Buckets:
| { | |
| "description": "Where the flag object sits in the rendered image, in both pixel space and VLM merged-image-token space. All 84 images in images/ share this exact geometry (single fixed template render: 336x336 canvas, flag placed at (112,140) sized 112x56) -- this is one shared config, not per-image, because only the flag's pixel content varies between images, never its position or size.", | |
| "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" | |
| }, | |
| "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 flag -- verified by diffing pixel_values between two images differing only inside the flag region." | |
| }, | |
| "object_token_indices": { | |
| "entity_only": { | |
| "description": "The 8 merged tokens the flag itself 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] | |
| } | |
| }, | |
| "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/animals/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": "Smallest whole-cell 16x16-grid box CONTAINING object_bbox_px (112,140,112,56) after the 336->448 resize (rows 6-9, cols 5-10) -- not a pixel-exact cover, same caveat as animals' InternVL3 entry (336/448 resize doesn't align with this grid's 28px-in-resized-space cell stride). Verified by resizing two 336x336 synthetic images differing only inside object_bbox_px through the real InternVL3-8B-hf image processor and diffing pixel_values: changed pixels span [185,262]x[147,300] in the resized 448x448 image, matching the predicted [186.67,261.33)x[149.33,298.67) up to the bicubic resample filter's edge blur. No separate entity_only variant, same reasoning as animals.", | |
| "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:
- 3.37 kB
- Xet hash:
- 56e87728c460e39fd59544b8da0fa1dbde3e6fdf8fc91eaf16a726b726120dda
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.