MIRAGE

MIRAGE (Multimodal Information-seeking and Reasoning in AGricultural Expert-Guided conversations) is a benchmark for evaluating vision-language models on expert-level reasoning and decision-making in real-world agricultural consultations. It is derived from over 200,000 raw interactions between real users and certified agronomy experts on AskExtension, distilled to 37,512 high-quality user-expert conversations spanning more than 7,000 unique biological entities (plants, pests, diseases) across seven agronomic categories.

This dataset has been standardized to the AgML image_text_to_text format with a single conversational messages schema, converted to Parquet with image bytes embedded directly. It is indexed on https://project-agml.github.io/ as part of the AgML python library.

Benchmark Structure

MIRAGE evaluates two distinct challenges:

• MIRAGE-MMST (Multimodal Single-Turn) — Given a user query and associated image(s), can a model identify key biological entities, reason about causal symptoms, and produce actionable management recommendations?
• MIRAGE-MMMT (Multimodal Multi-Turn) — In an ongoing conversation, can a model decide whether to seek clarification or respond, and generate the appropriate follow-up utterance? Even top models achieve only ~63% decision accuracy on this task.

Available Subsets

Config	Description
MMST_Standard	Single-turn query + image(s) → expert response. Largest subset.
MMST_Contextual	Single-turn, with additional annotated location/time-dependence reasoning fields
MMMT_Direct	Multi-turn clarify-or-respond decision task, using only the raw dialogue context (the realistic/harder setting)
MMMT_Decomp	Same multi-turn task, but with the conversation's goal-state pre-decomposed into known_goal/missing_goal fields (a semi-privileged/easier setting) — giving models this decomposition raises decision accuracy from ~70% to ~89% in the original paper's ablation

Usage

from datasets import load_dataset

ds = load_dataset("Project-AgML/MIRAGE", "MMST_Standard")
first = ds["train"][0]

# Access an image — decoded to PIL automatically
img = first["images"][0]
img.show()

Schema

Every record shares the SAME columns so heterogeneous AgML datasets concatenate cleanly: id, images (embedded image bytes), messages, answer_type, and raw_metadata.

• answer_type — "mcqa" if the question/answer shows a discrete lettered/numbered choice structure, "oeqa" (open-ended) otherwise. MIRAGE's underlying task is entirely free-form expert consultation, so this is expected to be "oeqa" for the overwhelming majority (if not all) rows.
• raw_metadata is a JSON-encoded string holding source fields not folded into messages. Restore it with json.loads(row["raw_metadata"]).

For MMST_Standard / MMST_Contextual: original_id, subset, split, category, entity_type, entity_name, entity_scientific_name, entity_common_names, state, county, asked_time — and for MMST_Contextual only, additionally location_related, time_related, location_related_analysis, time_related_analysis.

For MMMT_Direct / MMMT_Decomp: original_id, subset, split, decision (<Respond> or <Clarify> — the expert's actual clarify-or-respond decision at that point in the conversation), dialog_turns, known_goal, missing_goal (the latter two are populated for MMMT_Decomp and null for MMMT_Direct, per the difference described above).

Image placeholders in messages align 1:1 with the images column; multiple images per row are supported (up to 3 in the source data).

Citation

@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}
}

Dongre, Vardhan; Gui, Chi; Garg, Shubham; Nayyeri, Hooshang; Tur, Gokhan; Hakkani-Tür, Dilek; Adve, Vikram S. (2025), "MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations", arXiv:2506.20100

License

Released by the original authors under CC BY-SA 4.0. This license information is for reference only and does not constitute legal advice — refer to the original repository for authoritative license terms: https://huggingface.co/datasets/MIRAGE-Benchmark/MIRAGE.