stneric commited on
Commit
ad9956e
·
verified ·
1 Parent(s): f5f7f05

v1: Qwen3.5-2B router distilled from Haiku 5.5 (checkpoint 350), training code, synthetic data, v2 plan

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +3 -0
  2. LICENSE +202 -0
  3. Modelfile +11 -0
  4. README.md +115 -0
  5. chat_template.jinja +154 -0
  6. config.json +96 -0
  7. generation_config.json +6 -0
  8. merges.txt +0 -0
  9. model.safetensors +3 -0
  10. preprocessor_config.json +21 -0
  11. tokenizer.json +3 -0
  12. tokenizer_config.json +305 -0
  13. training/build_prompts.py +106 -0
  14. training/data/labels.jsonl +0 -0
  15. training/data/prompts.jsonl +3 -0
  16. training/data/relabel_syn_labels.jsonl +150 -0
  17. training/data/synthetic_questions.jsonl +0 -0
  18. training/data/train.jsonl +3 -0
  19. training/data/valid.jsonl +0 -0
  20. training/eval/dev-syn-2b-v1-s350.jsonl +150 -0
  21. training/eval/dev-syn-2b-v1-s700.jsonl +0 -0
  22. training/eval/dump_prompt.py +22 -0
  23. training/eval/eval.py +135 -0
  24. training/eval/eval_dev.py +102 -0
  25. training/eval/make_testset.py +311 -0
  26. training/eval/report.py +21 -0
  27. training/eval/rescore.py +12 -0
  28. training/eval/results-2b-v1-s350.jsonl +0 -0
  29. training/eval/results-2b-v1-s700.jsonl +0 -0
  30. training/eval/results-haiku-5.5-think.jsonl +0 -0
  31. training/eval/results-haiku-5.5.jsonl +0 -0
  32. training/eval/results-qwen3.5-0.8b.jsonl +0 -0
  33. training/eval/results-qwen3.5-2b-q4_K_M.jsonl +0 -0
  34. training/eval/results-qwen3.5-2b.jsonl +0 -0
  35. training/eval/results-qwen3.5-4b.jsonl +0 -0
  36. training/eval/schema.json +65 -0
  37. training/eval/system_prompt.txt +220 -0
  38. training/eval/testset.jsonl +105 -0
  39. training/export_ckpt.py +42 -0
  40. training/extract_real.py +147 -0
  41. training/gen_questions.py +259 -0
  42. training/label.py +72 -0
  43. training/make_dataset.py +132 -0
  44. training/orclient.py +60 -0
  45. training/pyproject.toml +24 -0
  46. training/run_train.sh +17 -0
  47. training/to_ollama.sh +18 -0
  48. training/train.py +268 -0
  49. training/uv.lock +971 -0
  50. v2_plan.md +93 -0
.gitattributes CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
+ training/data/prompts.jsonl filter=lfs diff=lfs merge=lfs -text
38
+ training/data/train.jsonl filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ Apache License
3
+ Version 2.0, January 2004
4
+ http://www.apache.org/licenses/
5
+
6
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
7
+
8
+ 1. Definitions.
9
+
10
+ "License" shall mean the terms and conditions for use, reproduction,
11
+ and distribution as defined by Sections 1 through 9 of this document.
12
+
13
+ "Licensor" shall mean the copyright owner or entity authorized by
14
+ the copyright owner that is granting the License.
15
+
16
+ "Legal Entity" shall mean the union of the acting entity and all
17
+ other entities that control, are controlled by, or are under common
18
+ control with that entity. For the purposes of this definition,
19
+ "control" means (i) the power, direct or indirect, to cause the
20
+ direction or management of such entity, whether by contract or
21
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
22
+ outstanding shares, or (iii) beneficial ownership of such entity.
23
+
24
+ "You" (or "Your") shall mean an individual or Legal Entity
25
+ exercising permissions granted by this License.
26
+
27
+ "Source" form shall mean the preferred form for making modifications,
28
+ including but not limited to software source code, documentation
29
+ source, and configuration files.
30
+
31
+ "Object" form shall mean any form resulting from mechanical
32
+ transformation or translation of a Source form, including but
33
+ not limited to compiled object code, generated documentation,
34
+ and conversions to other media types.
35
+
36
+ "Work" shall mean the work of authorship, whether in Source or
37
+ Object form, made available under the License, as indicated by a
38
+ copyright notice that is included in or attached to the work
39
+ (an example is provided in the Appendix below).
40
+
41
+ "Derivative Works" shall mean any work, whether in Source or Object
42
+ form, that is based on (or derived from) the Work and for which the
43
+ editorial revisions, annotations, elaborations, or other modifications
44
+ represent, as a whole, an original work of authorship. For the purposes
45
+ of this License, Derivative Works shall not include works that remain
46
+ separable from, or merely link (or bind by name) to the interfaces of,
47
+ the Work and Derivative Works thereof.
