papers-data commited on
Commit
340e0ec
·
verified ·
1 Parent(s): 4571ad0

Rendered twin sets, per-scene model answers and dataset card

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. README.md +91 -0
  2. colorshape_twins.zip +3 -0
  3. colorshape_twins_hard.zip +3 -0
  4. gamebind_twins.zip +3 -0
  5. gamebind_twins_ext.zip +3 -0
  6. gamebind_twins_ext_api.zip +3 -0
  7. gamebind_twins_full.zip +3 -0
  8. per_scene/README.md +73 -0
  9. per_scene/colorshape/gpt-4.1__hard.csv.gz +3 -0
  10. per_scene/colorshape/internvl3_5-8b-hf__easy.csv.gz +3 -0
  11. per_scene/colorshape/internvl3_5-8b-hf__hard.csv.gz +3 -0
  12. per_scene/colorshape/llava-onevision-qwen2-7b-ov-hf__easy.csv.gz +3 -0
  13. per_scene/colorshape/llava-onevision-qwen2-7b-ov-hf__hard.csv.gz +3 -0
  14. per_scene/colorshape/qwen2.5-vl-7b-instruct__easy.csv.gz +3 -0
  15. per_scene/colorshape/qwen2.5-vl-7b-instruct__hard.csv.gz +3 -0
  16. per_scene/colorshape/qwen3-vl-32b-instruct__easy.csv.gz +3 -0
  17. per_scene/colorshape/qwen3-vl-32b-instruct__hard.csv.gz +3 -0
  18. per_scene/colorshape/qwen3-vl-8b-instruct__easy.csv.gz +3 -0
  19. per_scene/colorshape/qwen3-vl-8b-instruct__hard.csv.gz +3 -0
  20. per_scene/gpt-4.1__api_subset.csv.gz +3 -0
  21. per_scene/gpt-4o__api_subset.csv.gz +3 -0
  22. per_scene/gpt-5.5__api_subset.csv.gz +3 -0
  23. per_scene/internvl3_5-38b-hf__api_subset.csv.gz +3 -0
  24. per_scene/internvl3_5-38b-hf__full.csv.gz +3 -0
  25. per_scene/internvl3_5-8b-hf__full.csv.gz +3 -0
  26. per_scene/lfm2-vl-3b__full.csv.gz +3 -0
  27. per_scene/llava-onevision-qwen2-7b-ov-hf__full.csv.gz +3 -0
  28. per_scene/mistral-small-3.1-24b-instruct-2503__api_subset.csv.gz +3 -0
  29. per_scene/mistral-small-3.1-24b-instruct-2503__full.csv.gz +3 -0
  30. per_scene/models.csv +30 -0
  31. per_scene/molmo2-8b__full.csv.gz +3 -0
  32. per_scene/phi-4-multimodal-instruct__full.csv.gz +3 -0
  33. per_scene/photo_twins/gpt-4.1__keyframe.csv.gz +3 -0
  34. per_scene/photo_twins/internvl3_5-38b-hf__keyframe.csv.gz +3 -0
  35. per_scene/photo_twins/internvl3_5-8b-hf__keyframe.csv.gz +3 -0
  36. per_scene/photo_twins/mistral-small-3.1-24b-instruct-2503__keyframe.csv.gz +3 -0
  37. per_scene/photo_twins/qwen2-vl-7b-instruct__keyframe.csv.gz +3 -0
  38. per_scene/photo_twins/qwen2.5-vl-32b-instruct__keyframe.csv.gz +3 -0
  39. per_scene/photo_twins/qwen2.5-vl-72b-instruct__keyframe.csv.gz +3 -0
  40. per_scene/photo_twins/qwen2.5-vl-7b-instruct__keyframe.csv.gz +3 -0
  41. per_scene/photo_twins/qwen3-vl-32b-instruct__keyframe.csv.gz +3 -0
  42. per_scene/photo_twins/qwen3-vl-8b-instruct__blind.csv.gz +3 -0
  43. per_scene/photo_twins/qwen3-vl-8b-instruct__keyframe.csv.gz +3 -0
  44. per_scene/qwen2-vl-2b-instruct__full.csv.gz +3 -0
  45. per_scene/qwen2-vl-72b-instruct__full.csv.gz +3 -0
  46. per_scene/qwen2-vl-7b-instruct__api_subset.csv.gz +3 -0
  47. per_scene/qwen2-vl-7b-instruct__full.csv.gz +3 -0
  48. per_scene/qwen2.5-vl-32b-instruct__full.csv.gz +3 -0
  49. per_scene/qwen2.5-vl-3b-instruct__full.csv.gz +3 -0
  50. per_scene/qwen2.5-vl-72b-instruct__api_subset.csv.gz +3 -0
README.md ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - visual-question-answering
5
+ language:
6
+ - en
7
+ pretty_name: GameBind counterfactual twins
8
+ size_categories:
9
+ - 10K<n<100K
10
+ ---
11
+
12
+ # GameBind counterfactual twins
13
+
14
+ Rendered two-agent scenes and their counterfactual twins, released with the anonymous submission "Counterfactual
15
+ Twins Identify Perception and Binding Errors in Vision-Language Models".
