vh
ce09ac4fa6
surface_test.py's vision check is one image and one word. It proves the tower loads; it does not prove the tower works. This battery uses generated images with known ground truth so every answer is objectively gradeable. Against orcarouter NVFP4-mixed on the `gen` alias: T1 OCR, 5 lines incl. one at 18px PASS all 5 exact T2 counting + attribute binding PASS 7 circles / 3 triangles / 1 square T3 bar chart, 6 values + max/min PASS 6/6 exact T4b occlusion, star behind rectangle PASS T4c aspect ratio of a 160x140 rectangle FAIL called it taller than wide T5 two images, which has text PASS T6 four images, the seat's cap PASS all four named T7 five images, one over the cap PASS rejected with HTTP 400 No <think> leak on any vision call. The single miss is fine-grained relative-dimension estimation on a near-square shape, and it reproduced across two runs (the longer T4 called the same rectangle "equal width and height"). Counting, OCR, chart values and occlusion ordering are all solid, so this is a precise-geometry weakness, not a broken tower. Recorded so nobody builds a feature on this model judging relative sizes. T7 earns its place separately: it confirms the per-prompt image cap fails loudly with a 400 rather than silently dropping the extra image.
38 lines
2.1 KiB
Markdown
38 lines
2.1 KiB
Markdown
# Vision battery for a VL gen seat
|
|
|
|
The `surface_test.py` vision check is one image and one word ("Blue"). It proves
|
|
the tower loads; it does not prove the tower *works*. This battery does, against
|
|
images generated with known ground truth so every answer is objectively gradeable.
|
|
|
|
Generate the fixtures with the PIL snippet in this repo's history (or any images
|
|
whose content you know exactly), then:
|
|
|
|
uv run --with requests python vistest.py <image-dir> gen
|
|
uv run --with requests python vis2.py <image-dir>
|
|
|
|
## Results — orcarouter NVFP4-mixed on the `gen` alias, 2026-08-21
|
|
|
|
| test | what it exercises | result |
|
|
|---|---|---|
|
|
| T1 OCR | 5 lines, mixed case, digits, punctuation, one line at 18px | **PASS** — all 5 exact |
|
|
| T2 counting + attribute binding | 7 yellow circles / 3 purple triangles / 1 red square, scattered | **PASS** — 7/3/1 |
|
|
| T3 chart reading | 6 labelled bars, plus highest/lowest | **PASS** — 6/6 values, max/min right |
|
|
| T4b occlusion | a green star partly behind a grey rectangle | **PASS** |
|
|
| T4c aspect ratio | a 160x140 rectangle: wider, taller, or square? | **FAIL** — said taller |
|
|
| T5 two images | describe each, say which has text | **PASS** |
|
|
| T6 four images | the seat's `--limit-mm-per-prompt` ceiling | **PASS** — all 4 named |
|
|
| T7 five images | one over the cap | **PASS** — rejected with HTTP 400 |
|
|
|
|
7 of 8. No `<think>` leak on any vision call.
|
|
|
|
**The one miss is worth keeping.** T4c is fine-grained aspect-ratio estimation on a
|
|
near-square shape (160x140, a 14% difference); the model called it taller, and in
|
|
the longer T4 run it called the same shape "a rectangle with equal width and
|
|
height". Counting, OCR down to 18px, chart values and occlusion ordering are all
|
|
solid, so this is a precise-geometry weakness rather than a broken tower. **Do not
|
|
build a feature on this model's estimate of relative dimensions** — ask it what
|
|
shapes are present and where, not how big they are relative to each other.
|
|
|
|
T7 is worth keeping for a different reason: it confirms the per-prompt image cap
|
|
fails **loudly** with a 400 rather than silently dropping the fifth image.
|