagnosis consultations — curated down from 218,000 raw interactions (2012–2025) to 37,512 high-quality ones, spanning 7,600+ distinct biological entities (plants, pests, diseases). Unlike closed-set VQA benchmarks, MIRAGE's scenarios are underspecified and open-world by design, requiring models to recognize when they need more context rather than guessing. In the paper's evaluation, 20+ frontier vision-language models were tested, with the best (GPT-4.1) reaching only 43.9% identification accuracy.

Configs

MIRAGE ships as four configs, corresponding to the paper's task families:

config	description	splits
MMST_Standard	Single-turn, self-contained questions answerable from the given text + image(s) alone — entity identification, causal explanation, visually-grounded answers.	train, test
MMST_Contextual	Single-turn questions with a latent information gap — the model must recognize missing context (location, time, agricultural background) rather than answer directly.	test
MMMT_Direct (default)	Multi-turn dialogues; at each turn the model must decide whether to ask a clarifying question or answer, and generate that utterance directly.	train, dev, test
MMMT_Decomp	The same underlying multi-turn dialogues as MMMT_Direct, reformulated as a decomposed reasoning task (hence the matching unique-image counts between the two configs below).	train, dev, test

MMMT_Direct is set as the default config.

Layout

Each config is a self-contained folder — its own metadata parquet(s) plus its own images/ folder of zip shards — following the same storage format across every AgML-standardized dataset:

MIRAGE_P/
  MMST_Standard/
    train-0000-of-0001.parquet   # 17,532 rows
    test-0000-of-0001.parquet    #  8,184 rows
    images/
      image-mmst_standard-000-of-003.zip ... 002   # 46,874 unique images, 6.92GB
      path_to_shard.parquet
  MMST_Contextual/
    test-0000-of-0001.parquet    #  3,934 rows
    images/
      image-mmst_contextual-000-of-001.zip         #  8,032 unique images, 1.33GB
      path_to_shard.parquet
  MMMT_Direct/
    train-0000-of-0001.parquet   #  3,876 rows
    dev-0000-of-0001.parquet     #    878 rows
    test-0000-of-0001.parquet    #    861 rows
    images/
      image-mmmt_direct-000-of-001.zip             #  7,995 unique images, 1.32GB
      path_to_shard.parquet
  MMMT_Decomp/
    train-0000-of-0001.parquet   #  3,876 rows
    dev-0000-of-0001.parquet     #    878 rows
    test-0000-of-0001.parquet    #    861 rows
    images/
      image-mmmt_decomp-000-of-001.zip             #  7,995 unique images, 1.32GB
      path_to_shard.parquet

Every row has: images, id, messages, raw_metadata. images holds {"bytes": None, "path": ...} per referenced image — but unlike AgroOmni, path here is a SHA256 content hash, not a filesystem path: MIRAGE's source data has no path metadata at all (images exist only as embedded bytes per row), so the hash is what makes dedup and shard lookup possible. path_to_shard.parquet maps each unique hash to the shard filename holding it. answer_type (mcqa/oeqa) lives inside raw_metadata rather than as its own column, keeping the schema consistent with every other AgML dataset.

Not auto-loadable via load_dataset(). Requires the companion loader (load_agml_dataset()), which reads images lazily and directly out of the zip shards — no extraction step.

Why content hashing instead of paths

The companion loader looks up each image independently via zipfile.ZipFile.open(member_name) — true random access via zip's central directory, no extraction needed — so, as with AgroOmni, there's no correctness requirement that a row's co-referenced images share a shard. Since MIRAGE has no natural path identifiers, a SHA256 hash of each image's raw bytes is used instead, which has the added benefit of deduplicating any images that repeat across rows within a config.

Usage

from agml import loadImageTextToTextDataset

ds = loadImageTextToTextDataset(
      "Project-AgML/MIRAGE",
      config="MMMT_Direct",
      token=HF_TOKEN
)
print(ds)                 # DatasetDict({'train': ..., 'dev': ..., 'test': ...})
ds["train"][0]            # images decoded lazily on access

Citation

If you use this dataset, please cite the original MIRAGE paper:

@misc{mirage2025,
  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.},
  year   = {2025},
  eprint = {2506.20100},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  note   = {NeurIPS 2025},
  url    = {https://arxiv.org/abs/2506.20100}
}

Project page: https://mirage-benchmark.github.io/ Code: https://github.com/MIRAGE-Benchmark/MIRAGE-Benchmark

---

This dataset is indexed and structured on https://project-agml.github.io/ as part of the AgML python library. 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.