nardi commited on
Commit
8f94540
·
0 Parent(s):
.gitattributes ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ LFM Open License v1.0
2
+
3
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
4
+
5
+ 1. Definitions.
6
+
7
+ "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by this document.
8
+
9
+ "Licensor" shall mean Liquid AI, Inc.
10
+
11
+ "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
12
+
13
+ "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
14
+
15
+ "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
16
+
17
+ "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
18
+
19
+ "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work.
20
+
21
+ "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
22
+
23
+ "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
24
+
25
+ "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
26
+
27
+ "Commercial Use" shall mean any use of the Work for direct or indirect commercial advantage or monetary compensation.
28
+
29
+ "Qualified Non-Profit Organization" shall mean a Legal Entity that is organized and operated exclusively for religious, charitable, scientific, testing for public safety, literary, or educational purposes, and which is exempt from federal income tax under Section 501(c)(3) of the United States Internal Revenue Code of 1986, as amended, or any equivalent non-profit or charitable organization in a foreign jurisdiction.
30
+
31
+ "Non-Commercial or Research Purposes" shall mean purposes that do not involve any use of the Work or a Derivative Work for Commercial Use.
32
+
33
+ "Threshold" shall mean annual revenue of 10 million United States dollars ($10,000,000) or more.
34
+
35
+ 2. Grant of Copyright License. Subject to the terms and conditions of this License, including the Commercial Use limitation set forth in Section 5, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
36
+
37
+ 3. Grant of Patent License. Subject to the terms and conditions of this License, including the Commercial Use limitation set forth in Section 5, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
38
+
39
+ 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
40
+
41
+ (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and
42
+
43
+ (b) You must cause any modified files to carry prominent notices stating that You changed the files; and
44
+
45
+ (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
46
+
47
+ (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
48
+
49
+ You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
50
+
51
+ 5. Commercial Use Limitation.
52
+
53
+ (a) The rights granted under this License for Commercial Use are conditioned upon You or Your Legal Entity not exceeding the Threshold.
54
+
55
+ (b) Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement.
56
+
57
+ (c) The Threshold shall not apply to a Qualified Non-Profit Organization's use of the Work or a Derivative Work for Non-Commercial or Research Purposes.
58
+
59
+ 6. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
60
+
61
+ 7. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except for the reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
62
+
63
+ 8. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
64
+
65
+ 9. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
66
+
67
+ 10. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
68
+
69
+ 11. Termination. This License will terminate automatically and immediately if You fail to comply with any of its terms and conditions. Upon termination, You must cease all use of the Work and any Derivative Works and delete all copies in Your possession.
70
+
71
+ END OF TERMS AND CONDITIONS
LICENSE-Apache-2.0.txt ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright 2018- The Hugging Face team. All rights reserved.
2
+
3
+ Apache License
4
+ Version 2.0, January 2004
5
+ http://www.apache.org/licenses/
6
+
7
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
8
+
9
+ 1. Definitions.
10
+
11
+ "License" shall mean the terms and conditions for use, reproduction,
12
+ and distribution as defined by Sections 1 through 9 of this document.
13
+
14
+ "Licensor" shall mean the copyright owner or entity authorized by
15
+ the copyright owner that is granting the License.
16
+
17
+ "Legal Entity" shall mean the union of the acting entity and all
18
+ other entities that control, are controlled by, or are under common
19
+ control with that entity. For the purposes of this definition,
20
+ "control" means (i) the power, direct or indirect, to cause the
21
+ direction or management of such entity, whether by contract or
22
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
23
+ outstanding shares, or (iii) beneficial ownership of such entity.
24
+
25
+ "You" (or "Your") shall mean an individual or Legal Entity
26
+ exercising permissions granted by this License.
27
+
28
+ "Source" form shall mean the preferred form for making modifications,
29
+ including but not limited to software source code, documentation
30
+ source, and configuration files.
31
+
32
+ "Object" form shall mean any form resulting from mechanical
33
+ transformation or translation of a Source form, including but
34
+ not limited to compiled object code, generated documentation,
35
+ and conversions to other media types.
36
+
37
+ "Work" shall mean the work of authorship, whether in Source or
38
+ Object form, made available under the License, as indicated by a
39
+ copyright notice that is included in or attached to the work
40
+ (an example is provided in the Appendix below).
41
+
42
+ "Derivative Works" shall mean any work, whether in Source or Object
43
+ form, that is based on (or derived from) the Work and for which the
44
+ editorial revisions, annotations, elaborations, or other modifications
45
+ represent, as a whole, an original work of authorship. For the purposes
46
+ of this License, Derivative Works shall not include works that remain
47
+ separable from, or merely link (or bind by name) to the interfaces of,
48
+ the Work and Derivative Works thereof.
49
+
50
+ "Contribution" shall mean any work of authorship, including
51
+ the original version of the Work and any modifications or additions
52
+ to that Work or Derivative Works thereof, that is intentionally
53
+ submitted to Licensor for inclusion in the Work by the copyright owner
54
+ or by an individual or Legal Entity authorized to submit on behalf of
55
+ the copyright owner. For the purposes of this definition, "submitted"
56
+ means any form of electronic, verbal, or written communication sent
57
+ to the Licensor or its representatives, including but not limited to
58
+ communication on electronic mailing lists, source code control systems,
59
+ and issue tracking systems that are managed by, or on behalf of, the
60
+ Licensor for the purpose of discussing and improving the Work, but
61
+ excluding communication that is conspicuously marked or otherwise
62
+ designated in writing by the copyright owner as "Not a Contribution."
63
+
64
+ "Contributor" shall mean Licensor and any individual or Legal Entity
65
+ on behalf of whom a Contribution has been received by Licensor and
66
+ subsequently incorporated within the Work.
67
+
68
+ 2. Grant of Copyright License. Subject to the terms and conditions of
69
+ this License, each Contributor hereby grants to You a perpetual,
70
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
71
+ copyright license to reproduce, prepare Derivative Works of,
72
+ publicly display, publicly perform, sublicense, and distribute the
73
+ Work and such Derivative Works in Source or Object form.
74
+
75
+ 3. Grant of Patent License. Subject to the terms and conditions of
76
+ this License, each Contributor hereby grants to You a perpetual,
77
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
78
+ (except as stated in this section) patent license to make, have made,
79
+ use, offer to sell, sell, import, and otherwise transfer the Work,
80
+ where such license applies only to those patent claims licensable
81
+ by such Contributor that are necessarily infringed by their
82
+ Contribution(s) alone or by combination of their Contribution(s)
83
+ with the Work to which such Contribution(s) was submitted. If You
84
+ institute patent litigation against any entity (including a
85
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
86
+ or a Contribution incorporated within the Work constitutes direct
87
+ or contributory patent infringement, then any patent licenses
88
+ granted to You under this License for that Work shall terminate
89
+ as of the date such litigation is filed.
90
+
91
+ 4. Redistribution. You may reproduce and distribute copies of the
92
+ Work or Derivative Works thereof in any medium, with or without
93
+ modifications, and in Source or Object form, provided that You
94
+ meet the following conditions:
95
+
96
+ (a) You must give any other recipients of the Work or
97
+ Derivative Works a copy of this License; and
98
+
99
+ (b) You must cause any modified files to carry prominent notices
100
+ stating that You changed the files; and
101
+
102
+ (c) You must retain, in the Source form of any Derivative Works
103
+ that You distribute, all copyright, patent, trademark, and
104
+ attribution notices from the Source form of the Work,
105
+ excluding those notices that do not pertain to any part of
106
+ the Derivative Works; and
107
+
108
+ (d) If the Work includes a "NOTICE" text file as part of its
109
+ distribution, then any Derivative Works that You distribute must
110
+ include a readable copy of the attribution notices contained
111
+ within such NOTICE file, excluding those notices that do not
112
+ pertain to any part of the Derivative Works, in at least one
113
+ of the following places: within a NOTICE text file distributed
114
+ as part of the Derivative Works; within the Source form or
115
+ documentation, if provided along with the Derivative Works; or,
116
+ within a display generated by the Derivative Works, if and
117
+ wherever such third-party notices normally appear. The contents
118
+ of the NOTICE file are for informational purposes only and
119
+ do not modify the License. You may add Your own attribution
120
+ notices within Derivative Works that You distribute, alongside
121
+ or as an addendum to the NOTICE text from the Work, provided
122
+ that such additional attribution notices cannot be construed
123
+ as modifying the License.
124
+
125
+ You may add Your own copyright statement to Your modifications and
126
+ may provide additional or different license terms and conditions
127
+ for use, reproduction, or distribution of Your modifications, or
128
+ for any such Derivative Works as a whole, provided Your use,
129
+ reproduction, and distribution of the Work otherwise complies with
130
+ the conditions stated in this License.
131
+
132
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
133
+ any Contribution intentionally submitted for inclusion in the Work
134
+ by You to the Licensor shall be under the terms and conditions of
135
+ this License, without any additional terms or conditions.
136
+ Notwithstanding the above, nothing herein shall supersede or modify
137
+ the terms of any separate license agreement you may have executed
138
+ with Licensor regarding such Contributions.
139
+
140
+ 6. Trademarks. This License does not grant permission to use the trade
141
+ names, trademarks, service marks, or product names of the Licensor,
142
+ except as required for reasonable and customary use in describing the
143
+ origin of the Work and reproducing the content of the NOTICE file.
144
+
145
+ 7. Disclaimer of Warranty. Unless required by applicable law or
146
+ agreed to in writing, Licensor provides the Work (and each
147
+ Contributor provides its Contributions) on an "AS IS" BASIS,
148
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
149
+ implied, including, without limitation, any warranties or conditions
150
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
151
+ PARTICULAR PURPOSE. You are solely responsible for determining the
152
+ appropriateness of using or redistributing the Work and assume any
153
+ risks associated with Your exercise of permissions under this License.
154
+
155
+ 8. Limitation of Liability. In no event and under no legal theory,
156
+ whether in tort (including negligence), contract, or otherwise,
157
+ unless required by applicable law (such as deliberate and grossly
158
+ negligent acts) or agreed to in writing, shall any Contributor be
159
+ liable to You for damages, including any direct, indirect, special,
160
+ incidental, or consequential damages of any character arising as a
161
+ result of this License or out of the use or inability to use the
162
+ Work (including but not limited to damages for loss of goodwill,
163
+ work stoppage, computer failure or malfunction, or any and all
164
+ other commercial damages or losses), even if such Contributor
165
+ has been advised of the possibility of such damages.
166
+
167
+ 9. Accepting Warranty or Additional Liability. While redistributing
168
+ the Work or Derivative Works thereof, You may choose to offer,
169
+ and charge a fee for, acceptance of support, warranty, indemnity,
170
+ or other liability obligations and/or rights consistent with this
171
+ License. However, in accepting such obligations, You may act only
172
+ on Your own behalf and on Your sole responsibility, not on behalf
173
+ of any other Contributor, and only if You agree to indemnify,
174
+ defend, and hold each Contributor harmless for any liability
175
+ incurred by, or claims asserted against, such Contributor by reason
176
+ of your accepting any such warranty or additional liability.
177
+
178
+ END OF TERMS AND CONDITIONS
179
+
180
+ APPENDIX: How to apply the Apache License to your work.
181
+
182
+ To apply the Apache License to your work, attach the following
183
+ boilerplate notice, with the fields enclosed by brackets "[]"
184
+ replaced with your own identifying information. (Don't include
185
+ the brackets!) The text should be enclosed in the appropriate
186
+ comment syntax for the file format. We also recommend that a
187
+ file or class name and description of purpose be included on the
188
+ same "printed page" as the copyright notice for easier
189
+ identification within third-party archives.
190
+
191
+ Copyright [yyyy] [name of copyright owner]
192
+
193
+ Licensed under the Apache License, Version 2.0 (the "License");
194
+ you may not use this file except in compliance with the License.
195
+ You may obtain a copy of the License at
196
+
197
+ http://www.apache.org/licenses/LICENSE-2.0
198
+
199
+ Unless required by applicable law or agreed to in writing, software
200
+ distributed under the License is distributed on an "AS IS" BASIS,
201
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
202
+ See the License for the specific language governing permissions and
203
+ limitations under the License.
NOTICE ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ LFM3-TS
2
+ Copyright 2026 Nardi Xhepi
3
+
4
+ LFM3-TS is distributed under the LFM Open License v1.0. A copy of the licence is in the file LICENSE.
5
+
6
+ This is an independent project. It is not affiliated with, sponsored by or endorsed by Liquid AI, Inc. The names
7
+ "Liquid AI", "LFM" and "LFM3" belong to Liquid AI, Inc. and are used here only to describe the origin of the work.
8
+
9
+
10
+ 1. LFM3-1.7B (Liquid AI, Inc.)
11
+
12
+ The LFM3-TS model weights are a Derivative Work of LFM3-1.7B by Liquid AI, Inc.
13
+ (https://huggingface.co/LiquidAI/LFM3-1.7B), which is licensed under the LFM Open License v1.0.
