PuzzleComm/Puzzle_Perception

Puzzle Perception — Segmentation + pVQA

A single table over two tasks on synthetic puzzle images:

• Segmentation — per-pixel masks over chess, maze and tower-of-hanoi under one unified 30-class label space.
• pVQA — multiple-choice perception probes over chess and N-Queens boards.

Every row carries the same 13 columns; the type column says which task it belongs to, and columns that do not apply are null.

from datasets import load_dataset

ds = load_dataset("PuzzleBench/Puzzle_Perception", split="test")
pvqa = ds.filter(lambda r: r["type"] == "pvqa")
seg = ds.filter(lambda r: r["type"] == "segmentation")

row = seg[0]
row["image"]                 # PIL.Image, 512x512 RGB — the puzzle photo
row["mask"]                  # PIL.Image, 512x512 RGB — colorized label map
row["segmentation_labels"]   # e.g. ['black_square', 'white_queen', ...]

Columns

Column	Type	Segmentation rows	pVQA rows
id	int64	unique, contiguous from 0	unique, contiguous from 0
puzzle	string	chess / maze / hanoi	chess / nqueens
image	Image	512×512 RGB — the puzzle photo	512×512 RGB
image_path	string	source-tree path (provenance)	source-tree path (provenance)
mask	Image	512×512 RGB — colorized label map	null
mask_path	string	path to that same colorized PNG in this repo	null
raw_mask	string	path to the raw single-channel PNG, values 0..29	null
type	string	segmentation	pvqa
segmentation_labels	list<string>	class names present in the mask	null
question	string	null	probe question text
answer	string	null	correct option, as text
question_id	string	null	q1…q8, joins pvqa_questions.yaml
options	list<string>	null	the full answer space

image and mask are both embedded in the Parquet as PNG bytes, so both render in the viewer and load_dataset hands back a PIL.Image for each. mask is a colorized render — one fixed color per class (see the palette below) — not the raw class-id values, which are almost black on their own and uninformative to look at directly.

Training needs the raw values, not the colors, so they are published separately as paths to real files under raw_masks/ — resolvable, but never decoded inline:

from huggingface_hub import hf_hub_download
import numpy as np
from PIL import Image

p = hf_hub_download("PuzzleBench/Puzzle_Perception", row["raw_mask"], repo_type="dataset")
mask = np.array(Image.open(p))        # (512, 512) uint8, values in [0, 29]

Splits

Split	Total	Segmentation	pVQA	Per-puzzle
train	6000	6000	0	seg: chess 2000, hanoi 2000, maze 2000
val	1500	1500	0	seg: chess 500, hanoi 500, maze 500
test	2700	1500	1200	seg: chess 500, hanoi 500, maze 500 · pVQA: chess 800, nqueens 400
total	10200	9000	1200

pVQA appears only in test, and its rows are written before the segmentation rows of that split.

Segmentation labels

raw_mask files are 8-bit PNGs whose pixel values are these class ids — no remapping at load time. mask recolors those same values one-for-one using the Color column below (also shipped as palette.yaml): each id's hue is stepped by the golden angle (~137.5°) around the wheel, deterministically, so a rebuild always reproduces the same colors — and so that classes with adjacent ids (e.g. the five Hanoi disks) land far apart in hue instead of clustering, since linear spacing would otherwise squeeze every class of one puzzle into a narrow slice of the wheel (ids are grouped contiguously by puzzle).

Puzzle	Class ids	Count
maze	0–7	8
chess	8–22	15
hanoi	23–29	7

Class id	Name	Puzzle	Color
0	wall	maze	#f25555
1	path	maze	#55f283
2	start	maze	#b155f2
3	dest_a	maze	#f2df55
4	dest_b	maze	#55d8f2
5	dest_c	maze	#f255aa
6	dest_d	maze	#7cf255
7	dest_e	maze	#5b55f2
8	background	chess	#f28955
9	white_square	chess	#55f2b7
10	black_square	chess	#e555f2
11	white_pawn	chess	#d1f255
12	white_knight	chess	#55a3f2
13	white_bishop	chess	#f25575
14	white_rook	chess	#55f262
15	white_queen	chess	#9055f2
16	white_king	chess	#f2be55
17	black_pawn	chess	#55f2ec
18	black_knight	chess	#f255cb
19	black_bishop	chess	#9df255
20	black_rook	chess	#556ff2
21	black_queen	chess	#f26955
22	black_king	chess	#55f297
23	background	hanoi	#c555f2
24	peg	hanoi	#f2f255
25	disk_1	hanoi	#55c4f2
26	disk_2	hanoi	#f25596
27	disk_3	hanoi	#68f255
28	disk_4	hanoi	#7055f2
29	disk_5	hanoi	#f29e55