16
+
17
+ Each base scene shows a target agent and a competitor agent in an 8x8 tile world. Each agent performs one action
18
+ (pick up, push, hit or go to) on one object. Every action leaves a visible end state: a picked-up object is carried,
19
+ a pushed object slides and leaves skid marks, a hit object cracks and leaves debris, and go to changes nothing. A twin
20
+ re-renders the same scene with one controlled change, and all other pixels stay the same. Comparing a model's answers
21
+ across the twins of a scene separates two causes of a wrong answer: missing the target's action and guessing, or
22
+ seeing it and attributing the competitor's action to the target.
23
+
24
+ ## Files
25
+
26
+ | file | contents | clips | questions | images |
27
+ |---|---|---|---|---|
28
+ | `gamebind_twins_full.zip` | 1,000 base scenes and 300 near-condition scenes with their main twins | 5,362 | 16,086 | 42,896 |
29
+ | `gamebind_twins_ext.zip` | further twins of the same 1,000 scenes | 2,318 | 6,954 | 18,544 |
30
+ | `gamebind_twins.zip` | the API subset: a separate draw of 300 base scenes and 150 near-condition scenes | 1,717 | 5,151 | 13,736 |
31
+ | `gamebind_twins_ext_api.zip` | further twins of the API subset | 677 | 2,031 | 5,416 |
32
+ | `colorshape_twins.zip` | colour-shape displays, 600 base scenes (easy display) | 4,200 | 4,200 | 3,601 |
33
+ | `colorshape_twins_hard.zip` | colour-shape displays, 600 base scenes (hard display) | 4,200 | 4,200 | 3,601 |
34
+ | `per_scene/` | every model's answers per scene and twin, from which all results of the paper are computed | | | |
35
+
36
+ Unzip all files in one folder. Each zip creates `data/<set>/` with:
37
+
38
+ - `clips.json`: one entry per clip, with the base scene, the twin, the events, the agents and the paths of its frames;
39
+ - `queries.json`: the questions asked on each clip, with their options and the correct answer;
40
+ - `checks.json`: the pixel checks of every twin;
41
+ - `summary.json`: the generator settings and the check summary;
42
+ - `frames/`: the images.
43
+
44
+ The frame paths in `clips.json` are relative to the folder where you unzip. A GameBind clip has 8 frames
45
+ (`f0.png` to `f7.png`, 512x512 pixels), and the models in the paper saw the last frame, `f7.png`. A colour-shape clip
46
+ is one 512x512 image.
47
+
48
+ ## Twins
49
+
50
+ | twin | change |
51
+ |---|---|
52
+ | `S` | base scene |
53
+ | `minus_c` | the competitor stays idle at its start tile |
54
+ | `minus_cue` | the target moves as in S, but its action leaves no trace |
55
+ | `minus_cue_minus_c` | both changes |
56
+ | `minus_ccue` | only the competitor's trace is deleted, the competitor still moves |
57
+ | `degraded_cue` | the target's trace is weakened |
58
+ | `swap` | the two agents exchange their events |
59
+ | `placebo` | an irrelevant change to the floor pixels |
60
+ | `S_repro` | a fresh rendering of S, used to check that rendering is exact |
61
+
62
+ In the colour-shape displays, the twins grey out the competitor shape (`minus_c`), the target shape (`minus_cue`) or
63
+ both, add a grey shape elsewhere (`placebo`), or exchange the two colours (`swap`). A blind grey image (`blind`) is
64
+ also included.
