aneuwirth commited on
Commit
037fdd3
·
verified ·
1 Parent(s): 7f1d501
Files changed (10) hide show
  1. LICENSE +70 -0
  2. NOTICE +40 -0
  3. README.md +104 -0
  4. calibration.json +6 -0
  5. config.json +366 -0
  6. marker_calibration.json +42 -0
  7. model.safetensors +3 -0
  8. option_marker.pt +3 -0
  9. tokenizer.json +0 -0
  10. tokenizer_config.json +24 -0
LICENSE ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, and distribution as defined by Sections 1 through 9 of this document.
10
+
11
+ "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
12
+
13
+ "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
14
+
15
+ "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
16
+
17
+ "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
18
+
19
+ "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
20
+
21
+ "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).
22
+
23
+ "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
24
+
25
+ "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
26
+
27
+ "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
28
+
29
+ 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
30
+
31
+ 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
32
+
33
+ 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
34
+
35
+ (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and
36
+
37
+ (b) You must cause any modified files to carry prominent notices stating that You changed the files; and
38
+
39
+ (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
40
+
41
+ (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
42
+
43
+ You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
44
+
45
+ 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
46
+
47
+ 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
48
+
49
+ 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
50
+
51
+ 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or exemplary damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
52
+
53
+ 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
54
+
55
+ END OF TERMS AND CONDITIONS
56
+
57
+ Copyright 2026 The Tars Authors
58
+ Portions Copyright 2026 Victor Hugo Panisa (Von, https://github.com/wfzyx/von)
59
+
60
+ Licensed under the Apache License, Version 2.0 (the "License");
61
+ you may not use this file except in compliance with the License.
62
+ You may obtain a copy of the License at
63
+
64
+ http://www.apache.org/licenses/LICENSE-2.0
65
+
66
+ Unless required by applicable law or agreed to in writing, software
67
+ distributed under the License is distributed on an "AS IS" BASIS,
68
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
69
+ See the License for the specific language governing permissions and
70
+ limitations under the License.
NOTICE ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Tars
2
+ Copyright 2026 The Tars Authors
3
+ https://github.com/andrewneuwirth/Tars
4
+
5
+ This product is based on Von:
6
+ Von, the open-source System One decision model
7
+ Copyright 2026 Victor Hugo Panisa (Victor Hugo)
8
+ https://github.com/wfzyx/von
9
+ Licensed under the Apache License, Version 2.0
10
+
11
+ Von's model uses the ModernBERT encoder:
12
+ ModernBERT, Answer.AI and LightOn (https://github.com/AnswerDotAI/ModernBERT)
13
+ Licensed under the Apache License, Version 2.0
14
+
15
+ What Tars uses from Von
16
+ - Von's model class (von.models.option_marker.OptionMarkerModel), its training
17
+ collator and loss (training/train_option_marker.py) and its synthetic
18
+ general-reasoning generator (training/generate_synthetic_decisions.py).
19
+ These are imported from a Von checkout at run time, not copied.
20
+ - Von 1.2's weights are the starting point of every Tars fine-tune.
21
+ - pipeline/train.py and tars/model.py adapt parts of Von's training loop and
22
+ model loading; each file's header says which parts.
23
+
24
+ What Tars changed or added
25
+ - Three decision questions for commands, built from a skill catalog:
26
+ split (does a new step start here?), route (which family?) and skill
27
+ (which skill in that family?) (tars/questions.py).
28
+ - A planner that chains the three questions into steps (tars/planner.py)
29
+ and an HTTP server (tars/server.py).
30
+ - A full fine-tuning pipeline: label generation, dataset build with speech
31
+ recognition noise and general-reasoning replay, time-budgeted training on
32
+ Apple GPUs or CUDA, and a promote/keep evaluation (pipeline/).
33
+ - A macOS app for defining a catalog, collecting labels from your own AI
34
+ provider or any MCP client, reviewing them, training and comparing models (app/).
35
+
36
+ Training labels
37
+ Tars's own training phrasings were written by an LLM (Anthropic's Claude,
38
+ through the Claude Code CLI). Labels you make with the app come from
39
+ whatever provider you choose. Check your provider's terms before
40
+ redistributing generated datasets.
