Release Jet v6: full merged Qwen3.5-4B BF16 weights and native typed inference
Browse files- LICENSE +202 -0
- README.md +110 -107
- TRAINING_HISTORY.md +0 -71
- calibration.json +4 -4
- chat_template.jinja +106 -37
- config.json +81 -28
- evaluation.json +371 -0
- export-provenance.json +0 -12
- export-validation.txt +0 -1
- format.py +143 -0
- golden.json +0 -0
- inference.py +41 -0
- jet.py +41 -0
- merge-provenance.json +11 -0
- merge-validation.json +255 -0
- model.safetensors → model-00001-of-00009.safetensors +2 -2
- onnx/model_q8.onnx → model-00002-of-00009.safetensors +2 -2
- model-00003-of-00009.safetensors +3 -0
- model-00004-of-00009.safetensors +3 -0
- model-00005-of-00009.safetensors +3 -0
- model-00006-of-00009.safetensors +3 -0
- model-00007-of-00009.safetensors +3 -0
- model-00008-of-00009.safetensors +3 -0
- model-00009-of-00009.safetensors +3 -0
- model.safetensors.index.json +432 -317
- release-manifest.json +35 -25
- requirements.txt +5 -0
- runtime.py +36 -0
- tokenizer.json +2 -2
- tokenizer_config.json +20 -4
LICENSE
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
Apache License
|
| 3 |
+
Version 2.0, January 2004
|
| 4 |
+
http://www.apache.org/licenses/
|
| 5 |
+
|
| 6 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 7 |
+
|
| 8 |
+
1. Definitions.
|
| 9 |
+
|
| 10 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 11 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 12 |
+
|
| 13 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 14 |
+
the copyright owner that is granting the License.
|
| 15 |
+
|
| 16 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 17 |
+
other entities that control, are controlled by, or are under common
|
| 18 |
+
control with that entity. For the purposes of this definition,
|
| 19 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 20 |
+
direction or management of such entity, whether by contract or
|
| 21 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 22 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 23 |
+
|
| 24 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 25 |
+
exercising permissions granted by this License.
|
| 26 |
+
|
| 27 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 28 |
+
including but not limited to software source code, documentation
|
| 29 |
+
source, and configuration files.
|
| 30 |
+
|
| 31 |
+
"Object" form shall mean any form resulting from mechanical
|
| 32 |
+
transformation or translation of a Source form, including but
|
| 33 |
+
not limited to compiled object code, generated documentation,
|
| 34 |
+
and conversions to other media types.
|
| 35 |
+
|
| 36 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 37 |
+
Object form, made available under the License, as indicated by a
|
| 38 |
+
copyright notice that is included in or attached to the work
|
| 39 |
+
(an example is provided in the Appendix below).
|
| 40 |
+
|
| 41 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 42 |
+
form, that is based on (or derived from) the Work and for which the
|
| 43 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 44 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 45 |
+
of this License, Derivative Works shall not include works that remain
|
| 46 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 47 |
+
the Work and Derivative Works thereof.
|
| 48 |
+
|
| 49 |
+
"Contribution" shall mean any work of authorship, including
|
| 50 |
+
the original version of the Work and any modifications or additions
|
| 51 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 52 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 53 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 54 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 55 |
+
means any form of electronic, verbal, or written communication sent
|
| 56 |
+
to the Licensor or its representatives, including but not limited to
|
| 57 |
+
communication on electronic mailing lists, source code control systems,
|
| 58 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 59 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 60 |
+
excluding communication that is conspicuously marked or otherwise
|
| 61 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 62 |
+
|
| 63 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 64 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 65 |
+
subsequently incorporated within the Work.
|
| 66 |
+
|
| 67 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 68 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 69 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 70 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 71 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 72 |
+
Work and such Derivative Works in Source or Object form.
|
| 73 |
+
|
| 74 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 75 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 76 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 77 |
+
(except as stated in this section) patent license to make, have made,
|
| 78 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 79 |
+
where such license applies only to those patent claims licensable
|
| 80 |
+
by such Contributor that are necessarily infringed by their
|
| 81 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 82 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 83 |
+
institute patent litigation against any entity (including a
|
| 84 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 85 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 86 |
+
or contributory patent infringement, then any patent licenses
|
| 87 |
+
granted to You under this License for that Work shall terminate
|
| 88 |
+
as of the date such litigation is filed.
|
| 89 |
+
|
| 90 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 91 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 92 |
+
modifications, and in Source or Object form, provided that You
|
| 93 |
+
meet the following conditions:
|
| 94 |
+
|
| 95 |
+
(a) You must give any other recipients of the Work or
|
| 96 |
+
Derivative Works a copy of this License; and
|
| 97 |
+
|
| 98 |
+
(b) You must cause any modified files to carry prominent notices
|
| 99 |
+
stating that You changed the files; and
|
| 100 |
+
|
| 101 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 102 |
+
that You distribute, all copyright, patent, trademark, and
|
| 103 |
+
attribution notices from the Source form of the Work,
|
| 104 |
+
excluding those notices that do not pertain to any part of
|
| 105 |
+
the Derivative Works; and
|
| 106 |
+
|
| 107 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 108 |
+
distribution, then any Derivative Works that You distribute must
|
| 109 |
+
include a readable copy of the attribution notices contained
|
| 110 |
+
within such NOTICE file, excluding those notices that do not
|
| 111 |
+
pertain to any part of the Derivative Works, in at least one
|
| 112 |
+
of the following places: within a NOTICE text file distributed
|
| 113 |
+
as part of the Derivative Works; within the Source form or
|
| 114 |
+
documentation, if provided along with the Derivative Works; or,
|
| 115 |
+
within a display generated by the Derivative Works, if and
|
| 116 |
+
wherever such third-party notices normally appear. The contents
|
| 117 |
+
of the NOTICE file are for informational purposes only and
|
| 118 |
+
do not modify the License. You may add Your own attribution
|
| 119 |
+
notices within Derivative Works that You distribute, alongside
|
| 120 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 121 |
+
that such additional attribution notices cannot be construed
|
| 122 |
+
as modifying the License.
|
| 123 |
+
|
| 124 |
+
You may add Your own copyright statement to Your modifications and
|
| 125 |
+
may provide additional or different license terms and conditions
|
| 126 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 127 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 128 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 129 |
+
the conditions stated in this License.
|
| 130 |
+
|
| 131 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 132 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 133 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 134 |
+
this License, without any additional terms or conditions.
|
| 135 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 136 |
+
the terms of any separate license agreement you may have executed
|
| 137 |
+
with Licensor regarding such Contributions.
|
| 138 |
+
|
| 139 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 140 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 141 |
+
except as required for reasonable and customary use in describing the
|
| 142 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 143 |
+
|
| 144 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 145 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 146 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 147 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 148 |
+
implied, including, without limitation, any warranties or conditions
|
| 149 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 150 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 151 |
+
appropriateness of using or redistributing the Work and assume any
|
| 152 |
+
risks associated with Your exercise of permissions under this License.
|
| 153 |
+
|
| 154 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 155 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 156 |
+
unless required by applicable law (such as deliberate and grossly
|
| 157 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 158 |
+
liable to You for damages, including any direct, indirect, special,
|
| 159 |
+
incidental, or consequential damages of any character arising as a
|
| 160 |
+
result of this License or out of the use or inability to use the
|
| 161 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 162 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 163 |
+
other commercial damages or losses), even if such Contributor
|
| 164 |
+
has been advised of the possibility of such damages.
|
| 165 |
+
|
| 166 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 167 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 168 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 169 |
+
or other liability obligations and/or rights consistent with this
|
| 170 |
+
License. However, in accepting such obligations, You may act only
|
| 171 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 172 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 173 |
+
defend, and hold each Contributor harmless for any liability
|
| 174 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 175 |
+
of your accepting any such warranty or additional liability.
|
| 176 |
+
|
| 177 |
+
END OF TERMS AND CONDITIONS
|
| 178 |
+
|
| 179 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 180 |
+
|
| 181 |
+
To apply the Apache License to your work, attach the following
|
| 182 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 183 |
+
replaced with your own identifying information. (Don't include
|
| 184 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 185 |
+
comment syntax for the file format. We also recommend that a
|
| 186 |
+
file or class name and description of purpose be included on the
|
| 187 |
+
same "printed page" as the copyright notice for easier
|
| 188 |
+
identification within third-party archives.
|
| 189 |
+
|
| 190 |
+
Copyright 2026 Alibaba Cloud
|
| 191 |
+
|
| 192 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 193 |
+
you may not use this file except in compliance with the License.
|
| 194 |
+
You may obtain a copy of the License at
|
| 195 |
+
|
| 196 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 197 |
+
|
| 198 |
+
Unless required by applicable law or agreed to in writing, software
|
| 199 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 200 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 201 |
+
See the License for the specific language governing permissions and
|
| 202 |
+
limitations under the License.
|
README.md
CHANGED
|
@@ -1,130 +1,133 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
-
base_model: Qwen/Qwen3
|
| 4 |
-
library_name:
|
| 5 |
language:
|
| 6 |
- en
|
| 7 |
tags:
|
| 8 |
- decision-model
|
| 9 |
-
-
|
| 10 |
-
-
|
| 11 |
-
-
|
| 12 |
-
- lora
|
| 13 |
---
|
| 14 |
|
| 15 |
# Jet
|
| 16 |
|
| 17 |
-
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
|
| 20 |
-
typed questions; it returns choices, scores, and probabilities without generating
|
| 21 |
-
free-form text. Answers always follow the requested type, but decisions can still
|
| 22 |
-
be wrong.
|
| 23 |
|
| 24 |
-
|
| 25 |
-
[Hugging Face demo](https://huggingface.co/spaces/michaljach/jet) ·
|
| 26 |
-
[Training history](TRAINING_HISTORY.md)
|
| 27 |
-
|
| 28 |
-
| Type | Criteria | Answer |
|
| 29 |
|---|---|---|
|
| 30 |
-
| `choice` | 2–255 named options | Selected key
|
| 31 |
-
| `score` | 2–10 ordered levels |
|
| 32 |
-
| `noul` |
|
| 33 |
|
| 34 |
-
|
|
|
|
| 35 |
|
| 36 |
-
|
| 37 |
-
git clone https://github.com/jach-labs/jet
|
| 38 |
-
cd jet
|
| 39 |
-
uv sync # Apple Silicon / Metal
|
| 40 |
-
# Linux with NVIDIA: uv sync --extra cuda
|
| 41 |
-
JET_API_KEY=secret uv run jet-serve --base-model michaljach/jet
|
| 42 |
-
```
|
| 43 |
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
-d '{"state":"I was charged twice this month.","questions":{"topic":{"type":"choice","instructions":"What is the primary issue?","criteria":{"billing":"billing or payment problem","bug":"the product is broken"}},"escalate":{"type":"noul","instructions":"Does this require human support?"}}}'
|
| 49 |
-
```
|
| 50 |
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
calibration temperatures, and returns typed answers. Serving truncates long states
|
| 54 |
-
in the middle; the Decision Index adapter instead requires complete inputs.
|
| 55 |
-
|
| 56 |
-
## How it works
|
| 57 |
-
|
| 58 |
-
Each answer option maps to a single token. Jet reads the next-token logits only
|
| 59 |
-
for those labels and applies softmax with a temperature fitted on held-out data.
|
| 60 |
-
Standard serving uses one forward pass per question, without sampling.
|
| 61 |
-
The fused bf16 weights include the trained LoRA updates.
|
| 62 |
-
|
| 63 |
-
The benchmark package also provides `decision_index_ensemble:TwoOrderJetEngine`.
|
| 64 |
-
It averages original and reversed option-order probabilities using exact option
|
| 65 |
-
keys. This approximately doubles forward-pass work and is separate from the
|
| 66 |
-
standard serving API.
|
| 67 |
-
|
| 68 |
-
## Training and evaluation
|
| 69 |
-
|
| 70 |
-
The current checkpoint was trained locally on an RTX 4080 Super with MLX CUDA.
|
| 71 |
-
The final training stage used 15,997 examples, rank-16 LoRA, learning rate `1e-5`,
|
| 72 |
-
gradient checkpointing, and one epoch. Selection chose step 3,500 of 3,570.
|
| 73 |
-
Public training partitions and deterministic generators cover relevance,
|
| 74 |
-
entailment, stance, sarcasm, routing, arithmetic, Boolean rules, code behavior,
|
| 75 |
-
commonsense completion, and tool-response preference. Calibration uses a separate
|
| 76 |
-
held-out split.
|
| 77 |
-
|
| 78 |
-
| Held-out set | Rows | Accuracy |
|
| 79 |
-
|---|---:|---:|
|
| 80 |
-
| General typed decisions | 3,683 | 66.66% |
|
| 81 |
-
| Intent, entailment, commonsense and tool-response preference | 600 | 73.50% |
|
| 82 |
-
| Transfer tasks | 720 | 67.50% |
|
| 83 |
-
| Domain-held-out tool routing | 240 | 85.00% |
|
| 84 |
-
|
| 85 |
-
These are single-order adapter measurements. They are not a Decision Index
|
| 86 |
-
leaderboard score. Partial public benchmark reconstructions do not cover the
|
| 87 |
-
complete current suite. Cross-dataset text overlap inherited during training
|
| 88 |
-
also prevents a certified decontamination claim. Performance varies by task;
|
| 89 |
-
see the [evaluation report](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5/summary.md) for full methodology,
|
| 90 |
-
limitations, and the separate two-order results.
|
| 91 |
-
|
| 92 |
-
To reproduce training, source acquisition, audits, and evaluation, use the
|
| 93 |
-
[recorded protocol](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5/protocol.md) and scripts under `scripts/`.
|
| 94 |
-
Dataset sources, revisions, exclusion checks, configuration, and checkpoint hashes
|
| 95 |
-
are recorded under `docs/training/`.
|
| 96 |
|
| 97 |
```sh
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
|
|
|
| 103 |
```
|
| 104 |
|
| 105 |
-
|
| 106 |
-
`--backend claude-code` mode. It is not required for serving or reproducing the
|
| 107 |
-
public-source training pipeline; hosted generation can incur charges.
|
| 108 |
-
|
| 109 |
-
## Deployment
|
| 110 |
-
|
| 111 |
-
The Hugging Face model package includes fused bf16 weights, tokenizer,
|
| 112 |
-
calibration, golden reference vectors, and a q8 ONNX export. The Gradio Space
|
| 113 |
-
uses a pinned model revision and exposes `/decide`; its availability depends on
|
| 114 |
-
Hugging Face's free hosting quota. See [deployment instructions](https://github.com/jach-labs/jet/blob/main/deploy/huggingface/README.md).
|
| 115 |
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
- `src/train.py`, `src/evaluate.py`, `src/fuse.py`: training, calibration, evaluation and fusion
|
| 126 |
-
- `src/data/`, `scripts/`: data builders and reproducible experiments
|
| 127 |
-
- `src/decision_index_engine.py`, `src/decision_index_ensemble.py`: benchmark adapters
|
| 128 |
-
- `deploy/huggingface/`: hosted demo and API deployment
|
| 129 |
|
| 130 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
base_model: Qwen/Qwen3.5-4B
|
| 4 |
+
library_name: transformers
|
| 5 |
language:
|
| 6 |
- en
|
| 7 |
tags:
|
| 8 |
- decision-model
|
| 9 |
+
- qwen3_5
|
| 10 |
+
- transformers
|
| 11 |
+
- calibrated
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
# Jet
|
| 15 |
|
| 16 |
+
Jet is a typed decision model based on Qwen3.5-4B. It selects from supplied
|
| 17 |
+
options and returns calibrated label probabilities, without generating free-form answers.
|
| 18 |
+
This release contains the **full merged BF16 model**. It loads directly without a separate base-model download or LoRA adapter.
|
| 19 |
|
| 20 |
+
**Version:** v6.0.0 · **Checkpoint:** step 3750 · **Released:** 2026-09-24.
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
+
| Question type | Input | Output |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
|---|---|---|
|
| 24 |
+
| `choice` | 2–255 named options | Selected key and probabilities |
|
| 25 |
+
| `score` | 2–10 ordered levels | Expected zero-based score, selected level and probabilities |
|
| 26 |
+
| `noul` | Yes/no question | Probability of yes |
|
| 27 |
|
| 28 |
+
The model is trained for text decisions. It is not a general chat or image model.
|
| 29 |
+
The previous Qwen3-0.6B release is preserved under the `qwen3-0.6b-final` tag in this repository.
|
| 30 |
|
| 31 |
+
## Run
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
+
Tested backend: Python 3.12 on Linux, NVIDIA CUDA, PyTorch 2.11.0 + CUDA 12.8,
|
| 34 |
+
Transformers 5.17.0 for the included optimized runtime. Training used FP32 LoRA parameters; this
|
| 35 |
+
release merges them into BF16 weights. The original base revision is
|
| 36 |
+
`851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`.
