mlboydaisuke commited on
Commit
d1e9435
·
verified ·
1 Parent(s): 18f95df

Add files using upload-large-folder tool

Browse files
.gitattributes CHANGED
@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ macos/fp32-s1024/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
+ macos/fp32-s512/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
38
+ macos/fp32-s512/julia1_fp32_s512.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
39
+ macos/fp32-s1024/julia1_fp32_s1024.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS
177
+
178
+ APPENDIX: How to apply the Apache License to your work.
179
+
180
+ To apply the Apache License to your work, attach the following
181
+ boilerplate notice, with the fields enclosed by brackets "[]"
182
+ replaced with your own identifying information. (Don't include
183
+ the brackets!) The text should be enclosed in the appropriate
184
+ comment syntax for the file format. We also recommend that a
185
+ file or class name and description of purpose be included on the
186
+ same "printed page" as the copyright notice for easier
187
+ identification within third-party archives.
188
+
189
+ Copyright [yyyy] [name of copyright owner]
190
+
191
+ Licensed under the Apache License, Version 2.0 (the "License");
192
+ you may not use this file except in compliance with the License.
193
+ You may obtain a copy of the License at
194
+
195
+ http://www.apache.org/licenses/LICENSE-2.0
196
+
197
+ Unless required by applicable law or agreed to in writing, software
198
+ distributed under the License is distributed on an "AS IS" BASIS,
199
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200
+ See the License for the specific language governing permissions and
201
+ limitations under the License.
LICENSE-NOTE.md ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # License and changes
2
+
3
+ Supersonic Labs declares Apache-2.0 for `SupersonicLabs/Julia-1` in the upstream model card at the pinned revision
4
+ `a85b127321d580d65176c89ced8273f305745d85` ("The model artifacts are licensed under Apache 2.0"); no standalone LICENSE or NOTICE file is
5
+ listed at that revision. The standard Apache-2.0 text is included as `LICENSE`. The upstream base encoder is
6
+ `jhu-clsp/mmBERT-small` (MIT).
7
+
8
+ Changes relative to the upstream checkpoint:
9
+ - re-authored as a static Core AI graph from the raw weights (no transformers or `julia` package code in the
10
+ graph), batch 1, one function `main` (encoder, type embedding, the two head layers and the scorer at every
11
+ position); the checkpoint's `act_head` and `temperature` buffer, which the upstream inference API does not
12
+ read, are not exported;
13
+ - `fp32` folders keep the weights as published (F32); `wfp16` folders store them in fp16 (rounded) and
14
+ compute in fp32;
15
+ - `tokenizer/` holds the checkpoint's `tokenizer.json` and `tokenizer_config.json`, unmodified;
16
+ `config.json` (the upstream `encoder/config.json`) and `julia_config.json` are the checkpoint's own files,
17
+ unmodified.
18
+ Source: `SupersonicLabs/Julia-1@a85b127321d580d65176c89ced8273f305745d85`.
README.md ADDED
@@ -0,0 +1,259 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: coreai
3
+ license: apache-2.0
4
+ base_model: SupersonicLabs/Julia-1
5
+ tags:
6
+ - coreai
7
+ - apple-silicon
8
+ - on-device
9
+ - modernbert
10
+ - decision-model
11
+ - text-classification
12
+ - multilingual
13
+ language:
14
+ - multilingual
15
+ pipeline_tag: text-classification
16
+ ---
17
+
18
+ Core AI is Apple's on-device ML runtime in iOS 27 / macOS 27 and the successor to Core ML: PyTorch models are exported with Apple's `coreai-torch` (LLMs: `coreai.llm.export`) into `.aimodel` bundles that run on the GPU or the Neural Engine, e.g. Qwen3-8B 4-bit decodes at 94 tok/s on an M4 Max GPU, MLX 90 under the same protocol ([apple-silicon-llm-bench](https://github.com/john-rocky/apple-silicon-llm-bench), macOS 27 beta, 2026-06).
19
+
20
+ <!-- gen-cards:devicemark begin (managed by scripts/gen-cards + tools/devicemark_row.py — edit cards.json, not this block) -->
21
+ This model has no row on [DeviceMark](https://devicemark.github.io/), the on-device LLM leaderboard.
22
+ <!-- gen-cards:devicemark end -->
23
+
24
+ # Julia-1 — Core AI export
25
+
26
+ Zoo card, recipe and gate transcript: [coreai-model-zoo/models/julia-1](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/julia-1/README.md).
27
+
28
+ [🤗 mlboydaisuke/Julia-1-CoreAI](https://huggingface.co/mlboydaisuke/Julia-1-CoreAI) · Apache-2.0 · source [SupersonicLabs/Julia-1](https://huggingface.co/SupersonicLabs/Julia-1/tree/a85b127321d580d65176c89ced8273f305745d85) (revision `a85b127`) · base jhu-clsp/mmBERT-small
29
+
30
+ This is Supersonic Labs' Julia-1, a 144M typed decision model, as a Core AI `.aimodel` converted with Apple's `coreai-torch`; on the GPU of an Apple M4 Max it takes 16.00 ms per decision at the 1,024-token window, tokenization not included (2026-09-30).
31
+
32
+ Give it a state (a message or a JSON record) and a question with 2 to 20 options you write: pick one (`choice`), place the state on an ordered rubric (`score`), or answer false or true (`noul`). It returns a probability for every option from one forward pass and generates no text.
33
+
34
+ ## Graph contract
35
+
36
+ ```
37
+ function "main"
38
+ input "input_ids" [1, S] int32 the row, right-padded with PAD 0
39
+ input "attention_mask" [1, S] int32 1 over real tokens, 0 over padding
40
+ input "qtype_onehot" [1, 3] fp32 choice / score / noul
41
+ output "token_logits" [1, S] fp32 the scorer at every position, read at the option markers
42
+ S = 512 or 1024, one bundle per window; batch 1; fp32 weights and compute
43
+ ```
44
+
45
+ Inside `main` is the mmBERT-small encoder in the ModernBERT layout: 22 layers, hidden 384, 6 heads × 64,
46
+ GLU MLP 1152, global attention on layers 0, 3, …, 21 and a sliding window of inclusive radius 64 elsewhere,
47
+ RoPE θ 160,000, vocabulary 256,000. After it come the type embedding, two pre-norm transformer layers (384
48
+ wide, 6 heads, ReLU FFN 1536) and the scorer (LayerNorm → Linear → GELU → Linear to one logit).
