| { |
| "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 |
| ] |
| } |
| } |
|
|