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
Download the versioned repository, install dependencies in a dedicated environment,
|
| 39 |
+
then run from that directory. Inference uses only the downloaded model files.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
```sh
|
| 42 |
+
hf download michaljach/jet --revision v6.0.0 --local-dir jet
|
| 43 |
+
cd jet
|
| 44 |
+
python -m pip install -r requirements.txt
|
| 45 |
+
python jet.py <<'JSON'
|
| 46 |
+
{"state":"I was charged twice this month.","questions":{"topic":{"type":"choice","instructions":"What is the primary issue?","criteria":{"billing":"billing or payment problem","bug":"the product is broken"}}}}
|
| 47 |
+
JSON
|
| 48 |
```
|
| 49 |
|
| 50 |
+
Or use the included Python wrapper:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
+
```python
|
| 53 |
+
from jet import Jet
|
| 54 |
+
model = Jet()
|
| 55 |
+
result = model.decide(
|
| 56 |
+
"I was charged twice this month.",
|
| 57 |
+
{"billing": {"type": "noul", "instructions": "Is this a billing issue?"}},
|
| 58 |
+
)
|
| 59 |
+
print(result)
|
| 60 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
+
Use the native prompt and restricted label-token readout in these files. Generic
|
| 63 |
+
text-generation pipelines do not implement Jet's typed inference or calibration.
|
| 64 |
+
The wrapper rejects any complete question prompt over 8192 tokens rather than
|
| 65 |
+
truncating it. Questions are processed sequentially without shared-prefix caching.
|
| 66 |
+
`confidence` is normalized inverse entropy, not a separate correctness probability.
|
| 67 |
+
The FLA kernel is enabled; causal convolution uses the PyTorch fallback.
|
| 68 |
+
The previous Qwen3-0.6B MLX/ONNX exports belong to the archived release. This
|
| 69 |
+
release provides BF16 Transformers weights; its native typed API is shown above.
|
| 70 |
+
|
| 71 |
+
## Training
|
| 72 |
+
|
| 73 |
+
- 15,997 `train_v5_r2` examples; one epoch, 4,000 optimizer updates.
|
| 74 |
+
- Fresh LoRA rank/alpha 16, dropout 0.05, learning rate 1e-4, batch 1 × accumulation 4.
|
| 75 |
+
- Soft-target objective, label smoothing 0.02, ordinal loss weight 2; seed 240924.
|
| 76 |
+
- Selected step 3750 by lowest NLL on a separate 1,400-row selection split.
|
| 77 |
+
- Training ran on one RTX 4080 SUPER 16 GB, about 2 h 26 min; peak allocated memory 10.39 GB.
|
| 78 |
+
- Calibration uses a separate 1,400-row split and per-type temperature scaling.
|
| 79 |
+
|
| 80 |
+
Training and evaluation inherit Jet v5's public-source classification, ordinal and
|
| 81 |
+
yes/no mixtures and earlier source-exposure limitations. Source overlap has not
|
| 82 |
+
been comprehensively ruled out. Dataset files are not included in this release.
|
| 83 |
+
The backbone and pilot were selected using earlier validation results.
|
| 84 |
+
|
| 85 |
+
## Evaluation at release
|
| 86 |
+
|
| 87 |
+
| Evaluation | Result |
|
| 88 |
+
|---|---:|
|
| 89 |
+
| Selection accuracy, 1,400 rows | 87.93% |
|
| 90 |
+
| Selection NLL | 0.3561 |
|
| 91 |
+
| Independent local test accuracy, 600 rows | 94.00% |
|
| 92 |
+
| Local test NLL, raw / calibrated | 0.2176 / 0.2199 |
|
| 93 |
+
| Sampled benchmark requests | 1,900 across 15 datasets |
|
| 94 |
+
| Errors / unsupported requests in that diagnostic | 0 / 0 |
|
| 95 |
+
| Official Decision Index 0.2 overall | **Not measured** |
|
| 96 |
+
|
| 97 |
+
Calibration slightly worsened NLL on the independent local test; it was not retuned
|
| 98 |
+
on that test. These local results are not the overall Decision Index.
|
| 99 |
+
|
| 100 |
+
Selected diagnostic results (different case sets from the public leaderboard):
|
| 101 |
+
|
| 102 |
+
| Benchmark | Sample size | Metric | Jet score |
|
| 103 |
+
|---|---:|---|---:|
|
| 104 |
+
| GSM8K | 112 | accuracy | 76.79% |
|
| 105 |
+
| NLI4CT | 55 | macro-F1 | 84.01% |
|
| 106 |
+
| ContractNLI | 56 | macro-F1 | 74.04% |
|
| 107 |
+
| CRUXEval | 55 | accuracy | 38.18% |
|
| 108 |
+
| WinoGrande | 200 | accuracy | 64.50% |
|
| 109 |
+
| HellaSwag | 200 | accuracy | 91.00% |
|
| 110 |
+
|
| 111 |
+
The expanded full-benchmark evaluation is in progress at release. The current
|
| 112 |
+
[Decision Index](https://huggingface.co/spaces/multimodalart/jev-decision-index)
|
| 113 |
+
requires 40 benchmarks across five areas with chance-corrected scoring. Access to
|
| 114 |
+
HLE and the complete matching release-v2 corpus/recipes is still required. No
|
| 115 |
+
subset average is presented as an overall score or rank.
|
| 116 |
+
|
| 117 |
+
Aggregate diagnostic metrics are in `evaluation.json`; `merge-provenance.json`
|
| 118 |
+
records the selected checkpoint and merge. `release-manifest.json` records file hashes.
|
| 119 |
+
`merge-validation.json` measures merged-versus-adapter behavior; the scores above
|
| 120 |
+
were measured on the unmerged adapter. In 36 fixed comparison cases (12 per
|
| 121 |
+
question type), 35 argmax answers matched; one ordinal answer changed. The largest
|
| 122 |
+
absolute calibrated-probability difference was 0.04373 (4.37 percentage points).
|
| 123 |
+
This is the expected distinction between separate LoRA inference and merging into
|
| 124 |
+
BF16 weights; the merged model is not claimed to be bit-identical.
|
| 125 |
+
Decisions can be wrong, and calibration on this mixture does not guarantee
|
| 126 |
+
calibration on another domain. Exact-state checks found no diagnostic/training
|
| 127 |
+
matches but are not comprehensive contamination certification.
|
| 128 |
+
|
| 129 |
+
## License
|
| 130 |
+
|
| 131 |
+
Apache-2.0, consistent with the base model and previous Jet release. Upstream data
|
| 132 |
+
sources retain their own terms. This repository does not redistribute training
|
| 133 |
+
or benchmark datasets.
|
TRAINING_HISTORY.md
DELETED
|
@@ -1,71 +0,0 @@
|
|
| 1 |
-
# Jet training history
|
| 2 |
-
|
| 3 |
-
All runs below use Qwen3-0.6B unless stated otherwise. Results depend on the
|
| 4 |
-
specified splits and inference backend; partial benchmark samples are not official
|
| 5 |
-
Decision Index scores. Historical reports use “released Jet” to mean the model
|
| 6 |
-
published at the time of that experiment, not necessarily today's release.
|
| 7 |
-
|
| 8 |
-
## Initial experiments and V2
|
| 9 |
-
|
| 10 |
-
Public-source training introduced choice, score and yes/no decisions, followed
|
| 11 |
-
by varied-scale ordinal questions. V2 used two epochs and 2,910 steps, selecting
|
| 12 |
-
step 2,750 by validation NLL. The fused release reached about 67.7% accuracy on
|
| 13 |
-
the 3,683-row general test. A Qwen3-1.7B comparison was evaluated but not retained.
|
| 14 |
-
|
| 15 |
-
The complete original tables, earlier 1,500-row experiment, ordinal evaluation,
|
| 16 |
-
and Kev transfer evaluation are preserved in [initial results](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/initial-results.md).
|
| 17 |
-
|
| 18 |
-
## V3 — broader decision training
|
| 19 |
-
|
| 20 |
-
Added relevance, entailment, stance, sarcasm, tool selection, Boolean rules,
|
| 21 |
-
arithmetic and code behavior. Trained on 47,313 rows for two epochs / 4,888
|
| 22 |
-
steps with rank-16 LoRA and learning rate `1e-4`.
|
| 23 |
-
New-family accuracy reached 80.6%; general-test accuracy regressed to 65.2%.
|
| 24 |
-
|
| 25 |
-
[Experiment report](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v3/summary.md)
|
| 26 |
-
|
| 27 |
-
## Warm start and V4 — transfer and routing
|
| 28 |
-
|
| 29 |
-
A lower-learning-rate warm start was followed by transfer and routing training.
|
| 30 |
-
V4 reached 66.4% general-test accuracy and 93.3% on domain-held-out tool routing.
|
| 31 |
-
It improved several partial language/relevance diagnostics but regressed on code.
|
| 32 |
-
A CPU-fused bf16/q8 release package was prepared separately.
|
| 33 |
-
|
| 34 |
-
[Training and comparisons](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-next/summary.md) ·
|
| 35 |
-
[Export preparation](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v4-release.md)
|
| 36 |
-
|
| 37 |
-
## V5 — wider intent and adversarial language tasks
|
| 38 |
-
|
| 39 |
-
Continued from V4 with 7,997 retained examples and 2,000 each from native training
|
| 40 |
-
partitions of CLINC150, ANLI, HellaSwag and When2Call preference data.
|
| 41 |
-
The completed R2 run used 15,997 rows, one epoch / 3,570 steps, rank 16,
|
| 42 |
-
learning rate `1e-5`, and selected step 3,500. Calibration was separate.
|
| 43 |
-
|
| 44 |
-
The first attempt was stopped after identifying retained utterances shared with
|
| 45 |
-
native evaluation content. R2 removes three such utterances from continuation
|
| 46 |
-
training, but does not erase inherited cross-dataset overlap in its initialization.
|
| 47 |
-
The checkpoint is not certified decontaminated.
|
| 48 |
-
|
| 49 |
-
| Held-out accuracy | V4 | V5 |
|
| 50 |
-
|---|---:|---:|
|
| 51 |
-
| General test | 66.44% | 66.66% |
|
| 52 |
-
| New families | 59.00% | 73.50% |
|
| 53 |
-
| Transfer tasks | 65.97% | 67.50% |
|
| 54 |
-
| Domain-held-out tool routing | 93.33% | 85.00% |
|
| 55 |
-
|
| 56 |
-
Selection-only evaluation chose original/reversed option-order averaging for the
|
| 57 |
-
benchmark adapter. This increases inference work and is not part of standard
|
| 58 |
-
serving. Gains and regressions on fixed public benchmark samples are recorded
|
| 59 |
-
separately; no overall leaderboard result or win has been established.
|
| 60 |
-
|
| 61 |
-
[Results](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5/summary.md) ·
|
| 62 |
-
[Protocol](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5/protocol.md) ·
|
| 63 |
-
[Completion audit](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5/completion-audit.json)
|
| 64 |
-
|
| 65 |
-
## Release — 2026-09-24
|
| 66 |
-
|
| 67 |
-
The current release is the V5 R2 selected checkpoint, fused with native MLX CUDA,
|
| 68 |
-
with a calibrated bf16 model and validated q8 ONNX export. Export and publication
|
| 69 |
-
receipts are recorded in [release notes](https://huggingface.co/michaljach/jet/blob/ee7d33ecae2b8bee30891657dd57c92d34e0ce53/docs/training/jet-v5-release.md).
|
| 70 |
-
|
| 71 |
-
Detailed experiment files are preserved in the [training archive](https://huggingface.co/michaljach/jet/tree/training-archive-20260924/docs/training).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
calibration.json
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"choice": 1.
|
| 3 |
-
"score": 1.
|
| 4 |
-
"noul": 1.
|
| 5 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"choice": 1.1224620483093728,
|
| 3 |
+
"score": 1.155352696872273,
|
| 4 |
+
"noul": 1.5422108254079405
|
| 5 |
+
}
|
chat_template.jinja
CHANGED
|
@@ -1,85 +1,154 @@
|
|
| 1 |
-
{%-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
{%- endif %}
|
| 6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
{%- for tool in tools %}
|
| 8 |
{{- "\n" }}
|
| 9 |
{{- tool | tojson }}
|
| 10 |
{%- endfor %}
|
| 11 |
-
{{- "\n</tools>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
{%- else %}
|
| 13 |
{%- if messages[0].role == 'system' %}
|
| 14 |
-
{
|
|
|
|
| 15 |
{%- endif %}
|
| 16 |
{%- endif %}
|
| 17 |
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
{%- for message in messages[::-1] %}
|
| 19 |
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
-
{%- if ns.multi_step_tool and message.role == "user"
|
| 21 |
-
{%- set
|
| 22 |
-
{%-
|
|
|
|
|
|
|
|
|
|
| 23 |
{%- endif %}
|
| 24 |
{%- endfor %}
|
|
|
|
|
|
|
|
|
|
| 25 |
{%- for message in messages %}
|
| 26 |
-
{%-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
{%- elif message.role == "assistant" %}
|
| 29 |
-
{%- set content = message.content %}
|
| 30 |
{%- set reasoning_content = '' %}
|
| 31 |
-
{%- if message.reasoning_content is
|
| 32 |
{%- set reasoning_content = message.reasoning_content %}
|
| 33 |
{%- else %}
|
| 34 |
-
{%- if '</think>' in
|
| 35 |
-
{%- set
|
| 36 |
-
{%- set
|
| 37 |
{%- endif %}
|
| 38 |
{%- endif %}
|
|
|
|
| 39 |
{%- if loop.index0 > ns.last_query_index %}
|
| 40 |
-
{
|
| 41 |
-
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 42 |
-
{%- else %}
|
| 43 |
-
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 44 |
-
{%- endif %}
|
| 45 |
{%- else %}
|
| 46 |
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 47 |
{%- endif %}
|
| 48 |
-
{%- if message.tool_calls %}
|
| 49 |
{%- for tool_call in message.tool_calls %}
|
| 50 |
-
{%- if
|
| 51 |
-
{{- '\n' }}
|
| 52 |
-
{%- endif %}
|
| 53 |
-
{%- if tool_call.function %}
|
| 54 |
{%- set tool_call = tool_call.function %}
|
| 55 |
{%- endif %}
|
| 56 |
-
{
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
|
|
|
| 61 |
{%- else %}
|
| 62 |
-
{{- tool_call.