14
+
15
+ NOTICE OF CHANGES. The weights of LFM3-1.7B were modified by Nardi Xhepi:
16
+
17
+ - the weights of the language model were fine-tuned on time-series question answering, except the token
18
+ embedding matrix (shared with the output layer), which is unchanged;
19
+ - a time-series encoder (`model.ts_tower`) and a projector (`model.multi_modal_projector`) were added; these
20
+ weights are new and are not part of LFM3-1.7B;
21
+ - the tokenizer gained the name `<|ts|>` for a previously unused token id (125033), and the chat template gained
22
+ a `timeseries` content part.
23
+
24
+ The distribution of LFM3-1.7B this work started from contains no NOTICE file.
25
+
26
+
27
+ 2. Hugging Face Transformers
28
+
29
+ The files
30
+
31
+ src/lfm3_ts/configuration_lfm3_ts.py
32
+ src/lfm3_ts/modeling_lfm3_ts.py
33
+ src/lfm3_ts/processing_lfm3_ts.py
34
+
35
+ (shipped with the model checkpoint as configuration_lfm3_ts.py, modeling_lfm3_ts.py and processing_lfm3_ts.py) are
36
+ adapted from `transformers.models.lfm3_vl` of Hugging Face Transformers
37
+ (https://github.com/huggingface/transformers), Copyright 2026 The HuggingFace Team, licensed under the Apache
38
+ License, Version 2.0. They were changed by Nardi Xhepi: the vision tower and the image inputs are replaced by a
39
+ time-series encoder and time-series inputs. A copy of the Apache License, Version 2.0 is in
40
+ LICENSES/Apache-2.0.txt.
41
+
42
+
43
+ 3. Method and data
44
+
45
+ The method follows ChatTS (Xie et al., "ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced
46
+ Understanding and Reasoning", arXiv:2412.03104). No ChatTS code or weights are included.
47
+
48
+ Training used the ChatTS synthetic training data (https://huggingface.co/datasets/ChatTSRepo/ChatTS-Training-Dataset),
49
+ windows of real series from TSB-AD-U (https://github.com/TheDatumOrg/TSB-AD) and prompts from databricks-dolly-15k
50
+ (https://huggingface.co/datasets/databricks/databricks-dolly-15k). None of these datasets is redistributed here.
51
+
52
+ The real series were gathered by TSB-AD from their original sources. Those used in training, with their licences as
53
+ stated by each source: the MIT-BIH Supraventricular Arrhythmia, Long-Term ECG and Arrhythmia databases of PhysioNet
54
+ (Open Data Commons Attribution License v1.0); OPPORTUNITY and Daphnet of the UCI Machine Learning Repository and
55
+ CATS (CC BY 4.0); the Server Machine Dataset of OmniAnomaly (MIT); Exathlon (CC BY-NC-SA 4.0); WSD, SED and Power
56
+ (no licence stated).
README.md ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: lfm1.0
4
+ license_link: LICENSE
5
+ base_model: LiquidAI/LFM3-1.7B
6
+ base_model_relation: finetune
7
+ library_name: transformers
8
+ pipeline_tag: text-generation
9
+ language:
10
+ - en
11
+ tags:
12
+ - time-series
13
+ - anomaly-detection
14
+ - time-series-question-answering
15
+ - multimodal
16
+ - lfm3
17
+ - liquid
18
+ - chatts
19
+ - custom_code
20
+ datasets:
21
+ - ChatTSRepo/ChatTS-Training-Dataset
22
+ - databricks/databricks-dolly-15k
23
+ ---
24
+
25
+ # LFM3-TS
26
+
27
+ A small language model that reads time series. You give it a series and a question, and it answers in text: is
28
+ there an anomaly, where, and what does it look like.
29
+
30
+ LFM3-TS is [LFM3-1.7B](https://huggingface.co/LiquidAI/LFM3-1.7B) from Liquid AI with a time-series encoder in front,
31
+ built on the ideas and the synthetic data of [ChatTS](https://arxiv.org/abs/2412.03104). It has 1.8 billion
32
+ parameters.
33
+
34
+ - **Code and training recipe:** https://github.com/nardi-xhepi/lfm3-ts
35
+ - **Blog post:** https://nardi-xhepi.github.io/blog/posts/lfm3-ts.html
36
+ - **Author:** Nardi Xhepi
37
+
38
+ This is an independent project, not affiliated with Liquid AI.
39
+
40
+ ## How to use
41
+
42
+ ```bash
43
+ pip install -U "transformers>=5.23"
44
+ ```
45
+
46
+ You also need PyTorch and NumPy. LFM3 was added to `transformers` in version 5.23; on Colab, which already has
47
+ PyTorch, run only the line above and restart the runtime.
48
+
49
+ ```python
50
+ import numpy as np
51
+ from transformers import AutoModelForCausalLM, AutoProcessor
52
+
53
+ model = AutoModelForCausalLM.from_pretrained("nardi/LFM3-TS", trust_remote_code=True)
54
+ processor = AutoProcessor.from_pretrained("nardi/LFM3-TS", trust_remote_code=True)
55
+
56
+ series = np.sin(np.arange(512) / 8)
57
+ series[300] += 4 # a spike
58
+
59
+ print(model.chat(processor, series, "Is there an anomaly in this time series? If so, where is it and what does it look like?"))
60
+ ```
61
+
62
+ `series` is a 1D array or list (NaN marks a missing value). Points are numbered from 0. The model was trained on
63
+ series of up to 2,048 points, in English.
64
+
65
+ ## Model
66
+
67
+ | | |
68
+ |---|---|
69
+ | Language model | LFM3-1.7B, every weight fine-tuned except the token embeddings |
70
+ | Time-series encoder | patches of 8 points, 4 feed-forward and 2 attention layers, width 1024 |
71
+ | Projector | MLP to the width of the language model (2048) |
72
+ | Tokens per series | one per patch, plus 2: a series of 512 points costs 66 tokens |
73
+ | Parameters | 1.8 billion |
74
+
75
+ The series is standardized for the encoder, and its offset and scale are written in the prompt as text, so the
76
+ answers can give real values.
77
+
78
+ ## Results
79
+
80
+ **The ChatTS benchmark** ([sets A and B](https://zenodo.org/records/14349206)), scored with the scoring code of the
81
+ ChatTS repository:
82
+
83
+ | | Parameters | Set A, categorical | Set A, numerical | Set B, categorical | Set B, numerical |
84
+ |---|---|---|---|---|---|
85
+ | **LFM3-TS** | 1.8B | 82.0% | 77.5% | 79.5% | 76.8% |
86
+ | ChatTS-8B | 8B | 83.8% | 82.3% | 84.6% | 82.9% |
87
+
88
+ **Anomalies in real series**: 234 windows of 512 points from seven sources of
89
+ [TSB-AD-U](https://github.com/TheDatumOrg/TSB-AD) that neither model was trained on (105 with a labelled anomaly, 129
90
+ without). An answer is correct when it says "yes" and names a point within 16 points of the labelled stretch.
91
+
92
+ | | Parameters | Precision | Recall | F1 | False alarms |
93
+ |---|---|---|---|---|---|
94
+ | **LFM3-TS** | 1.8B | 66% | 64% | 0.65 | 28 of 129 |
95
+ | ChatTS-8B | 8B | 40% | 53% | 0.46 | 61 of 129 |
96
+
97
+ ## Training
98
+
99
+ Three stages, each starting from the weights of the one before: generated questions that can only be answered by
100
+ reading the series; the same questions on series of up to 2,048 points; then a mixture of ChatTS synthetic data
101
+ (48%), generated questions (19%), plain chat (15%), real series with labelled anomalies (10%) and real series with
102
+ an event put in (8%). The commands are in the code repository.
103
+
104
+ ## Licence
105
+
106
+ LFM3-TS is released under the [LFM Open License v1.0](LICENSE), the licence of LFM3-1.7B, which the weights derive
107
+ from. See `NOTICE` for what was changed and for the sources and licences of the training data.
108
+
109
+ ## Credits
110
+
111
+ [LFM3-1.7B](https://huggingface.co/LiquidAI/LFM3-1.7B) (Liquid AI) ·
112
+ [ChatTS](https://github.com/NetManAIOps/ChatTS) (Xie et al.) ·
113
+ [TSB-AD](https://github.com/TheDatumOrg/TSB-AD) (Liu and Paparrizos) ·
114
+ [databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) ·
115
+ [Hugging Face Transformers](https://github.com/huggingface/transformers)
chat_template.jinja ADDED
@@ -0,0 +1,284 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{- bos_token -}}
2
+ {%- set preserve_thinking = preserve_thinking | default(false) -%}
3
+ {%- set enable_thinking = enable_thinking | default(true) -%}
4
+
5
+ {%- macro render_value(v) -%}
6
+ {%- if v is string -%}
7
+ {{- v -}}
8
+ {%- elif v is boolean -%}
9
+ {{- "true" if v else "false" -}}
10
+ {%- elif v is none -%}
11
+ {{- "null" -}}
12
+ {%- elif v is mapping or (v is sequence and v is not string) -%}
13
+ {{- v | tojson -}}
14
+ {%- else -%}
15
+ {{- v -}}
16
+ {%- endif -%}
17
+ {%- endmacro -%}
18
+
19
+ {%- macro parse_content(content) -%}
20
+ {%- if content is string -%}
21
+ {{- content -}}
22
+ {%- elif content is mapping and content.get("type") == "image" -%}
23
+ {{- "<|vision_start|><|image|><|vision_end|>" -}}
24
+ {%- elif content is mapping and content.get("type") == "video" -%}
25
+ {{- "<|vision_start|><|video|><|vision_end|>" -}}
26
+ {%- elif content is mapping and content.get("type") == "timeseries" -%}
27
+ {{- "<|ts|>" -}}
28
+ {%- elif content is mapping and content.get("type") == "text" -%}
29
+ {{- (content.get("text") or "") | string -}}
30
+ {%- elif content is mapping -%}
31
+ {{- content | tojson -}}
32
+ {%- elif content is iterable -%}
33
+ {%- set _ns = namespace(result="") -%}
34
+ {%- for item in content -%}
35
+ {%- if item is string -%}
36
+ {%- set _ns.result = _ns.result + item -%}
37
+ {%- elif item is mapping and item.get("type") == "image" -%}
38
+ {%- set _ns.result = _ns.result + "<|vision_start|><|image|><|vision_end|>" -%}
39
+ {%- elif item is mapping and item.get("type") == "video" -%}
40
+ {%- set _ns.result = _ns.result + "<|vision_start|><|video|><|vision_end|>" -%}
41
+ {%- elif item is mapping and item.get("type") == "timeseries" -%}
42
+ {%- set _ns.result = _ns.result + "<|ts|>" -%}
43
+ {%- elif item is mapping and item.get("type") == "text" -%}
44
+ {%- set _ns.result = _ns.result + ((item.get("text") or "") | string) -%}
45
+ {%- else -%}
46
+ {%- set _ns.result = _ns.result + (item | tojson) -%}
47
+ {%- endif -%}
48
+ {%- endfor -%}
49
+ {{- _ns.result -}}
50
+ {%- elif content is not none -%}
51
+ {{- content | tojson -}}
52
+ {%- endif -%}
53
+ {%- endmacro -%}
54
+
55
+ {%- macro render_tool_calls(tool_calls) -%}
56
+ {%- for tool_call in tool_calls -%}
57
+ {%- if not loop.first -%}{{- "\n\n" -}}{%- endif -%}
58
+ {%- set func = tool_call["function"] if "function" in tool_call else tool_call -%}
59
+ {%- set func_args = func.get("arguments") -%}
60
+ {%- if func_args is string and (func_args | trim) not in ["", "{}", "null"] -%}
61
+ {{- raise_exception("Tool call arguments must be a mapping, got a JSON-encoded string: parse arguments with json.loads() before applying the chat template") -}}
62
+ {%- endif -%}
63
+ {{- "<tool_call>\n<function=" + func["name"] + ">\n" -}}
64
+ {%- if func_args is mapping -%}
65
+ {%- for arg_name, arg_value in func_args.items() -%}
66
+ {{- "<parameter=" + arg_name + ">\n" + render_value(arg_value) + "\n</parameter>\n" -}}
67
+ {%- endfor -%}
68
+ {%- endif -%}
69
+ {{- "</function>\n</tool_call>" -}}
70
+ {%- endfor -%}
71
+ {%- endmacro -%}
72
+
73
+ {%- macro tool_to_json(tool) -%}
74
+ {%- set fn = tool.function if tool.function is defined else tool -%}
75
+ {%- set ns_t = namespace(first=true) -%}
76
+ {{- "{" -}}
77
+ {%- for k, v in fn.items() -%}
78
+ {%- if k != "defer_loading" -%}
79
+ {%- if not ns_t.first -%}{{- ", " -}}{%- endif -%}
80
+ {%- set ns_t.first = false -%}
81
+ {{- k | tojson -}}{{- ": " -}}{{- v | tojson -}}
82
+ {%- endif -%}
83
+ {%- endfor -%}
84
+ {{- "}" -}}
85
+ {%- endmacro -%}
86
+
87
+ {%- macro is_tool_reference(item) -%}
88
+ {{- "1" if item is mapping and item.get("type") == "tool_reference" else "" -}}
89
+ {%- endmacro -%}
90
+
91
+ {%- macro referenced_tool_to_json(name) -%}
92
+ {%- set _ns = namespace(found=none) -%}
93
+ {%- for tool in tools or [] -%}
94