49
+
50
+ The checkpoint also holds an act head (Linear 388 → 256, GELU, Linear 256 → 2) and a temperature buffer of
51
+ [1, 1, 1]. Neither is in the graph, because the publisher's inference API reads neither: it calls
52
+ `forward(..., return_actions=False)`, uses no temperature, and `inference-policy.json` says `calibration: null`.
53
+
54
+ ## Host recipe
55
+
56
+ The row and the readout are the publisher's (`julia/data.py` `sequence()`, `julia/typed.py`). `metadata.json`
57
+ in each folder records them with the window, the head budget and the token ids.
58
+
59
+ - Row: `[CLS 2] tok("{type} question: {instructions}") [SEP 1] ([MASK 4] tok(" " + option)[:48])… [SEP 1]
60
+ tok(state) [SEP 1]`. Each piece is tokenized alone, without special tokens, by the checkpoint's
61
+ `tokenizer.json`. A dict or list state becomes `json.dumps(state, ensure_ascii=False)`. An option's marker is
62
+ the position of its `[MASK]`.
63
+ - Options: the criteria descriptions as given. `choice` uses the mapping's values in order, and the answer is
64
+ the caller's id. `score` uses the rubric in order. `noul` uses `[criteria.false, criteria.true]`, or the
65
+ literal words `false` / `true` when the question has no criteria.
66
+ - Readout: the softmax of the raw marker logits at T = 1; no calibration exists. `choice` = the winning id,
67
+ `score` = the expected rubric index Σ i·p_i (not rounded), `noul` = p[true].
68
+ - Strict encoding: 2–20 options, each at most 48 tokens, and no literal `<mask>` in any text (the publisher's
69
+ non-strict builder turns it into a space). A question that would need any cut is refused, never truncated.
70
+ - Head budget (question plus options): 512 tokens at S = 1024, the publisher's protocol, and 256 at S = 512.
71
+ Every row that fits 512 gets the same ids with either budget (2,065 of 2,065). All 2,000 typed-decisions
72
+ test questions fit 1,024 tokens (the longest is 607); 1,965 fit 512.
73
+
74
+ The option text is where Julia differs from [laya-multilingual](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/laya-multilingual/README.md), the same
75
+ design on mmBERT-base with the same graph inputs. laya hosts render `label: description`, `level i: …` and
76
+ `false: …` / `true: …`, so a laya host builds other rows for Julia and gets other answers.
77
+
78
+ ## Measured (Mac)
79
+
80
+ Apple M4 Max, macOS 27.0 build 26A428, coreai-build 3600.83.1, coreai-core 1.0.0b2, 2026-09-30. Each bundle
81
+ ran every row of its window: at S = 1024 the 2,000 typed-decisions test questions, the publisher's 100 parity
82
+ requests and 20 window-filling rows; at S = 512 the 1,965 questions that fit, the same 100 and 35
83
+ window-filling rows. The GPU rows ran alone (the machine-wide GPU lock held, no other GPU job seen).
84
+ `cpu_only` is the parity option; its milliseconds were taken while other jobs used the CPU and are reference
85
+ values. A question's time covers the NumPy inputs, the graph and the marker gather, not the tokenization.
86
+
87
+ | Variant | S | Compute | Rows | Argmax | Max \|Δp\| | Marker max \|Δ\| | Repeat drift | Load | Warm median |
88
+ |---|---:|---|---:|---:|---:|---:|---:|---:|---:|
89
+ | fp32 | 1024 | GPU | 2,120 | 2,120/2,120 | 8.58e-5 | 5.71e-4 | 0 | 641 ms | 16.00 ms |
90
+ | fp32 | 512 | GPU | 2,100 | 2,100/2,100 | 7.97e-5 | 6.17e-4 | 0 | 702 ms | 8.28 ms |
91
+ | fp32 | 1024 | cpu_only | 2,120 | 2,120/2,120 | 5.91e-5 | 7.58e-4 | 0 | 648 ms | 109.7 ms |
92
+ | fp32 | 512 | cpu_only | 2,100 | 2,100/2,100 | 5.91e-5 | 7.58e-4 | 0 | 621 ms | 81.0 ms |
93
+
94
+ The GPU median at S = 1024 is over 60 warm questions (p10 15.89 ms, p90 16.16 ms); at S = 512, p10 is 8.20 ms
95
+ and p90 8.42 ms. The question right after load took 601 ms at S = 1024 and 118 ms at S = 512. Repeat drift is
96
+ the largest change in the marker logits when a row runs 3 times (212 rows at S = 1024).
97
+
98
+ On the typed-decisions test split, the bundle's own argmax (GPU, S = 1024) gives the same three totals as the
99
+ publisher's CPU FP32 run. The split is `LocalLLaMA/typed-decisions` at `c76749e` (400 cases, 2,000
100
+ questions); the publisher's run used torch 2.14.0, transformers 5.0.0, max_length 1024, head_length 512 and
101
+ strict encoding.
102
+
103
+ | | choice | score | noul |
104
+ |---|---:|---:|---:|
105
+ | Supersonic Labs, PyTorch CPU FP32 | 426/600 | 542/800 | 483/600 |
106
+ | this bundle, Mac GPU, S = 1024 | 426/600 | 542/800 | 483/600 |
107
+
108
+ From text to answer, the Python host (`julia_coreai.JuliaCoreAI`, the S = 1024 bundle, GPU) gives the
109
+ publisher's `engine.predict` answers on all 2,000 named questions: argmax 2,000/2,000, max |Δp| 8.58e-5,
110
+ score and noul values within 8.6e-5, 0 questions refused, and 426/542/483 when re-scored. The 2,000 questions
111
+ took 32.9 s, tokenization included.
112
+
113
+ The Swift host (`JuliaDecisions.swift` over the system CoreAI framework, built with Xcode 27.0 RC) was
114
+ checked on the 200 strict rows of the fixture below. Its row builder gives the publisher's ids and markers on
115
+ 200/200 rows, and its answers through the bundle match the publisher's: argmax 200/200, max |Δp| 4.3e-5 on the
116
+ GPU and 3.3e-5 with cpu_only. That check read the publisher's token ids per text piece from a table, so a
117
+ Swift tokenizer is not part of it.
118
+
119
+ ## Numerics gate
120
+
121
+ The oracle is the publisher's own `julia` package at the pinned revision (torch 2.14.0, transformers 5.0.0,
122
+ CPU fp32, 4 threads, one question per forward). On this Mac it reproduces 426/600, 542/800 and 483/600, and
123
+ its 2,000 answers are identical to those of the publisher's `scripts/reproduce_typed.py`, run unchanged here.