|
| 63 |
{%- endif %}
|
| 64 |
-
{
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
{%- endfor %}
|
| 66 |
{%- endif %}
|
| 67 |
{{- '<|im_end|>\n' }}
|
| 68 |
{%- elif message.role == "tool" %}
|
| 69 |
-
{%- if loop.
|
| 70 |
{{- '<|im_start|>user' }}
|
| 71 |
{%- endif %}
|
| 72 |
{{- '\n<tool_response>\n' }}
|
| 73 |
-
{{-
|
| 74 |
{{- '\n</tool_response>' }}
|
| 75 |
-
{%- if loop.last
|
|
|
|
|
|
|
| 76 |
{{- '<|im_end|>\n' }}
|
| 77 |
{%- endif %}
|
|
|
|
|
|
|
| 78 |
{%- endif %}
|
| 79 |
{%- endfor %}
|
| 80 |
{%- if add_generation_prompt %}
|
| 81 |
{{- '<|im_start|>assistant\n' }}
|
| 82 |
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 83 |
{{- '<think>\n\n</think>\n\n' }}
|
|
|
|
|
|
|
| 84 |
{%- endif %}
|
| 85 |
{%- endif %}
|
|
|
|
| 1 |
+
{%- set image_count = namespace(value=0) %}
|
| 2 |
+
{%- set video_count = namespace(value=0) %}
|
| 3 |
+
{%- macro render_content(content, do_vision_count, is_system_content=false) %}
|
| 4 |
+
{%- if content is string %}
|
| 5 |
+
{{- content }}
|
| 6 |
+
{%- elif content is iterable and content is not mapping %}
|
| 7 |
+
{%- for item in content %}
|
| 8 |
+
{%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
|
| 9 |
+
{%- if is_system_content %}
|
| 10 |
+
{{- raise_exception('System message cannot contain images.') }}
|
| 11 |
+
{%- endif %}
|
| 12 |
+
{%- if do_vision_count %}
|
| 13 |
+
{%- set image_count.value = image_count.value + 1 %}
|
| 14 |
+
{%- endif %}
|
| 15 |
+
{%- if add_vision_id %}
|
| 16 |
+
{{- 'Picture ' ~ image_count.value ~ ': ' }}
|
| 17 |
+
{%- endif %}
|
| 18 |
+
{{- '<|vision_start|><|image_pad|><|vision_end|>' }}
|
| 19 |
+
{%- elif 'video' in item or item.type == 'video' %}
|
| 20 |
+
{%- if is_system_content %}
|
| 21 |
+
{{- raise_exception('System message cannot contain videos.') }}
|
| 22 |
+
{%- endif %}
|
| 23 |
+
{%- if do_vision_count %}
|
| 24 |
+
{%- set video_count.value = video_count.value + 1 %}
|
| 25 |
+
{%- endif %}
|
| 26 |
+
{%- if add_vision_id %}
|
| 27 |
+
{{- 'Video ' ~ video_count.value ~ ': ' }}
|
| 28 |
+
{%- endif %}
|
| 29 |
+
{{- '<|vision_start|><|video_pad|><|vision_end|>' }}
|
| 30 |
+
{%- elif 'text' in item %}
|
| 31 |
+
{{- item.text }}
|
| 32 |
+
{%- else %}
|
| 33 |
+
{{- raise_exception('Unexpected item type in content.') }}
|
| 34 |
+
{%- endif %}
|
| 35 |
+
{%- endfor %}
|
| 36 |
+
{%- elif content is none or content is undefined %}
|
| 37 |
+
{{- '' }}
|
| 38 |
+
{%- else %}
|
| 39 |
+
{{- raise_exception('Unexpected content type.') }}
|
| 40 |
{%- endif %}
|
| 41 |
+
{%- endmacro %}
|
| 42 |
+
{%- if not messages %}
|
| 43 |
+
{{- raise_exception('No messages provided.') }}
|
| 44 |
+
{%- endif %}
|
| 45 |
+
{%- if tools and tools is iterable and tools is not mapping %}
|
| 46 |
+
{{- '<|im_start|>system\n' }}
|
| 47 |
+
{{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
|
| 48 |
{%- for tool in tools %}
|
| 49 |
{{- "\n" }}
|
| 50 |
{{- tool | tojson }}
|
| 51 |
{%- endfor %}
|
| 52 |
+
{{- "\n</tools>" }}
|
| 53 |
+
{{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
|
| 54 |
+
{%- if messages[0].role == 'system' %}
|
| 55 |
+
{%- set content = render_content(messages[0].content, false, true)|trim %}
|
| 56 |
+
{%- if content %}
|
| 57 |
+
{{- '\n\n' + content }}
|
| 58 |
+
{%- endif %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{{- '<|im_end|>\n' }}
|
| 61 |
{%- else %}
|
| 62 |
{%- if messages[0].role == 'system' %}
|
| 63 |
+
{%- set content = render_content(messages[0].content, false, true)|trim %}
|
| 64 |
+
{{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
|
| 65 |
{%- endif %}
|
| 66 |
{%- endif %}
|
| 67 |
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 68 |
{%- for message in messages[::-1] %}
|
| 69 |
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 70 |
+
{%- if ns.multi_step_tool and message.role == "user" %}
|
| 71 |
+
{%- set content = render_content(message.content, false)|trim %}
|
| 72 |
+
{%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
|
| 73 |
+
{%- set ns.multi_step_tool = false %}
|
| 74 |
+
{%- set ns.last_query_index = index %}
|
| 75 |
+
{%- endif %}
|
| 76 |
{%- endif %}
|
| 77 |
{%- endfor %}
|
| 78 |
+
{%- if ns.multi_step_tool %}
|
| 79 |
+
{{- raise_exception('No user query found in messages.') }}
|
| 80 |
+
{%- endif %}
|
| 81 |
{%- for message in messages %}
|
| 82 |
+
{%- set content = render_content(message.content, true)|trim %}
|
| 83 |
+
{%- if message.role == "system" %}
|
| 84 |
+
{%- if not loop.first %}
|
| 85 |
+
{{- raise_exception('System message must be at the beginning.') }}
|
| 86 |
+
{%- endif %}
|
| 87 |
+
{%- elif message.role == "user" %}
|
| 88 |
+
{{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
|
| 89 |
{%- elif message.role == "assistant" %}
|
|
|
|
| 90 |
{%- set reasoning_content = '' %}
|
| 91 |
+
{%- if message.reasoning_content is string %}
|
| 92 |
{%- set reasoning_content = message.reasoning_content %}
|
| 93 |
{%- else %}
|
| 94 |
+
{%- if '</think>' in content %}
|
| 95 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 96 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 97 |
{%- endif %}
|
| 98 |
{%- endif %}
|
| 99 |
+
{%- set reasoning_content = reasoning_content|trim %}
|
| 100 |
{%- if loop.index0 > ns.last_query_index %}
|
| 101 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
{%- else %}
|
| 103 |
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 104 |
{%- endif %}
|
| 105 |
+
{%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
|
| 106 |
{%- for tool_call in message.tool_calls %}
|
| 107 |
+
{%- if tool_call.function is defined %}
|
|
|
|
|
|
|
|
|
|
| 108 |
{%- set tool_call = tool_call.function %}
|
| 109 |
{%- endif %}
|
| 110 |
+
{%- if loop.first %}
|
| 111 |
+
{%- if content|trim %}
|
| 112 |
+
{{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 113 |
+
{%- else %}
|
| 114 |
+
{{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 115 |
+
{%- endif %}
|
| 116 |
{%- else %}
|
| 117 |
+
{{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 118 |
{%- endif %}
|
| 119 |
+
{%- if tool_call.arguments is defined %}
|
| 120 |
+
{%- for args_name, args_value in tool_call.arguments|items %}
|
| 121 |
+
{{- '<parameter=' + args_name + '>\n' }}
|
| 122 |
+
{%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
|
| 123 |
+
{{- args_value }}
|
| 124 |
+
{{- '\n</parameter>\n' }}
|
| 125 |
+
{%- endfor %}
|
| 126 |
+
{%- endif %}
|
| 127 |
+
{{- '</function>\n</tool_call>' }}
|
| 128 |
{%- endfor %}
|
| 129 |
{%- endif %}
|
| 130 |
{{- '<|im_end|>\n' }}
|
| 131 |
{%- elif message.role == "tool" %}
|
| 132 |
+
{%- if loop.previtem and loop.previtem.role != "tool" %}
|
| 133 |
{{- '<|im_start|>user' }}
|
| 134 |
{%- endif %}
|
| 135 |
{{- '\n<tool_response>\n' }}
|
| 136 |
+
{{- content }}
|
| 137 |
{{- '\n</tool_response>' }}
|
| 138 |
+
{%- if not loop.last and loop.nextitem.role != "tool" %}
|
| 139 |
+
{{- '<|im_end|>\n' }}
|
| 140 |
+
{%- elif loop.last %}
|
| 141 |
{{- '<|im_end|>\n' }}
|
| 142 |
{%- endif %}
|
| 143 |
+
{%- else %}
|
| 144 |
+
{{- raise_exception('Unexpected message role.') }}
|
| 145 |
{%- endif %}
|
| 146 |
{%- endfor %}
|
| 147 |
{%- if add_generation_prompt %}
|
| 148 |
{{- '<|im_start|>assistant\n' }}
|
| 149 |
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 150 |
{{- '<think>\n\n</think>\n\n' }}
|
| 151 |
+
{%- else %}
|
| 152 |
+
{{- '<think>\n' }}
|
| 153 |
{%- endif %}
|
| 154 |
{%- endif %}
|
config.json
CHANGED
|
@@ -1,30 +1,83 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
],
|
| 5 |
-
"
|
| 6 |
-
"
|
| 7 |
-
"
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
"max_position_embeddings": 40960,
|
| 15 |
-
"max_window_layers": 28,
|
| 16 |
-
"model_type": "qwen3",
|
| 17 |
-
"num_attention_heads": 16,
|
| 18 |
-
"num_hidden_layers": 28,
|
| 19 |
-
"num_key_value_heads": 8,
|
| 20 |
-
"rms_norm_eps": 1e-06,
|
| 21 |
-
"rope_scaling": null,
|
| 22 |
-
"rope_theta": 1000000,
|
| 23 |
-
"sliding_window": null,
|
| 24 |
-
"tie_word_embeddings": true,
|
| 25 |
-
"torch_dtype": "bfloat16",
|
| 26 |
-
"transformers_version": "4.51.0",
|
| 27 |
-
"use_cache": true,
|
| 28 |
-
"use_sliding_window": false,
|
| 29 |
-
"vocab_size": 151936
|
| 30 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3_5ForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"attn_output_gate": true,
|
| 8 |
+
"bos_token_id": null,
|
| 9 |
+
"dtype": "bfloat16",
|
| 10 |
+
"eos_token_id": 248044,
|
| 11 |
+
"full_attention_interval": 4,
|
| 12 |
+
"head_dim": 256,
|
| 13 |
+
"hidden_act": "silu",
|
| 14 |
+
"hidden_size": 2560,
|
| 15 |
+
"initializer_range": 0.02,
|
| 16 |
+
"intermediate_size": 9216,
|
| 17 |
+
"layer_types": [
|
| 18 |
+
"linear_attention",
|
| 19 |
+
"linear_attention",
|
| 20 |
+
"linear_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"linear_attention",
|
| 23 |
+
"linear_attention",
|
| 24 |
+
"linear_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"linear_attention",
|
| 27 |
+
"linear_attention",
|
| 28 |
+
"linear_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"linear_attention",
|
| 31 |
+
"linear_attention",
|
| 32 |
+
"linear_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"linear_attention",
|
| 35 |
+
"linear_attention",
|
| 36 |
+
"linear_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"linear_attention",
|
| 39 |
+
"linear_attention",
|
| 40 |
+
"linear_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"linear_attention",
|
| 43 |
+
"linear_attention",
|
| 44 |
+
"linear_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"linear_attention",
|
| 47 |
+
"linear_attention",
|
| 48 |
+
"linear_attention",
|
| 49 |
+
"full_attention"
|
| 50 |
+
],
|
| 51 |
+
"linear_conv_kernel_dim": 4,
|
| 52 |
+
"linear_key_head_dim": 128,
|
| 53 |
+
"linear_num_key_heads": 16,
|
| 54 |
+
"linear_num_value_heads": 32,
|
| 55 |
+
"linear_value_head_dim": 128,
|
| 56 |
+
"mamba_ssm_dtype": "float32",
|
| 57 |
+
"max_position_embeddings": 262144,
|
| 58 |
+
"mlp_only_layers": [],
|
| 59 |
+
"model_type": "qwen3_5_text",
|
| 60 |
+
"mtp_num_hidden_layers": 1,
|
| 61 |
+
"mtp_use_dedicated_embeddings": false,
|
| 62 |
+
"num_attention_heads": 16,
|
| 63 |
+
"num_hidden_layers": 32,
|
| 64 |
+
"num_key_value_heads": 4,
|
| 65 |
+
"pad_token_id": null,
|
| 66 |
+
"partial_rotary_factor": 0.25,
|
| 67 |
+
"rms_norm_eps": 1e-06,
|
| 68 |
+
"rope_parameters": {
|
| 69 |
+
"mrope_interleaved": true,
|
| 70 |
+
"mrope_section": [
|
| 71 |
+
11,
|
| 72 |
+
11,
|
| 73 |
+
10
|
| 74 |
],
|
| 75 |
+
"partial_rotary_factor": 0.25,
|
| 76 |
+
"rope_theta": 10000000,
|
| 77 |
+
"rope_type": "default"
|
| 78 |
+
},
|
| 79 |
+
"tie_word_embeddings": true,
|
| 80 |
+
"transformers_version": "5.17.0",
|
| 81 |
+
"use_cache": true,
|
| 82 |
+
"vocab_size": 248320
|
| 83 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
evaluation.json
ADDED
|
@@ -0,0 +1,371 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"engine": "trained",
|
| 3 |
+
"counts": {
|
| 4 |
+
"ok": 1900
|
| 5 |
+
},
|
| 6 |
+
"successful_request_latency_ms": {
|
| 7 |
+
"median": 67.33474849897902,
|
| 8 |
+
"p95": 309.1855669990764,
|
| 9 |
+
"mean": 348.8505987489407
|
| 10 |
+
},
|
| 11 |
+
"benchmarks": [
|
| 12 |
+
{
|
| 13 |
+
"catalog_id": 11,
|
| 14 |
+
"dataset": "ContractNLI",