+ {%- set fn = tool.function if tool.function is defined else tool -%}
95
+ {%- if fn.name == name -%}
96
+ {%- set _ns.found = fn -%}
97
+ {%- endif -%}
98
+ {%- endfor -%}
99
+ {%- if _ns.found is none -%}
100
+ {{- raise_exception("Tool reference '" + name | string + "' not found in available tools. Every tool referenced by a tool_reference must be in `tools` on every request.") -}}
101
+ {%- endif -%}
102
+ {{- tool_to_json(_ns.found) -}}
103
+ {%- endmacro -%}
104
+
105
+ {%- macro render_tool_response(content) -%}
106
+ {%- if content is string and "<tool_response>" in content -%}
107
+ {{- content -}}
108
+ {%- else -%}
109
+ {{- "<tool_response>\n" -}}
110
+ {%- if content is string -%}
111
+ {{- content -}}
112
+ {%- elif content is mapping -%}
113
+ {{- parse_content(content) -}}
114
+ {%- elif content is iterable -%}
115
+ {%- for item in content -%}
116
+ {%- if not loop.first -%}{{- "\n" -}}{%- endif -%}
117
+ {%- if is_tool_reference(item) -%}
118
+ {%- if loop.first or not is_tool_reference(loop.previtem) -%}
119
+ {{- "<additional_tools>\n" -}}
120
+ {%- endif -%}
121
+ {{- referenced_tool_to_json(item.name) -}}
122
+ {%- if loop.last or not is_tool_reference(loop.nextitem) -%}
123
+ {{- "\n</additional_tools>" -}}
124
+ {%- endif -%}
125
+ {%- elif item is mapping and item.result is defined and item.result is mapping and item.result.tools is defined -%}
126
+ {{- "<additional_tools>\n" -}}
127
+ {%- set _tl = item.result.tools -%}
128
+ {%- if _tl is string -%}
129
+ {{- _tl -}}
130
+ {%- elif _tl is mapping -%}
131
+ {{- tool_to_json(_tl) -}}
132
+ {%- else -%}
133
+ {%- for t in _tl -%}
134
+ {%- if not loop.first -%}{{- "\n" -}}{%- endif -%}
135
+ {{- tool_to_json(t) -}}
136
+ {%- endfor -%}
137
+ {%- endif -%}
138
+ {{- "\n</additional_tools>" -}}
139
+ {%- else -%}
140
+ {{- parse_content(item) -}}
141
+ {%- endif -%}
142
+ {%- endfor -%}
143
+ {%- elif content is not none -%}
144
+ {{- content | tojson -}}
145
+ {%- endif -%}
146
+ {{- "\n</tool_response>" -}}
147
+ {%- endif -%}
148
+ {%- endmacro -%}
149
+
150
+ {%- set ns = namespace(system_prompt="", last_user_index=-1) -%}
151
+ {%- if messages and messages[0]["role"] == "system" -%}
152
+ {%- if messages[0].get("content") -%}
153
+ {%- set ns.system_prompt = parse_content(messages[0]["content"]) -%}
154
+ {%- endif -%}
155
+ {%- set messages = messages[1:] -%}
156
+ {%- endif -%}
157
+ {%- if tools -%}
158
+ {%- set tools_ns = namespace(lines=[]) -%}
159
+ {%- for tool in tools -%}
160
+ {%- set fn = tool.function if tool.function is defined else tool -%}
161
+ {%- set _dyn = (fn.defer_loading is defined and fn.defer_loading) or (tool.defer_loading is defined and tool.defer_loading) -%}
162
+ {%- if not _dyn -%}
163
+ {%- set tools_ns.lines = tools_ns.lines + [tool_to_json(fn)] -%}
164
+ {%- endif -%}
165
+ {%- endfor -%}
166
+ {%- if tools_ns.lines -%}
167
+ {%- set ns.system_prompt = ns.system_prompt + ("\n\n" if ns.system_prompt else "") + "<tools>\n" + (tools_ns.lines | join("\n")) + "\n</tools>" -%}
168
+ {%- endif -%}
169
+ {%- endif -%}
170
+ {%- if ns.system_prompt -%}
171
+ {{- "<|im_start|>system\n" + ns.system_prompt + "<|im_end|>\n" -}}
172
+ {%- endif -%}
173
+ {%- for message in messages -%}
174
+ {%- if message["role"] == "user" -%}
175
+ {%- set ns.last_user_index = loop.index0 -%}
176
+ {%- endif -%}
177
+ {%- endfor -%}
178
+ {%- set carry = namespace(pending=[]) -%}
179
+ {%- for message in messages -%}
180
+ {%- set _next = messages[loop.index0 + 1] if not loop.last else none -%}
181
+ {%- set _merge = message.role == "tool" and not message.get("additional_tools")
182
+ and _next is not none and _next.role == "tool" and not _next.get("additional_tools")
183
+ and message.get("tool_call_id") and _next.get("tool_call_id") == message.get("tool_call_id") -%}
184
+ {%- set _items = [] -%}
185
+ {%- if message.role == "tool" and not message.get("additional_tools") -%}
186
+ {%- set _c = message.get("content") -%}
187
+ {%- if _c is string -%}
188
+ {%- set _items = [{"type": "text", "text": _c}] if _c else [] -%}
189
+ {%- elif _c is mapping -%}
190
+ {%- set _items = [_c] -%}
191
+ {%- elif _c is iterable -%}
192
+ {%- set _items = _c | list -%}
193
+ {%- elif _c is not none -%}
194
+ {%- set _items = [_c] -%}
195
+ {%- endif -%}
196
+ {%- endif -%}
197
+ {%- if _merge -%}
198
+ {%- set carry.pending = carry.pending + _items -%}
199
+ {%- else -%}
200
+ {{- "<|im_start|>" + message.role + "\n" -}}
201
+ {%- if message.role == "assistant" -%}
202
+ {%- generation -%}
203
+ {%- set thinking = message.thinking if message.thinking is defined else message.reasoning if message.reasoning is defined else message.reasoning_content -%}
204
+ {%- set keep_thinking = preserve_thinking or loop.index0 > ns.last_user_index -%}
205
+ {%- if thinking is defined and keep_thinking -%}
206
+ {{- "<think>" + (thinking or "") + "</think>" -}}
207
+ {%- elif keep_thinking and "</think>" not in (message.get("content") if message.get("content") is string else "") -%}
208
+ {{- "<think></think>" -}}
209
+ {%- endif -%}
210
+ {%- set _cfm_tag = "CONTINUE_FINAL_MESSAGE_TAG " -%}
211
+ {%- set _has_cfm = false -%}
212
+ {%- set content = "" -%}
213
+ {%- if message.get("content") -%}
214
+ {%- set content = parse_content(message.content) -%}
215
+ {%- endif -%}
216
+ {%- if not keep_thinking -%}
217
+ {%- if "</think>" in content -%}
218
+ {%- set content = content.split("</think>")[-1] | trim -%}
219
+ {%- endif -%}
220
+ {%- set content = "<think></think>" + content -%}
221
+ {%- endif -%}
222
+ {%- if content.endswith(_cfm_tag) -%}
223
+ {%- set _has_cfm = true -%}
224
+ {%- set _trunc_len = (content | length) - (_cfm_tag | length) -%}
225
+ {%- set content = content[:_trunc_len] -%}
226
+ {%- endif -%}
227
+ {{- content -}}
228
+ {%- if message.tool_calls -%}
229
+ {{- render_tool_calls(message.tool_calls) -}}
230
+ {%- endif -%}
231
+ {%- if _has_cfm -%}
232
+ {{- _cfm_tag -}}
233
+ {%- endif -%}
234
+ {{- "<|im_end|>\n" -}}
235
+ {%- endgeneration -%}
236
+ {%- elif message.role == "tool" -%}
237
+ {%- if message.get("additional_tools") -%}
238
+ {%- set _at = message["additional_tools"] -%}
239
+ {%- if _at is mapping or (_at is sequence and _at is not string) -%}
240
+ {%- set _src = _at -%}
241
+ {%- else -%}
242
+ {%- set _src = message.get("content") -%}
243
+ {%- endif -%}
244
+ {%- if _src is string and "<tool_response>" in _src -%}
245
+ {{- _src -}}
246
+ {%- elif _src is string and "<additional_tools>" in _src -%}
247
+ {{- "<tool_response>\n" + _src + "\n</tool_response>" -}}
248
+ {%- elif _src -%}
249
+ {{- "<tool_response>\n<additional_tools>\n" -}}
250
+ {%- if _src is string -%}
251
+ {{- _src -}}
252
+ {%- elif _src is mapping -%}
253
+ {{- tool_to_json(_src) -}}
254
+ {%- else -%}
255
+ {%- for t in _src -%}
256
+ {%- if not loop.first -%}{{- "\n" -}}{%- endif -%}
257
+ {{- tool_to_json(t) -}}
258
+ {%- endfor -%}
259
+ {%- endif -%}
260
+ {{- "\n</additional_tools>\n</tool_response>" -}}
261
+ {%- endif -%}
262
+ {%- elif carry.pending -%}
263
+ {{- render_tool_response(carry.pending + _items) -}}
264
+ {%- set carry.pending = [] -%}
265
+ {%- else -%}
266
+ {{- render_tool_response("" if message.get("content") is none else message.get("content")) -}}
267
+ {%- endif -%}
268
+ {{- "<|im_end|>\n" -}}
269
+ {%- else %}
270
+ {%- if message.get("content") -%}
271
+ {{- parse_content(message["content"]) -}}
272
+ {%- endif -%}
273
+ {{- "<|im_end|>\n" -}}
274
+ {%- endif %}
275
+ {%- endif -%}
276
+ {%- endfor -%}
277
+ {%- if add_generation_prompt -%}
278
+ {{- "<|im_start|>assistant\n" -}}
279
+ {%- if enable_thinking -%}
280
+ {{- "<think>" -}}
281
+ {%- else -%}
282
+ {{- "<think></think>" -}}
283
+ {%- endif -%}
284
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Lfm3TsForConditionalGeneration"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_lfm3_ts.Lfm3TsConfig",
7
+ "AutoModelForCausalLM": "modeling_lfm3_ts.Lfm3TsForConditionalGeneration"
8
+ },
9
+ "dtype": "bfloat16",
10
+ "model_type": "lfm3_ts",
11
+ "projector_bias": true,
12
+ "projector_hidden_act": "gelu",
13
+ "projector_hidden_size": 4096,
14
+ "projector_layer_norm_eps": 1e-06,
15
+ "projector_output_scale": 0.020457027480006218,
16
+ "projector_position_dim": 0,
17
+ "text_config": {
18
+ "_name_or_path": "LiquidAI/LFM3-1.7B",
19
+ "architectures": [
20
+ "Lfm3ForCausalLM"
21
+ ],
22
+ "attention_bias": false,
23
+ "attention_dropout": 0.0,
24
+ "auto_map": {
25
+ "AutoConfig": "configuration_lfm3.Lfm3Config",
26
+ "AutoModel": "modeling_lfm3.Lfm3Model",
27
+ "AutoModelForCausalLM": "modeling_lfm3.Lfm3ForCausalLM"
28
+ },
29
+ "bos_token_id": 124894,
30
+ "chunk_size_feed_forward": 0,
31
+ "conv_L_cache": 7,
32
+ "conv_bias": false,
33
+ "dtype": "bfloat16",
34
+ "eos_token_id": 124900,
35
+ "head_dim": 128,
36
+ "hidden_size": 2048,
37
+ "id2label": {
38
+ "0": "LABEL_0",
39
+ "1": "LABEL_1"
40
+ },
41
+ "initializer_range": 0.02,
42
+ "intermediate_size": 7680,
43
+ "is_encoder_decoder": false,
44
+ "label2id": {
45
+ "LABEL_0": 0,
46
+ "LABEL_1": 1
47
+ },
48
+ "layer_types": [
49
+ "conv",
50
+ "conv",
51
+ "full_attention",
52
+ "conv",
53
+ "sliding_attention",
54
+ "conv",
55
+ "full_attention",
56
+ "conv",
57
+ "sliding_attention",
58
+ "conv",
59
+ "full_attention",
60
+ "conv",
61
+ "sliding_attention",
62
+ "conv",
63
+ "full_attention",
64
+ "conv",
65
+ "sliding_attention",
66
+ "conv",
67
+ "full_attention",
68
+ "conv",
69
+ "sliding_attention",
70
+ "conv",
71
+ "full_attention",
72
+ "conv"
73
+ ],
74
+ "max_position_embeddings": 262144,
75
+ "model_type": "lfm3",
76
+ "norm_eps": 1e-05,
77
+ "num_attention_heads": 16,
78
+ "num_hidden_layers": 24,
79
+ "num_key_value_heads": 4,
80
+ "output_attentions": false,
81
+ "output_hidden_states": false,
82
+ "pad_token_id": 124893,
83
+ "problem_type": null,
84
+ "return_dict": true,
85
+ "rope_parameters": {
86
+ "full_attention": {
87
+ "partial_rotary_factor": 0.5,
88
+ "rope_theta": 1000000.0,
89
+ "rope_type": "default"
90
+ },
91
+ "sliding_attention": {
92
+ "rope_theta": 10000,
93
+ "rope_type": "default"
94
+ }
95
+ },
96
+ "sliding_window": 256,
97
+ "tie_word_embeddings": true,
98
+ "use_attention_gate": true,
99
+ "use_cache": true,
100
+ "use_sandwich_norm": true,
101
+ "vocab_size": 128000
102
+ },
103
+ "tie_word_embeddings": true,
104
+ "transformers_version": "5.23.0",
105
+ "ts_config": {
106
+ "_name_or_path": "",
107
+ "architectures": null,
108
+ "chunk_size_feed_forward": 0,
109
+ "dtype": "bfloat16",
110
+ "hidden_act": "gelu",
111
+ "hidden_size": 1024,
112
+ "id2label": {
113
+ "0": "LABEL_0",
114
+ "1": "LABEL_1"
115
+ },
116
+ "initializer_range": 0.02,
117
+ "intermediate_size": 2560,
118
+ "is_encoder_decoder": false,
119
+ "label2id": {
120
+ "LABEL_0": 0,
121
+ "LABEL_1": 1
122
+ },
123
+ "model_type": "lfm3_ts_encoder",
124
+ "norm_eps": 1e-05,
125
+ "num_attention_heads": 8,
126
+ "num_attention_layers": 2,
127
+ "num_hidden_layers": 4,
128
+ "output_attentions": false,
129
+ "output_hidden_states": false,
130
+ "patch_size": 8,
131
+ "position_embedding_dim": 16,
132
+ "problem_type": null,
133
+ "return_dict": true,
134
+ "use_differences": false,
135
+ "value_compression": "none"
136
+ },
137
+ "ts_token_id": 125033
138
+ }
configuration_lfm3_ts.py ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2026 The HuggingFace Team. All rights reserved.