124
+ On the publisher's 100 WebGPU parity requests (`parity-cases.json` in `SupersonicLabs/Julia-1-ONNX` at
125
+ `82a2fad`) it matches the PyTorch logits the publisher recorded: argmax 100/100, max |Δ logit| 1.03e-4 (theirs
126
+ in padded batches of 4, ours one per forward).
127
+
128
+ Bars at every stage: the argmax equal to the oracle's on every row, max |Δp| ≤ 1e-3 over the options at
129
+ T = 1, and marker logits within 1e-3. The rows include near-ties: 58 typed questions have a top probability
130
+ below 0.6, 46 a top-two gap below 0.2 and 9 below 0.05 (the smallest 0.0098). The window-filling rows have no
131
+ padding: the 35 questions over 512 tokens, cut to exactly 512 by the publisher's non-strict builder (head
132
+ budget 256), and 20 synthetic states (consecutive test states in one JSON list) cut to exactly 1,024.
133
+
134
+ - Authoring: the graph re-authored in plain PyTorch from `model.safetensors` (no transformers or `julia` code,
135
+ all 170 tensors loaded strictly), run in the zoo venv with torch 2.9.0 on every row of each window.
136
+ S = 1024: argmax 2,120/2,120, max |Δp| 6.56e-5, marker 7.95e-4, typed totals 426/542/483. S = 512: argmax
137
+ 2,100/2,100, max |Δp| 6.56e-5, marker 7.95e-4.
138
+ - Layer gate on the hidden states: absolute ≤ 1e-4 on the embeddings and layers 0–9 (max 4.8e-5 at S = 1024,
139
+ 3.5e-5 at S = 512); relative ≤ 5e-4 from layer 10 through the head (max 2.07e-4 and 3.42e-4).
140
+ - Negative controls, caught at both windows (S = 1024 / 512): all-global attention 5.09 / 3.59 at layer 1;
141
+ all-local 3.72 / 4.20 at layer 0; radius 63 instead of 64, 1.29 / 1.45 at layer 1; padding ignored,
142
+ 29.5 / 29.4 at layer 0; no type embedding, caught at the head input (marker logits moved by 0.81 at
143
+ S = 512). Random ids in every pad position leave every real position bit-identical (11 rows per window).
144
+ On the bundle, judging each row against another row's outputs fails on 1,724 of 2,120 pairs (S = 1024).
145
+ - Export: the `torch.export` graph after `coreai_torch` decomposition, gated before conversion on every 4th
146
+ row plus the layer-gate rows and every window-filling row. S = 1024: 552 rows, argmax 552/552, max |Δp|
147
+ 5.81e-5. S = 512: 560 rows, argmax 560/560, max |Δp| 6.56e-5.
148
+ - Runtime: the table above.
149
+ - Host: the NumPy host (tokenizers 0.23.0rc0 on the checkpoint's `tokenizer.json`) rebuilds all 2,155 oracle
150
+ rows from text with identical ids and markers. It renders the 2,000 named questions the publisher's way and
151
+ reproduces the publisher's 2,000 answer dictionaries from the same logits (max difference 8.9e-16).
152
+
153
+ The relative bar is twice the publisher's own SDPA-vs-eager distance on this checkpoint (2.87e-4 at the final
154
+ norm, S = 1024), rounded down; laya's 2e-4, tried before it, failed at the final norm. Under the publisher's
155
+ torch 2.14.0 the graph equals the publisher's eager-attention path to 1.3e-7 relative at every hidden state
156
+ and 1.9e-6 on the marker logits (13 rows per window).
157
+
158
+ [`fixtures-julia-1.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/julia-1/fixtures-julia-1.json) in the zoo repository (543 KB) holds the 100 parity requests, every 20th typed question and
159
+ 5 window-filling rows per window, each with the request, token ids, markers and the publisher's raw logits.
160
+
161
+ ## Precision
162
+
163
+ fp32 ships: the checkpoint's F32 weights as published, computed in fp32. fp16 weight storage with fp32 compute
164
+ was measured at S = 1024 and fails the 1e-3 bar, so it does not ship. It keeps argmax 2,120/2,120 and the
165
+ typed totals 426/542/483, but max |Δp| is 0.0234 (over 1e-3 on 325 rows, over 1e-2 on 32); a score's expected
166
+ index moves by up to 0.029 and a noul's p[true] by up to 0.023.
167
+
168
+ Neural Engine: not used. With a Neural Engine preference the Mac returns the GPU's results bit for bit at the same speed (844 rows, 2026-09-30), and `coreai-build compile --preferred-compute neural-engine` produces no Neural Engine region for macOS (h16c) or the iPhone 18 Pro (h19p): the graph computes in fp32, which the Neural Engine does not take.
169
+
170
+ ## Bundle
171
+
172
+ This repository holds one folder per window, each self-contained: the bundle,
173
+ `tokenizer/` (the checkpoint's files, unmodified), `metadata.json` (the decision contract), `reference.json`
174
+ (the window's rows with token ids, markers and the publisher's logits) and `provenance/` (the export manifest
175
+ and the export and runtime gate records).
176
+
177
+ | Folder | Platform | Format | Bundle | Bytes |
178
+ |---|---|---|---|---:|
179
+ | `macos/fp32-s1024/` | macOS 27 | JIT `.aimodel` | `julia1_fp32_s1024.aimodel` | 581,765,544 |
180
+ | `macos/fp32-s512/` | macOS 27 | JIT `.aimodel` | `julia1_fp32_s512.aimodel` | 578,357,637 |
181
+
182
+ The whole folders are 627,543,671 and 623,850,816 bytes.
183
+
184
+ Convert yourself: [`conversion/julia/`](https://github.com/john-rocky/coreai-model-zoo/blob/main/conversion/julia/README.md), the staged scripts in order;
185
+ [`recipe.toml`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/julia-1/recipe.toml) names the commands.
186
+
187
+ ## Use it
188
+
189
+ Python, with `conversion/julia/julia_coreai.py` (the reference host; it needs `coreai-core`, `numpy` and
190
+ `tokenizers`):
191
+
192
+ ```python
193
+ import sys
194
+ sys.path.insert(0, "conversion/julia") # the folder with julia_coreai.py and _julia_host.py
195
+ from julia_coreai import JuliaCoreAI
196
+
197
+ julia = JuliaCoreAI("Julia-1-CoreAI/macos/fp32-s1024") # a download of the HF repo; GPU by default
198
+ result = julia.predict(
199
+ state="I was charged twice for the same order.",
200
+ questions={
201
+ "team": {
202
+ "type": "choice",
203
+ "instructions": "Which team should handle this request?",
204
+ "criteria": {
205
+ "billing": "Billing and payment disputes",
206
+ "shipping": "Shipping and delivery",
207
+ "access": "Account access and login",
208
+ },
209
+ },
210
+ },
211
+ )
212
+ print(result["answers"]["team"]["choice"])
213
+ print(result["answers"]["team"]["probabilities"])
214
+ ```
215
+
216
+ The call and the returned dictionary follow the publisher's `engine.predict(state=..., questions=...)`.