65
+
66
+ ## Questions
67
+
68
+ On every GameBind clip three questions are asked, each with four options. The joint question asks what the target
69
+ agent did, and its options are the true event, two answers that mix the target's and the competitor's action and
70
+ object, and the competitor's whole event. The object question and the action question ask for one part. Options are
71
+ shown under rotated orders, and all twins of a base scene share the same orders. The colour-shape question asks for
72
+ the colour of the target shape, with the target's colour, the competitor's colour and two absent colours as options.
73
+
74
+ ## Model answers
75
+
76
+ `per_scene/` holds one gzip CSV per model and scene set: 19 open models on all scenes, and GPT-4.1, GPT-4o, GPT-5.5
77
+ and seven open models on the API subset. Each row is one question about one twin of one base scene, with the answer
78
+ probability of every option. The subfolders `colorshape/` and `photo_twins/` hold
79
+ the answers on the colour-shape displays and on the real-photo twins (numbers only). `per_scene/README.md` describes
80
+ the columns.
81
+
82
+ ## Notes
83
+
84
+ - All images are rendered. The dataset contains no photographs and no personal data.
85
+ - The renderer regenerates every frame bit for bit. The code is linked from the paper.
86
+ - The API subset is a separate draw from the 1,000 scenes, although the scene names repeat. A scene is identified by
87
+ its set and its `base_id`.
88
+
89
+ ## Licence
90
+
91
+ CC BY 4.0.
colorshape_twins.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:33d9ca57f6090d5395f33c41126e2daf001c5d1e6e0bdb132b613dae841e68b9
3
+ size 18174964
colorshape_twins_hard.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b07cf497bbf0a2f5dee3369b1b386a7814b3a80b506da4729a6d8237ec1044bb
3
+ size 27647293
gamebind_twins.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92ee9813dc4638a9477ae04030dfb49c020326d494e26146571fe72009e3d4a4
3
+ size 90380758
gamebind_twins_ext.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f104cbbfa9edc08ed5414229c54d0f320cd04ac34abd7f58f037061a6dc5e477
3
+ size 120573842
gamebind_twins_ext_api.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c363d94bef38febd1b8031885ff6c571db1b31f2e5e75698e61f38611bc7439
3
+ size 35016804
gamebind_twins_full.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ffa074d1a93c4491df0f2dcbe6dc59cb6eb99835b89e4816c127e8990d8aa1c5
3
+ size 283697956
per_scene/README.md ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Per-scene rates
2
+
3
+ One gzip CSV per model and scene set (`<slug>__<scene_set>.csv.gz`) and an index (`models.csv`). Every row is one
4
+ scored question about one twin of one base scene: model x base scene x twin x question type. These rows are
5
+ everything Table 2, the data figures, the appendix tables, `numbers.tex` (the numbers quoted in the text), the
6
+ sensitivity analysis, the heterogeneity, scene-level and extra checks are computed from (`reproduce/reproduce.py`).
7
+ They were exported from the raw scoring outputs by `code/scripts/export_per_scene.py`
8
+ (`export_per_scene_photo.py`, `export_per_scene_colorshape.py` for the two subdirectories).
9
+
10
+ **Scene sets.** `full`: 1,000 main + 300 near base scenes (seed 424242). `api_subset`: 300 main + 150 near base
11
+ 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
12
+ scenes, although the base-scene names (`tw_main_g_0000`, ...) overlap: a scene is identified by
13
+ (`scene_set`, `base_id`), and every file holds exactly one scene set.