|
| 15 |
+
"requests": 56,
|
| 16 |
+
"answered": 56,
|
| 17 |
+
"unsupported": 0,
|
| 18 |
+
"errors": 0,
|
| 19 |
+
"abstained": 0,
|
| 20 |
+
"pending": 0,
|
| 21 |
+
"scored_requests": 56,
|
| 22 |
+
"metric": "macro-F1",
|
| 23 |
+
"score": 0.7404310060948458,
|
| 24 |
+
"reference_same_cases": null,
|
| 25 |
+
"median_ms": 6274.808131498503,
|
| 26 |
+
"detail": {
|
| 27 |
+
"field_accuracy": 0.8392857142857143,
|
| 28 |
+
"case_exact_accuracy": 0.07142857142857142,
|
| 29 |
+
"macro_f1": 0.7404310060948458,
|
| 30 |
+
"custom_metrics": {}
|
| 31 |
+
}
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"catalog_id": 30,
|
| 35 |
+
"dataset": "GSM8K",
|
| 36 |
+
"requests": 112,
|
| 37 |
+
"answered": 112,
|
| 38 |
+
"unsupported": 0,
|
| 39 |
+
"errors": 0,
|
| 40 |
+
"abstained": 0,
|
| 41 |
+
"pending": 0,
|
| 42 |
+
"scored_requests": 112,
|
| 43 |
+
"metric": "accuracy",
|
| 44 |
+
"score": 0.7678571428571429,
|
| 45 |
+
"reference_same_cases": null,
|
| 46 |
+
"median_ms": 67.91422949754633,
|
| 47 |
+
"tracks": {
|
| 48 |
+
"GSM8K-10choice": {
|
| 49 |
+
"metric": "accuracy",
|
| 50 |
+
"score": 0.7321428571428571,
|
| 51 |
+
"scored_requests": 56
|
| 52 |
+
},
|
| 53 |
+
"GSM8K-4choice": {
|
| 54 |
+
"metric": "accuracy",
|
| 55 |
+
"score": 0.8035714285714286,
|
| 56 |
+
"scored_requests": 56
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
"detail": {
|
| 60 |
+
"field_accuracy": 0.7678571428571429,
|
| 61 |
+
"case_exact_accuracy": 0.625,
|
| 62 |
+
"custom_metrics": {}
|
| 63 |
+
}
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"catalog_id": 40,
|
| 67 |
+
"dataset": "iSarcasmEval",
|
| 68 |
+
"requests": 56,
|
| 69 |
+
"answered": 56,
|
| 70 |
+
"unsupported": 0,
|
| 71 |
+
"errors": 0,
|
| 72 |
+
"abstained": 0,
|
| 73 |
+
"pending": 0,
|
| 74 |
+
"scored_requests": 56,
|
| 75 |
+
"metric": "see separate subtask tracks",
|
| 76 |
+
"score": null,
|
| 77 |
+
"reference_same_cases": null,
|
| 78 |
+
"median_ms": 70.44221549949725,
|
| 79 |
+
"tracks": {
|
| 80 |
+
"iSarcasmEval-A-Ar": {
|
| 81 |
+
"metric": "see separate subtask tracks",
|
| 82 |
+
"score": null,
|
| 83 |
+
"scored_requests": 12
|
| 84 |
+
},
|
| 85 |
+
"iSarcasmEval-A-En": {
|
| 86 |
+
"metric": "see separate subtask tracks",
|
| 87 |
+
"score": null,
|
| 88 |
+
"scored_requests": 19
|
| 89 |
+
},
|
| 90 |
+
"iSarcasmEval-B-En": {
|
| 91 |
+
"metric": "see separate subtask tracks",
|
| 92 |
+
"score": null,
|
| 93 |
+
"scored_requests": 23
|
| 94 |
+
},
|
| 95 |
+
"iSarcasmEval-C-Ar": {
|
| 96 |
+
"metric": "see separate subtask tracks",
|
| 97 |
+
"score": null,
|
| 98 |
+
"scored_requests": 1
|
| 99 |
+
},
|
| 100 |
+
"iSarcasmEval-C-En": {
|
| 101 |
+
"metric": "see separate subtask tracks",
|
| 102 |
+
"score": null,
|
| 103 |
+
"scored_requests": 1
|
| 104 |
+
}
|
| 105 |
+
},
|
| 106 |
+
"detail": {
|
| 107 |
+
"field_accuracy": 0.7777777777777778,
|
| 108 |
+
"case_exact_accuracy": 0.5535714285714286,
|
| 109 |
+
"macro_f1": 0.7693247126436782,
|
| 110 |
+
"custom_metrics": {},
|
| 111 |
+
"positive_f1_by_field": {
|
| 112 |
+
"sarcastic": 0.2857142857142857,
|
| 113 |
+
"sarcasm": 0.2222222222222222,
|
| 114 |
+
"irony": 0.25,
|
| 115 |
+
"satire": 0.2857142857142857,
|
| 116 |
+
"understatement": 0.0,
|
| 117 |
+
"overstatement": 0.0,
|
| 118 |
+
"rhetorical_question": 0.0
|
| 119 |
+
},
|
| 120 |
+
"category_macro_f1": 0.14909297052154194
|
| 121 |
+
}
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"catalog_id": 41,
|
| 125 |
+
"dataset": "VAST",
|
| 126 |
+
"requests": 56,
|
| 127 |
+
"answered": 56,
|
| 128 |
+
"unsupported": 0,
|
| 129 |
+
"errors": 0,
|
| 130 |
+
"abstained": 0,
|
| 131 |
+
"pending": 0,
|
| 132 |
+
"scored_requests": 56,
|
| 133 |
+
"metric": "macro-F1",
|
| 134 |
+
"score": 0.43621013133208253,
|
| 135 |
+
"reference_same_cases": null,
|
| 136 |
+
"median_ms": 67.72124700000859,
|
| 137 |
+
"detail": {
|
| 138 |
+
"field_accuracy": 0.5357142857142857,
|
| 139 |
+
"case_exact_accuracy": 0.5357142857142857,
|
| 140 |
+
"macro_f1": 0.43621013133208253,
|
| 141 |
+
"custom_metrics": {}
|
| 142 |
+
}
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"catalog_id": 42,
|
| 146 |
+
"dataset": "NLI4CT",
|
| 147 |
+
"requests": 55,
|
| 148 |
+
"answered": 55,
|
| 149 |
+
"unsupported": 0,
|
| 150 |
+
"errors": 0,
|
| 151 |
+
"abstained": 0,
|
| 152 |
+
"pending": 0,
|
| 153 |
+
"scored_requests": 55,
|
| 154 |
+
"metric": "macro-F1",
|
| 155 |
+
"score": 0.8401162790697674,
|
| 156 |
+
"reference_same_cases": null,
|
| 157 |
+
"median_ms": 98.04200899816351,
|
| 158 |
+
"detail": {
|
| 159 |
+
"field_accuracy": 0.8909090909090909,
|
| 160 |
+
"case_exact_accuracy": 0.8909090909090909,
|
| 161 |
+
"macro_f1": 0.8401162790697674,
|
| 162 |
+
"custom_metrics": {}
|
| 163 |
+
}
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"catalog_id": 43,
|
| 167 |
+
"dataset": "CRUXEval",
|
| 168 |
+
"requests": 55,
|
| 169 |
+
"answered": 55,
|
| 170 |
+
"unsupported": 0,
|
| 171 |
+
"errors": 0,
|
| 172 |
+
"abstained": 0,
|
| 173 |
+
"pending": 0,
|
| 174 |
+
"scored_requests": 55,
|
| 175 |
+
"metric": "accuracy",
|
| 176 |
+
"score": 0.38181818181818183,
|
| 177 |
+
"reference_same_cases": null,
|
| 178 |
+
"median_ms": 67.40817399986554,
|
| 179 |
+
"detail": {
|
| 180 |
+
"field_accuracy": 0.38181818181818183,
|
| 181 |
+
"case_exact_accuracy": 0.38181818181818183,
|
| 182 |
+
"custom_metrics": {}
|
| 183 |
+
}
|
| 184 |
+
},
|
| 185 |
+
{
|
| 186 |
+
"catalog_id": 44,
|
| 187 |
+
"dataset": "CLadder",
|
| 188 |
+
"requests": 55,
|
| 189 |
+
"answered": 55,
|
| 190 |
+
"unsupported": 0,
|
| 191 |
+
"errors": 0,
|
| 192 |
+
"abstained": 0,
|
| 193 |
+
"pending": 0,
|
| 194 |
+
"scored_requests": 55,
|
| 195 |
+
"metric": "accuracy",
|
| 196 |
+
"score": 0.6545454545454545,
|
| 197 |
+
"reference_same_cases": null,
|
| 198 |
+
"median_ms": 67.2220330015989,
|
| 199 |
+
"detail": {
|
| 200 |
+
"field_accuracy": 0.6545454545454545,
|
| 201 |
+
"case_exact_accuracy": 0.6545454545454545,
|
| 202 |
+
"custom_metrics": {}
|
| 203 |
+
}
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"catalog_id": 50,
|
| 207 |
+
"dataset": "Habermas Machine",
|
| 208 |
+
"requests": 55,
|
| 209 |
+
"answered": 55,
|
| 210 |
+
"unsupported": 0,
|
| 211 |
+
"errors": 0,
|
| 212 |
+
"abstained": 0,
|
| 213 |
+
"pending": 0,
|
| 214 |
+
"scored_requests": 55,
|
| 215 |
+
"metric": "accuracy",
|
| 216 |
+
"score": 0.34545454545454546,
|
| 217 |
+
"reference_same_cases": null,
|
| 218 |
+
"median_ms": 188.9903100018273,
|
| 219 |
+
"detail": {
|
| 220 |
+
"field_accuracy": 0.34545454545454546,
|
| 221 |
+
"case_exact_accuracy": 0.34545454545454546,
|
| 222 |
+
"custom_metrics": {
|
| 223 |
+
"normalized_rank_regret": 0.2393939393939394
|
| 224 |
+
}
|
| 225 |
+
}
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"catalog_id": 5,
|
| 229 |
+
"dataset": "CLINC150+OOS",
|
| 230 |
+
"requests": 200,
|
| 231 |
+
"answered": 200,
|
| 232 |
+
"unsupported": 0,
|
| 233 |
+
"errors": 0,
|
| 234 |
+
"abstained": 0,
|
| 235 |
+
"pending": 0,
|
| 236 |
+
"scored_requests": 200,
|
| 237 |
+
"metric": "macro-F1",
|
| 238 |
+
"score": 0.852693430018526,
|
| 239 |
+
"reference_same_cases": null,
|
| 240 |
+
"median_ms": 295.12597599932633,
|
| 241 |
+
"detail": {
|
| 242 |
+
"field_accuracy": 0.9,
|
| 243 |
+
"case_exact_accuracy": 0.9,
|
| 244 |
+
"macro_f1": 0.852693430018526,
|
| 245 |
+
"custom_metrics": {}
|
| 246 |
+
}
|
| 247 |
+
},
|
| 248 |
+
{
|
| 249 |
+
"catalog_id": 12,
|
| 250 |
+
"dataset": "ANLI",
|
| 251 |
+
"requests": 200,
|
| 252 |
+
"answered": 200,
|
| 253 |
+
"unsupported": 0,
|
| 254 |
+
"errors": 0,
|
| 255 |
+
"abstained": 0,
|
| 256 |
+
"pending": 0,
|
| 257 |
+
"scored_requests": 200,
|
| 258 |
+
"metric": "macro-F1",
|
| 259 |
+
"score": 0.5958562735439004,
|
| 260 |
+
"reference_same_cases": null,
|
| 261 |
+
"median_ms": 67.09086350201687,
|
| 262 |
+
"detail": {
|
| 263 |
+
"field_accuracy": 0.595,
|
| 264 |
+
"case_exact_accuracy": 0.595,
|
| 265 |
+
"macro_f1": 0.5958562735439004,
|
| 266 |
+
"custom_metrics": {}
|
| 267 |
+
}
|
| 268 |
+
},
|
| 269 |
+
{
|
| 270 |
+
"catalog_id": 24,
|
| 271 |
+
"dataset": "MMLU",
|
| 272 |
+
"requests": 200,
|
| 273 |
+
"answered": 200,
|
| 274 |
+
"unsupported": 0,
|
| 275 |
+
"errors": 0,
|
| 276 |
+
"abstained": 0,
|
| 277 |
+
"pending": 0,
|
| 278 |
+
"scored_requests": 200,
|
| 279 |
+
"metric": "accuracy",
|
| 280 |
+
"score": 0.74,
|
| 281 |
+
"reference_same_cases": null,
|
| 282 |
+
"median_ms": 66.66580449928006,
|
| 283 |
+
"detail": {
|
| 284 |
+
"field_accuracy": 0.74,
|
| 285 |
+
"case_exact_accuracy": 0.74,
|
| 286 |
+
"custom_metrics": {}
|
| 287 |
+
}
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"catalog_id": 26,
|
| 291 |
+
"dataset": "ARC-Easy",
|
| 292 |
+
"requests": 200,
|
| 293 |
+
"answered": 200,
|
| 294 |
+
"unsupported": 0,
|
| 295 |
+
"errors": 0,
|
| 296 |
+
"abstained": 0,
|
| 297 |
+
"pending": 0,
|
| 298 |
+
"scored_requests": 200,
|
| 299 |
+
"metric": "accuracy",
|
| 300 |
+
"score": 0.98,
|
| 301 |
+
"reference_same_cases": null,
|
| 302 |
+
"median_ms": 66.54907899974205,
|
| 303 |
+
"detail": {
|
| 304 |
+
"field_accuracy": 0.98,
|
| 305 |
+
"case_exact_accuracy": 0.98,
|
| 306 |
+
"custom_metrics": {}
|
| 307 |
+
}
|
| 308 |
+
},
|
| 309 |
+
{
|
| 310 |
+
"catalog_id": 27,
|
| 311 |
+
"dataset": "ARC-Challenge",
|
| 312 |
+
"requests": 200,
|
| 313 |
+
"answered": 200,
|
| 314 |
+
"unsupported": 0,
|
| 315 |
+
"errors": 0,
|
| 316 |
+
"abstained": 0,
|
| 317 |
+
"pending": 0,
|
| 318 |
+
"scored_requests": 200,
|
| 319 |
+
"metric": "accuracy",
|
| 320 |
+
"score": 0.94,
|
| 321 |
+
"reference_same_cases": null,
|
| 322 |
+
"median_ms": 66.42713149994961,
|
| 323 |
+
"detail": {
|
| 324 |
+
"field_accuracy": 0.94,
|
| 325 |
+
"case_exact_accuracy": 0.94,
|
| 326 |
+
"custom_metrics": {}
|
| 327 |
+
}
|
| 328 |
+
},
|
| 329 |
+
{
|
| 330 |
+
"catalog_id": 28,
|
| 331 |
+
"dataset": "WinoGrande",
|
| 332 |
+
"requests": 200,
|
| 333 |
+
"answered": 200,
|
| 334 |
+
"unsupported": 0,
|
| 335 |
+
"errors": 0,
|
| 336 |
+
"abstained": 0,
|
| 337 |
+
"pending": 0,
|
| 338 |
+
"scored_requests": 200,
|
| 339 |
+
"metric": "accuracy",
|
| 340 |
+
"score": 0.645,
|
| 341 |
+
"reference_same_cases": null,
|
| 342 |
+
"median_ms": 66.51633900037268,
|
| 343 |
+
"detail": {
|
| 344 |
+
"field_accuracy": 0.645,
|
| 345 |
+
"case_exact_accuracy": 0.645,
|
| 346 |
+
"custom_metrics": {}
|
| 347 |
+
}
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"catalog_id": 29,
|
| 351 |
+
"dataset": "HellaSwag",
|
| 352 |
+
"requests": 200,
|
| 353 |
+
"answered": 200,
|
| 354 |
+
"unsupported": 0,
|
| 355 |
+
"errors": 0,
|
| 356 |
+
"abstained": 0,
|
| 357 |
+
"pending": 0,
|
| 358 |
+
"scored_requests": 200,
|
| 359 |
+
"metric": "accuracy",
|
| 360 |
+
"score": 0.91,
|
| 361 |
+
"reference_same_cases": null,
|
| 362 |
+
"median_ms": 67.15036900095583,
|
| 363 |
+
"detail": {
|
| 364 |
+
"field_accuracy": 0.91,
|
| 365 |
+
"case_exact_accuracy": 0.91,
|
| 366 |
+
"custom_metrics": {}
|
| 367 |
+
}
|
| 368 |
+
}
|
| 369 |
+
],
|
| 370 |
+
"note": "Native benchmark metrics on complete supported case groups. Unsupported cases excluded from accuracy but retained in coverage."