2
+ # Copyright 2026 Nardi Xhepi.
3
+ #
4
+ # NOTICE OF CHANGES: this file is part of LFM3-TS. It is adapted from `transformers.models.lfm3_vl`
5
+ # (Hugging Face Transformers, Apache License 2.0) and was changed by Nardi Xhepi: the vision tower and image
6
+ # inputs are replaced by a time-series encoder and time-series inputs.
7
+ #
8
+ # The portions taken from Transformers remain available under the Apache License, Version 2.0; a copy is in
9
+ # LICENSES/Apache-2.0.txt. LFM3-TS as a whole is distributed under the LFM Open License v1.0: see LICENSE and
10
+ # NOTICE at the root of the repository.
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, this software is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ """LFM3-TS configuration."""
15
+
16
+ from huggingface_hub.dataclasses import strict
17
+ from transformers.configuration_utils import PreTrainedConfig
18
+ from transformers.models.auto import CONFIG_MAPPING, AutoConfig
19
+
20
+
21
+ @strict
22
+ class Lfm3TsEncoderConfig(PreTrainedConfig):
23
+ r"""
24
+ Configuration of the LFM3-TS time-series encoder.
25
+
26
+ Args:
27
+ patch_size (`int`, *optional*, defaults to 8):
28
+ The number of consecutive time steps packed into one patch, i.e. one language token.
29
+ position_embedding_dim (`int`, *optional*, defaults to 16):
30
+ The size of the sinusoidal encoding of each time step's index, given to the encoder next to its value so
31
+ that a patch knows where in the series it sits. Must be even; 0 disables it.
32
+ use_differences (`bool`, *optional*, defaults to `False`):
33
+ Whether each time step also carries its change from the previous step, which makes spikes, shifts and
34
+ slopes explicit.
35
+ value_compression (`str`, *optional*, defaults to `"none"`):
36
+ `"asinh"` passes the standardized values (and differences) through `asinh` before the patch embedding:
37
+ close to the identity for ordinary values, logarithmic for outliers. A spike of 15 standard deviations
38
+ otherwise dominates its patch and drowns the position of the patch along with it.
39
+ hidden_size (`int`, *optional*, defaults to 1024):
40
+ The width of the encoder.
41
+ intermediate_size (`int`, *optional*, defaults to 2560):
42
+ The width of the feed-forward layer inside each encoder layer.
43
+ num_hidden_layers (`int`, *optional*, defaults to 4):
44
+ The number of residual feed-forward layers applied to each patch.
45
+ num_attention_layers (`int`, *optional*, defaults to 0):
46
+ The number of bidirectional self-attention layers, with rotary position embeddings, applied across the
47
+ patches of a series after the feed-forward layers. 0 keeps the patches independent, as in ChatTS.
48
+ num_attention_heads (`int`, *optional*, defaults to 8):
49
+ The number of heads of the attention layers.
50
+ hidden_act (`str`, *optional*, defaults to `"gelu"`):
51
+ The activation of the encoder layers.
52
+ norm_eps (`float`, *optional*, defaults to 1e-05):
53
+ The epsilon used by the normalization layers.
54
+ initializer_range (`float`, *optional*, defaults to 0.02):
55
+ The standard deviation of the weight initializer.
56
+ """
57
+
58
+ model_type = "lfm3_ts_encoder"
59
+ base_config_key = "ts_config"
60
+
61
+ patch_size: int = 8
62
+ position_embedding_dim: int = 16
63
+ use_differences: bool = False
64
+ value_compression: str = "none"
65
+ hidden_size: int = 1024
66
+ intermediate_size: int = 2560
67
+ num_hidden_layers: int = 4
68
+ num_attention_layers: int = 0
69
+ num_attention_heads: int = 8
70
+ hidden_act: str = "gelu"
71
+ norm_eps: float = 1e-5
72
+ initializer_range: float = 0.02
73
+
74
+ def validate_architecture(self):
75
+ """Part of ``@strict``-powered validation."""
76
+ if self.patch_size < 1:
77
+ raise ValueError(f"`patch_size` must be positive, got {self.patch_size}.")
78
+ if self.position_embedding_dim < 0 or self.position_embedding_dim % 2:
79
+ raise ValueError(f"`position_embedding_dim` must be even and >= 0, got {self.position_embedding_dim}.")
80
+ if self.value_compression not in ("none", "asinh"):
81
+ raise ValueError(f'`value_compression` must be "none" or "asinh", got {self.value_compression!r}.')
82
+ if self.num_attention_layers and (self.hidden_size // self.num_attention_heads) % 2:
83
+ raise ValueError("`hidden_size // num_attention_heads` must be even for rotary position embeddings.")
84
+ if self.num_attention_layers and self.hidden_size % self.num_attention_heads:
85
+ raise ValueError("`hidden_size` must be a multiple of `num_attention_heads`.")
86
+
87
+
88
+ @strict
89
+ class Lfm3TsConfig(PreTrainedConfig):
90
+ r"""
91
+ Configuration of an LFM3-TS model: an LFM3 language model reading time series through a patch encoder.
92
+
93
+ Args:
94
+ text_config (`dict` or `PreTrainedConfig`, *optional*):
95
+ The configuration of the LFM3 language model. Defaults to `Lfm3Config()`.
96
+ ts_config (`dict` or `Lfm3TsEncoderConfig`, *optional*):
97
+ The configuration of the time-series encoder. Defaults to `Lfm3TsEncoderConfig()`.
98
+ ts_token_id (`int`, *optional*, defaults to 125033):
99
+ The id of the placeholder token whose embeddings are replaced by time-series features. 125033 is the
100
+ first unused row of the LFM3 embedding matrix, so adding the token does not resize the embeddings.
101
+ projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
102
+ The activation of the multimodal projector.
103
+ projector_hidden_size (`int`, *optional*, defaults to 4096):
104
+ The hidden width of the multimodal projector.
105
+ projector_bias (`bool`, *optional*, defaults to `True`):
106
+ Whether the projector's linear layers have a bias.
107
+ projector_layer_norm_eps (`float`, *optional*, defaults to 1e-06):
108
+ The epsilon of the layer normalizations of the projector.
109
+ projector_position_dim (`int`, *optional*, defaults to 0):
110
+ The size of a sinusoidal encoding of each patch's first index, projected and added to the patch's token
111
+ after the content path and its normalization. Whatever the patch holds, its token then states where it
112
+ sits. Must be even; 0 disables it.
113
+ projector_output_scale (`float`, *optional*, defaults to 0.02):
114
+ The per-dimension scale the projected features start at. It should equal that of the language model's
115
+ token embeddings (0.0207 for LFM3-1.7B), which enter the residual stream unnormalized: features a few
116
+ times larger drown the layers' updates at their positions.
117
+ """
118
+
119
+ model_type = "lfm3_ts"
120
+ sub_configs = {"text_config": AutoConfig, "ts_config": Lfm3TsEncoderConfig}
121
+
122
+ text_config: dict | PreTrainedConfig | None = None
123
+ ts_config: dict | PreTrainedConfig | None = None
124
+ ts_token_id: int = 125_033
125
+ projector_hidden_act: str = "gelu"
126
+ projector_hidden_size: int = 4096
127
+ projector_bias: bool = True
128
+ projector_layer_norm_eps: float = 1e-6
129
+ projector_position_dim: int = 0
130
+ projector_output_scale: float = 0.02
131
+ tie_word_embeddings: bool = True
132
+
133
+ def __post_init__(self, **kwargs):
134
+ if "lfm3" not in CONFIG_MAPPING:
135
+ import transformers
136
+
137
+ raise ImportError(
138
+ f"LFM3-TS needs transformers 5.23 or later, the first release that includes LFM3, and found "
139
+ f"{transformers.__version__}. Run `pip install -U \"transformers>=5.23\"` and restart the runtime."
140
+ )
141
+ if isinstance(self.ts_config, dict):
142
+ self.ts_config = self.sub_configs["ts_config"](**self.ts_config)
143
+ elif self.ts_config is None:
144
+ self.ts_config = self.sub_configs["ts_config"]()
145
+
146
+ if isinstance(self.text_config, dict):
147
+ self.text_config = {"model_type": "lfm3", **self.text_config}
148
+ self.text_config = CONFIG_MAPPING[self.text_config["model_type"]](**self.text_config)
149
+ elif self.text_config is None:
150
+ self.text_config = CONFIG_MAPPING["lfm3"]()
151
+ self.validate_ts_fields()
152
+
153
+ super().__post_init__(**kwargs)
154
+
155
+ def validate_ts_fields(self):
156
+ """Part of ``@strict``-powered validation, so `validate()` and `save_pretrained` refuse later edits too."""
157
+ # an `inputs_embeds` forward embeds the placeholder id to find it
158
+ vocab_size = self.text_config.vocab_size
159
+ if self.projector_position_dim < 0 or self.projector_position_dim % 2:
160
+ raise ValueError(f"`projector_position_dim` must be even and >= 0, got {self.projector_position_dim}.")
161
+ if not 0 <= self.ts_token_id < vocab_size:
162
+ raise ValueError(f"ts_token_id={self.ts_token_id} must lie in [0, text_config.vocab_size={vocab_size})")
163
+
164
+
165
+ __all__ = ["Lfm3TsConfig", "Lfm3TsEncoderConfig"]
feature_extraction_lfm3_ts.py ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2026 Nardi Xhepi.
2
+ #
3
+ # This file is part of LFM3-TS, distributed under the LFM Open License v1.0: see LICENSE and NOTICE at the root
4
+ # of the repository. Unless required by applicable law or agreed to in writing, this software is distributed on
5
+ # an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
6
+ """Feature extractor for LFM3-TS: normalizes each series and cuts it into patches."""
7
+
8
+ from dataclasses import dataclass
9
+
10
+ import numpy as np
11
+ from transformers.feature_extraction_utils import BatchFeature, FeatureExtractionMixin
12
+
13
+
14
+ @dataclass
15
+ class Lfm3TsSeriesStats:
16
+ """What normalization removed from a series; the processor writes it into the prompt so values stay recoverable."""
17
+
18
+ offset: float
19
+ scale: float
20
+ length: int
21
+ num_patches: int
22
+ minimum: float
23
+ maximum: float
24
+ first: float
25
+ last: float
26
+
27
+
28
+ class Lfm3TsFeatureExtractor(FeatureExtractionMixin):
29
+ r"""
30
+ Turns univariate series into the `input_values` and `input_values_mask` of an LFM3-TS model.
31
+
32
+ Each series is standardized with its own mean and standard deviation over its observed values, then cut into
33
+ patches of `patch_size` time steps. NaN marks a missing value: it is zeroed and masked out, as is the padding
34
+ that fills the last patch and aligns series of different lengths.
35
+
36
+ Args:
37
+ patch_size (`int`, *optional*, defaults to 8):
38
+ The number of time steps per patch. Must equal the model's `ts_config.patch_size`.
39
+ eps (`float`, *optional*, defaults to 1e-06):
40
+ Series whose standard deviation is below `eps` times their magnitude are treated as constant and left
41
+ unscaled.
42
+ """
43
+
44
+ model_input_names = ["input_values", "input_values_mask"]
45
+
46
+ def __init__(self, patch_size: int = 8, eps: float = 1e-6, **kwargs):
47
+ super().__init__(**kwargs)
48
+ self.patch_size = patch_size
49
+ self.eps = eps
50
+
51
+ def get_num_patches(self, length: int) -> int:
52
+ """Number of patches covering `length` time steps."""
53
+ return -(-length // self.patch_size)
54
+
55
+ def preprocess(self, timeseries: list) -> tuple[np.ndarray, np.ndarray, list[Lfm3TsSeriesStats]]:
56
+ """Normalizes and patches a list of 1D series.