14
+
15
+ | file | scene set | twins | rows |
16
+ |---|---|---|---|
17
+ | 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 |
18
+ | `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 |
19
+ | `gpt-4o__api_subset.csv.gz` | `api_subset` | as GPT-4.1 | 6,486 |
20
+ | `gpt-5.5__api_subset.csv.gz` | 250 of the 300 main API-subset scenes | S, minus_c (binding question only) | 500 |
21
+ | 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 |
22
+ | `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 |
23
+ | `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 |
24
+
25
+ ## Columns
26
+
27
+ | column | meaning |
28
+ |---|---|
29
+ | `model` | display name used in the paper |
30
+ | `scene_set` | `full` (`data/gamebind_twins_full` + `data/gamebind_twins_ext`) or `api_subset` (`data/gamebind_twins` + `data/gamebind_twins_ext_api`) |
31
+ | `condition` | `main` (the two events far apart) or `near` (the target's two candidate objects 1-2 tiles apart) |
32
+ | `base_id` | base scene S (e.g. `tw_main_g_0000`; `g`/`t` = grid / textured skin) |
33
+ | `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) |
34
+ | `qtype` | `binding` ("What did the red agent do?"), `object`, `action` |
35
+ | `skin`, `target_agent` | skin of the scene; the queried (target) agent, `red` or `blue` |
36
+ | `target_action`, `target_object` | the target's event (a, o) in S |
37
+ | `competitor_action`, `competitor_object` | the competitor's event (a', o') in S |
38
+ | `primary` | 1 if neither agent's action is `go to`: the a priori primary subset of the processing tree |
39
+ | `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')) |
40
+ | `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] |
41
+ | `pred` | canonical index of the argmax of `p0..p3` (hard response) |
42
+ | `p_competitor` | P(answer names the competitor's action a'): `p2+p3` for binding, `p1` for action; for object P(answer names o') = `p1` |
43
+ | `correct` | 1 if `pred == answer` (empty where no answer is defined) |
44
+ | `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) |
45
+
46
+ Probabilities are written with full double precision so that the reproduction is bit-identical; the rows of each
47
+ model come from one scoring run (one GPU type, one batch size). The per-rotation probabilities are not released (only
48
+ their mean `p0..p3`); the analysis uses the mean and the argmax only. The photo-twin files have the same columns plus
49
+ `protocol`, `dataset` (vcoco / vg) and `photo_id`, without `scene_set`, `condition`, `skin` and `primary`.
50
+
51
+ The colour-shape files (`colorshape/`, question "What color is the <target shape>?") have the columns `model`,
52
+ `display` (`easy`: `data/colorshape_twins`, large saturated shapes, 1-2 grey distractors; `hard`:
53
+ `data/colorshape_twins_hard`, small shapes among 14 coloured distractors, colours at 15% saturation), `base_id`,
54
+ `twin`, `distance` (near / far), `target_shape`, `competitor_shape`, `target_color`, `competitor_color`, `answer`
55
+ (empty for minus_cue, minus_cue_minus_c and blind), `p0..p3` (canonical order [target colour, competitor colour,
56
+ absent, absent], mean over 4 rotations), `pred`, `lm0..lm3` (probability mass on the four option letters before
57
+ renormalisation, per rotation) and `n_rotations`.
