ovindu-a/counterfact-data / data /brands /object_location.json
ovindu-a's picture
download
raw
4.87 kB
{
"description": "Where the logo object sits in the rendered image, in both pixel space and VLM merged-image-token space. All 128 images in images/ share this exact geometry (single fixed template render: 336x336 canvas, logo mark fitted inside a 112x56 box placed at (112,140)) -- one shared config, not per-image, because only the mark's pixel content varies between images, never its box position or size. This is FLAGS' box, adopted on 2026-09-05 in place of the 168x168 box brands previously shared with animals: an interchange intervention copies token positions, so the box size IS the intervention size, and 36 tokens (25% of the image) is a far blunter edit than flags' 8 (5.6%) -- enough that Cause was already high in early layers. Matching flags makes the two entities' Cause/Iso curves comparable. Aspect ratio is PRESERVED inside the box, so a wide wordmark renders short (median mark height 33 px, minimum 8 px) and some box cells are background for a given brand; per-image rendered sizes are in build/aligned_boxes_flagbox.json. The previous 168-box images are kept in build/_archive_box168/images/.",
"canvas_px": {
"width": 336,
"height": 336
},
"object_bbox_px": {
"x": 112,
"y": 140,
"width": 112,
"height": 56,
"note": "top-left origin, [x, x+width) x [y, y+height) half-open. Longest side of the logo mark is fit inside this box; remainder padded with the canvas's neutral gray (128, 128, 128), not stretched. Verified: no non-background pixel in any of the 128 images falls outside this box."
},
"vlm_token_grid": {
"model": "Qwen/Qwen2.5-VL-7B-Instruct",
"patch_size_px": 14,
"merge_size": 2,
"merged_patch_size_px": 28,
"grid_rows": 12,
"grid_cols": 12,
"n_tokens": 144,
"flat_index_formula": "row * grid_cols + col",
"note": "object_bbox_px falls exactly on merged-patch-grid boundaries (112/28=4, 224/28=8, 140/28=5, 196/28=7), so the token sets below are an exact, non-partial cover of the object's box -- identical indices to flags."
},
"object_token_indices": {
"entity_only": {
"description": "The 8 merged tokens the entity box occupies (rows 5-6, cols 4-7).",
"rows": [
5,
6
],
"cols": [
4,
5,
6,
7
],
"flat": [
64,
65,
66,
67,
76,
77,
78,
79
]
},
"entity_ring1": {
"description": "entity_only dilated by one merged-token cell in every direction (rows 4-7, cols 3-8) -- entity_only is a strict subset.",
"rows": [
4,
5,
6,
7
],
"cols": [
3,
4,
5,
6,
7,
8
],
"flat": [
51,
52,
53,
54,
55,
56,
63,
64,
65,
66,
67,
68,
75,
76,
77,
78,
79,
80,
87,
88,
89,
90,
91,
92
]
}
},
"blank_control_image": "images/_BLANK.png",
"models": {
"InternVL3-8B-hf": {
"vlm_token_grid": {
"model": "OpenGVLab/InternVL3-8B-hf",
"patch_size_px": 14,
"merge_size": 2,
"merged_patch_size_px": 28,
"grid_rows": 16,
"grid_cols": 16,
"n_tokens": 256,
"flat_index_formula": "row * grid_cols + col",
"note": "Same model-constant-grid rationale as data/flags/object_location.json's InternVL3-8B-hf entry: this model resizes every image to a fixed 448x448 tile (crop_to_patches=False) before patchifying, so the grid doesn't scale with this entity's 336x336 canvas the way Qwen2.5-VL's does."
},
"object_token_indices": {
"entity_ring1": {
"description": "Same derivation as flags's InternVL3-8B-hf entry (identical canvas and object_bbox_px (112,140,112,56)): smallest whole-cell 16x16-grid box CONTAINING object_bbox_px after the 336->448 resize (rows 6-9, cols 5-10). Not a pixel-exact cover; no separate entity_only variant. Copied from flags rather than re-verified against the image processor.",
"rows": [
6,
7,
8,
9
],
"cols": [
5,
6,
7,
8,
9,
10
],
"flat": [
101,
102,
103,
104,
105,
106,
117,
118,
119,
120,
121,
122,
133,
134,
135,
136,
137,
138,
149,
150,
151,
152,
153,
154
]
}
}
}
}
}

Xet Storage Details

Size:
4.87 kB
·
Xet hash:
fe4403fb684d0216e9ee4f26c12bd32301058bf6cc396c18770a638a8f35ae24

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.