|
| 371 |
+
}
|
export-provenance.json
DELETED
|
@@ -1,12 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"adapter": "jet-v5-panel02-r2-20260924",
|
| 3 |
-
"adapter_sha256": "6416adc20cc1517dc5d9cf3c0a2655da8a899296c5640dfc6d5551170824d0cb",
|
| 4 |
-
"fusion": "Native MLX CUDA module.fuse(), 196 linear layers",
|
| 5 |
-
"fusion_validation": {
|
| 6 |
-
"cases": 36,
|
| 7 |
-
"answer_flips": [],
|
| 8 |
-
"max_probability_difference": 0.02581135243415833,
|
| 9 |
-
"reference": "unfused MLX CUDA vs native MLX CUDA fused bf16"
|
| 10 |
-
},
|
| 11 |
-
"onnx_validation": "36 golden cases: worst |dlogit| 0.254, worst |dprob| 0.0249, answers flipped 0 (178 tokens/s on 4 threads)"
|
| 12 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
export-validation.txt
DELETED
|
@@ -1 +0,0 @@
|
|
| 1 |
-
36 golden cases: worst |dlogit| 0.254, worst |dprob| 0.0249, answers flipped 0 (178 tokens/s on 4 threads)
|
|
|
|
|
|
format.py
ADDED
|
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Prompt format shared by training and inference.
|
| 2 |
+
|
| 3 |
+
Every question is reduced to "pick one label token". The model is run once over
|
| 4 |
+
the prompt and we read the next-token logits restricted to the label tokens:
|
| 5 |
+
|
| 6 |
+
- choice: one label per option (A..Z, then single-token AA, AB, ...), up to 255 options
|
| 7 |
+
- score: one digit per level (0..9), 2-10 levels, ordered low -> high
|
| 8 |
+
- noul: "no" / "yes"
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from __future__ import annotations
|
| 12 |
+
|
| 13 |
+
import itertools
|
| 14 |
+
import json
|
| 15 |
+
import string
|
| 16 |
+
from dataclasses import dataclass
|
| 17 |
+
from typing import Any, Literal
|
| 18 |
+
|
| 19 |
+
QuestionType = Literal["choice", "score", "noul"]
|
| 20 |
+
|
| 21 |
+
MAX_CHOICE_OPTIONS = 255
|
| 22 |
+
MIN_SCORE_LEVELS, MAX_SCORE_LEVELS = 2, 10
|
| 23 |
+
|
| 24 |
+
SYSTEM_PROMPT = (
|
| 25 |
+
"You are Jet, a decision model. Read the state and the question, then answer "
|
| 26 |
+
"with exactly one label from the allowed labels."
|
| 27 |
+
)
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
@dataclass(frozen=True)
|
| 31 |
+
class Question:
|
| 32 |
+
type: QuestionType
|
| 33 |
+
instructions: str
|
| 34 |
+
# choice: {key: description}; score: [level descriptions, low -> high]; noul: None
|
| 35 |
+
criteria: dict[str, str] | list[str] | None = None
|
| 36 |
+
|
| 37 |
+
@property
|
| 38 |
+
def keys(self) -> list[str]:
|
| 39 |
+
"""Answer keys in label order (option keys, level descriptions, or no/yes)."""
|
| 40 |
+
if self.type == "choice":
|
| 41 |
+
return list(self.criteria)
|
| 42 |
+
if self.type == "score":
|
| 43 |
+
return list(self.criteria)
|
| 44 |
+
return ["no", "yes"]
|
| 45 |
+
|
| 46 |
+
@classmethod
|
| 47 |
+
def from_dict(cls, d: dict[str, Any]) -> Question:
|
| 48 |
+
q = cls(type=d["type"], instructions=d["instructions"], criteria=d.get("criteria"))
|
| 49 |
+
q.validate()
|
| 50 |
+
return q
|
| 51 |
+
|
| 52 |
+
def to_dict(self) -> dict[str, Any]:
|
| 53 |
+
d: dict[str, Any] = {"type": self.type, "instructions": self.instructions}
|
| 54 |
+
if self.criteria is not None:
|
| 55 |
+
d["criteria"] = self.criteria
|
| 56 |
+
return d
|
| 57 |
+
|
| 58 |
+
def validate(self) -> None:
|
| 59 |
+
if self.type not in ("choice", "score", "noul"):
|
| 60 |
+
raise ValueError(f"unknown question type {self.type!r}")
|
| 61 |
+
if not isinstance(self.instructions, str) or not self.instructions.strip():
|
| 62 |
+
raise ValueError("instructions must be a non-empty string")
|
| 63 |
+
if self.type == "choice":
|
| 64 |
+
if not isinstance(self.criteria, dict) or not 2 <= len(self.criteria) <= MAX_CHOICE_OPTIONS:
|
| 65 |
+
raise ValueError(f"choice criteria must be an object with 2-{MAX_CHOICE_OPTIONS} options")
|
| 66 |
+
elif self.type == "score":
|
| 67 |
+
if not isinstance(self.criteria, list) or not MIN_SCORE_LEVELS <= len(self.criteria) <= MAX_SCORE_LEVELS:
|
| 68 |
+
raise ValueError(f"score criteria must be a list of {MIN_SCORE_LEVELS}-{MAX_SCORE_LEVELS} levels")
|
| 69 |
+
elif self.criteria is not None:
|
| 70 |
+
raise ValueError("noul questions take no criteria")
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
_CHOICE_LABELS: dict[int, list[str]] = {}
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def choice_labels(tokenizer, n: int) -> list[str]:
|
| 77 |
+
"""A..Z, then the two-letter labels AA, AB, ... that are single tokens for this tokenizer."""
|
| 78 |
+
key = id(tokenizer)
|
| 79 |
+
if key not in _CHOICE_LABELS:
|
| 80 |
+
letters = string.ascii_uppercase
|
| 81 |
+
pairs = ("".join(p) for p in itertools.product(letters, repeat=2))
|
| 82 |
+
candidates = list(letters) + [p for p in pairs if len(tokenizer.encode(p, add_special_tokens=False)) == 1]
|
| 83 |
+
if len(candidates) < MAX_CHOICE_OPTIONS:
|
| 84 |
+
raise ValueError(f"tokenizer only has {len(candidates)} single-token choice labels")
|
| 85 |
+
_CHOICE_LABELS[key] = candidates[:MAX_CHOICE_OPTIONS]
|
| 86 |
+
return _CHOICE_LABELS[key][:n]
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def labels_for(tokenizer, q: Question) -> list[str]:
|
| 90 |
+
if q.type == "choice":
|
| 91 |
+
return choice_labels(tokenizer, len(q.criteria))
|
| 92 |
+
if q.type == "score":
|
| 93 |
+
return [str(i) for i in range(len(q.criteria))]
|
| 94 |
+
return ["no", "yes"]
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def render_state(state: Any) -> str:
|
| 98 |
+
if isinstance(state, str):
|
| 99 |
+
return state
|
| 100 |
+
if isinstance(state, list) and all(isinstance(s, str) for s in state):
|
| 101 |
+
return "\n\n".join(state)
|
| 102 |
+
return json.dumps(state, ensure_ascii=False, indent=2)
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def render_user(state: Any, q: Question, labels: list[str]) -> str:
|
| 106 |
+
parts = [f"<state>\n{render_state(state)}\n</state>", ""]
|
| 107 |
+
if q.type == "choice":
|
| 108 |
+
opts = "\n".join(f"{lab}: {key}: {desc}" for lab, (key, desc) in zip(labels, q.criteria.items()))
|
| 109 |
+
parts += [
|
| 110 |
+
f"Question: {q.instructions}",
|
| 111 |
+
f"Options:\n{opts}",
|
| 112 |
+
f"Answer with the label of the best option ({labels[0]}-{labels[-1]}).",
|
| 113 |
+
]
|
| 114 |
+
elif q.type == "score":
|
| 115 |
+
levels = "\n".join(f"{lab}: {desc}" for lab, desc in zip(labels, q.criteria))
|
| 116 |
+
parts += [
|
| 117 |
+
f"Question: {q.instructions}",
|
| 118 |
+
f"Scale (lowest to highest):\n{levels}",
|
| 119 |
+
f"Answer with the level number (0-{len(labels) - 1}).",
|
| 120 |
+
]
|
| 121 |
+
else:
|
| 122 |
+
parts += [f"Question: {q.instructions}", "Answer yes or no."]
|
| 123 |
+
return "\n".join(parts)
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
def build_prompt(tokenizer, state: Any, q: Question) -> str:
|
| 127 |
+
return tokenizer.apply_chat_template(
|
| 128 |
+
[{"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": render_user(state, q, labels_for(tokenizer, q))}],
|
| 129 |
+
add_generation_prompt=True,
|
| 130 |
+
tokenize=False,
|
| 131 |
+
enable_thinking=False,
|
| 132 |
+
)
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
def _label_token_id(tokenizer, label: str) -> int:
|
| 136 |
+
ids = tokenizer.encode(label, add_special_tokens=False)
|
| 137 |
+
if len(ids) != 1:
|
| 138 |
+
raise ValueError(f"label {label!r} is not a single token for this tokenizer")
|
| 139 |
+
return ids[0]
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
def label_token_ids(tokenizer, q: Question) -> list[int]:
|
| 143 |
+
return [_label_token_id(tokenizer, lab) for lab in labels_for(tokenizer, q)]
|
golden.json
DELETED
|
The diff for this file is too large to render.
See raw diff
|
|
|
inference.py
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Backend-independent Jet prompt encoding and typed answers."""
|
| 2 |
+
from __future__ import annotations
|
| 3 |
+
import math
|
| 4 |
+
from typing import Any
|
| 5 |
+
import numpy as np
|
| 6 |
+
from format import Question, build_prompt, render_state
|
| 7 |
+
|
| 8 |
+
DEFAULT_MAX_STATE_TOKENS = 4096
|
| 9 |
+
|
| 10 |
+
def encode(tokenizer, state: Any, q: Question, max_state_tokens: int = DEFAULT_MAX_STATE_TOKENS) -> list[int]:
|
| 11 |
+
"""Tokenize a prompt, truncating the middle of an over-long state."""
|
| 12 |
+
text = render_state(state)
|
| 13 |
+
ids = tokenizer.encode(text, add_special_tokens=False)
|
| 14 |
+
if len(ids) > max_state_tokens:
|
| 15 |
+
half = max_state_tokens // 2
|
| 16 |
+
text = tokenizer.decode(ids[:half]) + "\n[...]\n" + tokenizer.decode(ids[-half:])
|
| 17 |
+
return tokenizer.encode(build_prompt(tokenizer, text, q), add_special_tokens=False)
|
| 18 |
+
|
| 19 |
+
def summarize(q: Question, probs: np.ndarray) -> dict[str, Any]:
|
| 20 |
+
"""Turn a label distribution into the typed answer for the question."""
|
| 21 |
+
probs = probs.astype(np.float64)
|
| 22 |
+
k = len(probs)
|
| 23 |
+
entropy = -float(np.sum(probs * np.log(np.clip(probs, 1e-12, 1.0))))
|
| 24 |
+
confidence = round(1.0 - entropy / math.log(k), 4)
|
| 25 |
+
if q.type == "choice":
|
| 26 |
+
keys = q.keys
|
| 27 |
+
return {
|
| 28 |
+
"type": "choice",
|
| 29 |
+
"choice": keys[int(np.argmax(probs))],
|
| 30 |
+
"probabilities": {key: round(float(p), 4) for key, p in zip(keys, probs)},
|
| 31 |
+
"confidence": confidence,
|
| 32 |
+
}
|
| 33 |
+
if q.type == "score":
|
| 34 |
+
return {
|
| 35 |
+
"type": "score",
|
| 36 |
+
"score": round(float(np.dot(np.arange(k), probs)), 4),
|
| 37 |
+
"level": q.criteria[int(np.argmax(probs))],
|
| 38 |
+
"probabilities": [round(float(p), 4) for p in probs],
|
| 39 |
+
"confidence": confidence,
|
| 40 |
+
}
|
| 41 |
+
return {"type": "noul", "probability": round(float(probs[1]), 4), "confidence": confidence}
|
jet.py
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Typed inference for the merged Jet model. Linux + NVIDIA CUDA, Python 3.12."""
|
| 2 |
+
import json
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
import torch
|
| 5 |
+
from format import Question, label_token_ids
|
| 6 |
+
from inference import encode, summarize
|
| 7 |
+
from runtime import load_model, label_logits
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class Jet:
|
| 11 |
+
def __init__(self, model_path=None, max_tokens=8192):
|
| 12 |
+
self.path = Path(model_path) if model_path else Path(__file__).resolve().parent
|
| 13 |
+
if not 0 < max_tokens <= 8192:
|
| 14 |
+
raise ValueError('This release supports complete prompts up to 8192 tokens.')
