nace-ai/nace-grounding-v2-4b

Prompt (variant B)

Extract the answer text in this image crop for:
"{focus}"

Return exactly one field named "answer" with three keys: value, raw_text, bbox.

Output:

{
  "answer": {
    "value": "normalized_answer_text",
    "raw_text": "exact_text_as_it_appears_in_image",
    "bbox": [x1, y1, x2, y2]
  }
}

Rules:
- "bbox": [x1, y1, x2, y2] in range 0..1000 relative to this image crop.
- If the answer is not in the visible crop, return null for value, raw_text, and bbox.
- Do not infer or guess missing data.

({focus} = target phrase for the crop.)

Eval (handin bench, n=22, prompt B)

Model	IoU ≥ 0.5	Mean IoU
Qwen3-VL-4B (base)	10/22	0.469
nace-grounding-v2-4b	17/22	0.702
VGSE-32B	16/22	0.607