README.md ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: wfzyx/von
4
+ library_name: transformers
5
+ language: en
6
+ tags:
7
+ - decision-model
8
+ - intent-routing
9
+ - modernbert
10
+ ---
11
+ # Model card: Tars 0.5
12
+
13
+ ## Summary
14
+
15
+ Tars is a 395M-parameter decision model that takes a spoken command and decides what to do with it: where one
16
+ request ends and the next begins, which part of the assistant handles each request, and which skill to run. It
17
+ does not generate text. It picks one option from a list you supply and gives a probability for it.
18
+
19
+ Tars 0.5 is trained only on public, synthetic data: labels for the made-up example catalog in the code repository.
20
+ It is meant as a starting point. Fine-tune it on your own catalog (the Tars Mac app's "Start from Tars" does this),
21
+ since it only knows the example catalog's skills well.
22
+
23
+ | | |
24
+ |---|---|
25
+ | Base model | [Von](https://github.com/wfzyx/von) 1.2 by Victor Hugo (Apache-2.0), a ModernBERT option-marker decision model |
26
+ | Encoder | [ModernBERT](https://github.com/AnswerDotAI/ModernBERT) (Answer.AI and LightOn, Apache-2.0) |
27
+ | Fine-tuning | Full fine-tune (encoder and scoring head) |
28
+ | Size | 395M parameters, about 2.9 GB on disk with tokenizer and calibration |
29
+ | License | Apache-2.0 |
30
+ | Code, labels and benchmark | https://github.com/andrewneuwirth/Tars |
31
+
32
+ ## Intended use
33
+
34
+ - Deciding which skill, tool or app should handle a short command, given a catalog of skills you write.
35
+ - Splitting a sentence that asks for several things into steps.
36
+ - A starting point for fine-tuning on your own catalog.
37
+ - Running locally on a laptop as a fast first pass. Act on Tars's decision above a probability threshold and fall
38
+ back to your own rules or a larger model below it.
39
+
40
+ It is not a chat model, cannot answer questions, and cannot take actions by itself.
41
+
42
+ ## Training data
43
+
44
+ - **Synthetic labels for the public example catalog** (`catalogs/example.json`: made-up apps such as Tunebox and
45
+ Inkpad, 10 families, 31 skills), written by Anthropic's Claude through `pipeline/generate.py`: messy single-skill
46
+ phrasings in 13 speaking styles (terse, casual, speech-to-text errors with no punctuation, rambling, and more),
47
+ 2 to 5 step chained commands, near-miss pairs between similar skills, and not-a-command chatter (`none`). Every
48
+ label was checked against the catalog and deduplicated. 8,432 labeled sentences, published in the code
49
+ repository as `labels/example.jsonl`. Speech recognition noise is added in the build step.
50
+ - **General-reasoning replay**: 13,000 rows from Von's own synthetic decision generator, mixed in so fine-tuning
51
+ does not erase what the base model already did.
52
+ - No personal data: no logged commands, no personal catalog.
53
+ - 76,542 training rows (63,542 split, route and skill rows plus 13,000 replay), 3,124 optimizer steps (batch 8 x 4,
54
+ learning rate 2e-5), 151 minutes on an Apple M5 Pro. The checkpoint is the one from step 3,000, picked on
55
+ validation plus general replay, never on the benchmark.
56
+
57
+ ## Evaluation
58
+
59
+ The public benchmark (`benchmark/public/benchmark.jsonl` in the code repository) has 477 hard multi-step commands
60
+ for the example catalog, 2 to 5 steps each. It was generated separately from the training labels, in three speaking
61
+ styles never used in training (formal dictation, gamer talk, a non-technical parent). No training sentence has 80%
62
+ or more word overlap with a benchmark sentence. A sentence counts only if every step's words, family and skill are
63
+ right.
64
+
65
+ | Model | End to end | Step count right | Split exact | Families right | Skills right | Median latency |
66
+ |---|---|---|---|---|---|---|
67
+ | Von 1.2 (base, no fine-tune) | 0.0% | 1.7% | 0.4% | 0.0% | 0.0% | 791 ms |
68
+ | **Tars 0.5** | **69.2%** | **77.8%** | **71.3%** | **74.8%** | **74.8%** | **280 ms** |
69
+
70
+ Latency is the whole plan (every split, route and skill question), in-process on an Apple M5 Pro (MPS).
