scripts/results-vision-everything-512k.json

{
 "model": "glm-max",
 "correct": 10,
 "total": 10,
 "gate": 8,
 "verdict": "PASS",
 "probes": [
  {
   "file": "img01.png",
   "question": "What color is the circle in this image? Answer with one word.",
   "expect": [
    "red"
   ],
   "answer": "Red",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 5.7
  },
  {
   "file": "img02.png",
   "question": "What word is written in this image? Answer with just the word.",
   "expect": [
    "saturn"
   ],
   "answer": "SATURN",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 0.3
  },
  {
   "file": "img03.png",
   "question": "How many squares are in this image? Answer with just the number.",
   "expect": [
    "4",
    "four"
   ],
   "answer": "4",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 1.0
  },
  {
   "file": "img04.png",
   "question": "What shape is shown in this image? Answer with one word.",
   "expect": [
    "triangle"
   ],
   "answer": "Triangle",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 283,
   "wall_s": 0.6
  },
  {
   "file": "img05.png",
   "question": "What number is written in this image? Answer with just the digits.",
   "expect": [
    "7318"
   ],
   "answer": "7318",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 0.3
  },
  {
   "file": "img06.png",
   "question": "How many circles (not squares) are in this image? Answer with just the number.",
   "expect": [
    "3",
    "three"
   ],
   "answer": "3",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 288,
   "wall_s": 1.2
  },
  {
   "file": "img07.png",
   "question": "Which is larger in this image, the circle or the square? Answer with one word.",
   "expect": [
    "circle"
   ],
   "answer": "Circle",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 288,
   "wall_s": 0.7
  },
  {
   "file": "img08.png",
   "question": "What word is written in this image? Answer with just the word.",
   "expect": [
    "mango"
   ],
   "answer": "MANGO",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 0.3
  },
  {
   "file": "img09.png",
   "question": "How many dots are in this image? Answer with just the number.",
   "expect": [
    "5",
    "five"
   ],
   "answer": "5",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 284,
   "wall_s": 0.8
  },
  {
   "file": "img10.png",
   "question": "Is the black square on the left side or the right side of this image? Answer left or right.",
   "expect": [
    "left"
   ],
   "answer": "The black square is on the **left** side of the image.",
   "correct": true,
   "image_tokens": 258,
   "prompt_tokens": 291,
   "wall_s": 0.5
  }
 ]
}