{
  "kind": "decision",
  "format": "encoder",
  "model": "convaiinnovations/laya",
  "subfolder": "multilingual",
  "revision": "1c5edc17a7acd8701df6fc341c0d179f1c62c982",
  "variants": {
    "macos": {
      "path": "macos/wfp16-s256",
      "sizeMB": 681,
      "format": "aimodel"
    },
    "macos-s512": {
      "path": "macos/wfp16-s512",
      "sizeMB": 682,
      "format": "aimodel"
    },
    "ios": {
      "path": "ios/wfp16-s256",
      "sizeMB": 680,
      "format": "aimodel",
      "minOS": "27.0"
    },
    "ios-s512": {
      "path": "ios/wfp16-s512",
      "sizeMB": 681,
      "format": "aimodel",
      "minOS": "27.0"
    },
    "ios-h18p": {
      "path": "ios-h18p/wfp16-s256",
      "sizeMB": 680,
      "format": "aimodelc",
      "minOS": "27.0",
      "architecture": "h18p"
    },
    "ios-h18p-s512": {
      "path": "ios-h18p/wfp16-s512",
      "sizeMB": 681,
      "format": "aimodelc",
      "minOS": "27.0",
      "architecture": "h18p"
    },
    "macos-fp32": {
      "path": "macos/fp32-s256",
      "sizeMB": 1324,
      "format": "aimodel",
      "reference": true
    },
    "macos-fp32-s512": {
      "path": "macos/fp32-s512",
      "sizeMB": 1325,
      "format": "aimodel",
      "reference": true
    }
  },
  "host": {
    "layout": "laya",
    "tokenizer": "tokenizer/tokenizer.json",
    "cls_id": 2,
    "sep_id": 1,
    "pad_id": 0,
    "mask_id": 4,
    "head_max_len": 256,
    "option_text_tokens": 48,
    "source_max_len": 1024,
    "sequence": "[CLS] <type> question: <instructions> [SEP] ([MASK] <option text, <=48 tokens>)... [SEP] <state, right-truncated> [SEP]",
    "readout": "token_logits at the marker positions; softmax of the raw logits -> the four act features -> act; answer = softmax(logits / T), T by bucket (type:K) first, then by type",
    "temperatures": "metadata.json decision block: source_temperature [1,1,1] and the fitted calibration",
    "compute_units": "gpu (a Neural Engine preference is refused: its answers fall outside the bar and vary run to run)"
  },
  "graph": {
    "functions": {
      "main": {
        "inputs": {
          "input_ids": {
            "dtype": "int32",
            "shape": [
              1,
              256
            ],
            "padding": "right, PAD 0"
          },
          "attention_mask": {
            "dtype": "int32",
            "shape": [
              1,
              256
            ],
            "values": "1 real token, 0 padding"
          },
          "qtype_onehot": {
            "dtype": "float32",
            "shape": [
              1,
              3
            ],
            "order": [
              "choice",
              "score",
              "noul"
            ]
          }
        },
        "outputs": {
          "token_logits": {
            "dtype": "float32",
            "shape": [
              1,
              256
            ],
            "read": "at the option marker positions"
          },
          "pooled_cls": {
            "dtype": "float32",
            "shape": [
              1,
              768
            ],
            "read": "input of the act function"
          }
        }
      },
      "act": {
        "inputs": {
          "pooled_cls": {
            "dtype": "float32",
            "shape": [
              1,
              768
            ]
          },
          "feats": {
            "dtype": "float32",
            "shape": [
              1,
              4
            ],
            "order": [
              "top1",
              "top1_minus_top2",
              "entropy_over_ln_max_k_2",
              "max_k_2_over_255"
            ],
            "from": "softmax of the RAW marker logits (no temperature)"
          }
        },
        "outputs": {
          "act_logits": {
            "dtype": "float32",
            "shape": [
              1,
              2
            ],
            "class_0": "direct answer"
          }
        }
      }
    },
    "windows": [
      256,
      512
    ]
  }
}