57
+
58
+ Returns:
59
+ The values and the observation mask, both of shape `(num_series, max_patches, patch_size)`, and the
60
+ statistics of each series.
61
+ """
62
+ normalized, stats = [], []
63
+ for index, series in enumerate(timeseries):
64
+ series = np.asarray(series, dtype=np.float64)
65
+ if series.ndim != 1:
66
+ raise ValueError(f"Series {index} must be one-dimensional, got shape {series.shape}.")
67
+ observed = np.isfinite(series)
68
+ if not observed.any():
69
+ raise ValueError(f"Series {index} has no observed value.")
70
+ # trailing missing values carry no information: the series ends at its last observed step
71
+ length = int(np.flatnonzero(observed)[-1]) + 1
72
+ series, observed = series[:length], observed[:length]
73
+
74
+ offset = float(series[observed].mean())
75
+ scale = float(series[observed].std())
76
+ if scale <= self.eps * max(1.0, abs(offset)):
77
+ scale = 1.0
78
+ values = np.where(observed, (series - offset) / scale, 0.0)
79
+ normalized.append((values, observed))
80
+ seen = series[observed]
81
+ stats.append(
82
+ Lfm3TsSeriesStats(
83
+ offset=offset,
84
+ scale=scale,
85
+ length=length,
86
+ num_patches=self.get_num_patches(length),
87
+ minimum=float(seen.min()),
88
+ maximum=float(seen.max()),
89
+ first=float(seen[0]),
90
+ last=float(seen[-1]),
91
+ )
92
+ )
93
+
94
+ max_patches = max(stat.num_patches for stat in stats)
95
+ input_values = np.zeros((len(stats), max_patches * self.patch_size), dtype=np.float32)
96
+ input_values_mask = np.zeros_like(input_values)
97
+ for index, (values, observed) in enumerate(normalized):
98
+ input_values[index, : len(values)] = values
99
+ input_values_mask[index, : len(values)] = observed
100
+ shape = (len(stats), max_patches, self.patch_size)
101
+ return input_values.reshape(shape), input_values_mask.reshape(shape), stats
102
+
103
+ def __call__(self, timeseries: list, return_tensors: str | None = None, **kwargs) -> BatchFeature:
104
+ """
105
+ Args:
106
+ timeseries (`list`):
107
+ One-dimensional series (arrays or lists of numbers), possibly of different lengths.
108
+ return_tensors (`str`, *optional*):
109
+ The tensor type to return, e.g. `"pt"`.
110
+ """
111
+ input_values, input_values_mask, _ = self.preprocess(timeseries)
112
+ data = {"input_values": input_values, "input_values_mask": input_values_mask}
113
+ return BatchFeature(data=data, tensor_type=return_tensors)
114
+
115
+
116
+ __all__ = ["Lfm3TsFeatureExtractor", "Lfm3TsSeriesStats"]
generation_config.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 124894,
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 124900
7
+ ],
8
+ "output_attentions": false,
9
+ "output_hidden_states": false,
10
+ "pad_token_id": 124893,
11
+ "repetition_penalty": 1.1,
12
+ "temperature": 0.1,
13
+ "top_k": 50,
14
+ "transformers_version": "5.23.0",
15
+ "use_cache": true
16
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e11577553e71d7be7e6c6c96d787642bf6838ae5efb41f30b810c6ab77fe0843
3
+ size 3562909848
modeling_lfm3_ts.py ADDED
@@ -0,0 +1,629 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2026 The HuggingFace Team. All rights reserved.
2
+ # Copyright 2026 Nardi Xhepi.
3
+ #
4
+ # NOTICE OF CHANGES: this file is part of LFM3-TS. It is adapted from `transformers.models.lfm3_vl`
5
+ # (Hugging Face Transformers, Apache License 2.0) and was changed by Nardi Xhepi: the vision tower and image
6
+ # inputs are replaced by a time-series encoder and time-series inputs.
7
+ #
8
+ # The portions taken from Transformers remain available under the Apache License, Version 2.0; a copy is in
9
+ # LICENSES/Apache-2.0.txt. LFM3-TS as a whole is distributed under the LFM Open License v1.0: see LICENSE and
10
+ # NOTICE at the root of the repository.
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, this software is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ """LFM3-TS model: a time-series tower and a projector in front of the LFM3 language model."""
15
+
16
+ import math
17
+ from dataclasses import dataclass
18
+
19
+ import numpy as np
20
+ import torch
21
+ import torch.nn.functional as F
22
+ from torch import nn
23
+ from transformers import initialization as init
24
+ from transformers.activations import ACT2FN
25
+ from transformers.cache_utils import Cache
26
+ from transformers.generation import GenerationMixin
27
+ from transformers.modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, CausalLMOutputWithPast
28
+ from transformers.modeling_utils import PreTrainedModel
29
+ from transformers.models.lfm3.modeling_lfm3 import Lfm3Model
30
+ from transformers.utils import can_return_tuple, torch_compilable_check
31
+
32
+ from .configuration_lfm3_ts import Lfm3TsConfig, Lfm3TsEncoderConfig
33
+
34
+
35
+ def get_ts_patch_counts(input_values_mask: torch.Tensor) -> torch.Tensor:
36
+ """Number of patches each series occupies: up to and including its last patch holding an observed time step.
37
+
38
+ Args:
39
+ input_values_mask (`torch.Tensor` of shape `(num_series, max_patches, patch_size)`):
40
+ 1 for observed time steps, 0 for padding and missing values.
41
+
42
+ Returns:
43
+ `torch.LongTensor` of shape `(num_series,)`.
44
+ """
45
+ observed = input_values_mask.bool().any(-1)
46
+ index = torch.arange(1, observed.shape[1] + 1, device=observed.device)
47
+ return (observed * index).amax(-1)
48
+
49
+
50
+ def _as_series_list(series) -> list[np.ndarray]:
51
+ """One 1D array per series, from a 1D array or list of numbers, a list of them, or a 2D array of rows."""
52
+ if hasattr(series, "detach"):
53
+ series = series.detach().cpu().numpy()
54
+ if hasattr(series, "to_numpy"):
55
+ series = series.to_numpy()
56
+ if isinstance(series, np.ndarray):
57
+ items = [series] if series.ndim == 1 else list(series)
58
+ elif isinstance(series, (list, tuple)) and len(series) and all(
59
+ isinstance(value, (int, float, np.number)) for value in series
60
+ ):
61
+ items = [series]
62
+ elif isinstance(series, (list, tuple)):
63
+ items = list(series)
64
+ else:
65
+ raise ValueError(f"`series` must be an array or a list of numbers, or a list of them, got {type(series)}.")
66
+ arrays = []
67
+ for index, item in enumerate(items):
68
+ if hasattr(item, "detach"):
69
+ item = item.detach().cpu().numpy()
70
+ array = np.asarray(item, dtype=np.float64)
71
+ if array.ndim != 1 or not array.size:
72
+ raise ValueError(f"Series {index} must be one-dimensional and not empty, got shape {array.shape}.")
73
+ arrays.append(array)
74
+ if not arrays:
75
+ raise ValueError("`series` is empty.")
76
+ return arrays
77
+
78
+
79
+ def build_chat_messages(series, question: str) -> list[dict]:
80
+ """The conversation `Lfm3TsForConditionalGeneration.chat` sends: one user message holding the series and the
81
+ question, in the wording the released model was trained on for one series
82
+ (`"This is a time series of length N: "`, the series, `". "`, the question).
83
+
84
+ Args:
85
+ series:
86
+ A 1D array or list of numbers; or several series, as a list of them or as the rows of a 2D array; or
87
+ `None` for a question without a series. NaN marks a missing value.
88
+ question (`str`):
89
+ What to ask.
90
+ """
91
+ question = question.strip()
92
+ if series is None:
93
+ return [{"role": "user", "content": [{"type": "text", "text": question}]}]
94
+ arrays = _as_series_list(series)
95
+ if len(arrays) == 1:
96
+ content = [
97
+ {"type": "text", "text": f"This is a time series of length {len(arrays[0])}: "},
98
+ {"type": "timeseries", "timeseries": arrays[0]},
99
+ ]
100
+ else:
101
+ content = [{"type": "text", "text": f"There are {len(arrays)} time series. "}]
102
+ for index, array in enumerate(arrays):
103
+ if index:
104
+ content.append({"type": "text", "text": "; "})
105
+ content.append({"type": "text", "text": f"Series {index + 1}, of length {len(array)}: "})
106
+ content.append({"type": "timeseries", "timeseries": array})
107
+ content.append({"type": "text", "text": f". {question}"})
108
+ return [{"role": "user", "content": content}]
109
+
110
+
111
+ class Lfm3TsEncoderLayer(nn.Module):
112
+ """A pre-norm residual feed-forward layer applied to each patch on its own."""
113
+
114
+ def __init__(self, config: Lfm3TsEncoderConfig):
115
+ super().__init__()
116
+ self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps)
117
+ self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
118
+ self.act = ACT2FN[config.hidden_act]
119
+ self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
120
+
121
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
122
+ return hidden_states + self.fc2(self.act(self.fc1(self.norm(hidden_states))))
123
+
124
+
125
+ def apply_rotary_position_embedding(states: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
126
+ """Rotates each pair of channels of `(batch, heads, patches, head_dim)` states by its patch's angle."""
127
+ even, odd = states[..., 0::2], states[..., 1::2]
128
+ return torch.stack([even * cos - odd * sin, even * sin + odd * cos], dim=-1).flatten(-2)
129
+
130
+
131
+ class Lfm3TsAttentionLayer(nn.Module):
132
+ """A pre-norm bidirectional self-attention layer over the patches of one series, with rotary position
133
+ embeddings, so that a patch can read its context: a period or a trend spans several patches."""
134
+
135
+ def __init__(self, config: Lfm3TsEncoderConfig):
136
+ super().__init__()
137
+ self.num_heads = config.num_attention_heads
138
+ self.head_dim = config.hidden_size // config.num_attention_heads
139
+ self.attention_norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps)
140
+ self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
141
+ self.k_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
142
+ self.v_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
143
+ self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
144
+ self.feed_forward = Lfm3TsEncoderLayer(config)
145
+
146
+ def forward(
147
+ self, hidden_states: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, attention_mask: torch.Tensor
148
+ ) -> torch.Tensor:
149
+ num_series, num_patches, hidden_size = hidden_states.shape
150
+ states = self.attention_norm(hidden_states)
151
+ split = lambda tensor: tensor.view(num_series, num_patches, self.num_heads, self.head_dim).transpose(1, 2)
152
+ queries = apply_rotary_position_embedding(split(self.q_proj(states)), cos, sin)
153
+ keys = apply_rotary_position_embedding(split(self.k_proj(states)), cos, sin)
154
+ attended = F.scaled_dot_product_attention(queries, keys, split(self.v_proj(states)), attn_mask=attention_mask)
155
+ attended = attended.transpose(1, 2).reshape(num_series, num_patches, hidden_size)
156
+ return self.feed_forward(hidden_states + self.o_proj(attended))
157
+
158
+
159
+ class Lfm3TsPreTrainedModel(PreTrainedModel):
160
+ config: Lfm3TsConfig
161
+ base_model_prefix = "model"
162
+ supports_gradient_checkpointing = True
163
+ _skip_keys_device_placement = ["past_key_values"]
164
+
165
+ _supports_flash_attn = True
166
+ _supports_sdpa = True
167
+ _supports_flex_attn = True
168
+ _supports_attention_backend = True
169
+
170
+ def _init_weights(self, module):
171
+ super()._init_weights(module)
172
+ if isinstance(module, Lfm3TsMultiModalProjector):
173
+ scale = module.output_scale
174
+ init.normal_(module.boundary_embeddings, mean=0.0, std=scale)
175
+ init.constant_(module.output_norm.weight, scale)
176
+ init.zeros_(module.output_norm.bias)
177
+ if module.position_embed is not None:
178
+ # sinusoids have a mean square of 1/2: the position starts at the scale of the content
179
+ init.normal_(module.position_embed.weight, mean=0.0, std=scale / math.sqrt(module.position_dim / 2))
180
+
181
+
182
+ class Lfm3TsEncoder(Lfm3TsPreTrainedModel):
183
+ """The LFM3-TS time-series encoder: a linear patch embedding over the values, observation mask and position of
184
+ each time step of a patch, followed by residual feed-forward layers. With `num_attention_layers=0` patches are
185
+ encoded independently and only the language model relates them; attention layers let them read each other."""
186
+
187
+ config: Lfm3TsEncoderConfig
188
+ main_input_name = "input_values"
189
+ _input_embed_layer = "patch_embed"
190
+
191
+ def __init__(self, config: Lfm3TsEncoderConfig):
192
+ super().__init__(config)
193
+ # each time step carries its value, its observation mask and the encoding of its index in the series
194
+ channels = 2 + int(config.use_differences) + config.position_embedding_dim
195
+ self.patch_embed = nn.Linear(channels * config.patch_size, config.hidden_size)
196
+ self.layers = nn.ModuleList([Lfm3TsEncoderLayer(config) for _ in range(config.num_hidden_layers)])
197
+ self.attention_layers = nn.ModuleList(
198
+ [Lfm3TsAttentionLayer(config) for _ in range(config.num_attention_layers)]
199
+ )
200
+ self.post_init()
201
+
202
+ def forward(
203
+ self, input_values: torch.Tensor, input_values_mask: torch.Tensor | None = None, **kwargs
204
+ ) -> BaseModelOutputWithPooling:
205
+ r"""
206
+ input_values (`torch.FloatTensor` of shape `(num_series, max_patches, patch_size)`):
207
+ The normalized values of every series, cut into patches.
