87cc2s/ans-benchmark

ANS Benchmark

A cognitively-grounded benchmark for numerical cognition in vision-language models, covering exact counting and approximate numerosity discrimination (ANS).

Each row is a self-contained trial: the stimulus image is embedded inline, and prompt holds the exact question text sent to the model during evaluation.

Configs (6)

Config	Task	Rows	Stimuli
exp1_counting	Count items in one image	1,994	Abstract shapes + real-world (FSC-133), small-n & large-n
exp2_discrimination	Which side has more? (abstract, random sizes)	480	Pre-stitched pairs, 6 ratio conditions
exp3_controlled	Which side has more? (numerosity decorrelated from area/spacing)	2,500	Pre-stitched pairs, 3 shapes × 4 colors
exp4_realworld_naming	Name the object category	864	Single real-world panels
exp4_realworld_discrimination	Which side has more? (real-world)	600	Pre-stitched same/diff-category pairs
exp5_synthetic_matched	Which side has more? (abstract matched to real-world)	1,164	Pre-stitched pairs, counts/coords matched to Exp4

Every paired discrimination image is left | 10 px white divider | right.

Common columns (all configs)

Column	Type	Description
image	Image	Stimulus image (embedded bytes, loads as PIL)
prompt	string	Exact question text used in evaluation
task	string	Task type — see values below
tag	string	Analysis tag (ans, ans2, ans3, ans4, ans5)
eval_id	int	Global sequential trial index within each tag
answer / correct_answer	string	Ground-truth label (left/right/tie or integer count or category name)

task values:

Value	Description
counting	Exact count of objects
discrimination_abstract	Which side has more? — abstract shapes, unconstrained
discrimination_controlled	Which side has more? — numerosity decorrelated from surface area
discrimination_realworld	Which side has more? — real-world object images
discrimination_synthetic	Which side has more? — synthetic shapes matched to real-world stimuli
naming	Identify the object category

Config-specific columns

exp1_counting — subset, stimulus_type {abstract, realworld}, n_range {small_n, large_n}, shape, category, count

exp2_discrimination — pair_id, shape {dots, rectangles}, anchor_n, ratio_condition, ratio_value, left_count, right_count, correct_answer

exp3_controlled — pair_num, pair_shape {dots, ellipse, rectangle}, pair_color, pair_ratio, left_numerosity, right_numerosity, left_TSA, right_TSA, left_FA, right_FA, correct_answer (TSA = total surface area; FA = field area / convex hull)

exp4_realworld_naming — category, count, original_image_name

exp4_realworld_discrimination — pair_num, condition {same_cat, diff_cat}, left_category, right_category, left_count, right_count, num_ratio, answer

exp5_synthetic_matched — pair_num, condition, shape_type {dots, rectangles}, left_category, right_category, left_count, right_count, num_ratio, answer

Prompt templates

5 semantically-equivalent templates per task, assigned one per stimulus deterministically (md5 hash of stimulus id). Each image is tested exactly once — prompts rotate across stimuli, not within.

• Counting: "How many {item} are there in the image? Please answer with an integer directly." (+ 4 variants)
• Discrimination abstract: "Two images of {desc} appear side by side. Which image contains more shapes? Respond with 'left' or 'right'." (+ 4 variants)
• Discrimination same-category: "Look at these two images. Which image has more {cat}? … Answer with 'left' or 'right'." (+ 2 variants)
• Discrimination diff-category: "Left image contains {lcat}, right image contains {rcat}. Which image has more items? … Answer with 'left' or 'right'." (+ 2 variants)
• Naming: "What is(are) the object(s) in the image called? Please answer with one object category name directly." (+ 4 variants)

Quick start

from datasets import load_dataset

# Load one config
ds = load_dataset("87cc2s/ans-benchmark", "exp1_counting")
row = ds["train"][0]
row["image"]   # PIL Image
row["prompt"]  # "How many dots are there in the image? ..."
row["answer"]  # "1"
row["task"]    # "counting"

# Load discrimination
ds2 = load_dataset("87cc2s/ans-benchmark", "exp2_discrimination")
row2 = ds2["train"][0]
row2["prompt"]          # "Two images of white dots appear side by side ..."
row2["correct_answer"]  # "right"
row2["task"]            # "discrimination_abstract"

Provenance & notes

• Derived from original Experiment1–Experiment4, Experiment1_1, Experiment2_1, Experiment3_1 stimulus sets.
• Excluded: Exp1 random_largen-in-small-n bucket, real-world FSC_sampled/FSC_sampled_largen, Exp3 ellipses_white/rectangles_white (rendered incorrectly — replaced by Exp3_1 re-renders). 36 synthetic pairs with missing source images and 1 exact-duplicate counting row were dropped.
• License: Real-world images derive from FSC-133 (Ranjan et al., 2021) / FSC-147. Abstract stimuli are procedurally generated and freely usable. Respect upstream FSC licensing.