71
+
72
+ Each question alone, given the correct steps (`pipeline/evaluate.py`, same 477 sentences): split 44.2% → 98.6%,
73
+ route 29.0% → 98.8%, skill 74.7% → 100.0%, whole sentence 0.0% → 90.1%. General reasoning on 600 of Von's replay
74
+ rows went from 84.3% to 91.2%.
75
+
76
+ Tars 0.4's numbers (65.0% end to end) came from a private benchmark for a personal catalog and can't be compared
77
+ directly with these.
78
+
79
+ ## Limits
80
+
81
+ - It knows the example catalog's skills. Your own catalog needs a fine-tune; this model is the starting point.
82
+ - Most errors are extra splits: a filler word or aside at the start or middle ("yo", "please", "cause the raid
83
+ starts then") becomes its own `none` step. 92 of the 147 failed benchmark sentences have too many steps.
84
+ - Labels written by an LLM carry that LLM's mistakes and its idea of how people talk. A few benchmark labels are
85
+ debatable too.
86
+ - English only. Tested on short commands; long dictation is out of scope.
87
+ - Probabilities come from Von 1.2's calibration and are not refit after fine-tuning. On the benchmark they are close
88
+ to 1.0 for almost every step, including steps in sentences it gets wrong, so they are a weak signal. Check them on
89
+ your own data before choosing a threshold.
90
+
91
+ ## Use
92
+
93
+ ```python
94
+ from tars import Catalog, Planner # https://github.com/andrewneuwirth/Tars
95
+ planner = Planner(Catalog.load("catalogs/example.json"), "path/to/this/folder")
96
+ planner.plan("open tunebox and play some jazz then mute")
97
+ ```
98
+
99
+ The Tars Mac app downloads this checkpoint to fine-tune from it ("Start from Tars 0.5" in Settings).
100
+
101
+ ## Attribution
102
+
103
+ Tars is a fine-tune of [Von](https://github.com/wfzyx/von) 1.2 by Victor Hugo, Apache-2.0. See NOTICE. The training
104
+ labels and benchmark were written by Anthropic's Claude.
calibration.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "wfzyx/von-1.0",
3
+ "temperature": 1.1692,
4
+ "best_val_accuracy": 0.9643,
5
+ "timestamp": 1789856665.9293928
6
+ }
config.json ADDED
@@ -0,0 +1,366 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "ModernBertModel"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 50281,
8
+ "classifier_activation": "gelu",
9
+ "classifier_bias": false,
10
+ "classifier_dropout": 0.0,
11
+ "classifier_pooling": "mean",
12
+ "classifiers_size": [
13
+ 3,
14
+ 2,
15
+ 2,
16
+ 2,
17
+ 2,
18
+ 2,
19
+ 1,
20
+ 2,
21
+ 3,
22
+ 2,
23
+ 2,
24
+ 2,
25
+ 3,
26
+ 3,
27
+ 3,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 2,
32
+ 2,
33
+ 3,
34
+ 2,
35
+ 2,
36
+ 2,
37
+ 2,
38
+ 2,
39
+ 6,
40
+ 2,
41
+ 2,
42
+ 2,
43
+ 2,
44
+ 2,
45
+ 3,
46
+ 3,
47
+ 3,
48
+ 3,
49
+ 3,
50
+ 3,
51
+ 3,
52
+ 2,
53
+ 2,
54
+ 2,
55
+ 2,
56
+ 3,
57
+ 3,
58
+ 3,
59
+ 3,
60
+ 3,
61
+ 3,
62
+ 3,
63
+ 3,
64
+ 2,
65
+ 2,
66
+ 2,
67
+ 2,
68
+ 2,