58
+
59
+ ## Example: GPT-4.1's competitor effect and processing tree
60
+
61
+ ```python
62
+ import pandas as pd
63
+ d = pd.read_csv("gpt-4.1__api_subset.csv.gz")
64
+ d = d[d.condition == "main"]
65
+ b = d[d.qtype == "binding"].pivot(index="base_id", columns="twin", values="p_competitor")
66
+ print("kappa_A =", (b.S - b.minus_c).mean()) # all 300 scenes (Table 2: 15.5 pts)
67
+ prim = d[(d.qtype == "binding") & (d.primary == 1)].pivot(index="base_id", columns="twin", values="p_competitor")
68
+ prim = prim.dropna(subset=["S", "minus_c", "minus_cue", "minus_cue_minus_c"])
69
+ pi_S, pi_c, b1, b0 = (prim[t].mean() for t in ("S", "minus_c", "minus_cue", "minus_cue_minus_c"))
70
+ p_A = 1 - pi_c / b0
71
+ m_A = (pi_S - (1 - p_A) * b1) / p_A
72
+ 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
73
+ ```
per_scene/colorshape/gpt-4.1__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:670b58c5ecce1fef52f60a7bda987ff5e3fe11e53c13da7a5b62cae5426b4f93
3
+ size 198798
per_scene/colorshape/internvl3_5-8b-hf__easy.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be4855e41d87bc2084710353164bd85ae38a97761a927ce54cefe3ade5b372dd
3
+ size 308025
per_scene/colorshape/internvl3_5-8b-hf__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a44f58457025870237f52281a9fbcab63eac3f828d79be10d11d72e10e865ce
3
+ size 316102
per_scene/colorshape/llava-onevision-qwen2-7b-ov-hf__easy.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7b9e37a609be15d99c512854644f2e2c1cf7b8fe355cd7e28100e6e950e4069
3
+ size 321101
per_scene/colorshape/llava-onevision-qwen2-7b-ov-hf__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31e8bf11405976e613a367d4a716606222eb629db8d109a619df2b9b4def255b
3
+ size 325584
per_scene/colorshape/qwen2.5-vl-7b-instruct__easy.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd6824b9d980a7739e2d5243bf0584ee4645e0b1c6840818420db36aeb563bab
3
+ size 321084
per_scene/colorshape/qwen2.5-vl-7b-instruct__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32b54abc77491098c6bdd15fe28accadcadadaacb918fc4b326e8cda34683bf4
3
+ size 330099
per_scene/colorshape/qwen3-vl-32b-instruct__easy.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:74fb20fbabc1e91c101f53b7250aa334e5a28be843133147af110fe5fa331c1a
3
+ size 310898
per_scene/colorshape/qwen3-vl-32b-instruct__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5855b7796f24bd13491968acecb80b17680ceed2ec54d336388dc878553ebef
3
+ size 335732
per_scene/colorshape/qwen3-vl-8b-instruct__easy.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70a63c81d4d462a191f49974bc54e39f06a4a6c43f92f2bd46d3089869aa7b6e
3
+ size 319153
per_scene/colorshape/qwen3-vl-8b-instruct__hard.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:983913ed0defc96700613d8fdae2923d7706d1c38ff98ac53da5533c5c2f36bd
3
+ size 334924
per_scene/gpt-4.1__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f8619b242719af8e2872a734111985e32c81e62367b2c40048500ad8cd7391f1
3
+ size 328786
per_scene/gpt-4o__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab32526e3f8bbce5202d97230c2b370876bf68427c746853d5262bd66f51b036
3
+ size 341965
per_scene/gpt-5.5__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5268ef2901cf992a14f41d97a25792d37e250e9fde255c3effb0bcdd0a171c4f
3
+ size 3571
per_scene/internvl3_5-38b-hf__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31b45ba456e2e7033f7c303926c2e2f8d3fd38f92ea21408226ab0eafe61e103
3
+ size 341290
per_scene/internvl3_5-38b-hf__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13ce723b49169ca58687c663f477b8b630c3278067263719456c0832017dcea9
3
+ size 1201066
per_scene/internvl3_5-8b-hf__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:53bf699d92faae1e74c31d5f2b3ebe13558196ba389e7ba594a2db0c95c93503
3
+ size 1208168
