papers-data's picture
Rendered twin sets, per-scene model answers and dataset card
340e0ec verified
|
Raw History Blame Contribute Delete
6.43 kB

Per-scene rates

One gzip CSV per model and scene set (<slug>__<scene_set>.csv.gz) and an index (models.csv). Every row is one scored question about one twin of one base scene: model x base scene x twin x question type. These rows are everything Table 2, the data figures, the appendix tables, numbers.tex (the numbers quoted in the text), the sensitivity analysis, the heterogeneity, scene-level and extra checks are computed from (reproduce/reproduce.py). They were exported from the raw scoring outputs by code/scripts/export_per_scene.py (export_per_scene_photo.py, export_per_scene_colorshape.py for the two subdirectories).

Scene sets. full: 1,000 main + 300 near base scenes (seed 424242). api_subset: 300 main + 150 near base scenes (seed 31337), scored with the API models and seven open models. The API subset is a separate draw, not a subset of the 1,000 scenes, although the base-scene names (tw_main_g_0000, ...) overlap: a scene is identified by (scene_set, base_id), and every file holds exactly one scene set.

file scene set twins rows
19 open models (qwen2.5-vl-7b-instruct__full.csv.gz, ...) full main: S, minus_c, minus_cue, swap, placebo, minus_cue_minus_c, minus_ccue, degraded_cue, S_repro; near: S, minus_c 23,040 each
gpt-4.1__api_subset.csv.gz api_subset main: S, minus_c, minus_cue, swap, placebo, minus_cue_minus_c, minus_ccue; near: S, minus_c 6,486
gpt-4o__api_subset.csv.gz api_subset as GPT-4.1 6,486
gpt-5.5__api_subset.csv.gz 250 of the 300 main API-subset scenes S, minus_c (binding question only) 500
7 open models on the API subset (Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-8B, Qwen3-VL-32B, Mistral-Small-3.1-24B, InternVL3.5-38B, Qwen2.5-VL-72B: <slug>__api_subset.csv.gz) api_subset as GPT-4.1 6,486 each
photo_twins/*.csv.gz 463 real-photo target instances S, minus_c_grey, minus_c_blur, placebo_grey, placebo_blur 6,237 per model; blind run 1,389
colorshape/<slug>__{easy,hard}.csv.gz 600 colour-shape base scenes per display (5 open models) S, minus_c, minus_cue, minus_cue_minus_c, placebo, swap, blind 4,200 per model and display

Columns

column meaning
model display name used in the paper
scene_set full (data/gamebind_twins_full + data/gamebind_twins_ext) or api_subset (data/gamebind_twins + data/gamebind_twins_ext_api)
condition main (the two events far apart) or near (the target's two candidate objects 1-2 tiles apart)
base_id base scene S (e.g. tw_main_g_0000; g/t = grid / textured skin)
twin S (base scene), minus_c (competitor idle), minus_cue (target's action leaves no trace), swap (events exchanged between agents), placebo (task-irrelevant floor pixels), minus_cue_minus_c (double deletion), minus_ccue (competitor's action leaves no trace), degraded_cue (target's trace at reduced strength), S_repro (an exact re-rendered copy of S, scored in the extension jobs: a scoring-calibration check)
qtype binding ("What did the red agent do?"), object, action
skin, target_agent skin of the scene; the queried (target) agent, red or blue
target_action, target_object the target's event (a, o) in S
competitor_action, competitor_object the competitor's event (a', o') in S
primary 1 if neither agent's action is go to: the a priori primary subset of the processing tree
answer canonical index of the correct option in this twin (empty where no answer is defined: minus_cue, minus_cue_minus_c; swap: the target now performs (a', o'))
p0..p3 answer probabilities in canonical option order, averaged over the option rotations (soft response). binding: [(a,o), (a,o'), (a',o), (a',o')]; object: [o, o', fill, fill]; action: [a, a', other, other]
pred canonical index of the argmax of p0..p3 (hard response)
p_competitor P(answer names the competitor's action a'): p2+p3 for binding, p1 for action; for object P(answer names o') = p1
correct 1 if pred == answer (empty where no answer is defined)
n_rotations option rotations averaged: 4 (open models, next-token letter probabilities), 2 (GPT-4.1 and GPT-4o, top-20 log-probabilities of the answer letter; GPT-5.5, one parsed answer per rotation, so its soft response is a vote share)

Probabilities are written with full double precision so that the reproduction is bit-identical; the rows of each model come from one scoring run (one GPU type, one batch size). The per-rotation probabilities are not released (only their mean p0..p3); the analysis uses the mean and the argmax only. The photo-twin files have the same columns plus protocol, dataset (vcoco / vg) and photo_id, without scene_set, condition, skin and primary.

The colour-shape files (colorshape/, question "What color is the ?") have the columns model, display (easy: data/colorshape_twins, large saturated shapes, 1-2 grey distractors; hard: data/colorshape_twins_hard, small shapes among 14 coloured distractors, colours at 15% saturation), base_id, twin, distance (near / far), target_shape, competitor_shape, target_color, competitor_color, answer (empty for minus_cue, minus_cue_minus_c and blind), p0..p3 (canonical order [target colour, competitor colour, absent, absent], mean over 4 rotations), pred, lm0..lm3 (probability mass on the four option letters before renormalisation, per rotation) and n_rotations.

Example: GPT-4.1's competitor effect and processing tree

import pandas as pd
d = pd.read_csv("gpt-4.1__api_subset.csv.gz")
d = d[d.condition == "main"]
b = d[d.qtype == "binding"].pivot(index="base_id", columns="twin", values="p_competitor")
print("kappa_A =", (b.S - b.minus_c).mean())                       # all 300 scenes (Table 2: 15.5 pts)
prim = d[(d.qtype == "binding") & (d.primary == 1)].pivot(index="base_id", columns="twin", values="p_competitor")
prim = prim.dropna(subset=["S", "minus_c", "minus_cue", "minus_cue_minus_c"])
pi_S, pi_c, b1, b0 = (prim[t].mean() for t in ("S", "minus_c", "minus_cue", "minus_cue_minus_c"))
p_A = 1 - pi_c / b0
m_A = (pi_S - (1 - p_A) * b1) / p_A
print(f"n={len(prim)} p_A={p_A:.2f} m_A={m_A:.3f}")                # n=145 p_A=0.82 m_A=0.153