208
+ input_values_mask (`torch.Tensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
209
+ 1 for observed time steps, 0 for padding and missing values. Defaults to all observed.
210
+ """
211
+ if input_values_mask is None:
212
+ input_values_mask = torch.ones_like(input_values)
213
+ dtype = self.patch_embed.weight.dtype
214
+ # outliers are compressed so that no value outweighs the rest of its patch
215
+ compress = torch.asinh if self.config.value_compression == "asinh" else (lambda tensor: tensor)
216
+ patches = [compress(input_values.to(dtype)), input_values_mask.to(dtype)]
217
+ if self.config.use_differences:
218
+ # the change from the previous step, across patch borders; zero where either step is unobserved
219
+ values, observed = input_values.to(dtype).flatten(1), input_values_mask.to(dtype).flatten(1)
220
+ differences = torch.zeros_like(values)
221
+ differences[:, 1:] = (values[:, 1:] - values[:, :-1]) * observed[:, 1:] * observed[:, :-1]
222
+ patches.append(compress(differences).view_as(input_values))
223
+ if self.config.position_embedding_dim:
224
+ patches.append(self.get_position_features(input_values_mask).to(dtype))
225
+ hidden_states = self.patch_embed(torch.cat(patches, dim=-1))
226
+ for layer in self.layers:
227
+ hidden_states = layer(hidden_states)
228
+
229
+ if self.attention_layers:
230
+ num_patches = hidden_states.shape[1]
231
+ head_dim = self.config.hidden_size // self.config.num_attention_heads
232
+ device = hidden_states.device
233
+ frequencies = 1.0 / (10000.0 ** (torch.arange(0, head_dim, 2, device=device) / head_dim))
234
+ angles = torch.arange(num_patches, device=device)[:, None] * frequencies[None, :]
235
+ cos, sin = angles.cos().to(dtype), angles.sin().to(dtype)
236
+ # a patch attends to the patches of its own series, not to the padding that aligns the batch
237
+ is_patch = torch.arange(num_patches, device=device)[None, :] < get_ts_patch_counts(input_values_mask)[:, None]
238
+ attention_mask = is_patch[:, None, None, :]
239
+ for layer in self.attention_layers:
240
+ hidden_states = layer(hidden_states, cos, sin, attention_mask)
241
+ return BaseModelOutputWithPooling(last_hidden_state=hidden_states)
242
+
243
+
244
+ def get_position_features(self, input_values_mask: torch.Tensor) -> torch.Tensor:
245
+ """Sinusoidal encoding of the index of every observed time step, zero elsewhere.
246
+
247
+ Without it a patch holds no trace of where it sits, and the language model could only locate an event by
248
+ counting tokens. Sinusoids need no training and extend to any length.
249
+
250
+ Returns:
251
+ `torch.FloatTensor` of shape `(num_series, max_patches, patch_size * position_embedding_dim)`.
252
+ """
253
+ num_series, max_patches, patch_size = input_values_mask.shape
254
+ half = self.config.position_embedding_dim // 2
255
+ device = input_values_mask.device
256
+ positions = torch.arange(max_patches * patch_size, device=device, dtype=torch.float32)
257
+ frequencies = torch.exp(-math.log(10000.0) * torch.arange(half, device=device) / half)
258
+ angles = positions[:, None] * frequencies[None, :]
259
+ features = torch.cat([angles.sin(), angles.cos()], dim=-1).view(max_patches, patch_size, 2 * half)
260
+ features = features[None] * input_values_mask[..., None].to(features.dtype)
261
+ return features.flatten(2)
262
+
263
+
264
+ class Lfm3TsMultiModalProjector(nn.Module):
265
+ """Projects each patch to one language token and frames every series with a learned start and end embedding.
266
+
267
+ The boundaries live here, not in the vocabulary, so that every parameter introduced for time series sits outside
268
+ the language model and the backbone can stay frozen."""
269
+
270
+ def __init__(self, config: Lfm3TsConfig):
271
+ super().__init__()
272
+ self.layer_norm = nn.LayerNorm(config.ts_config.hidden_size, eps=config.projector_layer_norm_eps)
273
+ self.linear_1 = nn.Linear(
274
+ config.ts_config.hidden_size, config.projector_hidden_size, bias=config.projector_bias
275
+ )
276
+ self.act = ACT2FN[config.projector_hidden_act]
277
+ self.linear_2 = nn.Linear(
278
+ config.projector_hidden_size, config.text_config.hidden_size, bias=config.projector_bias
279
+ )
280
+ # brings the features to the scale of the token embeddings they sit among; its weight starts at that scale
281
+ self.output_norm = nn.LayerNorm(config.text_config.hidden_size, eps=config.projector_layer_norm_eps)
282
+ self.output_scale = config.projector_output_scale
283
+ # the position of a patch joins its token after the normalization, where the content cannot outweigh it
284
+ self.patch_size = config.ts_config.patch_size
285
+ self.position_dim = config.projector_position_dim
286
+ self.position_embed = (
287
+ nn.Linear(self.position_dim, config.text_config.hidden_size, bias=False) if self.position_dim else None
288
+ )
289
+ self.boundary_embeddings = nn.Parameter(torch.empty(2, config.text_config.hidden_size))
290
+
291
+ def forward(self, ts_features: torch.Tensor) -> torch.Tensor:
292
+ hidden_states = self.layer_norm(ts_features)
293
+ hidden_states = self.linear_1(hidden_states)
294
+ hidden_states = self.act(hidden_states)
295
+ hidden_states = self.linear_2(hidden_states)
296
+ hidden_states = self.output_norm(hidden_states)
297
+ if self.position_embed is not None:
298
+ hidden_states = hidden_states + self.position_embed(self.get_position_features(hidden_states))
299
+ return hidden_states
300
+
301
+ def get_position_features(self, hidden_states: torch.Tensor) -> torch.Tensor:
302
+ """Sinusoidal encoding of the index of the first time step of each patch, of shape `(num_patches, position_dim)`."""
303
+ half = self.position_dim // 2
304
+ device = hidden_states.device
305
+ starts = torch.arange(hidden_states.shape[1], device=device, dtype=torch.float32) * self.patch_size
306
+ frequencies = torch.exp(-math.log(10000.0) * torch.arange(half, device=device) / half)
307
+ angles = starts[:, None] * frequencies[None, :]
308
+ return torch.cat([angles.sin(), angles.cos()], dim=-1).to(hidden_states.dtype)
309
+
310
+
311
+ @dataclass
312
+ class Lfm3TsModelOutputWithPast(BaseModelOutputWithPast):
313
+ r"""
314
+ ts_hidden_states (`torch.FloatTensor`, *optional*):
315
+ A `torch.FloatTensor` of size `(num_ts_tokens, hidden_size)`: the projected time-series features that
316
+ replaced the placeholder tokens.
317
+ """
318
+
319
+ ts_hidden_states: torch.FloatTensor | None = None
320
+
321
+
322
+ @dataclass
323
+ class Lfm3TsCausalLMOutputWithPast(CausalLMOutputWithPast):
324
+ r"""
325
+ ts_hidden_states (`torch.FloatTensor`, *optional*):
326
+ A `torch.FloatTensor` of size `(num_ts_tokens, hidden_size)`: the projected time-series features that
327
+ replaced the placeholder tokens.
328
+ """
329
+
330
+ ts_hidden_states: torch.FloatTensor | None = None
331
+
332
+
333
+ class Lfm3TsModel(Lfm3TsPreTrainedModel):
334
+ """The LFM3-TS model: a time-series encoder and an LFM3 language model, without a language modeling head."""
335
+
336
+ def __init__(self, config: Lfm3TsConfig):
337
+ super().__init__(config)
338
+ self.ts_tower = Lfm3TsEncoder(config.ts_config)
339
+ self.multi_modal_projector = Lfm3TsMultiModalProjector(config)
340
+ self.language_model = Lfm3Model(config.text_config)
341
+ self.post_init()
342
+
343
+ def get_input_embeddings(self) -> nn.Module:
344
+ return self.language_model.get_input_embeddings()
345
+
346
+ def set_input_embeddings(self, value: nn.Module):
347
+ self.language_model.set_input_embeddings(value)
348
+
349
+ def get_ts_features(
350
+ self, input_values: torch.Tensor, input_values_mask: torch.Tensor | None = None, **kwargs
351
+ ) -> BaseModelOutputWithPooling:
352
+ r"""
353
+ Encodes and projects the series. `pooler_output` holds one tensor per series of shape
354
+ `(num_patches + 2, hidden_size)`: a start embedding, one token per patch, and an end embedding.
355
+
356
+ input_values (`torch.FloatTensor` of shape `(num_series, max_patches, patch_size)`):
357
+ The normalized values of every series, cut into patches.
358
+ input_values_mask (`torch.Tensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
359
+ 1 for observed time steps, 0 for padding and missing values. Defaults to all observed.
360
+ """
361
+ if input_values_mask is None:
362
+ input_values_mask = torch.ones_like(input_values)
363
+ ts_outputs = self.ts_tower(input_values, input_values_mask, **kwargs)
364
+ ts_features = self.multi_modal_projector(ts_outputs.last_hidden_state)
365
+ start, end = self.multi_modal_projector.boundary_embeddings.to(ts_features.dtype).split(1)
366
+ patch_counts = get_ts_patch_counts(input_values_mask).tolist()
367
+ ts_outputs.pooler_output = tuple(
368
+ torch.cat([start, features[:count], end]) for features, count in zip(ts_features, patch_counts)
369
+ )
370
+ return ts_outputs
371
+
372
+ def get_placeholder_mask(
373
+ self,
374
+ input_ids: torch.LongTensor | None,
375
+ inputs_embeds: torch.FloatTensor,
376
+ ts_features: torch.FloatTensor,
377
+ ) -> torch.Tensor:
378
+ """Placeholder mask from `input_ids` or `inputs_embeds`, checked against the feature count."""
379
+ if input_ids is None:
380
+ special_ts_mask = inputs_embeds == self.get_input_embeddings()(
381
+ torch.full((), self.config.ts_token_id, dtype=torch.long, device=inputs_embeds.device)
382
+ )
383
+ special_ts_mask = special_ts_mask.all(-1)
384
+ else:
385
+ special_ts_mask = input_ids == self.config.ts_token_id
386
+
387
+ n_ts_tokens = special_ts_mask.sum()
388
+ torch_compilable_check(
389
+ n_ts_tokens * inputs_embeds.shape[-1] == ts_features.numel(),
390
+ f"Time-series features and time-series tokens do not match, tokens: {n_ts_tokens}, "
391
+ f"features: {ts_features.shape[0]}",
392
+ )
393
+ return special_ts_mask.unsqueeze(-1).to(inputs_embeds.device)
394
+
395
+ @can_return_tuple
396
+ def forward(
397
+ self,
398
+ input_ids: torch.LongTensor | None = None,
399
+ attention_mask: torch.Tensor | None = None,
400
+ position_ids: torch.LongTensor | None = None,
401
+ past_key_values: Cache | None = None,
402
+ inputs_embeds: torch.FloatTensor | None = None,
403
+ input_values: torch.FloatTensor | None = None,
404
+ input_values_mask: torch.Tensor | None = None,
405
+ use_cache: bool | None = None,
406
+ **kwargs,
407
+ ) -> tuple | Lfm3TsModelOutputWithPast:
408
+ r"""
409
+ input_values (`torch.FloatTensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
410
+ The normalized values of every series in the batch, in the order their placeholders appear.
411
+ input_values_mask (`torch.Tensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
412
+ 1 for observed time steps, 0 for padding and missing values. Defaults to all observed.
413
+ """
414
+ if (input_ids is None) ^ (inputs_embeds is not None):
415
+ raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
416
+
417
+ if inputs_embeds is None:
418
+ inputs_embeds = self.get_input_embeddings()(input_ids)
419
+
420
+ ts_features = None
421
+ if input_values is not None:
422
+ ts_outputs = self.get_ts_features(input_values, input_values_mask)
423
+ ts_features = torch.cat(ts_outputs.pooler_output, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
424
+ ts_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds, ts_features=ts_features)
425
+ inputs_embeds = inputs_embeds.masked_scatter(ts_mask, ts_features)
426
+
427
+ outputs = self.language_model(
428
+ attention_mask=attention_mask,
429
+ position_ids=position_ids,
430
+ past_key_values=past_key_values,
431
+ inputs_embeds=inputs_embeds,
432
+ use_cache=use_cache,
433
+ **kwargs,
434
+ )
435
+
436
+ return Lfm3TsModelOutputWithPast(
437
+ last_hidden_state=outputs.last_hidden_state,
438
+ past_key_values=outputs.past_key_values,
439
+ hidden_states=outputs.hidden_states,
440
+ attentions=outputs.attentions,
441
+ ts_hidden_states=ts_features,
442
+ )
443
+
444
+
445
+ class Lfm3TsForConditionalGeneration(Lfm3TsPreTrainedModel, GenerationMixin):
446
+ """The LFM3-TS model: a time-series encoder and an LFM3 language model with a language modeling head."""