|
| 15 |
+
self.max_tokens = max_tokens
|
| 16 |
+
torch.set_num_threads(4)
|
| 17 |
+
self.temperatures = json.loads((self.path / 'calibration.json').read_text())
|
| 18 |
+
self.model, self.tokenizer = load_model(str(self.path))
|
| 19 |
+
|
| 20 |
+
@torch.no_grad()
|
| 21 |
+
def decide(self, state, questions):
|
| 22 |
+
if not isinstance(questions, dict) or not questions:
|
| 23 |
+
raise ValueError('questions must be a nonempty mapping')
|
| 24 |
+
prepared = []
|
| 25 |
+
for name, value in questions.items():
|
| 26 |
+
q = Question.from_dict(value)
|
| 27 |
+
ids = encode(self.tokenizer, state, q, 10**9)
|
| 28 |
+
if len(ids) > self.max_tokens:
|
| 29 |
+
raise ValueError(f'{name}: complete prompt exceeds {self.max_tokens} tokens; no truncation applied')
|
| 30 |
+
prepared.append((name, q, {'ids': ids, 'labels': label_token_ids(self.tokenizer, q)}))
|
| 31 |
+
answers = {}
|
| 32 |
+
for name, q, example in prepared:
|
| 33 |
+
z = label_logits(self.model, example)
|
| 34 |
+
probabilities = (z.double() / self.temperatures[q.type]).softmax(-1).cpu().numpy()
|
| 35 |
+
answers[name] = summarize(q, probabilities)
|
| 36 |
+
return {'model': 'michaljach/jet', 'answers': answers}
|
| 37 |
+
|
| 38 |
+
if __name__ == '__main__':
|
| 39 |
+
import sys
|
| 40 |
+
request = json.load(sys.stdin)
|
| 41 |
+
print(json.dumps(Jet().decide(request['state'], request['questions']), indent=2))
|
merge-provenance.json
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"base": "Qwen/Qwen3.5-4B",
|
| 3 |
+
"base_revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
|
| 4 |
+
"adapter_sha256": "e3d8be9c82190f1357d58380a6908e7dbf511b4e9e3cafbd9288f0c7dbb044e2",
|
| 5 |
+
"step": 3750,
|
| 6 |
+
"merge": "BF16 backbone plus FP32 B@A times alpha/r, rounded to BF16 (PEFT default merge arithmetic)",
|
| 7 |
+
"architecture": "complete Qwen3_5ForCausalLM text backbone; tied lm_head; unused vision tower omitted",
|
| 8 |
+
"merged_modules": 248,
|
| 9 |
+
"tensors": 426,
|
| 10 |
+
"bytes": 8411510272
|
| 11 |
+
}
|
merge-validation.json
ADDED
|
@@ -0,0 +1,255 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cases": 36,
|
| 3 |
+
"selection": "First 12 cases per type, scanning local test then calibration; no outcome selection",
|
| 4 |
+
"argmax_flips": 1,
|
| 5 |
+
"max_probability_difference": 0.04372840585277227,
|
| 6 |
+
"records": [
|
| 7 |
+
{
|
| 8 |
+
"test_index": 0,
|
| 9 |
+
"type": "choice",
|
| 10 |
+
"same_argmax": true,
|
| 11 |
+
"max_probability_difference": 0.0011152033659319743
|
| 12 |
+
},
|
| 13 |
+
{
|
| 14 |
+
"test_index": 1,
|
| 15 |
+
"type": "choice",
|
| 16 |
+
"same_argmax": true,
|
| 17 |
+
"max_probability_difference": 0.0
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"test_index": 2,
|
| 21 |
+
"type": "choice",
|
| 22 |
+
"same_argmax": true,
|
| 23 |
+
"max_probability_difference": 0.0108992070366863
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"test_index": 3,
|
| 27 |
+
"type": "choice",
|
| 28 |
+
"same_argmax": true,
|
| 29 |
+
"max_probability_difference": 0.0008074928979708608
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_index": 4,
|
| 33 |
+
"type": "choice",
|
| 34 |
+
"same_argmax": true,
|
| 35 |
+
"max_probability_difference": 0.04372840585277227
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"test_index": 5,
|
| 39 |
+
"type": "choice",
|
| 40 |
+
"same_argmax": true,
|
| 41 |
+
"max_probability_difference": 0.0012248826281491855
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"test_index": 6,
|
| 45 |
+
"type": "choice",
|
| 46 |
+
"same_argmax": true,
|
| 47 |
+
"max_probability_difference": 0.0009706432719255795
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_index": 7,
|
| 51 |
+
"type": "choice",
|
| 52 |
+
"same_argmax": true,
|
| 53 |
+
"max_probability_difference": 0.002213577468929806
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"test_index": 8,
|
| 57 |
+
"type": "choice",
|
| 58 |
+
"same_argmax": true,
|
| 59 |
+
"max_probability_difference": 0.0006500812415296542
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"test_index": 9,
|
| 63 |
+
"type": "choice",
|
| 64 |
+
"same_argmax": true,
|
| 65 |
+
"max_probability_difference": 0.0009609628007468674
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"test_index": 10,
|
| 69 |
+
"type": "choice",
|
| 70 |
+
"same_argmax": true,
|
| 71 |
+
"max_probability_difference": 0.00012381259397531777
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"test_index": 11,
|
| 75 |
+
"type": "choice",
|
| 76 |
+
"same_argmax": true,
|
| 77 |
+
"max_probability_difference": 0.0
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"test_index": 624,
|
| 81 |
+
"type": "score",
|
| 82 |
+
"same_argmax": true,
|
| 83 |
+
"max_probability_difference": 0.0
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"test_index": 627,
|
| 87 |
+
"type": "score",
|
| 88 |
+
"same_argmax": true,
|
| 89 |
+
"max_probability_difference": 0.000754577823111946
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"test_index": 628,
|
| 93 |
+
"type": "noul",
|
| 94 |
+
"same_argmax": true,
|
| 95 |
+
"max_probability_difference": 0.0
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"test_index": 633,
|
| 99 |
+
"type": "noul",
|
| 100 |
+
"same_argmax": true,
|
| 101 |
+
"max_probability_difference": 0.0
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"test_index": 640,
|
| 105 |
+
"type": "score",
|
| 106 |
+
"same_argmax": true,
|
| 107 |
+
"max_probability_difference": 0.0010637251609036114
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"test_index": 666,
|
| 111 |
+
"type": "noul",
|
| 112 |
+
"same_argmax": true,
|
| 113 |
+
"max_probability_difference": 0.0
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"test_index": 667,
|
| 117 |
+
"type": "noul",
|
| 118 |
+
"same_argmax": true,
|
| 119 |
+
"max_probability_difference": 0.0
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"test_index": 668,
|
| 123 |
+
"type": "score",
|
| 124 |
+
"same_argmax": true,
|
| 125 |
+
"max_probability_difference": 0.017339621180759268
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"test_index": 669,
|
| 129 |
+
"type": "noul",
|
| 130 |
+
"same_argmax": true,
|
| 131 |
+
"max_probability_difference": 0.0
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"test_index": 670,
|
| 135 |
+
"type": "score",
|
| 136 |
+
"same_argmax": false,
|
| 137 |
+
"max_probability_difference": 0.028077057361391045
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"test_index": 682,
|
| 141 |
+
"type": "noul",
|
| 142 |
+
"same_argmax": true,
|
| 143 |
+
"max_probability_difference": 0.009419094585836119
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"test_index": 685,
|
| 147 |
+
"type": "score",
|
| 148 |
+
"same_argmax": true,
|
| 149 |
+
"max_probability_difference": 0.007916961020786789
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"test_index": 688,
|
| 153 |
+
"type": "noul",
|
| 154 |
+
"same_argmax": true,
|
| 155 |
+
"max_probability_difference": 0.003166425585453325
|
| 156 |
+
},
|
| 157 |
+
{
|
| 158 |
+
"test_index": 693,
|
| 159 |
+
"type": "noul",
|
| 160 |
+
"same_argmax": true,
|
| 161 |
+
"max_probability_difference": 0.0
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"test_index": 713,
|
| 165 |
+
"type": "score",
|
| 166 |
+
"same_argmax": true,
|
| 167 |
+
"max_probability_difference": 0.0010361768955787776
|
| 168 |
+
},
|
| 169 |
+
{
|
| 170 |
+
"test_index": 719,
|
| 171 |
+
"type": "noul",
|
| 172 |
+
"same_argmax": true,
|
| 173 |
+
"max_probability_difference": 0.0037330092417818816
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"test_index": 726,
|
| 177 |
+
"type": "score",
|
| 178 |
+
"same_argmax": true,
|
| 179 |
+
"max_probability_difference": 0.02127082649737211
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"test_index": 727,
|
| 183 |
+
"type": "noul",
|
| 184 |
+
"same_argmax": true,
|
| 185 |
+
"max_probability_difference": 0.0
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"test_index": 729,
|
| 189 |
+
"type": "score",
|
| 190 |
+
"same_argmax": true,
|
| 191 |
+
"max_probability_difference": 0.0006679498517172117
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"test_index": 741,
|
| 195 |
+
"type": "score",
|
| 196 |
+
"same_argmax": true,
|
| 197 |
+
"max_probability_difference": 0.014585731531814888
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"test_index": 748,
|
| 201 |
+
"type": "score",
|
| 202 |
+
"same_argmax": true,
|
| 203 |
+
"max_probability_difference": 0.0
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"test_index": 767,
|
| 207 |
+
"type": "noul",
|
| 208 |
+
"same_argmax": true,
|
| 209 |
+
"max_probability_difference": 0.0
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"test_index": 769,
|
| 213 |
+
"type": "noul",
|
| 214 |
+
"same_argmax": true,
|
| 215 |
+
"max_probability_difference": 0.0
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"test_index": 845,
|
| 219 |
+
"type": "score",
|
| 220 |
+
"same_argmax": true,
|
| 221 |
+
"max_probability_difference": 0.0002716545718139925
|
| 222 |
+
}
|
| 223 |
+
],
|
| 224 |
+
"wrapper_smoke": {
|
| 225 |
+
"model": "michaljach/jet",
|
| 226 |
+
"answers": {
|
| 227 |
+
"category": {
|
| 228 |
+
"type": "choice",
|
| 229 |
+
"choice": "damage",
|
| 230 |
+
"probabilities": {
|
| 231 |
+
"damage": 0.9872,
|
| 232 |
+
"billing": 0.0128
|
| 233 |
+
},
|
| 234 |
+
"confidence": 0.901
|
| 235 |
+
},
|
| 236 |
+
"severity": {
|
| 237 |
+
"type": "score",
|
| 238 |
+
"score": 1.5647,
|
| 239 |
+
"level": "severe",
|
| 240 |
+
"probabilities": [
|
| 241 |
+
0.0099,
|
| 242 |
+
0.4154,
|
| 243 |
+
0.5747
|
| 244 |
+
],
|
| 245 |
+
"confidence": 0.3363
|
| 246 |
+
},
|
| 247 |
+
"damaged": {
|
| 248 |
+
"type": "noul",
|
| 249 |
+
"probability": 0.9487,
|
| 250 |
+
"confidence": 0.7082
|
| 251 |
+
}
|
| 252 |
+
}
|
| 253 |
+
},
|
| 254 |
+
"base_download_required": false
|
| 255 |
+
}
|
model.safetensors → model-00001-of-00009.safetensors
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:758c65ea6c124bbe96b68a2d5475f7e4ec0ff9b6a4dbba87f295144ace55f07d
|
| 3 |
+
size 1271398544
|
onnx/model_q8.onnx → model-00002-of-00009.safetensors
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e84de8d9a45219bbeb1a33392370979cce7702ceab4dea27b5f4b3b4e78deb0
|
| 3 |
+
size 990906704
|
model-00003-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d96476f63563f6db809675f551857e392102189c17eae2e2999bed0c4f967b22
|
| 3 |
+
size 990906704
|
model-00004-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bf5440f1558b2b9fc1e4ddec335cc77c04ac89621ce91a72ae3b3a542ad6eb25
|
| 3 |
+
size 990906712
|
model-00005-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:52a8e3c8c06d2d16554e6146ffadcb0830367180d4ed2d4725084c0b4b257967
|
| 3 |
+
size 986062968
|
model-00006-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0d560b820ad7edc2faad717f6655b3f638e8779b40fda4e333d6457bcb9ed47c
|
| 3 |
+
size 983687224
|
model-00007-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fff39c3f2152571b72d377c6095daca3c796da187d94e88d18292c8709aa75d3
|
| 3 |
+
size 999026496
|
model-00008-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:da8fa5e8e7bbcab3529e6f4014fdc3a42240a27c4b7fd123f3690abb35c96789
|
| 3 |
+
size 982888776
|
model-00009-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8d76d978760f6c15b9f5cb0c3b5eeada3513ff8a53104389cd895b5a5a092a94
|
| 3 |
+
size 215774320
|
model.safetensors.index.json
CHANGED
|
@@ -1,318 +1,433 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
"
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 8411510272
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"model.embed_tokens.weight": "model-00001-of-00009.safetensors",
|
| 7 |
+
"model.layers.1.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 8 |
+
"model.layers.1.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 9 |
+
"model.layers.1.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 10 |
+
"model.layers.10.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 11 |
+
"model.layers.10.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 12 |
+
"model.layers.10.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 13 |
+
"model.layers.11.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 14 |
+
"model.layers.11.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 15 |
+
"model.layers.11.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 16 |
+
"model.layers.12.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 17 |
+
"model.layers.12.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 18 |
+
"model.layers.12.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 19 |
+
"model.layers.13.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 20 |
+
"model.layers.13.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 21 |
+
"model.layers.13.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 22 |
+
"model.layers.14.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 23 |
+
"model.layers.14.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 24 |
+
"model.layers.14.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 25 |
+
"model.layers.15.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 26 |
+
"model.layers.15.mlp.gate_proj.weight": "model-00002-of-00009.safetensors",
|
| 27 |
+
"model.layers.15.mlp.up_proj.weight": "model-00002-of-00009.safetensors",
|
| 28 |
+
"model.layers.16.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 29 |
+
"model.layers.16.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 30 |
+
"model.layers.16.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 31 |
+
"model.layers.17.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 32 |
+
"model.layers.17.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 33 |
+
"model.layers.17.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 34 |
+
"model.layers.18.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 35 |
+
"model.layers.18.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 36 |
+
"model.layers.18.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 37 |
+
"model.layers.19.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 38 |
+
"model.layers.19.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 39 |
+
"model.layers.19.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 40 |
+
"model.layers.2.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 41 |
+
"model.layers.2.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 42 |
+
"model.layers.2.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 43 |
+
"model.layers.20.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 44 |
+
"model.layers.20.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 45 |
+
"model.layers.20.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 46 |
+
"model.layers.21.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 47 |
+
"model.layers.21.mlp.gate_proj.weight": "model-00003-of-00009.safetensors",
|
| 48 |
+
"model.layers.21.mlp.up_proj.weight": "model-00003-of-00009.safetensors",
|
| 49 |
+
"model.layers.22.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 50 |
+
"model.layers.22.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 51 |
+
"model.layers.22.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 52 |
+
"model.layers.23.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 53 |
+
"model.layers.23.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 54 |
+
"model.layers.23.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 55 |
+
"model.layers.24.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 56 |
+
"model.layers.24.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 57 |
+
"model.layers.24.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 58 |
+
"model.layers.25.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 59 |
+
"model.layers.25.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 60 |
+
"model.layers.25.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 61 |
+
"model.layers.26.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 62 |
+
"model.layers.29.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 63 |
+
"model.layers.3.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 64 |
+
"model.layers.3.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 65 |
+
"model.layers.3.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 66 |
+
"model.layers.30.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 67 |
+
"model.layers.30.mlp.gate_proj.weight": "model-00004-of-00009.safetensors",
|
| 68 |
+
"model.layers.30.mlp.up_proj.weight": "model-00004-of-00009.safetensors",
|
| 69 |
+
"model.layers.31.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 70 |
+
"model.layers.31.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 71 |
+
"model.layers.31.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 72 |
+
"model.layers.4.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 73 |
+
"model.layers.4.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 74 |
+
"model.layers.4.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 75 |
+
"model.layers.5.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 76 |
+
"model.layers.5.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 77 |
+
"model.layers.5.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 78 |
+
"model.layers.6.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 79 |
+
"model.layers.6.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 80 |
+
"model.layers.6.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 81 |
+
"model.layers.7.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 82 |
+
"model.layers.7.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 83 |
+
"model.layers.7.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 84 |
+
"model.layers.8.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 85 |
+
"model.layers.8.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 86 |
+
"model.layers.8.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 87 |
+
"model.layers.9.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 88 |
+
"model.layers.9.mlp.gate_proj.weight": "model-00005-of-00009.safetensors",
|
| 89 |
+
"model.layers.9.mlp.up_proj.weight": "model-00005-of-00009.safetensors",
|
| 90 |
+
