{
  "examples": 896,
  "source_video_count": 461,
  "qa_video_count": 461,
  "source_hashes": {
    "/home/jovyan/data-1/jusang/superdyn/motion_audit/outputs/temporal_geometry_fullscope/qa/final/qa.json": "4ed3eb2356fb055c6a5fdad1ba89ab233af2941dd343c4cbffe9f7bd9bb503ce"
  },
  "presentation": "All nine measurement RGB frames as ordered images, with frame labels and requested target contours only.",
  "targets": "Original QA A/B mapping retained; model receives only visual marks and original question/descriptions.",
  "excluded_from_model_input": [
    "reference answer",
    "answer index",
    "depth",
    "camera pose",
    "geometry evidence",
    "threshold-derived metrics"
  ],
  "generation": {
    "do_sample": false,
    "num_beams": 1,
    "max_new_tokens": 64
  },
  "group4_mapping": {
    "LLL": "LLL",
    "LLH": "LLH",
    "HLL": "HLL+LHL+HHL",
    "LHL": "HLL+LHL+HHL",
    "HHL": "HLL+LHL+HHL",
    "HLH": "HLH+LHH+HHH",
    "LHH": "HLH+LHH+HHH",
    "HHH": "HLH+LHH+HHH"
  },
  "group4_counts": {
    "HLH+LHH+HHH": 896
  },
  "scoring": "Deterministic MCQ matching; invalid/error predictions count as incorrect; denominator 896 per model.",
  "reference_status": "Provisional model-estimated geometry labels; no human review.",
  "model_input_policy": "Same marked images, question and option order for all models; native per-model visual preprocessing.",
  "scope": "Initial evaluation at nine frames; not the papers original benchmark protocol or a 16/32-frame setting.",
  "jobs": [
    {
      "model": "qwen25vl_3b",
      "gpu": 0,
      "shard": 0,
      "shards": 1
    },
    {
      "model": "qwen25vl_7b",
      "gpu": 1,
      "shard": 0,
      "shards": 1
    },
    {
      "model": "vgllm_qa_4b",
      "gpu": 2,
      "shard": 0,
      "shards": 1
    },
    {
      "model": "vgllm_qa_8b",
      "gpu": 3,
      "shard": 0,
      "shards": 2
    },
    {
      "model": "vgllm_qa_8b",
      "gpu": 4,
      "shard": 1,
      "shards": 2
    },
    {
      "model": "mllm4d_sft_8b",
      "gpu": 5,
      "shard": 0,
      "shards": 1
    },
    {
      "model": "4drgpt_8b",
      "gpu": 6,
      "shard": 0,
      "shards": 1
    },
    {
      "model": "dsr_suite_7b",
      "gpu": 7,
      "shard": 0,
      "shards": 1
    }
  ],
  "created_at": "2026-10-21T09:51:25.961939+00:00",
  "native_input_packaging": {
    "default": "nine ordered marked RGB images",
    "dsr_suite_7b": "Same nine marked RGB frames packaged as a video; native final-frame temporal padding and 32 GSM query tokens. fps=2 is an index encoding convention, not measured seconds."
  },
  "excluded_sources": [
    "dynamic_replica",
    "dynpose-100k",
    "point_odyssey",
    "spring"
  ],
  "additional_model_source": "https://github.com/TencentARC/DSR_Suite",
  "selection_rule": "QA audit decision=keep AND group in HLH/LHH/HHH (N=L)",
  "reused_evaluation": "/home/jovyan/data-1/jusang/superdyn/motion_audit/outputs/temporal_vlm_eval_fullscope"
}