447
+
448
+ _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
449
+
450
+ def __init__(self, config: Lfm3TsConfig):
451
+ super().__init__(config)
452
+ self.model = Lfm3TsModel(config)
453
+ self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
454
+ self.post_init()
455
+
456
+ def get_input_embeddings(self) -> nn.Module:
457
+ return self.model.get_input_embeddings()
458
+
459
+ def set_input_embeddings(self, value: nn.Module):
460
+ self.model.set_input_embeddings(value)
461
+
462
+ def get_output_embeddings(self) -> nn.Module:
463
+ return self.lm_head
464
+
465
+ def get_ts_features(
466
+ self, input_values: torch.Tensor, input_values_mask: torch.Tensor | None = None, **kwargs
467
+ ) -> BaseModelOutputWithPooling:
468
+ return self.model.get_ts_features(input_values, input_values_mask, **kwargs)
469
+
470
+ @classmethod
471
+ def from_text_pretrained(
472
+ cls, pretrained_model_name_or_path: str = "LiquidAI/LFM3-1.7B", ts_config=None, **kwargs
473
+ ) -> "Lfm3TsForConditionalGeneration":
474
+ """Builds an LFM3-TS model around a pretrained LFM3 text checkpoint.
475
+
476
+ The language model takes the checkpoint's weights; the time-series tower and projector are freshly
477
+ initialized and need training. The projector's output starts at the scale of the checkpoint's embeddings.
478
+ `kwargs` go to the text model's `from_pretrained` (`dtype`, `device_map`, ...), except those naming an
479
+ `Lfm3TsConfig` field (`ts_token_id`, `projector_hidden_size`, ...).
480
+ """
481
+ from transformers import AutoModelForCausalLM
482
+
483
+ config_fields = {
484
+ "ts_token_id",
485
+ "projector_hidden_act",
486
+ "projector_hidden_size",
487
+ "projector_bias",
488
+ "projector_position_dim",
489
+ "projector_output_scale",
490
+ }
491
+ config_kwargs = {key: kwargs.pop(key) for key in list(kwargs) if key in config_fields}
492
+ text_model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, **kwargs)
493
+ if "projector_output_scale" not in config_kwargs:
494
+ # start the time-series features at the scale of the embeddings they will sit among
495
+ embeddings = text_model.get_input_embeddings().weight
496
+ config_kwargs["projector_output_scale"] = float(embeddings.detach().float().pow(2).mean().sqrt())
497
+ config = Lfm3TsConfig(text_config=text_model.config, ts_config=ts_config, **config_kwargs)
498
+
499
+ model = cls(config)
500
+ # adopt the loaded module itself: casting a fresh copy with `.to(dtype)` would also cast the buffers that
501
+ # `from_pretrained` keeps in full precision, and the outputs would drift from the text model's
502
+ model.model.language_model = text_model.model
503
+ model.lm_head = text_model.lm_head
504
+ reference = text_model.get_input_embeddings().weight
505
+ model.model.ts_tower.to(device=reference.device, dtype=reference.dtype)
506
+ model.model.multi_modal_projector.to(device=reference.device, dtype=reference.dtype)
507
+ model.generation_config = text_model.generation_config
508
+ return model
509
+
510
+ @torch.no_grad()
511
+ def chat(self, processor, series, question: str, max_new_tokens: int = 160, **generate_kwargs) -> str:
512
+ r"""
513
+ Asks one question about one or several series and returns the answer as text.
514
+
515
+ Args:
516
+ processor (`Lfm3TsProcessor`):
517
+ The processor of the checkpoint.
518
+ series:
519
+ A 1D array or list of numbers; or several series, as a list of them or as the rows of a 2D array;
520
+ or `None` for a question without a series. NaN marks a missing value.
521
+ question (`str`):
522
+ What to ask, e.g. `"Is there an anomaly? If yes, where?"`.
523
+ max_new_tokens (`int`, *optional*, defaults to 160):
524
+ The longest answer, in tokens.
525
+ generate_kwargs:
526
+ Passed to `generate`. Decoding is greedy unless `do_sample` is given.
527
+
528
+ Example:
529
+
530
+ ```python
531
+ >>> series = np.sin(np.arange(512) / 8)
532
+ >>> series[300] += 4
533
+ >>> model.chat(processor, series, "Is there an anomaly? If yes, where?")
534
+ ```"""
535
+ messages = build_chat_messages(series, question)
536
+ inputs = processor.apply_chat_template(
537
+ messages, add_generation_prompt=True, tokenize=True, enable_thinking=False, return_tensors="pt"
538
+ ).to(self.device)
539
+ generate_kwargs.setdefault("do_sample", False)
540
+ output = self.generate(**inputs, max_new_tokens=max_new_tokens, **generate_kwargs)
541
+ return processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True).strip()
542
+
543
+ @can_return_tuple
544
+ def forward(
545
+ self,
546
+ input_ids: torch.LongTensor | None = None,
547
+ attention_mask: torch.Tensor | None = None,
548
+ position_ids: torch.LongTensor | None = None,
549
+ past_key_values: Cache | None = None,
550
+ inputs_embeds: torch.FloatTensor | None = None,
551
+ input_values: torch.FloatTensor | None = None,
552
+ input_values_mask: torch.Tensor | None = None,
553
+ labels: torch.LongTensor | None = None,
554
+ use_cache: bool | None = None,
555
+ logits_to_keep: int | torch.Tensor = 0,
556
+ **kwargs,
557
+ ) -> tuple | Lfm3TsCausalLMOutputWithPast:
558
+ r"""
559
+ input_values (`torch.FloatTensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
560
+ The normalized values of every series in the batch, in the order their placeholders appear.
561
+ input_values_mask (`torch.Tensor` of shape `(num_series, max_patches, patch_size)`, *optional*):
562
+ 1 for observed time steps, 0 for padding and missing values. Defaults to all observed.
563
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
564
+ Labels for the language modeling loss; positions set to `-100` are ignored.
565
+
566
+ Example:
567
+
568
+ ```python
569
+ >>> import numpy as np
570
+ >>> from lfm3_ts import Lfm3TsForConditionalGeneration, Lfm3TsProcessor
571
+
572
+ >>> model = Lfm3TsForConditionalGeneration.from_text_pretrained("LiquidAI/LFM3-1.7B")
573
+ >>> processor = Lfm3TsProcessor.from_text_pretrained("LiquidAI/LFM3-1.7B")
574
+
575
+ >>> conversation = [
576
+ ... {
577
+ ... "role": "user",
578
+ ... "content": [
579
+ ... {"type": "timeseries", "timeseries": np.sin(np.arange(256) / 8)},
580
+ ... {"type": "text", "text": "Where are the anomalies in this series?"},
581
+ ... ],
582
+ ... },
583
+ ... ]
584
+ >>> inputs = processor.apply_chat_template(
585
+ ... conversation, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
586
+ ... )
587
+ >>> outputs = model.generate(**inputs, max_new_tokens=64)
588
+ >>> processor.batch_decode(outputs, skip_special_tokens=True)[0]
589
+ ```"""
590
+ outputs = self.model(
591
+ input_ids=input_ids,
592
+ attention_mask=attention_mask,
593
+ position_ids=position_ids,
594
+ past_key_values=past_key_values,
595
+ inputs_embeds=inputs_embeds,
596
+ input_values=input_values,
597
+ input_values_mask=input_values_mask,
598
+ use_cache=use_cache,
599
+ **kwargs,
600
+ )
601
+
602
+ hidden_states = outputs[0]
603
+ # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
604
+ slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
605
+ logits = self.lm_head(hidden_states[:, slice_indices, :])
606
+
607
+ loss = None
608
+ if labels is not None:
609
+ loss = self.loss_function(
610
+ logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
611
+ )
612
+
613
+ return Lfm3TsCausalLMOutputWithPast(
614
+ loss=loss,
615
+ logits=logits,
616
+ past_key_values=outputs.past_key_values,
617
+ hidden_states=outputs.hidden_states,
618
+ attentions=outputs.attentions,
619
+ ts_hidden_states=outputs.ts_hidden_states,
620
+ )
621
+
622
+
623
+ __all__ = [
624
+ "build_chat_messages",
625
+ "Lfm3TsEncoder",
626
+ "Lfm3TsForConditionalGeneration",
627
+ "Lfm3TsModel",
628
+ "Lfm3TsPreTrainedModel",
629
+ ]
processing_lfm3_ts.py ADDED
@@ -0,0 +1,285 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2026 The HuggingFace Team. All rights reserved.
2
+ # Copyright 2026 Nardi Xhepi.
3
+ #
4
+ # NOTICE OF CHANGES: this file is part of LFM3-TS. It is adapted from `transformers.models.lfm3_vl`
5
+ # (Hugging Face Transformers, Apache License 2.0) and was changed by Nardi Xhepi: the vision tower and image
6
+ # inputs are replaced by a time-series encoder and time-series inputs.
7
+ #
8
+ # The portions taken from Transformers remain available under the Apache License, Version 2.0; a copy is in
9
+ # LICENSES/Apache-2.0.txt. LFM3-TS as a whole is distributed under the LFM Open License v1.0: see LICENSE and
10
+ # NOTICE at the root of the repository.
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, this software is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ """Processor for LFM3-TS."""
15
+
16
+ import numpy as np
17
+ from transformers.feature_extraction_utils import BatchFeature
18
+ from transformers.processing_utils import ProcessorMixin
19
+
20
+ from .feature_extraction_lfm3_ts import Lfm3TsFeatureExtractor, Lfm3TsSeriesStats
21
+
22
+
23
+ TS_TOKEN = "<|ts|>"
24
+
25
+ # The LFM3 chat template renders `image`, `video` and `text` content parts; these add a `timeseries` part. Each pair is
26
+ # (anchor found in the LFM3 template, branch inserted before it).
27
+ _CHAT_TEMPLATE_PATCHES = (
28
+ (
29
+ '{%- elif content is mapping and content.get("type") == "text" -%}',
30
+ '{%- elif content is mapping and content.get("type") == "timeseries" -%}\n'
31
+ ' {{- "' + TS_TOKEN + '" -}}\n ',
32
+ ),
33
+ (
34
+ '{%- elif item is mapping and item.get("type") == "text" -%}',
35
+ '{%- elif item is mapping and item.get("type") == "timeseries" -%}\n'
36
+ ' {%- set _ns.result = _ns.result + "' + TS_TOKEN + '" -%}\n ',
37
+ ),
38
+ )
39
+
40
+
41
+ def add_timeseries_to_chat_template(chat_template: str) -> str:
42
+ """Teaches an LFM3 chat template to render `{"type": "timeseries"}` content parts as the placeholder token."""
43
+ for anchor, branch in _CHAT_TEMPLATE_PATCHES:
44
+ if chat_template.count(anchor) != 1:
45
+ raise ValueError(f"Cannot add time series to this chat template: expected exactly one `{anchor}`.")
46
+ chat_template = chat_template.replace(anchor, branch + anchor)
47
+ return chat_template
48
+
49
+
50
+ def _is_array(item) -> bool:
51
+ return hasattr(item, "ndim") and hasattr(item, "shape")
52
+
53
+
54
+ def _flatten_timeseries(timeseries) -> list[np.ndarray]:
55
+ """Flat list of items, one per placeholder: 1D series or 2D `(num_channels, length)` multivariate series."""
56
+ if _is_array(timeseries):
57
+ return [np.asarray(timeseries, dtype=np.float64)]
58
+ if not isinstance(timeseries, (list, tuple)):
59
+ raise ValueError(f"`timeseries` must be arrays or lists of numbers, got {type(timeseries)}.")
60
+ if len(timeseries) > 0 and all(isinstance(value, (int, float, np.number)) for value in timeseries):
61
+ return [np.asarray(timeseries, dtype=np.float64)]
62
+ return [item for nested in timeseries for item in _flatten_timeseries(nested)]
63
+
64
+
65
+ class Lfm3TsProcessor(ProcessorMixin):
66
+ r"""
67
+ Wraps an `Lfm3TsFeatureExtractor` and an LFM3 tokenizer.
68
+
69
+ Each `<|ts|>` in the text stands for one item of `timeseries`. The processor replaces it with the item's
70
+ normalization statistics in plain text, followed by as many `<|ts|>` tokens as the model will fill with features:
71
+
72
+ ```text
73
+ [offset: 12.5000 | scale: 3.2000 | length: 256 | min: 4.1000 | max: 21.7000 | first: 9.8000 | last: 15.2000]<|ts|>...<|ts|>
74
+ ```
75
+
76
+ The statistics keep the original values recoverable (the encoder only sees standardized values), and the
77
+ extremes and end points spare the model the arithmetic for the values most often asked about. A 2D item of
78
+ shape `(num_channels, length)` expands to one such block per channel, each on its own line.
79
+
80
+ Args:
81
+ feature_extractor (`Lfm3TsFeatureExtractor`):
82
+ The feature extractor.
83
+ tokenizer (`PreTrainedTokenizerBase`):
84
+ An LFM3 tokenizer holding the `<|ts|>` token.
