| { |
| "format": "kev-pointer-head/v1", |
| "files": { |
| "safetensors": "head.safetensors", |
| "original": "head.pt" |
| }, |
| "head.pt_sha256": "ce6cd9ffc54db41c179b65a33d60973dc8280c29e886218eb3ec07dc28d6f28b", |
| "tensors": { |
| "q.weight": { |
| "shape": [ |
| 256, |
| 1024 |
| ], |
| "dtype": "float32" |
| }, |
| "q.bias": { |
| "shape": [ |
| 256 |
| ], |
| "dtype": "float32" |
| }, |
| "k.weight": { |
| "shape": [ |
| 256, |
| 1024 |
| ], |
| "dtype": "float32" |
| }, |
| "k.bias": { |
| "shape": [ |
| 256 |
| ], |
| "dtype": "float32" |
| }, |
| "temperature": { |
| "shape": [], |
| "dtype": "float32" |
| } |
| }, |
| "hidden_size": 1024, |
| "head_dim": 256, |
| "temperature": 1.0, |
| "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": "absent from head.pt; kev-src Meta default 1.0 (what Kev's loader applies)", |
| "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": 151659, |
| "role": "state" |
| }, |
| "<|fim_middle|>": { |
| "id": 151660, |
| "role": "question" |
| }, |
| "<|box_start|>": { |
| "id": 151648, |
| "role": "option_open" |
| }, |
| "<|box_end|>": { |
| "id": 151649, |
| "role": "option_close" |
| }, |
| "<|fim_suffix|>": { |
| "id": 151661, |
| "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/Qwen3-0.6B-Base", |
| "base_revision": "da87bfb608c14b7cf20ba1ce41287e8de496c0cd", |
| "lora": 16, |
| "holdout": [] |
| } |
| } |
|
|