"model.layers.0.input_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 91 |
+
"model.layers.0.linear_attn.A_log": "model-00005-of-00009.safetensors",
|
| 92 |
+
"model.layers.0.linear_attn.conv1d.weight": "model-00005-of-00009.safetensors",
|
| 93 |
+
"model.layers.0.linear_attn.dt_bias": "model-00005-of-00009.safetensors",
|
| 94 |
+
"model.layers.0.linear_attn.in_proj_a.weight": "model-00005-of-00009.safetensors",
|
| 95 |
+
"model.layers.0.linear_attn.in_proj_b.weight": "model-00005-of-00009.safetensors",
|
| 96 |
+
"model.layers.0.linear_attn.in_proj_qkv.weight": "model-00005-of-00009.safetensors",
|
| 97 |
+
"model.layers.0.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 98 |
+
"model.layers.0.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 99 |
+
"model.layers.0.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 100 |
+
"model.layers.0.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 101 |
+
"model.layers.0.mlp.gate_proj.weight": "model-00006-of-00009.safetensors",
|
| 102 |
+
"model.layers.0.mlp.up_proj.weight": "model-00006-of-00009.safetensors",
|
| 103 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 104 |
+
"model.layers.1.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 105 |
+
"model.layers.1.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 106 |
+
"model.layers.1.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 107 |
+
"model.layers.1.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 108 |
+
"model.layers.1.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 109 |
+
"model.layers.1.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 110 |
+
"model.layers.1.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 111 |
+
"model.layers.1.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 112 |
+
"model.layers.1.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 113 |
+
"model.layers.1.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 114 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 115 |
+
"model.layers.10.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 116 |
+
"model.layers.10.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 117 |
+
"model.layers.10.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 118 |
+
"model.layers.10.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 119 |
+
"model.layers.10.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 120 |
+
"model.layers.10.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 121 |
+
"model.layers.10.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 122 |
+
"model.layers.10.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 123 |
+
"model.layers.10.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 124 |
+
"model.layers.10.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 125 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 126 |
+
"model.layers.11.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 127 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 128 |
+
"model.layers.11.self_attn.k_norm.weight": "model-00006-of-00009.safetensors",
|
| 129 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00006-of-00009.safetensors",
|
| 130 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 131 |
+
"model.layers.11.self_attn.q_norm.weight": "model-00006-of-00009.safetensors",
|
| 132 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00006-of-00009.safetensors",
|
| 133 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00006-of-00009.safetensors",
|
| 134 |
+
"model.layers.12.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 135 |
+
"model.layers.12.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 136 |
+
"model.layers.12.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 137 |
+
"model.layers.12.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 138 |
+
"model.layers.12.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 139 |
+
"model.layers.12.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 140 |
+
"model.layers.12.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 141 |
+
"model.layers.12.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 142 |
+
"model.layers.12.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 143 |
+
"model.layers.12.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 144 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 145 |
+
"model.layers.13.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 146 |
+
"model.layers.13.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 147 |
+
"model.layers.13.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 148 |
+
"model.layers.13.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 149 |
+
"model.layers.13.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 150 |
+
"model.layers.13.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 151 |
+
"model.layers.13.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 152 |
+
"model.layers.13.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 153 |
+
"model.layers.13.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 154 |
+
"model.layers.13.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 155 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 156 |
+
"model.layers.14.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 157 |
+
"model.layers.14.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 158 |
+
"model.layers.14.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 159 |
+
"model.layers.14.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 160 |
+
"model.layers.14.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 161 |
+
"model.layers.14.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 162 |
+
"model.layers.14.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 163 |
+
"model.layers.14.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 164 |
+
"model.layers.14.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 165 |
+
"model.layers.14.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 166 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 167 |
+
"model.layers.15.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 168 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 169 |
+
"model.layers.15.self_attn.k_norm.weight": "model-00006-of-00009.safetensors",
|
| 170 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00006-of-00009.safetensors",
|
| 171 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 172 |
+
"model.layers.15.self_attn.q_norm.weight": "model-00006-of-00009.safetensors",
|
| 173 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00006-of-00009.safetensors",
|
| 174 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00006-of-00009.safetensors",
|
| 175 |
+
"model.layers.16.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 176 |
+
"model.layers.16.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 177 |
+
"model.layers.16.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 178 |
+
"model.layers.16.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 179 |
+
"model.layers.16.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 180 |
+
"model.layers.16.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 181 |
+
"model.layers.16.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 182 |
+
"model.layers.16.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 183 |
+
"model.layers.16.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 184 |
+
"model.layers.16.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 185 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 186 |
+
"model.layers.17.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 187 |
+
"model.layers.17.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 188 |
+
"model.layers.17.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 189 |
+
"model.layers.17.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 190 |
+
"model.layers.17.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 191 |
+
"model.layers.17.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 192 |
+
"model.layers.17.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 193 |
+
"model.layers.17.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 194 |
+
"model.layers.17.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 195 |
+
"model.layers.17.linear_attn.out_proj.weight": "model-00006-of-00009.safetensors",
|
| 196 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 197 |
+
"model.layers.18.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 198 |
+
"model.layers.18.linear_attn.A_log": "model-00006-of-00009.safetensors",
|
| 199 |
+
"model.layers.18.linear_attn.conv1d.weight": "model-00006-of-00009.safetensors",
|
| 200 |
+
"model.layers.18.linear_attn.dt_bias": "model-00006-of-00009.safetensors",
|
| 201 |
+
"model.layers.18.linear_attn.in_proj_a.weight": "model-00006-of-00009.safetensors",
|
| 202 |
+
"model.layers.18.linear_attn.in_proj_b.weight": "model-00006-of-00009.safetensors",
|
| 203 |
+
"model.layers.18.linear_attn.in_proj_qkv.weight": "model-00006-of-00009.safetensors",
|
| 204 |
+
"model.layers.18.linear_attn.in_proj_z.weight": "model-00006-of-00009.safetensors",
|
| 205 |
+
"model.layers.18.linear_attn.norm.weight": "model-00006-of-00009.safetensors",
|
| 206 |
+
"model.layers.18.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 207 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 208 |
+
"model.layers.19.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 209 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 210 |
+
"model.layers.19.self_attn.k_norm.weight": "model-00007-of-00009.safetensors",
|
| 211 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00007-of-00009.safetensors",
|
| 212 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 213 |
+
"model.layers.19.self_attn.q_norm.weight": "model-00007-of-00009.safetensors",
|
| 214 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00007-of-00009.safetensors",
|
| 215 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00007-of-00009.safetensors",
|
| 216 |
+
"model.layers.2.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 217 |
+
"model.layers.2.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 218 |
+
"model.layers.2.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 219 |
+
"model.layers.2.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 220 |
+
"model.layers.2.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 221 |
+
"model.layers.2.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 222 |
+
"model.layers.2.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 223 |
+
"model.layers.2.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 224 |
+
"model.layers.2.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 225 |
+
"model.layers.2.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 226 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 227 |
+
"model.layers.20.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 228 |
+
"model.layers.20.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 229 |
+
"model.layers.20.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 230 |
+
"model.layers.20.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 231 |
+
"model.layers.20.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 232 |
+
"model.layers.20.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 233 |
+
"model.layers.20.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 234 |
+
"model.layers.20.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 235 |
+
"model.layers.20.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 236 |
+
"model.layers.20.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 237 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 238 |
+
"model.layers.21.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 239 |
+
"model.layers.21.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 240 |
+
"model.layers.21.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 241 |
+
"model.layers.21.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 242 |
+
"model.layers.21.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 243 |
+
"model.layers.21.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 244 |
+
"model.layers.21.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 245 |
+
"model.layers.21.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 246 |
+
"model.layers.21.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 247 |
+
"model.layers.21.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 248 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 249 |
+
"model.layers.22.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 250 |
+
"model.layers.22.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 251 |
+
"model.layers.22.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 252 |
+
"model.layers.22.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 253 |
+
"model.layers.22.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 254 |
+
"model.layers.22.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 255 |
+
"model.layers.22.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 256 |
+
"model.layers.22.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 257 |
+
"model.layers.22.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 258 |
+
"model.layers.22.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 259 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 260 |
+
"model.layers.23.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 261 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 262 |
+
"model.layers.23.self_attn.k_norm.weight": "model-00007-of-00009.safetensors",
|
| 263 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00007-of-00009.safetensors",
|
| 264 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 265 |
+
"model.layers.23.self_attn.q_norm.weight": "model-00007-of-00009.safetensors",
|
| 266 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00007-of-00009.safetensors",
|
| 267 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00007-of-00009.safetensors",
|
| 268 |
+
"model.layers.24.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 269 |
+
"model.layers.24.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 270 |
+
"model.layers.24.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 271 |
+
"model.layers.24.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 272 |
+
"model.layers.24.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 273 |
+
"model.layers.24.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 274 |
+
"model.layers.24.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 275 |
+
"model.layers.24.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 276 |
+
"model.layers.24.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 277 |
+
"model.layers.24.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 278 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 279 |
+
"model.layers.25.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 280 |
+
"model.layers.25.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 281 |
+
"model.layers.25.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 282 |
+
"model.layers.25.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 283 |
+
"model.layers.25.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 284 |
+
"model.layers.25.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 285 |
+
"model.layers.25.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 286 |
+
"model.layers.25.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 287 |
+
"model.layers.25.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 288 |
+
"model.layers.25.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 289 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 290 |
+
"model.layers.26.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 291 |
+
"model.layers.26.linear_attn.A_log": "model-00007-of-00009.safetensors",
|
| 292 |
+
"model.layers.26.linear_attn.conv1d.weight": "model-00007-of-00009.safetensors",
|
| 293 |
+
"model.layers.26.linear_attn.dt_bias": "model-00007-of-00009.safetensors",
|
| 294 |
+
"model.layers.26.linear_attn.in_proj_a.weight": "model-00007-of-00009.safetensors",
|
| 295 |
+
"model.layers.26.linear_attn.in_proj_b.weight": "model-00007-of-00009.safetensors",
|
| 296 |
+
"model.layers.26.linear_attn.in_proj_qkv.weight": "model-00007-of-00009.safetensors",
|
| 297 |
+
"model.layers.26.linear_attn.in_proj_z.weight": "model-00007-of-00009.safetensors",
|
| 298 |
+
"model.layers.26.linear_attn.norm.weight": "model-00007-of-00009.safetensors",
|
| 299 |
+
"model.layers.26.linear_attn.out_proj.weight": "model-00007-of-00009.safetensors",
|
| 300 |
+
"model.layers.26.mlp.gate_proj.weight": "model-00007-of-00009.safetensors",
|
| 301 |
+
"model.layers.26.mlp.up_proj.weight": "model-00007-of-00009.safetensors",
|
| 302 |
+
"model.layers.26.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 303 |
+
"model.layers.27.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 304 |
+
"model.layers.27.mlp.down_proj.weight": "model-00007-of-00009.safetensors",
|
| 305 |
+
"model.layers.27.mlp.gate_proj.weight": "model-00007-of-00009.safetensors",
|
| 306 |
+
"model.layers.27.mlp.up_proj.weight": "model-00007-of-00009.safetensors",
|
| 307 |
+
"model.layers.27.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 308 |
+
"model.layers.27.self_attn.k_norm.weight": "model-00007-of-00009.safetensors",
|
| 309 |
+
"model.layers.27.self_attn.k_proj.weight": "model-00007-of-00009.safetensors",
|
| 310 |
+
"model.layers.27.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 311 |
+
"model.layers.27.self_attn.q_norm.weight": "model-00008-of-00009.safetensors",
|
| 312 |
+
"model.layers.27.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 313 |
+
"model.layers.27.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 314 |
+
"model.layers.28.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 315 |
+
"model.layers.28.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 316 |
+
"model.layers.28.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 317 |
+
"model.layers.28.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 318 |
+
"model.layers.28.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 319 |
+
"model.layers.28.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 320 |
+
