MIME — multi-source visual in-context learning episodes (5src)

The official training dataset of the paper Watch and Learn: Teaching Visual In-Context Learning by Removing Textual Shortcuts.

MIME is a task-agnostic visual in-context learning (vICL) training set without textual shortcuts. Each row is one episode: 1–5 demonstration images with their answers, then a query image and its answer. The instruction does not state the rule that links the demonstration images to their answers, so the model must infer the rule from the visual demonstrations.

Subsets

Subset	Episodes	Shots	Answer type	Image origin
puzzlevqa_polygon_sides	5000	2–4	integer	PuzzleVQA-style polygons
bongard_logo_shape_concept_binary	5000	2, 4	text (yes/no)	Bongard-LOGO
re_arc_transform_induction_named	5000	3–5	text (transform name)	RE-ARC grids
puzzlevqa_triangle_arithmetic	5000	2–4	integer	PuzzleVQA-style triangles
video_icl	3578	1, 2, 4	integer / text	CLEVRER frames
all	23578	1–5	—	all of the above

all reads the same files as the per-source subsets; it adds no extra storage.

Columns

Column	Type	Meaning
source	string	Subset name
episode_id	string	Unique episode id
rule_family	string	Rule family within the source
num_shots	int	Number of demonstrations
rule_spec	string (JSON)	The hidden rule of the episode
answer_type	string	integer or text
demo_images	list[Image]	Demonstration images, num_shots long
demo_labels	list[string]	Answer for each demonstration
demo_scene_attrs	string (JSON)	Per-demo generation metadata
query_image	Image	Query image
query_label	string	Target answer
query_scene_attrs	string (JSON)	Query generation metadata
source_videos	list[string]	CLEVRER video ids used (empty for non-video sources)

Usage

Requires datasets>=4.0.

from datasets import load_dataset

ds = load_dataset("NimrodShabtay1986/MIME", "all", split="train")
ep = ds[0]
ep["demo_images"]  # list of PIL images
ep["query_label"]

Citation

@article{shabtay2026watch,
  title   = {Watch and Learn: Teaching Visual In-Context Learning by Removing Textual Shortcuts},
  author  = {Shabtay, Nimrod and Yehudai, Asaf and Dekel, Avihu and Schwartz, Eli and Giryes, Raja},
  journal = {arXiv preprint arXiv:2610.30913},
  year    = {2026}
}