217
+ `JuliaCoreAI(folder, compute="cpu_only")` loads the parity option.
218
+
219
+ Swift, with `conversion/julia/swift/JuliaDecisions.swift` added to your target (the system CoreAI framework,
220
+ macOS 27). It needs a tokenizer for the folder's `tokenizer/tokenizer.json`, passed in as `encode:`; with
221
+ swift-transformers:
222
+
223
+ ```swift
224
+ import CoreAI
225
+ import Tokenizers // swift-transformers
226
+
227
+ let folder = URL(filePath: "Julia-1-CoreAI/macos/fp32-s1024") // a download of the HF repo
228
+ let tokenizer = try await AutoTokenizer.from(modelFolder: folder.appendingPathComponent("tokenizer"))
229
+ let julia = try await JuliaDecisions(folder: folder, encode: { text in
230
+ tokenizer.encode(text: text, addSpecialTokens: false).map(Int32.init)
231
+ })
232
+ let answers = try await julia.predict(state: "I was charged twice for the same order.", questions: [
233
+ ("team", .choice("Which team should handle this request?", [
234
+ ("billing", "Billing and payment disputes"), ("shipping", "Shipping and delivery"),
235
+ ("access", "Account access and login")])),
236
+ ])
237
+ print(answers["team"]!.choice!)
238
+ ```
239
+
240
+ `JuliaDecisions` loads the folder with `AIModel` itself; Julia-1 is not a CoreAIKit catalog model. A
241
+ structured state goes in as text, serialized the way Python's `json.dumps(state, ensure_ascii=False)` writes
242
+ it. The Swift check above did not run a tokenizer; `reference.json` and the fixture hold the ids a tokenizer
243
+ has to reproduce.
244
+
245
+ ## Other runtimes
246
+
247
+ Not compared here: the publisher's ONNX export with a WebGPU adapter, `SupersonicLabs/Julia-1-ONNX`; MLX,
248
+ `zainmerchan/Julia-1-MLX`; GGUF, `andrelucas/Julia-1-GGUF`.
249
+
250
+ ## License and limits
251
+
252
+ Apache-2.0, from the upstream card: its metadata declares it, and its README says "The model artifacts are
253
+ licensed under Apache 2.0." The upstream repository has no LICENSE file. The base, jhu-clsp/mmBERT-small, is
254
+ MIT. Julia-1 and its weights are the work of Supersonic Labs; these bundles carry the weights in fp32 as
255
+ published. The fixture's requests come from `LocalLLaMA/typed-decisions` and the Julia-1-ONNX
256
+ `parity-cases.json`, both Apache-2.0.
257
+
258
+ Not tested: iPhone (not measured), other Macs and OS builds, batched shapes, windows other than 512 and 1024,
259
+ the publisher's other benchmarks (not re-run here), and the publisher's `Router` for more than 20 options.
SHA256SUMS ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4 LICENSE
2
+ 4670f123942007bc60b1525ab2ab726a37ebd020178cf0312a23df636b97de47 LICENSE-NOTE.md
3
+ ba01823bd1e64d927602b063fa7d2bda8b84d58cee6ee7d650a226dc7810d1f6 README.md
4
+ c79cda42d42ddf777218254845ce92fb25a10dc35874be40d06abd707894523d config.json
5
+ b9b881646beeac5414eaa78b39d3fe89ff8a2c062731ad718432620f9aee6bf7 julia_config.json
6
+ 4e1ee201ce359a55aed90422bea275e200a7331d995cc6c4d4728f4bcb5e4b1a macos/fp32-s1024/julia1_fp32_s1024.aimodel/main.hash
7
+ ad29d6800776a1077cb50e0afa7ffe786c51db11bf61514a1577212075a540e8 macos/fp32-s1024/julia1_fp32_s1024.aimodel/main.mlirb
8
+ a3dda1a55670ebf047213938eee9dfabae83e19e0e022de905b91279c8772da5 macos/fp32-s1024/julia1_fp32_s1024.aimodel/metadata.json
9
+ debbcac4fb5873788c78bf6e2f35e4bc6b791d55a6014bcda182c78c7b5a0048 macos/fp32-s1024/metadata.json
10
+ 37978d7abf233adfaf81214d2485625bd9608079890594fcbcedc0bfccbf6ec4 macos/fp32-s1024/provenance/export-gate.json
11
+ f2beeb90ccddb5705ca706c0adf7b7cdbfdd2eeda2f0fc37fa4bfdca28f127f7 macos/fp32-s1024/provenance/export-manifest.json
12
+ de82f57d9b1b6afc59b347bd2f44f6a3e662f5c3b5e7c24d7a91e51e51712494 macos/fp32-s1024/provenance/runtime-gate.json
13
+ 1fbd7e516ac8b0828efd508d4251f12aa5d461d12fd6dc7bb5b1d4a1222e3445 macos/fp32-s1024/reference.json
14
+ 609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f macos/fp32-s1024/tokenizer/tokenizer.json
15
+ 6b069e57db0ce0794c22547725275f22618809c4ef4249307337e651a0dfef8c macos/fp32-s1024/tokenizer/tokenizer_config.json
16
+ e6493cb1f5cabeaa395d558628485d906f3216a0f612b52891cfd334ba3ce732 macos/fp32-s512/julia1_fp32_s512.aimodel/main.hash
17
+ f9287684c9e2d01fb38a7776aa0f5397c733f8f7d390b6cabe36cfb45d3c81d4 macos/fp32-s512/julia1_fp32_s512.aimodel/main.mlirb
18
+ e9ee198e153c93d1d4678df8751cab8f3f9811e69a329c9fdea4cef19bd64366 macos/fp32-s512/julia1_fp32_s512.aimodel/metadata.json
19
+ 800bb514522ecdaafef95650348df7ae23234e98f77fb90ffb4f0a706ade2d47 macos/fp32-s512/metadata.json
20
+ 77300fdd1c131f336943228fe11839fe37eb36969654fd2d38e05bd090eccad4 macos/fp32-s512/provenance/export-gate.json
21
+ c310db109e6bc20b8437a1fe46e68c456527f0f210b93422e114d07f62c5fbf5 macos/fp32-s512/provenance/export-manifest.json
22
+ 075799cc5b19bb79cf0f9008629652290df091bd0513c671f5a8da818cb6de02 macos/fp32-s512/provenance/runtime-gate.json
23
+ 65a2ed0f8435d5998e12461cb07a78a8917b381e00c88b5e31a6d4026992b725 macos/fp32-s512/reference.json
