{
  "artifact": "fusion-embedding-2-ember",
  "version": "0.1-preview",
  "pack": {
    "gate": "thermal",
    "rank": 384,
    "adapters": 28,
    "d_model": 2048,
    "parameters": 44154880,
    "dtype": "float32",
    "init": "zero-initialized up projection (identity at attach time)",
    "attach_point": "decoder layers (language_model.layers), forward hooks"
  },
  "attaches_to": {
    "base": "Qwen/Qwen3-VL-Embedding-2B",
    "fe2": "EximiusLabs/fusion-embedding-2-2b-preview",
    "fe2_revision": "v0.2-preview",
    "note": "the decoder is byte-frozen and shared between the base and fusion-embedding-2, so the pack attaches to either"
  },
  "readout": {
    "thermal_instruction": "Represent this thermal infrared image.",
    "doc_instruction": "Represent the user's input.",
    "chat_template": "<|im_start|>system\\n{instruction}<|im_end|>\\n<|im_start|>user\\n{input}<|im_end|>\\n<|im_start|>assistant\\n",
    "image_user_content": "<|vision_start|><|image_pad|><|vision_end|>",
    "pooling": "last non-pad token, L2-normalized",
    "thermal_input": "single-channel thermal replicated to 3 channels, vision path, max 1310720 pixels",
    "trained_logit_scale": 2.8418076038360596
  },
  "training": {
    "objective": "contrastive thermal->caption (InfoNCE, 1024 bank negatives)",
    "corpus": "IR-TD early access minus FLIR-signature images",
    "train_pairs": 61320,
    "steps": 3900,
    "batch": 16,
    "seed": 2
  },
  "provenance": {
    "flir_strip": {
      "rule": "size == 640x512",
      "excluded": 20964,
      "of": 84284,
      "list": "release_strip_640x512.json",
      "sha256": "65a870df5e4fe0008fd9bacd9fa81bd0c47f919779a28e398fea8d1bcabad09f"
    },
    "llvip": "evaluation only",
    "dedup": "phash hamming<=4 vs LLVIP test: 0 hits"
  }
}