69
+ 2,
70
+ 16,
71
+ 100,
72
+ 13,
73
+ 100,
74
+ 8,
75
+ 3,
76
+ 3,
77
+ 2,
78
+ 3,
79
+ 2,
80
+ 4,
81
+ 3,
82
+ 2,
83
+ 3,
84
+ 2,
85
+ 2,
86
+ 2,
87
+ 2,
88
+ 2,
89
+ 3,
90
+ 2,
91
+ 3,
92
+ 2,
93
+ 4,
94
+ 3,
95
+ 3,
96
+ 3,
97
+ 2,
98
+ 3,
99
+ 1,
100
+ 2,
101
+ 2,
102
+ 3,
103
+ 13,
104
+ 2,
105
+ 2,
106
+ 3,
107
+ 2,
108
+ 2,
109
+ 3,
110
+ 3,
111
+ 3,
112
+ 3,
113
+ 2,
114
+ 3,
115
+ 2,
116
+ 3,
117
+ 2,
118
+ 3,
119
+ 2,
120
+ 2,
121
+ 2,
122
+ 2,
123
+ 2,
124
+ 3,
125
+ 4,
126
+ 3,
127
+ 3,
128
+ 2,
129
+ 2,
130
+ 3,
131
+ 3,
132
+ 2,
133
+ 2,
134
+ 2,
135
+ 2,
136
+ 2,
137
+ 4,
138
+ 3,
139
+ 2,
140
+ 2,
141
+ 2,
142
+ 3,
143
+ 3,
144
+ 3,
145
+ 2,
146
+ 3
147
+ ],
148
+ "cls_token_id": 50281,
149
+ "decoder_bias": true,
150
+ "deterministic_flash_attn": false,
151
+ "dtype": "float32",
152
+ "embedding_dropout": 0.0,
153
+ "eos_token_id": 50282,
154
+ "global_attn_every_n_layers": 3,
155
+ "gradient_checkpointing": false,
156
+ "hidden_activation": "gelu",
157
+ "hidden_size": 1024,
158
+ "id2label": {
159
+ "0": "entailment",
160
+ "1": "neutral",
161
+ "2": "contradiction"
162
+ },
163
+ "initializer_cutoff_factor": 2.0,
164
+ "initializer_range": 0.02,
165
+ "intermediate_size": 2624,
166
+ "label2id": {
167
+ "contradiction": 2,
168
+ "entailment": 0,
169
+ "neutral": 1
170
+ },
171
+ "layer_norm_eps": 1e-05,
172
+ "layer_types": [
173
+ "full_attention",
174
+ "sliding_attention",
175
+ "sliding_attention",
176
+ "full_attention",
177
+ "sliding_attention",
178
+ "sliding_attention",
179
+ "full_attention",
180
+ "sliding_attention",
181
+ "sliding_attention",
182
+ "full_attention",
183
+ "sliding_attention",
184
+ "sliding_attention",
185
+ "full_attention",
186
+ "sliding_attention",
187
+ "sliding_attention",
188
+ "full_attention",
189
+ "sliding_attention",
190
+ "sliding_attention",
191
+ "full_attention",
192
+ "sliding_attention",
193
+ "sliding_attention",
194
+ "full_attention",
195
+ "sliding_attention",
196
+ "sliding_attention",
197
+ "full_attention",
198
+ "sliding_attention",
199
+ "sliding_attention",
200
+ "full_attention"
201
+ ],
202
+ "local_attention": 128,
203
+ "max_position_embeddings": 8192,
204
+ "mlp_bias": false,
205
+ "mlp_dropout": 0.0,
206
+ "model_type": "modernbert",
207
+ "norm_bias": false,
208
+ "norm_eps": 1e-05,
209
+ "num_attention_heads": 16,
210
+ "num_hidden_layers": 28,
211
+ "pad_token_id": 50283,
212
+ "position_embedding_type": "absolute",
213
+ "problem_type": "single_label_classification",
214
+ "rope_parameters": {
215
+ "full_attention": {
216
+ "rope_theta": 160000.0,
217
+ "rope_type": "default"
218
+ },
219
+ "sliding_attention": {
220
+ "rope_theta": 10000.0,
221
+ "rope_type": "default"
222
+ }
223
+ },
224
+ "sep_token_id": 50282,
225
+ "sparse_pred_ignore_index": -100,
226
+ "sparse_prediction": false,
227
+ "tasks": [
228
+ "glue/mnli",
229
+ "glue/qnli",