24
+ 609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f macos/fp32-s512/tokenizer/tokenizer.json
25
+ 6b069e57db0ce0794c22547725275f22618809c4ef4249307337e651a0dfef8c macos/fp32-s512/tokenizer/tokenizer_config.json
config.json ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "ModernBertForMaskedLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 2,
8
+ "classifier_activation": "gelu",
9
+ "classifier_bias": false,
10
+ "classifier_dropout": 0.0,
11
+ "classifier_pooling": "mean",
12
+ "cls_token_id": 1,
13
+ "decoder_bias": true,
14
+ "deterministic_flash_attn": false,
15
+ "dtype": "float32",
16
+ "embedding_dropout": 0.0,
17
+ "eos_token_id": 1,
18
+ "global_attn_every_n_layers": 3,
19
+ "gradient_checkpointing": false,
20
+ "hidden_activation": "gelu",
21
+ "hidden_size": 384,
22
+ "initializer_cutoff_factor": 2.0,
23
+ "initializer_range": 0.02,
24
+ "intermediate_size": 1152,
25
+ "layer_norm_eps": 1e-05,
26
+ "layer_types": [
27
+ "full_attention",
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "full_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention"
49
+ ],
50
+ "local_attention": 128,
51
+ "mask_token_id": 4,
52
+ "max_position_embeddings": 8192,
53
+ "mlp_bias": false,
54
+ "mlp_dropout": 0.0,
55
+ "model_type": "modernbert",
56
+ "norm_bias": false,
57
+ "norm_eps": 1e-05,
58
+ "num_attention_heads": 6,
59
+ "num_hidden_layers": 22,
60
+ "pad_token_id": 0,
61
+ "position_embedding_type": "sans_pos",
62
+ "repad_logits_with_grad": false,
63
+ "rope_parameters": {
64
+ "full_attention": {
65
+ "rope_theta": 160000,
66
+ "rope_type": "default"
67
+ },
68
+ "sliding_attention": {
69
+ "rope_theta": 160000,
70
+ "rope_type": "default"
71
+ }
72
+ },
73
+ "sep_token_id": 1,
74
+ "sparse_pred_ignore_index": -100,
75
+ "sparse_prediction": true,
76
+ "tie_word_embeddings": true,
77
+ "transformers_version": "5.0.0",
78
+ "vocab_size": 256000
79
+ }
julia_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "format_version": 1,
3
+ "architecture": "JuliaDecisionModel",
4
+ "weight_dtype": "float32",
5
+ "head_layers": 2,
6
+ "n_act": 2,
7
+ "dropout": 0.1
8
+ }
macos/fp32-s1024/julia1_fp32_s1024.aimodel/main.hash ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ �)րv�|�
2
+ ��xlQ��aQJw! u�@�
macos/fp32-s1024/julia1_fp32_s1024.aimodel/main.mlirb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad29d6800776a1077cb50e0afa7ffe786c51db11bf61514a1577212075a540e8
3
+ size 581765120
macos/fp32-s1024/julia1_fp32_s1024.aimodel/metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "author" : "mlboydaisuke (coreai-model-zoo); weights Supersonic Labs (Julia-1)",
3
+ "assetVersion" : "2.0",
4
+ "description" : "Julia-1 decision encoder (mmBERT-small + typed head); fp32; S=1024; function main; source SupersonicLabs\/Julia-1@a85b127321d580d65176c89ced8273f305745d85",
5
+ "producer" : "coreai-core 1.0.0b2",
6
+ "license" : "Apache-2.0",
7
+ "creationDate" : "20260930T004313Z"
8
+ }
macos/fp32-s1024/metadata.json ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata_version": "0.2",
3
+ "kind": "encoder",
4
+ "decision": {
5
+ "head": "encoder",
6
+ "layout": "julia",
7
+ "window": 1024,
8
+ "head_max_len": 512,
9
+ "option_text_tokens": 48,
10
+ "strict_encoding": true,
11
+ "cls_token_id": 2,
12
+ "sep_token_id": 1,
13
+ "pad_token_id": 0,
14
+ "mask_token_id": 4,
15
+ "functions": {
16
+ "main": "main"
17
+ },
18
+ "inputs": {
19
+ "input_ids": {
20
+ "dtype": "int32",
21
+ "shape": [
22
+ 1,
23
+ 1024
24
+ ],
25
+ "padding": "right, PAD 0"
26
+ },
27
+ "attention_mask": {
28
+ "dtype": "int32",
29
+ "shape": [
30
+ 1,
31
+ 1024
32
+ ],
33
+ "values": "1 real token, 0 padding"
34
+ },
35
+ "qtype_onehot": {
36
+ "dtype": "float32",
37
+ "shape": [
38
+ 1,
39
+ 3
40
+ ],
41
+ "order": [
42
+ "choice",
43
+ "score",
44
+ "noul"
45
+ ]
46
+ }
47
+ },
48
+ "outputs": {
49
+ "token_logits": {
50
+ "dtype": "float32",
51
+ "shape": [
52
+ 1,
53
+ 1024
54
+ ],
55
+ "read": "at the option marker positions"
56
+ }
57
+ },
58
+ "options": "the descriptions as given: choice = the criteria mapping's values in order (answer = its id), score = the rubric in order, noul = [criteria.false, criteria.true] or the literal ['false', 'true'] when a noul question has no criteria",
59
+ "sequence": "[CLS 2] tok('{type} question: {instructions}') [SEP 1] ([MASK 4] tok(' ' + option)[:48])... [SEP 1] tok(state) [SEP 1]; state = json.dumps(state, ensure_ascii=False) when not text; strict: a row that does not fit, or text holding the literal '<mask>', is refused, never cut or rewritten (the publisher's non-strict builder turns '<mask>' into ' ')",
60
+ "readout": "softmax of the raw marker logits (T = 1); choice = argmax id, score = sum(i * p_i), noul = p[true]",
61
+ "source_temperature": [
62
+ 1.0,
63
+ 1.0,
64
+ 1.0
65
+ ],
66
+ "calibration": null,
67
+ "precision": "fp32"
68
+ },
69
+ "source": {
70
+ "hf_model_id": "SupersonicLabs/Julia-1",
71
