{
  "name": "SatQuery Vision Encoder v1",
  "version": "1.0.0",
  "type": "remote-sensing-adapted vision-language encoder",
  "backbone": "OpenCLIP ViT-L-14",
  "pretrained_weights": "laion2b_s32b_b82k",
  "embedding_dimension": 768,
  "adaptation_dataset": "BigEarthNet.txt",
  "image_source": "BigEarthNet v2 Lithuania Summer",
  "training_pairs": 4008,
  "validation_pairs": 2291,
  "test_pairs": 2053,
  "input": {
    "mode": "RGB",
    "sentinel_2_bands": [
      "B04",
      "B03",
      "B02"
    ],
    "input_size": [
      224,
      224
    ]
  },
  "supported_uses": [
    "Earth-observation image embeddings",
    "image-text similarity",
    "caption reranking",
    "scene-prior evidence",
    "retrieval",
    "agent routing support"
  ],
  "unsupported_claims": [
    "pixel-level segmentation",
    "object grounding",
    "calibrated probability",
    "semantic change detection"
  ],
  "limitations": [
    "Training subset is geographically limited to Lithuania",
    "Training subset contains summer imagery only",
    "Version 1 supports RGB imagery only",
    "Retrieval metrics remain low in absolute terms",
    "Outputs must not be treated as ground truth"
  ],
  "validation_metrics": {
    "loss": 4.120570659637451,
    "image_to_text_r1": 0.005674378015100956,
    "text_to_image_r1": 0.006110868416726589,
    "image_to_text_r5": 0.022697512060403824,
    "text_to_image_r5": 0.03491925075650215,
    "image_to_text_r10": 0.050632912665605545,
    "text_to_image_r10": 0.05805325135588646
  },
  "generic_openclip_baseline": {
    "loss": 4.8016743659973145,
    "image_to_text_r1": 0.0021824531722813845,
    "text_to_image_r1": 0.0008729812107048929,
    "image_to_text_r5": 0.010039283894002438,
    "text_to_image_r5": 0.004364906344562769,
    "image_to_text_r10": 0.021388040855526924,
    "text_to_image_r10": 0.00654735928401351
  }
}