MMBU-public

The official public split of the released MMBU benchmark, serving as the public set for the MMBU challenge. Open-ended VQA only for full context MMBU questions. Ground-truth answer is included.

Contents: 3,147 open-ended rows across 3,147 images. For every task except object detection, options lists the paired closed-VQA choices for the same question (for LLM-judge grading). Object detection rows have an empty options list; those answers are scored by box IoU.

Released metadata columns: question, file_name, question_type, options, metadata_medical_domain, metadata_modality, metadata_submodality, metadata_specimen, metadata_body_part, metadata_topic, microns_per_pixel, metadata_stain, task, unique_id, VQA_type, answer.

There are 5 main categories of tasks present:

• object detection
• fine-grained classification (from detection)
• fine-grained classification (from segmentation)
• classification
• counting

Each task contains multiple topics stemming from the 5 main categories, delineated by '|'.

Prompt

Send the image with the question text. Do not send options or answer. options is only for the judge.

Route on task, then append that suffix after the question.

If task…	Suffix
starts with object detection	Object detection
contains from segmentation	Multi-label classification
contains from detection	Single-label classification
starts with counting, or any other classification…	Single-answer classification

from detection and from segmentation images already show the region (a drawn box or mask). Only object detection asks for coordinates.

System message

You are a medical vision-language assistant evaluating biomedical images.
Analyze the image and question carefully and provide a concise rationale.
After the rationale, put your final committed answer inside
<answer>...</answer>. Put only the answer inside the tags—no reasoning,
explanation, or surrounding text. Follow the answer-format instructions in
the user message.

User message: {question}, a blank line, then the suffix.

Single-answer classification — plain classification, classification | …, and counting | …. Counting uses this same suffix; the tag should contain only the integer, for example <answer>339</answer>.

Output format: After your rationale, output only the concise class label
inside <answer> tags, for example <answer>glioblastoma</answer>. Do not
include a letter, rationale, or additional explanation inside the tags.

Fine-grained classification from a bounding box (from detection) uses that same single-label suffix.

Fine-grained classification from a segmentation mask (from segmentation)

Output format: After your rationale, output the applicable class label or
labels inside <answer> tags. If multiple labels apply, separate them with a
comma and a space, for example <answer>effusion, cardiomegaly</answer>. Do
not include rationale or additional explanation inside the tags.

Object detection

Output format: After your rationale, output the bounding box or boxes inside
<answer> tags using pixel coordinates from the image shown and the format
[x, y, width, height]. Do not use normalized 0–1 coordinates. For one box,
use <answer>[x, y, width, height]</answer>. For multiple boxes, use
<answer>[[x1, y1, width1, height1], [x2, y2, width2, height2]]</answer>.
Do not include rationale or additional text inside the tags.

x, y is the top-left corner in pixels of the image as shown, then width and height.

Scoring

Parse the committed answer from <answer>...</answer> before scoring. If the tags are missing, score the raw reply.

Object detection (task starts with object detection) is not judged. Parse boxes as [x, y, width, height] in pixels. IoU is the intersection area divided by the union area. Compare every predicted box with every ground-truth box and keep the maximum IoU. A row is correct when that IoU is at least 0.5.

All other tasks — classification, fine-grained classification from detection or segmentation, and counting — use an LLM judge (open_per_dataset_v3). The judge scores 1 only when both gates pass. Fill the prompt from that row’s options:

• {label_set}: option texts with the leading A) / B. prefix removed, one - label per line.
• {question}: the row question.
• {true_answer}: the row answer.
• {model_answer}: the parsed model answer.

Judge prompt

You are a conservative biomedical answer-equivalence judge for open-ended VQA.

You are grading, not answering. Never supply an answer yourself. Grade only the text given.

Valid labels are the paired closed-VQA options for this same question:
{label_set}

Output rules:
- Output exactly one line.
- Format: Rationale: <brief rationale naming the grounded concept> [RESULT] <0 or 1>
- No markdown, JSON, or extra text.

TWO-GATE DECISION (both gates must pass for [RESULT] 1):

GATE 1 — Grounding (apply first):
Does the MODEL ANSWER actually and unambiguously assert ONE specific concept from the
label set above (or a clear, unambiguous synonym of exactly one listed label)?
- Empty, whitespace-only, or no answer -> [RESULT] 0
- Vague category only ("some pigmented lesion", "an eye disease", "abnormal scan") -> [RESULT] 0
- Off-taxonomy or unrelated to the label set -> [RESULT] 0
- Hedged across multiple labels ("melanoma or nevus", "A or B") -> [RESULT] 0
- Only "closest" to a label without stating that label -> [RESULT] 0
If Gate 1 fails, stop with [RESULT] 0. Do NOT force-map to the nearest label.

GATE 2 — Equivalence (only if Gate 1 passes):
Is the grounded concept the same as the TRUE ANSWER under the equivalence policy above
(same specificity)?
- Broader/narrower, parent/subtype, related-but-distinct -> [RESULT] 0
- Same specific label or unambiguous synonym -> [RESULT] 1

Examples:
TRUE: "melanoma"  MODEL: "some pigmented skin lesion"
Rationale: Vague; no committed label from set. [RESULT] 0
TRUE: "no tumor"  MODEL: "normal brain, no mass"
Rationale: Grounded to no-tumor side; equivalent under binary policy. [RESULT] 1
TRUE: "benign keratosis"  MODEL: "normal skin"
Rationale: Distinct multiclass labels; not equivalent. [RESULT] 0

Question/context:
{question}

TRUE ANSWER:
{true_answer}

MODEL ANSWER:
{model_answer}

Now output exactly one line:
Rationale:

A row is correct when that line ends in [RESULT] 1. Official scores use claude-sonnet-5 with this user prompt.

Loading

from datasets import load_dataset

ds = load_dataset("mmbu/mmbu-public")
print(ds)
print(ds["train"][0].keys())

Files

• metadata.jsonl: Hugging Face ImageFolder metadata.
• metadata.tsv: TSV metadata mirror.