+ "hf_revision": "a85b127321d580d65176c89ced8273f305745d85"
72
+ },
73
+ "assets": {
74
+ "main": "julia1_fp32_s1024.aimodel"
75
+ },
76
+ "language": {
77
+ "tokenizer": "SupersonicLabs/Julia-1/tokenizer",
78
+ "vocab_size": 256000,
79
+ "max_context_length": 1024
80
+ }
81
+ }
macos/fp32-s1024/provenance/export-gate.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s1024/provenance/export-manifest.json ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "PASS",
3
+ "measure_only": false,
4
+ "model": "SupersonicLabs/Julia-1",
5
+ "model_sha": "a85b127321d580d65176c89ced8273f305745d85",
6
+ "bundle": "julia1_fp32_s1024.aimodel",
7
+ "intended_runtime": "macos",
8
+ "format": "JIT .aimodel (portable; one function main)",
9
+ "aot": false,
10
+ "precision": "fp32",
11
+ "window": 1024,
12
+ "inputs": {
13
+ "input_ids": {
14
+ "dtype": "int32",
15
+ "shape": [
16
+ 1,
17
+ 1024
18
+ ],
19
+ "padding": "right, PAD 0"
20
+ },
21
+ "attention_mask": {
22
+ "dtype": "int32",
23
+ "shape": [
24
+ 1,
25
+ 1024
26
+ ],
27
+ "values": "1 real token, 0 padding"
28
+ },
29
+ "qtype_onehot": {
30
+ "dtype": "float32",
31
+ "shape": [
32
+ 1,
33
+ 3
34
+ ],
35
+ "order": [
36
+ "choice",
37
+ "score",
38
+ "noul"
39
+ ]
40
+ }
41
+ },
42
+ "outputs": {
43
+ "token_logits": {
44
+ "dtype": "float32",
45
+ "shape": [
46
+ 1,
47
+ 1024
48
+ ],
49
+ "read": "at the option marker positions"
50
+ }
51
+ },
52
+ "files": [
53
+ {
54
+ "path": "main.hash",
55
+ "bytes": 32,
56
+ "sha256": "4e1ee201ce359a55aed90422bea275e200a7331d995cc6c4d4728f4bcb5e4b1a"
57
+ },
58
+ {
59
+ "path": "main.mlirb",
60
+ "bytes": 581765120,
61
+ "sha256": "ad29d6800776a1077cb50e0afa7ffe786c51db11bf61514a1577212075a540e8"
62
+ },
63
+ {
64
+ "path": "metadata.json",
65
+ "bytes": 392,
66
+ "sha256": "a3dda1a55670ebf047213938eee9dfabae83e19e0e022de905b91279c8772da5"
67
+ }
68
+ ],
69
+ "op_counts": {
70
+ "builtin.module": 1,
71
+ "coreai.cast": 1,
72
+ "coreai.concat": 44,
73
+ "coreai.constant": 195,
74
+ "coreai.decomposable.broadcasting_add": 130,
75
+ "coreai.decomposable.broadcasting_batch_matmul": 147,
76
+ "coreai.decomposable.broadcasting_mul": 182,
77
+ "coreai.decomposable.broadcasting_sub": 2,
78
+ "coreai.gather_nd": 1,
79
+ "coreai.gelu": 23,
80
+ "coreai.graph": 2,
81
+ "coreai.invoke": 50,
82
+ "coreai.output": 2,
83
+ "coreai.reduce_mean": 2,
84
+ "coreai.relu": 2,
85
+ "coreai.reshape": 122,
86
+ "coreai.rsqrt": 1,
87
+ "coreai.slice": 72,
88
+ "coreai.softmax": 24,
89
+ "coreai.split": 66,
90
+ "coreai.transpose": 215
91
+ },
92
+ "seconds": {
93
+ "torch_export": 2.504598124884069,
94
+ "convert_total": 0.9458033749833703,
95
+ "optimize": 0.03253654204308987,
96
+ "save": 0.46211579092778265,
97
+ "total": 145.6902863329742
98
+ },
99
+ "torch_export_gate": {
100
+ "status": "PASS",
101
+ "record": "provenance/export-gate.json",
102
+ "rows_stride": 4,
103
+ "rows_in_window": 2120,
104
+ "rows": 552,
105
+ "argmax_identical": 552,
106
+ "max_probability_error": 5.805650219314895e-05,
107
+ "max_marker_abs_error": 0.0002903938293457031,
108
+ "typed_accuracy": {
109
+ "choice": {
110
+ "count": 151,
111
+ "correct": 103
112
+ },
113
+ "score": {
114
+ "count": 202,
115
+ "correct": 138
116
+ },
117
+ "noul": {
118
+ "count": 153,
119
+ "correct": 121
120
+ }
121
+ }
122
+ },
123
+ "authoring_records": [
124
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/results/authoring_fp32_s1024.json"
125
+ ],
126
+ "runtime_gate": "NOT RUN — gate_julia_runtime.py <this folder> --compute cpu_only|gpu",
127
+ "environment": {
128
+ "date": "2026-09-30T09:43:14+09:00",
129
+ "machine": "Apple M4 Max",
130
+ "macos_version": "27.0",
131
+ "macos_build": "26A428",
132
+ "coreai_build": "coreai-build 3600.83.1",
133
+ "python": "3.11.13",
134
+ "packages": {
135
+ "torch": "2.9.0",
136
+ "coreai-torch": "0.4.1",
137
+ "coreai-core": "1.0.0b2",
138
+ "numpy": "2.3.5",
139
+ "safetensors": "0.7.0"
140
+ },
141
+ "argv": [
142
+ "conversion/julia/export_julia.py",
143
+ "--window",
144
+ "1024",
145
+ "--dtype",
146
+ "fp32",
147
+ "--rows-stride",
148
+ "4"
149
+ ]
150
+ },
151
+ "input_hashes": {
152
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/results/authoring_fp32_s1024.json": "4d27cd1702f639bc1e275e7108937fa74c01ffda6709ab873eddcb5977b76d80",
153
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/oracle/rows.json": "c805348599fb0a87688246f0a290bb293e6f967c5fe938686d23fa8b3ea1377a",
154
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/export_julia.py": "606dc911e2c092e8ee94affc28a52796bb7254909e9f1a2474f8a60ab6f27726",
155
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_julia_model.py": "616f219c230e2b36c40ef0939e8402ec42ed8fae7ce81f0c1b79de91b75b4cfe",
156
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_julia_host.py": "c77e391110b6295efdbad6ecb148b36cda06c6678183fa2e4dd1fd24ad967786",
157
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_gate_metrics.py": "8425201abe9aa7d232d67f95ca8b34a37a33d26835c3540eddcad9606e258de2"