85
+ chat_template (`str`, *optional*):
86
+ A chat template that renders `{"type": "timeseries"}` content parts as `<|ts|>`.
87
+ position_marker_every (`int`, *optional*, defaults to 0):
88
+ Writes the index of a time step, as text like `<64>`, before the token of the patch that starts there,
89
+ every this many time steps. Must
90
+ be a multiple of the patch size; 0 writes none.
91
+ value_ruler (`bool`, *optional*, defaults to `False`):
92
+ Also writes, in the statistics, the values that stand at -2, -1, 0, +1 and +2 in the standardized units
93
+ the encoder sees. A checkpoint keeps the setting it was trained with.
94
+ """
95
+
96
+ def __init__(
97
+ self,
98
+ feature_extractor=None,
99
+ tokenizer=None,
100
+ chat_template=None,
101
+ position_marker_every: int = 0,
102
+ value_ruler: bool = False,
103
+ **kwargs,
104
+ ):
105
+ if position_marker_every % feature_extractor.patch_size:
106
+ raise ValueError(
107
+ f"position_marker_every={position_marker_every} must be a multiple of the patch size "
108
+ f"{feature_extractor.patch_size}."
109
+ )
110
+ self.position_marker_every = position_marker_every
111
+ self.value_ruler = value_ruler
112
+ self.ts_token = getattr(tokenizer, "ts_token", TS_TOKEN)
113
+ self.ts_token_id = tokenizer.convert_tokens_to_ids(self.ts_token)
114
+ super().__init__(feature_extractor, tokenizer, chat_template=chat_template)
115
+
116
+ @classmethod
117
+ def from_text_pretrained(
118
+ cls,
119
+ pretrained_model_name_or_path: str = "LiquidAI/LFM3-1.7B",
120
+ patch_size: int = 8,
121
+ position_marker_every: int = 0,
122
+ value_ruler: bool = False,
123
+ **kwargs,
124
+ ) -> "Lfm3TsProcessor":
125
+ """Builds a processor from an LFM3 text checkpoint: its tokenizer gains the `<|ts|>` token and its chat
126
+ template a `timeseries` content part. `kwargs` go to the tokenizer's `from_pretrained`."""
127
+ from transformers import AutoTokenizer
128
+
129
+ tokenizer = AutoTokenizer.from_pretrained(
130
+ pretrained_model_name_or_path, extra_special_tokens={"ts_token": TS_TOKEN}, **kwargs
131
+ )
132
+ chat_template = add_timeseries_to_chat_template(tokenizer.chat_template)
133
+ tokenizer.chat_template = chat_template
134
+ return cls(
135
+ Lfm3TsFeatureExtractor(patch_size=patch_size),
136
+ tokenizer,
137
+ chat_template=chat_template,
138
+ position_marker_every=position_marker_every,
139
+ value_ruler=value_ruler,
140
+ )
141
+
142
+ def format_series(self, stats: Lfm3TsSeriesStats) -> str:
143
+ """The text standing for one series: its statistics, then a start, per-patch and end placeholder."""
144
+ header = (
145
+ f"[offset: {stats.offset:.4f} | scale: {stats.scale:.4f} | length: {stats.length}"
146
+ f" | min: {stats.minimum:.4f} | max: {stats.maximum:.4f}"
147
+ f" | first: {stats.first:.4f} | last: {stats.last:.4f}"
148
+ )
149
+ if self.value_ruler:
150
+ marks = ", ".join(f"{level:+d}: {stats.offset + level * stats.scale:.4f}" for level in (-2, -1, 0, 1, 2))
151
+ header += f" | levels {marks}"
152
+ header += "]"
153
+ if not self.position_marker_every:
154
+ return header + self.ts_token * (stats.num_patches + 2)
155
+ patch_size = self.feature_extractor.patch_size
156
+ pieces = [header, self.ts_token]
157
+ for patch in range(stats.num_patches):
158
+ index = patch * patch_size
159
+ if patch and index % self.position_marker_every == 0:
160
+ pieces.append(f"<{index}>")
161
+ pieces.append(self.ts_token)
162
+ return "".join(pieces) + self.ts_token
163
+
164
+ def __call__(self, text=None, timeseries=None, return_tensors=None, **kwargs) -> BatchFeature:
165
+ """
166
+ Args:
167
+ text (`str` or `list[str]`):
168
+ The prompts, with one `<|ts|>` per item of `timeseries`.
169
+ timeseries (*optional*):
170
+ A series, or a list of them (or one list per prompt), in the order of their placeholders. A 1D array
171
+ or list of numbers is one series; a 2D array of shape `(num_channels, length)` is a multivariate
172
+ series. NaN marks a missing value.
173
+ return_tensors (`str`, *optional*):
174
+ The tensor type to return, e.g. `"pt"`.
175
+ kwargs:
176
+ Passed to the tokenizer (`padding`, `truncation`, `add_special_tokens`, ...).
177
+
178
+ Returns:
179
+ `BatchFeature` with `input_ids`, `attention_mask` and, when `timeseries` is given, `input_values` and
180
+ `input_values_mask` of shape `(num_series, max_patches, patch_size)`.
181
+ """
182
+ if text is None:
183
+ raise ValueError("You must provide `text`.")
184
+ prompts = [text] if isinstance(text, str) else list(text)
185
+ items = [] if timeseries is None else _flatten_timeseries(timeseries)
186
+
187
+ num_placeholders = sum(prompt.count(self.ts_token) for prompt in prompts)
188
+ if num_placeholders != len(items):
189
+ raise ValueError(
190
+ f"Found {num_placeholders} {self.ts_token} tokens in the text but received {len(items)} time series."
191
+ )
192
+
193
+ data = {}
194
+ if items:
195
+ channels, channels_per_item = [], []
196
+ for index, item in enumerate(items):
197
+ if item.ndim not in (1, 2):
198
+ raise ValueError(f"Time series {index} must be 1D or 2D (num_channels, length), got {item.shape}.")
199
+ item_channels = [item] if item.ndim == 1 else list(item)
200
+ channels.extend(item_channels)
201
+ channels_per_item.append(len(item_channels))
202
+
203
+ input_values, input_values_mask, stats = self.feature_extractor.preprocess(channels)
204
+ data.update(input_values=input_values, input_values_mask=input_values_mask)
205
+
206
+ blocks = iter(self.format_series(stat) for stat in stats)
207
+ replacements = iter("\n".join(next(blocks) for _ in range(count)) for count in channels_per_item)
208
+ expanded = []
209
+ for prompt in prompts:
210
+ pieces = prompt.split(self.ts_token)
211
+ expanded.append("".join(piece + next(replacements) for piece in pieces[:-1]) + pieces[-1])
212
+ prompts = expanded
213
+
214
+ text_inputs = self.tokenizer(prompts, **kwargs)
215
+ return BatchFeature(data={**text_inputs, **data}, tensor_type=return_tensors)
216
+
217
+ def apply_chat_template(
218
+ self,
219
+ conversation,
220
+ chat_template: str | None = None,
221
+ tokenize: bool = False,
222
+ return_dict: bool = True,
223
+ add_generation_prompt: bool = False,
224
+ **kwargs,
225
+ ):
226
+ """
227
+ Renders one conversation, or a batch of them, and optionally tokenizes it together with its time series.
228
+
229
+ A time series is a content part `{"type": "timeseries", "timeseries": array}`. Template arguments (`tools`,
230
+ `documents`, `continue_final_message`, `reasoning_effort`, `enable_thinking`, `preserve_thinking`) are
231
+ forwarded to the template; other `kwargs` go to `__call__` when `tokenize=True`.
232
+ """
233
+ is_batched = isinstance(conversation, (list, tuple)) and isinstance(conversation[0], (list, tuple))
234
+ conversations = conversation if is_batched else [conversation]
235
+
236
+ timeseries, rendered_conversations = [], []
237
+ for messages in conversations:
238
+ rendered_messages = []
239
+ for message in messages:
240
+ content = message.get("content")
241
+ if isinstance(content, (list, tuple)):
242
+ parts = []
243
+ for part in content:
244
+ if isinstance(part, dict) and part.get("type") == "timeseries":
245
+ timeseries.append(part["timeseries"])
246
+ part = {"type": "timeseries"}
247
+ parts.append(part)
248
+ message = {**message, "content": parts}
249
+ rendered_messages.append(message)
250
+ rendered_conversations.append(rendered_messages)
251
+
252
+ template_keys = (
253
+ "tools",
254
+ "documents",
255
+ "continue_final_message",
256
+ "reasoning_effort",
257
+ "enable_thinking",
258
+ "preserve_thinking",
259
+ )
260
+ template_kwargs = {key: kwargs.pop(key) for key in template_keys if key in kwargs}
261
+ prompts = [
262
+ self.tokenizer.apply_chat_template(
263
+ messages,
264
+ chat_template=chat_template or self.chat_template,
265
+ tokenize=False,
266
+ add_generation_prompt=add_generation_prompt,
267
+ **template_kwargs,
268
+ )
269
+ for messages in rendered_conversations
270
+ ]
271
+ if not tokenize:
272
+ return prompts if is_batched else prompts[0]
273
+
274
+ # the template writes the special tokens itself
275
+ kwargs.setdefault("add_special_tokens", False)
276
+ inputs = self(text=prompts, timeseries=timeseries or None, **kwargs)
277
+ return inputs if return_dict else inputs["input_ids"]
278
+
279
+ @property
280
+ def model_input_names(self):
281
+ names = self.tokenizer.model_input_names + self.feature_extractor.model_input_names
282
+ return list(dict.fromkeys(names))
283
+
284
+
285
+ __all__ = ["Lfm3TsProcessor", "add_timeseries_to_chat_template"]
processor_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "auto_map": {
3
+ "AutoProcessor": "processing_lfm3_ts.Lfm3TsProcessor"
4
+ },
5
+ "feature_extractor": {
6
+ "auto_map": {
7
+ "AutoFeatureExtractor": "feature_extraction_lfm3_ts.Lfm3TsFeatureExtractor",
8
+ "AutoProcessor": "processing_lfm3_ts.Lfm3TsProcessor"
9
+ },
10
+ "eps": 1e-06,
11
+ "feature_extractor_type": "Lfm3TsFeatureExtractor",
12
+ "patch_size": 8
13
+ },
14
+ "position_marker_every": 0,
15
+ "processor_class": "Lfm3TsProcessor",
16
+ "value_ruler": false
17
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:753a4f0070f119707185658554bcedf8eb43d05cc952b5f670c562de8cfd0aee
3
+ size 17908817
tokenizer_config.json ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audio_out_token": "<|audio_out|>",
3
+ "audio_token": "<|audio|>",
4
+ "auto_map": {
5
+ "AutoProcessor": "processing_lfm3_ts.Lfm3TsProcessor"
6
+ },
7
+ "backend": "tokenizers",
8
+ "bos_token": "<|startoftext|>",
9
+ "eos_token": "<|im_end|>",
10
+ "image_token": "<|image|>",
11
+ "is_local": true,
12
+ "local_files_only": false,
13
+ "model_max_length": 1000000000000000019884624838656,
14
+ "model_specific_special_tokens": {
15
+ "audio_out_token": "<|audio_out|>",
16
+ "audio_token": "<|audio|>",
17
+ "image_token": "<|image|>",
18
+ "ts_token": "<|ts|>",
19
+ "video_token": "<|video|>",
20
+ "vision_end_token": "<|vision_end|>",
21
+ "vision_start_token": "<|vision_start|>"
22
+ },
23
+ "pad_token": "<|pad|>",
24
+ "processor_class": "Lfm3TsProcessor",
25
+ "response_template": {
26
+ "defaults": {
27
+ "role": "assistant"
28
+ },
29
+ "fields": {
30
+ "content": {
31
+ "close_pattern": "<\\|im_end\\|>\\s*",
32
+ "content": "text"
33
+ },
34
+ "reasoning_content": {
35
+ "close_pattern": "</think>\\s*",
36
+ "content": "text",
37
+ "open_pattern": "<think>\\n?"
38
+ },
39
+ "tool_calls": {
40
+ "close_pattern": "</tool_call>\\s*",
41
+ "content": "xml-inline",
42
+ "content_args": {
43
+ "tag_pattern": "<parameter=(?P<key>[^>\\n]+)>\\n?(?P<value>.*?)\\n?</parameter>(?=\\s*(?:<parameter=|</function>\\s*\\Z|\\Z))",
44
+ "value_parser": {
45
+ "args": {
46
+ "strip": false
47
+ },
48
+ "name": "text"
49
+ }
50
+ },
51
+ "open_pattern": "<tool_call>\\s*<function=(?P<name>[^\\n>]+)>",
52
+ "repeats": true,
53
+ "transform": {
54
+ "function": {
55
+ "arguments": "{content}",
56
+ "name": "{name}"
57
+ },
58
+ "type": "function"
59
+ }
60
+ }
61
+ },
62
+ "start_anchor_pattern": "<\\|im_start\\|>assistant\\n"
63
+ },
64
+ "tokenizer_class": "TokenizersBackend",
65
+ "ts_token": "<|ts|>",
66
+ "video_token": "<|video|>",
67
+ "vision_end_token": "<|vision_end|>",
68
+ "vision_start_token": "<|vision_start|>"
69
+ }