230
+ "glue/rte",
231
+ "glue/wnli",
232
+ "glue/mrpc",
233
+ "glue/qqp",
234
+ "glue/stsb",
235
+ "super_glue/boolq",
236
+ "super_glue/cb",
237
+ "super_glue/multirc",
238
+ "super_glue/wic",
239
+ "super_glue/axg",
240
+ "anli/a1",
241
+ "anli/a2",
242
+ "anli/a3",
243
+ "sick/label",
244
+ "sick/entailment_AB",
245
+ "snli",
246
+ "scitail/snli_format",
247
+ "hans",
248
+ "WANLI",
249
+ "recast/recast_ner",
250
+ "recast/recast_sentiment",
251
+ "recast/recast_verbnet",
252
+ "recast/recast_megaveridicality",
253
+ "recast/recast_verbcorner",
254
+ "recast/recast_kg_relations",
255
+ "recast/recast_factuality",
256
+ "recast/recast_puns",
257
+ "probability_words_nli/reasoning_1hop",
258
+ "probability_words_nli/usnli",
259
+ "probability_words_nli/reasoning_2hop",
260
+ "nan-nli",
261
+ "nli_fever",
262
+ "breaking_nli",
263
+ "conj_nli",
264
+ "fracas",
265
+ "dialogue_nli",
266
+ "mpe",
267
+ "dnc",
268
+ "recast_white/fnplus",
269
+ "recast_white/sprl",
270
+ "recast_white/dpr",
271
+ "robust_nli/IS_CS",
272
+ "robust_nli/LI_LI",
273
+ "robust_nli/ST_WO",
274
+ "robust_nli/PI_SP",
275
+ "robust_nli/PI_CD",
276
+ "robust_nli/ST_SE",
277
+ "robust_nli/ST_NE",
278
+ "robust_nli/ST_LM",
279
+ "robust_nli_is_sd",
280
+ "robust_nli_li_ts",
281
+ "add_one_rte",
282
+ "paws/labeled_final",
283
+ "glue/cola",
284
+ "glue/sst2",
285
+ "pragmeval/pdtb",
286
+ "lex_glue/eurlex",
287
+ "lex_glue/scotus",
288
+ "lex_glue/ledgar",
289
+ "lex_glue/unfair_tos",
290
+ "dynasent/dynabench.dynasent.r1.all/r1",
291
+ "dynasent/dynabench.dynasent.r2.all/r2",
292
+ "cycic_classification",
293
+ "lingnli",
294
+ "monotonicity-entailment",
295
+ "scinli",
296
+ "naturallogic",
297
+ "dynahate",
298
+ "syntactic-augmentation-nli",
299
+ "autotnli",
300
+ "defeasible-nli/atomic",
301
+ "defeasible-nli/snli",
302
+ "help-nli",
303
+ "nli-veridicality-transitivity",
304
+ "lonli",
305
+ "dadc-limit-nli",
306
+ "folio",
307
+ "tomi-nli",
308
+ "puzzte",
309
+ "temporal-nli",
310
+ "counterfactually-augmented-snli",
311
+ "cnli",
312
+ "boolq-natural-perturbations",
313
+ "equate",
314
+ "chaos-mnli-ambiguity",
315
+ "logiqa-2.0-nli",
316
+ "mindgames",
317
+ "ConTRoL-nli",
318
+ "logical-fallacy",
319
+ "cladder",
320
+ "conceptrules_v2",
321
+ "zero-shot-label-nli",
322
+ "scone",
323
+ "monli",
324
+ "SpaceNLI",
325
+ "propsegment/nli",
326
+ "FLD.v2/default",
327
+ "FLD.v2/star",
328
+ "SDOH-NLI",
329
+ "scifact_entailment",
330
+ "feasibilityQA",
331
+ "AdjectiveScaleProbe-nli",
332
+ "resnli",
333
+ "semantic_fragments_nli",
334
+ "dataset_train_nli",
335
+ "nlgraph",
336
+ "ruletaker",
337
+ "PARARULE-Plus",
338
+ "logical-entailment",
339
+ "nope",
340
+ "LogicNLI",
341
+ "contract-nli/contractnli_a/seg",
342
+ "contract-nli/contractnli_b/full",
343
+ "nli4ct_semeval2024",
344
+ "biosift-nli",
345
+ "SIGA-nli",
346
+ "FOL-nli",
347