158
+ },
159
+ "bytes": 581765544
160
+ }
macos/fp32-s1024/provenance/runtime-gate.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s1024/reference.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s1024/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
3
+ size 34363188
macos/fp32-s1024/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<bos>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "cls_token": "<bos>",
6
+ "eos_token": "<eos>",
7
+ "extra_special_tokens": [
8
+ "<start_of_turn>",
9
+ "<end_of_turn>"
10
+ ],
11
+ "is_local": true,
12
+ "mask_token": "<mask>",
13
+ "max_length": 512,
14
+ "model_input_names": [
15
+ "input_ids",
16
+ "attention_mask"
17
+ ],
18
+ "model_max_length": 8192,
19
+ "pad_token": "<pad>",
20
+ "padding_side": "right",
21
+ "sep_token": "<eos>",
22
+ "spaces_between_special_tokens": false,
23
+ "stride": 0,
24
+ "tokenizer_class": "TokenizersBackend",
25
+ "truncation_side": "right",
26
+ "truncation_strategy": "longest_first",
27
+ "unk_token": "<unk>"
28
+ }
macos/fp32-s512/julia1_fp32_s512.aimodel/main.hash ADDED
@@ -0,0 +1 @@
 
 
1
+ �(v������wv�S��3��Ӑ�ʾ6ϴ]<��
macos/fp32-s512/julia1_fp32_s512.aimodel/main.mlirb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9287684c9e2d01fb38a7776aa0f5397c733f8f7d390b6cabe36cfb45d3c81d4
3
+ size 578357214
macos/fp32-s512/julia1_fp32_s512.aimodel/metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "creationDate" : "20260930T004000Z",
3
+ "description" : "Julia-1 decision encoder (mmBERT-small + typed head); fp32; S=512; function main; source SupersonicLabs\/Julia-1@a85b127321d580d65176c89ced8273f305745d85",
4
+ "producer" : "coreai-core 1.0.0b2",
5
+ "assetVersion" : "2.0",
6
+ "license" : "Apache-2.0",
7
+ "author" : "mlboydaisuke (coreai-model-zoo); weights Supersonic Labs (Julia-1)"
8
+ }
macos/fp32-s512/metadata.json ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata_version": "0.2",
3
+ "kind": "encoder",
4
+ "decision": {
5
+ "head": "encoder",
6
+ "layout": "julia",
7
+ "window": 512,
8
+ "head_max_len": 256,
9
+ "option_text_tokens": 48,
10
+ "strict_encoding": true,
11
+ "cls_token_id": 2,
12
+ "sep_token_id": 1,
13
+ "pad_token_id": 0,
14
+ "mask_token_id": 4,
15
+ "functions": {
16
+ "main": "main"
17
+ },
18
+ "inputs": {
19
+ "input_ids": {
20
+ "dtype": "int32",
21
+ "shape": [
22
+ 1,
23
+ 512
24
+ ],
25
+ "padding": "right, PAD 0"
26
+ },
27
+ "attention_mask": {
28
+ "dtype": "int32",
29
+ "shape": [
30
+ 1,
31
+ 512
32
+ ],
33
+ "values": "1 real token, 0 padding"
34
+ },
35
+ "qtype_onehot": {
36
+ "dtype": "float32",
37
+ "shape": [
38
+ 1,
39
+ 3
40
+ ],
41
+ "order": [
42
+ "choice",
43
+ "score",
44
+ "noul"
45
+ ]
46
+ }
47
+ },
48
+ "outputs": {
49
+ "token_logits": {
50
+ "dtype": "float32",
51
+ "shape": [
52
+ 1,
53
+ 512
54
+ ],
55
+ "read": "at the option marker positions"
56
+ }
57
+ },
58
+ "options": "the descriptions as given: choice = the criteria mapping's values in order (answer = its id), score = the rubric in order, noul = [criteria.false, criteria.true] or the literal ['false', 'true'] when a noul question has no criteria",
59
+ "sequence": "[CLS 2] tok('{type} question: {instructions}') [SEP 1] ([MASK 4] tok(' ' + option)[:48])... [SEP 1] tok(state) [SEP 1]; state = json.dumps(state, ensure_ascii=False) when not text; strict: a row that does not fit, or text holding the literal '<mask>', is refused, never cut or rewritten (the publisher's non-strict builder turns '<mask>' into ' ')",
60
+ "readout": "softmax of the raw marker logits (T = 1); choice = argmax id, score = sum(i * p_i), noul = p[true]",
61
+ "source_temperature": [
62
+ 1.0,
63
+ 1.0,
64
+ 1.0
65
+ ],
66
+ "calibration": null,
67
+ "precision": "fp32"
68
+ },
69
+ "source": {
70
+ "hf_model_id": "SupersonicLabs/Julia-1",
71
+ "hf_revision": "a85b127321d580d65176c89ced8273f305745d85"
72
+ },
73
+ "assets": {
74
+ "main": "julia1_fp32_s512.aimodel"
75
+ },
76
+ "language": {
77
+ "tokenizer": "SupersonicLabs/Julia-1/tokenizer",
78
+ "vocab_size": 256000,
79
+ "max_context_length": 512
80
+ }
81
+ }
macos/fp32-s512/provenance/export-gate.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s512/provenance/export-manifest.json ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "PASS",
3
+ "measure_only": false,
4
+ "model": "SupersonicLabs/Julia-1",
5
+ "model_sha": "a85b127321d580d65176c89ced8273f305745d85",
6
+ "bundle": "julia1_fp32_s512.aimodel",
7
+ "intended_runtime": "macos",
8
+ "format": "JIT .aimodel (portable; one function main)",
9
+ "aot": false,
10
+ "precision": "fp32",
11
+ "window": 512,
12
+ "inputs": {
13
+ "input_ids": {
14
+ "dtype": "int32",
15
+ "shape": [
16
+ 1,
17
+ 512
18
+ ],
19
+ "padding": "right, PAD 0"
20
+ },
21
+ "attention_mask": {
22
+ "dtype": "int32",
23
+ "shape": [
24
+ 1,
25
+ 512
26
+ ],
27
+ "values": "1 real token, 0 padding"
28
+ },
29
+ "qtype_onehot": {
30
+ "dtype": "float32",
31