Note background appears twice — chess id 8 and hanoi id 23 are distinct classes in this unified namespace, and get distinct colors — so segmentation_labels names should be read together with puzzle. Per-class loss weights live in classes.yaml, not here — they are training hyperparameters, not label definitions.

pVQA answer spaces

Puzzle	question_id	Question	options	Rows
chess	q1	Which quarter of the board is the white queen in? A=top-left B=top-right C=bottom-left D=bottom-right.	[A, B, C, D]	100
chess	q2	Which quarter of the board is the white bishop in? A=top-left B=top-right C=bottom-left D=bottom-right.	[A, B, C, D]	100
chess	q3	Is the leftmost white pawn in the top half or the bottom half of the board?	[top, bottom]	100
chess	q4	Is the white king above or below the black knight?	[above, below]	100
chess	q5	Are the white rook and the black king in the same row or the same column?	[Yes, No]	100
chess	q6	Are the white king and the black king on adjacent (touching) squares?	[Yes, No]	100
chess	q7	Is the white queen on the same rank, file, or diagonal as the black king?	[Yes, No]	100
chess	q8	Is there a white queen on the board?	[Yes, No]	100
nqueens	q1	Is the leftmost queen in the top half or the bottom half of the board?	[top, bottom]	100
nqueens	q2	Is the topmost queen in the left half or the right half of the board?	[left, right]	100
nqueens	q3	Is the leftmost queen above or below the rightmost queen?	[above, below]	100
nqueens	q4	Is the leftmost queen in the top, middle, or bottom third of the board?	[top, middle, bottom]	100

Answer spaces are not uniform: 2-way, 3-way, 4-way all occur, with chess/q1, chess/q2 and nqueens/q4 non-binary. That is why options is a column rather than an assumed [Yes, No].

Questions are deliberately coordinate-free ("the leftmost queen", not "the queen at row 4"), so a model must locate the referenced piece before it can answer. Full specs, including how each answer was derived from the board, are in pvqa_questions.yaml.

Notes and caveats

• The two tasks do not share images. pVQA chess boards are separately rendered, not the segmentation chess images, and N-Queens is not part of the 30-class label space. No image currently carries both a mask and a question.
• image_path repeats for N-Queens pVQA. Each of the 100 boards is asked 4 questions, so 4 rows share an image_path with distinct ids. Chess pVQA asks one question per board.
• Published N-Queens images are a derived render. The sources are 1600×1600 JPEG, downscaled here to 512×512 with PIL thumbnail(..., BILINEAR) — the same call the evaluation code applies, so these are the pixels models actually saw. The originals remain the reference copy in the project repository.
• Segmentation splits are sampled from the source datasets with a fixed seed (42) and are reproducible.
• manifest.csv maps unified_id to source task and split; classes.yaml carries the class map and loss weights.

Files

Path	Contents
data/*.parquet	the table — 10200 rows
masks/{train,val,test}/*.png	9000 colorized RGB renders — the mask_path targets
raw_masks/{train,val,test}/*.png	9000 raw label maps, 8-bit, values 0..29 — the raw_mask targets
classes.yaml	30-class map + per-class loss weights
manifest.csv	unified_id,source_task,source_id,split
pvqa_questions.yaml	published question specs, nested by task
palette.yaml	id → name → RGB/hex used to colorize mask

License

CC-BY-NC-SA 4.0. If you use this dataset, please cite the PuzzleBench project.

Citation

@article{patnala2026tddn,
      title={{TDDN}: {T}ext-aligned {D}iffused {D}INO {N}etwork for Puzzle Understanding}, 
      author={Harsha Patnala and Debopriyo Banerjee and Ayush Sunil Munot and Somak Aditya},
      year={2026},
      journal={arXiv:2609.07937}
      eprint={2609.07937},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.07937}, 
}