+ "doc-nli",
348
+ "mctest-nli",
349
+ "natural-language-satisfiability",
350
+ "idioms-nli",
351
+ "lifecycle-entailment",
352
+ "MSciNLI",
353
+ "hover-3way/nli",
354
+ "seahorse_summarization_evaluation",
355
+ "missing-item-prediction/contrastive",
356
+ "Pol_NLI",
357
+ "synthetic-retrieval-NLI/count",
358
+ "synthetic-retrieval-NLI/position",
359
+ "synthetic-retrieval-NLI/binary",
360
+ "babi_nli",
361
+ "gen_debiased_nli"
362
+ ],
363
+ "tie_word_embeddings": true,
364
+ "transformers_version": "5.17.0",
365
+ "vocab_size": 50368
366
+ }
marker_calibration.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_type": "option_marker",
3
+ "base_model": "von-1.2.0",
4
+ "model_id": "tars",
5
+ "independent_options": true,
6
+ "best_val_accuracy": 0.9472,
7
+ "temperature": 2.2,
8
+ "timestamp": 1790207664.1941526,
9
+ "calibration_map": {
10
+ "bias": 2.0151,
11
+ "entropy": -3.2369,
12
+ "log_tokens": 11.4916,
13
+ "n_options": -3.5574,
14
+ "lo": 0.3,
15
+ "hi": 12.0
16
+ },
17
+ "calibration_map_features": "T = bias + entropy*H_norm + log_tokens*log10(state_tokens)/4 + n_options*(K/8), clamped to [lo,hi]",
18
+ "calibration_fitted_on": "JevBench public easy+original+hard (231 items); objective 0.5*JevBench-Calibration + 0.5*per-tier honesty, best of 8 optimiser restarts",
19
+ "calibration_ece": {
20
+ "easy": 0.109,
21
+ "standard": 0.045,
22
+ "hard": 0.089
23
+ },
24
+ "noul_zero_shot_prior": {
25
+ "a": -0.5,
26
+ "b": 0.3
27
+ },
28
+ "noul_zero_shot_prior_fitted_on": "benchmarks/data/noul_zeroshot_dev.jsonl (150 boolq validation + 25 handwritten, never jabr/JevBench)",
29
+ "training": {
30
+ "kind": "tars fine-tune (full)",
31
+ "step": 3000,
32
+ "val": 0.9813423166623477,
33
+ "val_replay": 0.9116666666666666,
34
+ "baseline": {
35
+ "val": 0.45037574501166106,
36
+ "val_replay": 0.8433333333333334,
37
+ "test_style": null,
38
+ "test_real": null
39
+ },
40
+ "timestamp": 1794739177.795239
41
+ }
42
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fd7044673acf8eef39d6cd671737ec679e8e5357b40c71f143c99937d8fa260
3
+ size 1579143688
option_marker.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:224e95d7d3ef1638e6c718b259c3105c937c26599f30ca8309ac6249ff7124f1
3
+ size 1581297131
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "clean_up_tokenization_spaces": true,
4
+ "cls_token": "[CLS]",
5
+ "is_local": true,
6
+ "local_files_only": false,
7
+ "mask_token": "[MASK]",
8
+ "max_length": 2048,
9
+ "model_input_names": [
10
+ "input_ids",
11
+ "attention_mask"
12
+ ],
13
+ "model_max_length": 8192,
14
+ "pad_to_multiple_of": null,
15
+ "pad_token": "[PAD]",
16
+ "pad_token_type_id": 0,
17
+ "padding_side": "right",
18
+ "sep_token": "[SEP]",
19
+ "stride": 0,
20
+ "tokenizer_class": "TokenizersBackend",
21
+ "truncation_side": "right",
22
+ "truncation_strategy": "longest_first",
23
+ "unk_token": "[UNK]"
24
+ }