Publish English priority-only Core ML artifact (FP16, logits-only)

README.md

@@ -0,0 +1,63 @@
+---
+license: apache-2.0
+tags:
+  - coreml
+  - laya
+  - decision-model
+---
+
+# laya-priority-coreml (English, priority-only)
+
+Core ML conversion of `convaiinnovations/laya` (English checkpoint, revision
+`55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851`), restricted to a single fixed
+binary `choice` question used for one purpose: deciding whether a to-do item
+should be acted on now versus deferred. This is not a general-purpose export
+of Laya — only the two-option priority decision graph is included, and only
+`logits` (pre-temperature) is returned.
+
+Weights © Convai Innovations, Apache-2.0. Conversion graph and Core ML export
+by this repository's owner.
+
+## Known limitation — do not treat as validated for general use
+
+Zero-shot measurement on a 48-item balanced English/Korean-translated fixture
+(this checkpoint, priority-only question) scored 17/24 (70.8%) on English and
+15/24 (62.5%) on Korean text machine-translated to English, both below an
+80%-accuracy / 70%-recall gate. This artifact is shipped anyway at the
+consuming application's explicit request, with rule-based fallback preserved
+for degraded cases. Do not present its output as high-confidence.
+
+## Static IO contract
+
+- Inputs (`int32`): `input_ids [1,512]`, `attention_mask [1,512]`,
+  `marker_pos [1,2]`, `marker_mask [1,2]`, `qtype [1]`
+- Output (`float32`): `logits [1,2]` — `[now, later]`, pre-temperature raw
+  logits from the source `DecisionModel.scorer` head.
+- Special token ids: `pad=50283 cls=50281 sep=50282 mask=50284`
+- Sequence format follows upstream `laya.common.build_sequence` with
+  `max_len=512, head_max_len=192`, one fixed instruction/option pair (see
+  `laya_manifest.json`).
+
+## Parity evidence (from the exporter that produced this package)
+
+- Source PyTorch vs traced PyTorch: max abs error `0.0` (exact) across 24
+  fixture cases.
+- Native Core ML (`CPU_AND_NE` requested) vs source, 48 cases (24 English +
+  24 real Korean→English machine translations): argmax agreement 48/48; max
+  abs logit error `0.2274`; max abs option-margin error `0.2588`. Native
+  Swift `MLComputePlan` reports 1188 of ~1220 known ops preferred for the
+  Neural Engine (120 linear, 56 matmul, 61 layer_norm, 28 softmax), 32
+  preferred CPU — this is a compiler placement report, not a runtime device
+  trace.
+- An FP32 variant reproduces the source almost exactly (max abs error
+  `1.4e-5`) but the compiler prefers CPU for all ops in that variant, so it
+  is not an ANE artifact; it is not published here.
+
+## Files
+
+- `laya_english.mlpackage/` — compiled-at-load Core ML package (FP16,
+  `logits`-only output)
+- `tokenizer.json`, `tokenizer_config.json` — copied unmodified from the
+  source checkpoint's `tokenizer/`
+- `laya_manifest.json` — file list with SHA-256 and sizes, IO contract,
+  special token ids

laya_english.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f43cdcc47ddbbe67298bfe79419dc29368766470665af22cd62eec56b6d8868a
+size 678949

laya_english.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d0541cba4ba908c3ec95e14d90739cab023f6c97801ac08ef891f38669a8b97
+size 842357120

laya_english.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "04C2190A-819A-4A18-ACA1-AC98F13CB544": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "568EC863-4609-41CF-9375-8109302E29C6": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "04C2190A-819A-4A18-ACA1-AC98F13CB544"
+}

laya_manifest.json

@@ -0,0 +1,69 @@
+{
+  "format": "laya-priority-coreml-english",
+  "source": {
+    "repo": "convaiinnovations/laya",
+    "revision": "55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851"
+  },
+  "graph": "standard-fp16-logits-only",
+  "sequence_length": 512,
+  "inputs": {
+    "input_ids": [
+      1,
+      512
+    ],
+    "attention_mask": [
+      1,
+      512
+    ],
+    "marker_pos": [
+      1,
+      2
+    ],
+    "marker_mask": [
+      1,
+      2
+    ],
+    "qtype": [
+      1
+    ]
+  },
+  "outputs": {
+    "logits": [
+      1,
+      2
+    ]
+  },
+  "dtype": {
+    "inputs": "int32",
+    "outputs": "float32"
+  },
+  "special_tokens": {
+    "pad": 50283,
+    "cls": 50281,
+    "sep": 50282,
+    "mask": 50284
+  },
+  "model_package": "laya_english.mlpackage",
+  "files": {
+    "laya_english.mlpackage/Manifest.json": {
+      "sha256": "cd127062a620313512df5e38e55058c78edbabfea750f5c228f03f115ee6ce04",
+      "size": 617
+    },
+    "laya_english.mlpackage/Data/com.apple.CoreML/model.mlmodel": {
+      "sha256": "f43cdcc47ddbbe67298bfe79419dc29368766470665af22cd62eec56b6d8868a",
+      "size": 678949
+    },
+    "laya_english.mlpackage/Data/com.apple.CoreML/weights/weight.bin": {
+      "sha256": "5d0541cba4ba908c3ec95e14d90739cab023f6c97801ac08ef891f38669a8b97",
+      "size": 842357120
+    },
+    "tokenizer.json": {
+      "sha256": "6c8aaa9a542084f2457eab775d4eeb51f92a70c0fd9de28d5edb0ddec3c08d30",
+      "size": 3583228
+    },
+    "tokenizer_config.json": {
+      "sha256": "50044de60daaa73df97d262e15a40d4faf0160e7d742df64b377877a1320dd12",
+      "size": 308
+    }
+  }
+}

tokenizer_config.json

@@ -0,0 +1,14 @@
+{
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 8192,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}