"model.layers.28.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 321 |
+
"model.layers.28.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 322 |
+
"model.layers.28.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 323 |
+
"model.layers.28.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 324 |
+
"model.layers.28.mlp.down_proj.weight": "model-00008-of-00009.safetensors",
|
| 325 |
+
"model.layers.28.mlp.gate_proj.weight": "model-00008-of-00009.safetensors",
|
| 326 |
+
"model.layers.28.mlp.up_proj.weight": "model-00008-of-00009.safetensors",
|
| 327 |
+
"model.layers.28.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 328 |
+
"model.layers.29.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 329 |
+
"model.layers.29.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 330 |
+
"model.layers.29.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 331 |
+
"model.layers.29.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 332 |
+
"model.layers.29.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 333 |
+
"model.layers.29.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 334 |
+
"model.layers.29.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 335 |
+
"model.layers.29.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 336 |
+
"model.layers.29.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 337 |
+
"model.layers.29.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 338 |
+
"model.layers.29.mlp.gate_proj.weight": "model-00008-of-00009.safetensors",
|
| 339 |
+
"model.layers.29.mlp.up_proj.weight": "model-00008-of-00009.safetensors",
|
| 340 |
+
"model.layers.29.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 341 |
+
"model.layers.3.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 342 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 343 |
+
"model.layers.3.self_attn.k_norm.weight": "model-00008-of-00009.safetensors",
|
| 344 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 345 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 346 |
+
"model.layers.3.self_attn.q_norm.weight": "model-00008-of-00009.safetensors",
|
| 347 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 348 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 349 |
+
"model.layers.30.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 350 |
+
"model.layers.30.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 351 |
+
"model.layers.30.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 352 |
+
"model.layers.30.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 353 |
+
"model.layers.30.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 354 |
+
"model.layers.30.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 355 |
+
"model.layers.30.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 356 |
+
"model.layers.30.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 357 |
+
"model.layers.30.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 358 |
+
"model.layers.30.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 359 |
+
"model.layers.30.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 360 |
+
"model.layers.31.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 361 |
+
"model.layers.31.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 362 |
+
"model.layers.31.self_attn.k_norm.weight": "model-00008-of-00009.safetensors",
|
| 363 |
+
"model.layers.31.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 364 |
+
"model.layers.31.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 365 |
+
"model.layers.31.self_attn.q_norm.weight": "model-00008-of-00009.safetensors",
|
| 366 |
+
"model.layers.31.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 367 |
+
"model.layers.31.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 368 |
+
"model.layers.4.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 369 |
+
"model.layers.4.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 370 |
+
"model.layers.4.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 371 |
+
"model.layers.4.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 372 |
+
"model.layers.4.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 373 |
+
"model.layers.4.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 374 |
+
"model.layers.4.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 375 |
+
"model.layers.4.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 376 |
+
"model.layers.4.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 377 |
+
"model.layers.4.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 378 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 379 |
+
"model.layers.5.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 380 |
+
"model.layers.5.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 381 |
+
"model.layers.5.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 382 |
+
"model.layers.5.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 383 |
+
"model.layers.5.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 384 |
+
"model.layers.5.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 385 |
+
"model.layers.5.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 386 |
+
"model.layers.5.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 387 |
+
"model.layers.5.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 388 |
+
"model.layers.5.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 389 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 390 |
+
"model.layers.6.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 391 |
+
"model.layers.6.linear_attn.A_log": "model-00008-of-00009.safetensors",
|
| 392 |
+
"model.layers.6.linear_attn.conv1d.weight": "model-00008-of-00009.safetensors",
|
| 393 |
+
"model.layers.6.linear_attn.dt_bias": "model-00008-of-00009.safetensors",
|
| 394 |
+
"model.layers.6.linear_attn.in_proj_a.weight": "model-00008-of-00009.safetensors",
|
| 395 |
+
"model.layers.6.linear_attn.in_proj_b.weight": "model-00008-of-00009.safetensors",
|
| 396 |
+
"model.layers.6.linear_attn.in_proj_qkv.weight": "model-00008-of-00009.safetensors",
|
| 397 |
+
"model.layers.6.linear_attn.in_proj_z.weight": "model-00008-of-00009.safetensors",
|
| 398 |
+
"model.layers.6.linear_attn.norm.weight": "model-00008-of-00009.safetensors",
|
| 399 |
+
"model.layers.6.linear_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 400 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 401 |
+
"model.layers.7.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 402 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 403 |
+
"model.layers.7.self_attn.k_norm.weight": "model-00008-of-00009.safetensors",
|
| 404 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 405 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 406 |
+
"model.layers.7.self_attn.q_norm.weight": "model-00008-of-00009.safetensors",
|
| 407 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 408 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 409 |
+
"model.layers.8.input_layernorm.weight": "model-00009-of-00009.safetensors",
|
| 410 |
+
"model.layers.8.linear_attn.A_log": "model-00009-of-00009.safetensors",
|
| 411 |
+
"model.layers.8.linear_attn.conv1d.weight": "model-00009-of-00009.safetensors",
|
| 412 |
+
"model.layers.8.linear_attn.dt_bias": "model-00009-of-00009.safetensors",
|
| 413 |
+
"model.layers.8.linear_attn.in_proj_a.weight": "model-00009-of-00009.safetensors",
|
| 414 |
+
"model.layers.8.linear_attn.in_proj_b.weight": "model-00009-of-00009.safetensors",
|
| 415 |
+
"model.layers.8.linear_attn.in_proj_qkv.weight": "model-00009-of-00009.safetensors",
|
| 416 |
+
"model.layers.8.linear_attn.in_proj_z.weight": "model-00009-of-00009.safetensors",
|
| 417 |
+
"model.layers.8.linear_attn.norm.weight": "model-00009-of-00009.safetensors",
|
| 418 |
+
"model.layers.8.linear_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 419 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00009-of-00009.safetensors",
|
| 420 |
+
"model.layers.9.input_layernorm.weight": "model-00009-of-00009.safetensors",
|
| 421 |
+
"model.layers.9.linear_attn.A_log": "model-00009-of-00009.safetensors",
|
| 422 |
+
"model.layers.9.linear_attn.conv1d.weight": "model-00009-of-00009.safetensors",
|
| 423 |
+
"model.layers.9.linear_attn.dt_bias": "model-00009-of-00009.safetensors",
|
| 424 |
+
"model.layers.9.linear_attn.in_proj_a.weight": "model-00009-of-00009.safetensors",
|
| 425 |
+
"model.layers.9.linear_attn.in_proj_b.weight": "model-00009-of-00009.safetensors",
|
| 426 |
+
"model.layers.9.linear_attn.in_proj_qkv.weight": "model-00009-of-00009.safetensors",
|
| 427 |
+
"model.layers.9.linear_attn.in_proj_z.weight": "model-00009-of-00009.safetensors",
|
| 428 |
+
"model.layers.9.linear_attn.norm.weight": "model-00009-of-00009.safetensors",
|
| 429 |
+
"model.layers.9.linear_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 430 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00009-of-00009.safetensors",
|
| 431 |
+
"model.norm.weight": "model-00009-of-00009.safetensors"
|
| 432 |
+
}
|
| 433 |
+
}
|
release-manifest.json
CHANGED
|
@@ -1,30 +1,40 @@
|
|
| 1 |
{
|
| 2 |
-
"
|
| 3 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
"validation": {
|
| 5 |
-
"onnx_golden_passed": true,
|
| 6 |
"cases": 36,
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
"README.md": "62a746e114cc3365b2b1f00c03bda5bee234f30186a2529eb2e1e1f09262e015",
|
| 12 |
-
"TRAINING_HISTORY.md": "c94f2d2570078049f66ac89ec420b7bef7fba6a614d3e8add54444dfde092cde",
|
| 13 |
-
"calibration.json": "6d9ba9d874699bf1f48f0703a814075a175ccb460d65c2c4870a3e6fb37fdb2f",
|
| 14 |
-
"chat_template.jinja": "87a2728cb8dc9fe424d624542f6060ec05a1d285ebbec578bb078900e33396b5",
|
| 15 |
-
"config.json": "332d56de2e5cde28245ac8b1776639bf9f7ed55fc911dfd2399ae8872acee2a9",
|
| 16 |
-
"export-provenance.json": "a3f2a53d03816fe3e57ebcbdf850b84f98b52c3a135ad5cd581d96276eacc0d2",
|
| 17 |
-
"export-validation.txt": "154c0212c4d4f64500ec13ebdffd7f405180a0ab3be6d049127144b4bb162df0",
|
| 18 |
-
"golden.json": "98e895994a3544548a09ad72f68e599af908628ac8093ecd2eed58a679b69319",
|
| 19 |
-
"model.safetensors": "5fb52e23dbff62e4c9982420ba4b07375b582685c4c1d22d1754f260e16e0097",
|
| 20 |
-
"model.safetensors.index.json": "90681e31419773af1b89d8597c1f4f5153f092fec6e55d09ac268e62462a008e",
|
| 21 |
-
"onnx/model_q8.onnx": "7ee7a5bdda762f735c6aa9257d63f758afef546afbfc1a701ba698fc89ef67f0",
|
| 22 |
-
"tokenizer.json": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506",
|
| 23 |
-
"tokenizer_config.json": "d93ac1a2c7adb9ad022f354d822f1f0f57e32e62ef2989e902782d6ab896acfe"
|
| 24 |
},
|
| 25 |
-
"
|
| 26 |
-
"revision": "ee7d33ecae2b8bee30891657dd57c92d34e0ce53",
|
| 27 |
-
"tag": "training-archive-20260924",
|
| 28 |
-
"removed_from_main": 187
|
| 29 |
-
}
|
| 30 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"version": "v6.0.0",
|
| 3 |
+
"name": "Jet",
|
| 4 |
+
"format": "complete merged BF16 Qwen3_5ForCausalLM",
|
| 5 |
+
"files": {
|
| 6 |
+
"LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
|
| 7 |
+
"README.md": "43bcdd39d003a210b5ee4d669ca772ad846b3bdf46f19c26e7ebf96077c77e76",
|
| 8 |
+
"calibration.json": "c7b13d6446dfa455bc18c11a9f877994728e670ddb4c3323f1bd16eb4cde2bc7",
|
| 9 |
+
"chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
|
| 10 |
+
"config.json": "63f47812d0f11118e4d252d2b3ad488707eb9287a11589f4fd382a1d31182724",
|
| 11 |
+
"evaluation.json": "c6a6dcc897dfd3d406db5ef88e6ec848f1138dc30a473c6ad94239c3a09acf0e",
|
| 12 |
+
"format.py": "f9b86dd757dbebc1fa48effbf48b7ba6861dcd6def5a025cfa937866eed02828",
|
| 13 |
+
"inference.py": "38443ddac02367f5857557ebae72aaf3e4d840f407573fd0dcb3ba037a5bc19d",
|
| 14 |
+
"jet.py": "a053495e98ac3c10358d041004b2a661bfb7b56ebc539196b919e46f10f0e454",
|
| 15 |
+
"merge-provenance.json": "9c168012766e8dd4e97f441fd44ec198ca22a09cea0009e2a677c66b5f29b753",
|
| 16 |
+
"merge-validation.json": "344a62368b294cafbc9b4e5b2da60e64f9aec8c21af7d0301c829d0b69c4dde5",
|
| 17 |
+
"model-00001-of-00009.safetensors": "758c65ea6c124bbe96b68a2d5475f7e4ec0ff9b6a4dbba87f295144ace55f07d",
|
| 18 |
+
"model-00002-of-00009.safetensors": "8e84de8d9a45219bbeb1a33392370979cce7702ceab4dea27b5f4b3b4e78deb0",
|
| 19 |
+
"model-00003-of-00009.safetensors": "d96476f63563f6db809675f551857e392102189c17eae2e2999bed0c4f967b22",
|
| 20 |
+
"model-00004-of-00009.safetensors": "bf5440f1558b2b9fc1e4ddec335cc77c04ac89621ce91a72ae3b3a542ad6eb25",
|
| 21 |
+
"model-00005-of-00009.safetensors": "52a8e3c8c06d2d16554e6146ffadcb0830367180d4ed2d4725084c0b4b257967",
|
| 22 |
+
"model-00006-of-00009.safetensors": "0d560b820ad7edc2faad717f6655b3f638e8779b40fda4e333d6457bcb9ed47c",
|
| 23 |
+
"model-00007-of-00009.safetensors": "fff39c3f2152571b72d377c6095daca3c796da187d94e88d18292c8709aa75d3",
|
| 24 |
+
"model-00008-of-00009.safetensors": "da8fa5e8e7bbcab3529e6f4014fdc3a42240a27c4b7fd123f3690abb35c96789",
|
| 25 |
+
"model-00009-of-00009.safetensors": "8d76d978760f6c15b9f5cb0c3b5eeada3513ff8a53104389cd895b5a5a092a94",
|
| 26 |
+
"model.safetensors.index.json": "583cb9f6b27293828c3e8c5055cc5a534846585063b057c5643592ecb8220ead",
|
| 27 |
+
"requirements.txt": "2888712d1e07f61a0748da102183fbff30c7de6f4152f239625d17c864bf9dde",
|
| 28 |
+
"runtime.py": "42ecb8f12ea72d2c050d533f9038c8a5fc48778a16fe5b5e3e9affb9cb161109",
|
| 29 |
+
"tokenizer.json": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
|
| 30 |
+
"tokenizer_config.json": "9cf04fffe3d8c3b85e439fb35c7acad0761ab51c422a8c4256d9f887c3a0be7d"
|
| 31 |
+
},
|
| 32 |
"validation": {
|
|
|
|
| 33 |
"cases": 36,
|
| 34 |
+
"selection": "First 12 cases per type, scanning local test then calibration; no outcome selection",
|
| 35 |
+
"argmax_flips": 1,
|
| 36 |
+
"max_probability_difference": 0.04372840585277227,
|
| 37 |
+
"base_download_required": false
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
},
|
| 39 |
+
"official_decision_index": null
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
}
|
requirements.txt
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch==2.11.0
|
| 2 |
+
transformers==5.17.0
|
| 3 |
+
accelerate==1.15.0
|
| 4 |
+
numpy==2.5.3
|
| 5 |
+
flash-linear-attention==0.5.2
|
runtime.py
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""CUDA BF16 label readout for the merged Jet model."""
|
| 2 |
+
import torch
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from transformers import AutoTokenizer, Qwen3_5ForCausalLM
|
| 5 |
+
|
| 6 |
+
def install_linear_attention():
|
| 7 |
+
"""Use FLA's differentiable chunk kernel; leave convolution on native PyTorch."""
|
| 8 |
+
from fla.ops.gated_delta_rule import chunk_gated_delta_rule
|
| 9 |
+
from transformers.models.qwen3_5 import modeling_qwen3_5 as impl
|
| 10 |
+
def chunk(query, key, value, g, beta, chunk_size=64, initial_state=None,
|
| 11 |
+
output_final_state=False, use_qk_l2norm_in_kernel=False, **kwargs):
|
| 12 |
+
return chunk_gated_delta_rule(q=query,k=key,v=value,g=g,beta=beta,
|
| 13 |
+
initial_state=initial_state,output_final_state=output_final_state,
|
| 14 |
+
use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel)
|
| 15 |
+
impl.torch_chunk_gated_delta_rule = chunk
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def load_model(path):
|
| 19 |
+
install_linear_attention()
|
| 20 |
+
tokenizer = AutoTokenizer.from_pretrained(path)
|
| 21 |
+
model, info = Qwen3_5ForCausalLM.from_pretrained(path, dtype=torch.bfloat16,
|
| 22 |
+
device_map={'': 'cuda'}, attn_implementation='sdpa', output_loading_info=True)
|
| 23 |
+
if info.get('missing_keys') or info.get('mismatched_keys') or info.get('unexpected_keys'):
|
| 24 |
+
raise RuntimeError(f'Incomplete model load: {info}')
|
| 25 |
+
model.config.use_cache = False
|
| 26 |
+
model.eval()
|
| 27 |
+
return model, tokenizer
|
| 28 |
+
|
| 29 |
+
def label_logits(model, example):
|
| 30 |
+
base=model.get_base_model() if hasattr(model,'get_base_model') else model
|
| 31 |
+
ids=torch.tensor([example['ids']],device='cuda',dtype=torch.long)
|
| 32 |
+
hidden=base.model(input_ids=ids,use_cache=False).last_hidden_state[:, -1, :]
|
| 33 |
+
labels=torch.tensor(example['labels'],device='cuda',dtype=torch.long)
|
| 34 |
+
# Only materialize the requested rows of the frozen vocabulary head.
|
| 35 |
+
return F.linear(hidden,base.lm_head.weight.index_select(0,labels))[0].float()
|
| 36 |
+
|
tokenizer.json
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
|
| 3 |
+
size 19989325
|
tokenizer_config.json
CHANGED
|
@@ -1,16 +1,32 @@
|
|
| 1 |
{
|
| 2 |
"add_prefix_space": false,
|
|
|
|
|
|
|
|
|
|
| 3 |
"backend": "tokenizers",
|
| 4 |
"bos_token": null,
|
| 5 |
"clean_up_tokenization_spaces": false,
|
| 6 |
"eos_token": "<|im_end|>",
|
| 7 |
"errors": "replace",
|
| 8 |
-
"
|
|
|
|
| 9 |
"local_files_only": false,
|
| 10 |
-
"model_max_length":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
"pad_token": "<|endoftext|>",
|
|
|
|
| 12 |
"split_special_tokens": false,
|
| 13 |
"tokenizer_class": "Qwen2Tokenizer",
|
| 14 |
-
"
|
| 15 |
-
"
|
|
|
|
|
|
|
| 16 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"add_prefix_space": false,
|
| 3 |
+
"audio_bos_token": "<|audio_start|>",
|
| 4 |
+
"audio_eos_token": "<|audio_end|>",
|
| 5 |
+
"audio_token": "<|audio_pad|>",
|
| 6 |
"backend": "tokenizers",
|
| 7 |
"bos_token": null,
|
| 8 |
"clean_up_tokenization_spaces": false,
|
| 9 |
"eos_token": "<|im_end|>",
|
| 10 |
"errors": "replace",
|
| 11 |
+
"image_token": "<|image_pad|>",
|
| 12 |
+
"is_local": false,
|
| 13 |
"local_files_only": false,
|
| 14 |
+
"model_max_length": 262144,
|
| 15 |
+
"model_specific_special_tokens": {
|
| 16 |
+
"audio_bos_token": "<|audio_start|>",
|
| 17 |
+
"audio_eos_token": "<|audio_end|>",
|
| 18 |
+
"audio_token": "<|audio_pad|>",
|
| 19 |
+
"image_token": "<|image_pad|>",
|
| 20 |
+
"video_token": "<|video_pad|>",
|
| 21 |
+
"vision_bos_token": "<|vision_start|>",
|
| 22 |
+
"vision_eos_token": "<|vision_end|>"
|
| 23 |
+
},
|
| 24 |
"pad_token": "<|endoftext|>",
|
| 25 |
+
"pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
|
| 26 |
"split_special_tokens": false,
|
| 27 |
"tokenizer_class": "Qwen2Tokenizer",
|
| 28 |
+
"unk_token": null,
|
| 29 |
+
"video_token": "<|video_pad|>",
|
| 30 |
+
"vision_bos_token": "<|vision_start|>",
|
| 31 |
+
"vision_eos_token": "<|vision_end|>"
|
| 32 |
}
|