+ "shape": [
32
+ 1,
33
+ 3
34
+ ],
35
+ "order": [
36
+ "choice",
37
+ "score",
38
+ "noul"
39
+ ]
40
+ }
41
+ },
42
+ "outputs": {
43
+ "token_logits": {
44
+ "dtype": "float32",
45
+ "shape": [
46
+ 1,
47
+ 512
48
+ ],
49
+ "read": "at the option marker positions"
50
+ }
51
+ },
52
+ "files": [
53
+ {
54
+ "path": "main.hash",
55
+ "bytes": 32,
56
+ "sha256": "e6493cb1f5cabeaa395d558628485d906f3216a0f612b52891cfd334ba3ce732"
57
+ },
58
+ {
59
+ "path": "main.mlirb",
60
+ "bytes": 578357214,
61
+ "sha256": "f9287684c9e2d01fb38a7776aa0f5397c733f8f7d390b6cabe36cfb45d3c81d4"
62
+ },
63
+ {
64
+ "path": "metadata.json",
65
+ "bytes": 391,
66
+ "sha256": "e9ee198e153c93d1d4678df8751cab8f3f9811e69a329c9fdea4cef19bd64366"
67
+ }
68
+ ],
69
+ "op_counts": {
70
+ "builtin.module": 1,
71
+ "coreai.cast": 1,
72
+ "coreai.concat": 44,
73
+ "coreai.constant": 195,
74
+ "coreai.decomposable.broadcasting_add": 130,
75
+ "coreai.decomposable.broadcasting_batch_matmul": 147,
76
+ "coreai.decomposable.broadcasting_mul": 182,
77
+ "coreai.decomposable.broadcasting_sub": 2,
78
+ "coreai.gather_nd": 1,
79
+ "coreai.gelu": 23,
80
+ "coreai.graph": 2,
81
+ "coreai.invoke": 50,
82
+ "coreai.output": 2,
83
+ "coreai.reduce_mean": 2,
84
+ "coreai.relu": 2,
85
+ "coreai.reshape": 122,
86
+ "coreai.rsqrt": 1,
87
+ "coreai.slice": 72,
88
+ "coreai.softmax": 24,
89
+ "coreai.split": 66,
90
+ "coreai.transpose": 215
91
+ },
92
+ "seconds": {
93
+ "torch_export": 2.718504459131509,
94
+ "convert_total": 0.987051707925275,
95
+ "optimize": 0.05586250009946525,
96
+ "save": 0.4756990000605583,
97
+ "total": 53.41586508299224
98
+ },
99
+ "torch_export_gate": {
100
+ "status": "PASS",
101
+ "record": "provenance/export-gate.json",
102
+ "rows_stride": 4,
103
+ "rows_in_window": 2100,
104
+ "rows": 560,
105
+ "argmax_identical": 560,
106
+ "max_probability_error": 6.561606290444999e-05,
107
+ "max_marker_abs_error": 0.0003275871276855469,
108
+ "typed_accuracy": {
109
+ "choice": {
110
+ "count": 149,
111
+ "correct": 102
112
+ },
113
+ "score": {
114
+ "count": 198,
115
+ "correct": 135
116
+ },
117
+ "noul": {
118
+ "count": 150,
119
+ "correct": 128
120
+ }
121
+ }
122
+ },
123
+ "authoring_records": [
124
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/results/authoring_fp32_s512.json"
125
+ ],
126
+ "runtime_gate": "NOT RUN — gate_julia_runtime.py <this folder> --compute cpu_only|gpu",
127
+ "environment": {
128
+ "date": "2026-09-30T09:40:00+09:00",
129
+ "machine": "Apple M4 Max",
130
+ "macos_version": "27.0",
131
+ "macos_build": "26A428",
132
+ "coreai_build": "coreai-build 3600.83.1",
133
+ "python": "3.11.13",
134
+ "packages": {
135
+ "torch": "2.9.0",
136
+ "coreai-torch": "0.4.1",
137
+ "coreai-core": "1.0.0b2",
138
+ "numpy": "2.3.5",
139
+ "safetensors": "0.7.0"
140
+ },
141
+ "argv": [
142
+ "conversion/julia/export_julia.py",
143
+ "--window",
144
+ "512",
145
+ "--dtype",
146
+ "fp32",
147
+ "--rows-stride",
148
+ "4"
149
+ ]
150
+ },
151
+ "input_hashes": {
152
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/results/authoring_fp32_s512.json": "69bc0befb1922e7e93cb9937a9f6ee2386b5e090bae7435dd3a65904b66c5039",
153
+ "/Users/majimadaisuke/code/coreai/_julia1/julia-1/oracle/rows.json": "c805348599fb0a87688246f0a290bb293e6f967c5fe938686d23fa8b3ea1377a",
154
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/export_julia.py": "606dc911e2c092e8ee94affc28a52796bb7254909e9f1a2474f8a60ab6f27726",
155
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_julia_model.py": "616f219c230e2b36c40ef0939e8402ec42ed8fae7ce81f0c1b79de91b75b4cfe",
156
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_julia_host.py": "c77e391110b6295efdbad6ecb148b36cda06c6678183fa2e4dd1fd24ad967786",
157
+ "/Users/majimadaisuke/code/coreai/coreai-models-community-julia1-wt/conversion/julia/_gate_metrics.py": "8425201abe9aa7d232d67f95ca8b34a37a33d26835c3540eddcad9606e258de2"
158
+ },
159
+ "bytes": 578357637
160
+ }
macos/fp32-s512/provenance/runtime-gate.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s512/reference.json ADDED
The diff for this file is too large to render. See raw diff
 
macos/fp32-s512/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
3
+ size 34363188
macos/fp32-s512/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<bos>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "cls_token": "<bos>",
6
+ "eos_token": "<eos>",
7
+ "extra_special_tokens": [
8
+ "<start_of_turn>",
9
+ "<end_of_turn>"
10
+ ],
11
+ "is_local": true,
12
+ "mask_token": "<mask>",
13
+ "max_length": 512,
14
+ "model_input_names": [
15
+ "input_ids",
16
+ "attention_mask"
17
+ ],
18
+ "model_max_length": 8192,
19
+ "pad_token": "<pad>",
20
+ "padding_side": "right",
21
+ "sep_token": "<eos>",
22
+ "spaces_between_special_tokens": false,
23
+ "stride": 0,
24
+ "tokenizer_class": "TokenizersBackend",
25
+ "truncation_side": "right",
26
+ "truncation_strategy": "longest_first",
27
+ "unk_token": "<unk>"
28
+ }