per_scene/lfm2-vl-3b__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e31d81495b8c0e2ee9bc6a4a51c04c47bbe311740c5479fa5241821cbfe4cb82
3
+ size 1193759
per_scene/llava-onevision-qwen2-7b-ov-hf__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bd850c7085ec5077840932e763778e337268d66ee83b470e10615c1e4bd909a5
3
+ size 1192160
per_scene/mistral-small-3.1-24b-instruct-2503__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:764495f7f1529f5bfc5e9c4c69eb2c425332cbf37b1eb0cea101a603e4c6ca52
3
+ size 343144
per_scene/mistral-small-3.1-24b-instruct-2503__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99fe4f2b37734da308592d29af3e19930c706c43796daf38fbde91233b8a46f1
3
+ size 1205451
per_scene/models.csv ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model,slug,scene_set,file,n_rows,n_base_scenes,response
2
+ GPT-4.1,gpt-4.1,api_subset,gpt-4.1__api_subset.csv.gz,6486,450,next-token letter probabilities
3
+ GPT-4o,gpt-4o,api_subset,gpt-4o__api_subset.csv.gz,6486,450,next-token letter probabilities
4
+ Qwen2.5-VL-72B,qwen2.5-vl-72b-instruct,full,qwen2.5-vl-72b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
5
+ Qwen2-VL-72B,qwen2-vl-72b-instruct,full,qwen2-vl-72b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
6
+ Qwen3-VL-32B,qwen3-vl-32b-instruct,full,qwen3-vl-32b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
7
+ Qwen2.5-VL-32B,qwen2.5-vl-32b-instruct,full,qwen2.5-vl-32b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
8
+ Qwen3-VL-8B,qwen3-vl-8b-instruct,full,qwen3-vl-8b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
9
+ InternVL3.5-8B,internvl3_5-8b-hf,full,internvl3_5-8b-hf__full.csv.gz,23040,1300,next-token letter probabilities
10
+ Qwen2.5-VL-7B,qwen2.5-vl-7b-instruct,full,qwen2.5-vl-7b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
11
+ Qwen2-VL-7B,qwen2-vl-7b-instruct,full,qwen2-vl-7b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
12
+ LLaVA-OV-7B,llava-onevision-qwen2-7b-ov-hf,full,llava-onevision-qwen2-7b-ov-hf__full.csv.gz,23040,1300,next-token letter probabilities
13
+ Qwen2.5-VL-3B,qwen2.5-vl-3b-instruct,full,qwen2.5-vl-3b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
14
+ Qwen2-VL-2B,qwen2-vl-2b-instruct,full,qwen2-vl-2b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
15
+ SmolVLM2-2.2B,smolvlm2-2.2b-instruct,full,smolvlm2-2.2b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
16
+ InternVL3.5-38B,internvl3_5-38b-hf,full,internvl3_5-38b-hf__full.csv.gz,23040,1300,next-token letter probabilities
17
+ Qwen3-VL-30B-A3B,qwen3-vl-30b-a3b-instruct,full,qwen3-vl-30b-a3b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
18
+ Mistral-Small-3.1-24B,mistral-small-3.1-24b-instruct-2503,full,mistral-small-3.1-24b-instruct-2503__full.csv.gz,23040,1300,next-token letter probabilities
19
+ Qwen3-VL-4B,qwen3-vl-4b-instruct,full,qwen3-vl-4b-instruct__full.csv.gz,23040,1300,next-token letter probabilities
20
+ LFM2-VL-3B,lfm2-vl-3b,full,lfm2-vl-3b__full.csv.gz,23040,1300,next-token letter probabilities
21
+ Phi-4-multimodal,phi-4-multimodal-instruct,full,phi-4-multimodal-instruct__full.csv.gz,23040,1300,next-token letter probabilities
22
+ Molmo2-8B,molmo2-8b,full,molmo2-8b__full.csv.gz,23040,1300,next-token letter probabilities
23
+ GPT-5.5,gpt-5.5,api_subset,gpt-5.5__api_subset.csv.gz,500,250,vote (parsed answer per rotation)
24
+ Qwen2-VL-7B,qwen2-vl-7b-instruct,api_subset,qwen2-vl-7b-instruct__api_subset.csv.gz,6486,450,next-token letter probabilities
25
+ Qwen2.5-VL-7B,qwen2.5-vl-7b-instruct,api_subset,qwen2.5-vl-7b-instruct__api_subset.csv.gz,6486,450,next-token letter probabilities
26
+ Qwen3-VL-8B,qwen3-vl-8b-instruct,api_subset,qwen3-vl-8b-instruct__api_subset.csv.gz,6486,450,next-token letter probabilities
27
+ Qwen3-VL-32B,qwen3-vl-32b-instruct,api_subset,qwen3-vl-32b-instruct__api_subset.csv.gz,6486,450,next-token letter probabilities
28
