{
  "name": "PhyX",
  "release_date": "2025-01-01",
  "subsets": {
    "mc": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match",
        "llm-judge"
      ],
      "score_protocol": {
        "reference": "VLMEvalKit@vlmeval/dataset/utils/phyx.py:333-373 (PhyX_process_line_MC: regex letter extraction + case-insensitive string compare) and :214-253 (PhyX_auxeval_MC: rule match first; LLM semantic 1/0 judgement only for responses not resolved to a clean option letter). Authors' recommended flow (repo points to VLMEvalKit; --judge deepseek, image_vqa.py:3441-3448); lmms-eval@lmms_eval/tasks/phyx/phyx.yaml (quick_extract: false, eval_model_name deepseek-chat) uses the same PhyX_auxeval_MC path. [Re-verified in clone 2026-07-07: PhyX_process_line_MC at phyx.py:336-375, PhyX_auxeval_MC at :218-254 (letter mismatch scored 0 without judge at :236-238), evaluate valid_type assert at image_vqa.py:3448-3455, phyx.yaml:9-10.]",
        "note": "Official judge model is deepseek-chat. Rule-extracted letters that mismatch are scored 0 without a judge call; the judge only handles non-letter/verbose responses. mm-eval ships only the MC subset (phyx_mc + phyx_mini_mc splits); official OE subset and its text-only variants are not published. NOTE: phyx_mini_mc's 1000 ids are fully contained in phyx_mc's 3000 — pooling both splits double-counts the mini rows; evaluate one split at a time."
      },
      "prompt_template": "<image>{{ question }}",
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "index"
        },
        "question": {
          "from": "question"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "category": {
            "from": "category"
          },
          "subfield": {
            "from": "subfield"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "phyx_mc": "https://huggingface.co/datasets/Cloudriver/PhyX",
            "phyx_mini_mc": "https://huggingface.co/datasets/Cloudriver/PhyX"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/phyx/utils.py (phyx_doc_to_text — bare question; multi-choice variant)",
        "notes": "Tier 4: lmms-eval PhyX canonical evaluation prompt."
      }
    }
  }
}