{
  "name": "GQA-Spatial",
  "release_date": "2019-02-25",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "vqa",
      "prompt_template": "<image>{{ question }}\\nAnswer the question using a single word or phrase.",
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/gqa/utils.py (gqa_doc_to_text — canonical short-answer)",
        "notes": "Tier 4: lmms-eval GQA canonical evaluation prompt, applied unchanged to the spatial subset of GQA testdev_balanced. The SAT paper (arXiv 2412.07755) reports 'GQA-Sp (Hudson & Manning, 2019b)' without specifying the spatial filter; this subset is built by filtering official testdev_balanced rows whose types.semantic=='rel' OR types.detailed.startswith('position'), yielding 5952 spatial-relation rows. Filter rule and provenance documented here for auditability."
      },
      "mapping_from_source": {
        "media": {
          "from": "media",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "id"
        },
        "question": {
          "from": "question"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "imageId": {
            "from": "imageId"
          },
          "fullAnswer": {
            "from": "fullAnswer"
          },
          "structural": {
            "from": "structural"
          },
          "semantic_type": {
            "from": "semantic_type"
          },
          "detailed": {
            "from": "detailed"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "testdev": "https://huggingface.co/datasets/lmms-lab/GQA (testdev_balanced; spatial-relation filter applied — see prompt_template_source notes)"
          }
        }
      }
    }
  }
}