{
  "format": "kev-pointer-head/v1",
  "files": {
    "safetensors": "head.safetensors",
    "original": "head.pt"
  },
  "head.pt_sha256": "dd633435998ecc751ac538717a3742e32149500fabf7d7276287dbf0693f347c",
  "tensors": {
    "q.weight": {
      "shape": [
        256,
        2560
      ],
      "dtype": "float32"
    },
    "q.bias": {
      "shape": [
        256
      ],
      "dtype": "float32"
    },
    "k.weight": {
      "shape": [
        256,
        2560
      ],
      "dtype": "float32"
    },
    "k.bias": {
      "shape": [
        256
      ],
      "dtype": "float32"
    },
    "temperature": {
      "shape": [],
      "dtype": "float32"
    }
  },
  "hidden_size": 2560,
  "head_dim": 256,
  "temperature": 2.406050072164233,
  "temperature_note": "head.safetensors stores it as a 0-d float32 tensor (rounded); the exact float64 value is here and in the safetensors metadata key temperature_exact",
  "temperature_source": "head.pt['temperature']",
  "card_temperature_not_in_head_pt": null,
  "input": "last_hidden_state of the text backbone (after the final norm); no LM head is used",
  "readout": {
    "query": "h at the <|fim_suffix|> (decide) token of the question",
    "keys": "h at each option's closing <|box_end|> token",
    "formula": "logit_j = ((W_k h_opt_j + b_k) . (W_q h_decide + b_q)) / sqrt(head_dim) / temperature; p = softmax_j(logit_j) over the question's options",
    "compute_dtype": "float32 (kev/model.py PointerHead; hidden states cast to fp32)"
  },
  "token_layout": {
    "special_tokens": {
      "<|fim_prefix|>": {
        "id": 248060,
        "role": "state"
      },
      "<|fim_middle|>": {
        "id": 248061,
        "role": "question"
      },
      "<|box_start|>": {
        "id": 248049,
        "role": "option_open"
      },
      "<|box_end|>": {
        "id": 248050,
        "role": "option_close"
      },
      "<|fim_suffix|>": {
        "id": 248062,
        "role": "decide"
      }
    },
    "sequence": "[<|fim_prefix|> state...] then per question: <|fim_middle|> instruction... (<|box_start|> option... <|box_end|>)* <|fim_suffix|>",
    "tokenization": "each piece tokenized separately with add_special_tokens=False; user text matching <|name|> is rewritten to <¦name¦> first; no chat template, no BOS",
    "positions": "branch positions continue after the state; each question is its own causal row (state + its branch)",
    "reference": "github.com/jaredpalmer/kev@fe64b1274ea7f80d4095866df90666abb03e9cf6 kev/model.py encode(), rows_of(), PointerHead; kev/api.py to_record()"
  },
  "upstream_meta": {
    "base": "Qwen/Qwen3.5-4B-Base",
    "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
    "lora": 16,
    "weights_dtype": "fp32",
    "holdout": []
  }
}