+ Mistral-Small-3.1-24B,mistral-small-3.1-24b-instruct-2503,api_subset,mistral-small-3.1-24b-instruct-2503__api_subset.csv.gz,6486,450,next-token letter probabilities
29
+ InternVL3.5-38B,internvl3_5-38b-hf,api_subset,internvl3_5-38b-hf__api_subset.csv.gz,6486,450,next-token letter probabilities
30
+ Qwen2.5-VL-72B,qwen2.5-vl-72b-instruct,api_subset,qwen2.5-vl-72b-instruct__api_subset.csv.gz,6486,450,next-token letter probabilities
per_scene/molmo2-8b__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:296f72ec9ffa4bf6b46a6b36767b703479e87a83e579fe947fa171d250bdf43a
3
+ size 1202752
per_scene/phi-4-multimodal-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:470f903314f58ecd0bba36044d2acbaa058f249a25ca4c21f109d8e132064f1d
3
+ size 1199651
per_scene/photo_twins/gpt-4.1__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61682324703f730f212b4d0a3b95747c974fe48526966cda29e3cfce294769b3
3
+ size 121633
per_scene/photo_twins/internvl3_5-38b-hf__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7016fbb92791cd076abf1b27eb9ff3c225afffbd4ab2beff08ad8411f58d1466
3
+ size 331918
per_scene/photo_twins/internvl3_5-8b-hf__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48f76b740d9c41258dd2b4b69ec80b8223bf81b85adbc34e3baaac7f5d57f76d
3
+ size 334170
per_scene/photo_twins/mistral-small-3.1-24b-instruct-2503__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa190aafb3aaf5beec5becf36c95ed388846b9923938a0c07bfc69831b5e191c
3
+ size 331212
per_scene/photo_twins/qwen2-vl-7b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5d46a95b09ae6d61d2ffdd135a90a28defcb89353a4170b9a301c3d5ba1d243
3
+ size 327911
per_scene/photo_twins/qwen2.5-vl-32b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5411fabd8f3a6ba1df2f9436b7cc910ee7215545e205923e64fd188e07a5d9e1
3
+ size 330989
per_scene/photo_twins/qwen2.5-vl-72b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf1b01d3a857908618fe686b1e6d7919edaafb451b49f70c76fed58f6b34187c
3
+ size 331423
per_scene/photo_twins/qwen2.5-vl-7b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64caaf28aa56281ea6884137b799c81d6aa1e7ee673e0ed75e9c64604aba9413
3
+ size 330201
per_scene/photo_twins/qwen3-vl-32b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b3e53ae5e57dae368c9c0c66755e90d231d083e32e04fceeaed7a699ba6a6a75
3
+ size 335175
per_scene/photo_twins/qwen3-vl-8b-instruct__blind.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f02d6001dd2141e060fa3f9dd767b9f35514c6fbe2d26ff28f2083746b2caa1
3
+ size 76683
per_scene/photo_twins/qwen3-vl-8b-instruct__keyframe.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:833526fa72f19aa3b92229a7c1f8e3064f329637d63aa47a55c14555ba0f1791
3
+ size 335268
per_scene/qwen2-vl-2b-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7220c2f99999a4af45545a009f96ea4d5ad780191a839240ffd0c540cd4a6ede
3
+ size 1181417
per_scene/qwen2-vl-72b-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63dc6bc38d2757a98c5f0fba8274d61f7acdddd98bacb9dbf6e23d0e1b88a254
3
+ size 1189277
per_scene/qwen2-vl-7b-instruct__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bfa3daf5db81a614fc1576488d0c0bc2f803e8977749763fb5c4be0c90e80e61
3
+ size 337218
per_scene/qwen2-vl-7b-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2860760f5843b9cf842a71ae9d415accc26314fd05e2b44b99cac4c74b10d846
3
+ size 1186488
per_scene/qwen2.5-vl-32b-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d04bc45a2b4ad9d7fd161849cc8b344574b57667760b89b9aa65c68a112c192
3
+ size 1201006
per_scene/qwen2.5-vl-3b-instruct__full.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ba310baba3f94b844250cac16eef27755c5fb2c3a0a5cf29e1b0fc8cf93f73d
3
+ size 1184177
per_scene/qwen2.5-vl-72b-instruct__api_subset.csv.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8cc6aaa2729565e4e6e24c7c18c440d4708b9ab4979208d8a8b7d63e6e930ea6
3
+ size 339681