48
+
49
+ "Contribution" shall mean any work of authorship, including
50
+ the original version of the Work and any modifications or additions
51
+ to that Work or Derivative Works thereof, that is intentionally
52
+ submitted to Licensor for inclusion in the Work by the copyright owner
53
+ or by an individual or Legal Entity authorized to submit on behalf of
54
+ the copyright owner. For the purposes of this definition, "submitted"
55
+ means any form of electronic, verbal, or written communication sent
56
+ to the Licensor or its representatives, including but not limited to
57
+ communication on electronic mailing lists, source code control systems,
58
+ and issue tracking systems that are managed by, or on behalf of, the
59
+ Licensor for the purpose of discussing and improving the Work, but
60
+ excluding communication that is conspicuously marked or otherwise
61
+ designated in writing by the copyright owner as "Not a Contribution."
62
+
63
+ "Contributor" shall mean Licensor and any individual or Legal Entity
64
+ on behalf of whom a Contribution has been received by Licensor and
65
+ subsequently incorporated within the Work.
66
+
67
+ 2. Grant of Copyright License. Subject to the terms and conditions of
68
+ this License, each Contributor hereby grants to You a perpetual,
69
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
70
+ copyright license to reproduce, prepare Derivative Works of,
71
+ publicly display, publicly perform, sublicense, and distribute the
72
+ Work and such Derivative Works in Source or Object form.
73
+
74
+ 3. Grant of Patent License. Subject to the terms and conditions of
75
+ this License, each Contributor hereby grants to You a perpetual,
76
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
77
+ (except as stated in this section) patent license to make, have made,
78
+ use, offer to sell, sell, import, and otherwise transfer the Work,
79
+ where such license applies only to those patent claims licensable
80
+ by such Contributor that are necessarily infringed by their
81
+ Contribution(s) alone or by combination of their Contribution(s)
82
+ with the Work to which such Contribution(s) was submitted. If You
83
+ institute patent litigation against any entity (including a
84
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
85
+ or a Contribution incorporated within the Work constitutes direct
86
+ or contributory patent infringement, then any patent licenses
87
+ granted to You under this License for that Work shall terminate
88
+ as of the date such litigation is filed.
89
+
90
+ 4. Redistribution. You may reproduce and distribute copies of the
91
+ Work or Derivative Works thereof in any medium, with or without
92
+ modifications, and in Source or Object form, provided that You
93
+ meet the following conditions:
94
+
95
+ (a) You must give any other recipients of the Work or
96
+ Derivative Works a copy of this License; and
97
+
98
+ (b) You must cause any modified files to carry prominent notices
99
+ stating that You changed the files; and
100
+
101
+ (c) You must retain, in the Source form of any Derivative Works
102
+ that You distribute, all copyright, patent, trademark, and
103
+ attribution notices from the Source form of the Work,
104
+ excluding those notices that do not pertain to any part of
105
+ the Derivative Works; and
106
+
107
+ (d) If the Work includes a "NOTICE" text file as part of its
108
+ distribution, then any Derivative Works that You distribute must
109
+ include a readable copy of the attribution notices contained
110
+ within such NOTICE file, excluding those notices that do not
111
+ pertain to any part of the Derivative Works, in at least one
112
+ of the following places: within a NOTICE text file distributed
113
+ as part of the Derivative Works; within the Source form or
114
+ documentation, if provided along with the Derivative Works; or,
115
+ within a display generated by the Derivative Works, if and
116
+ wherever such third-party notices normally appear. The contents
117
+ of the NOTICE file are for informational purposes only and
118
+ do not modify the License. You may add Your own attribution
119
+ notices within Derivative Works that You distribute, alongside
120
+ or as an addendum to the NOTICE text from the Work, provided
121
+ that such additional attribution notices cannot be construed
122
+ as modifying the License.
123
+
124
+ You may add Your own copyright statement to Your modifications and
125
+ may provide additional or different license terms and conditions
126
+ for use, reproduction, or distribution of Your modifications, or
127
+ for any such Derivative Works as a whole, provided Your use,
128
+ reproduction, and distribution of the Work otherwise complies with
129
+ the conditions stated in this License.
130
+
131
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
132
+ any Contribution intentionally submitted for inclusion in the Work
133
+ by You to the Licensor shall be under the terms and conditions of
134
+ this License, without any additional terms or conditions.
135
+ Notwithstanding the above, nothing herein shall supersede or modify
136
+ the terms of any separate license agreement you may have executed
137
+ with Licensor regarding such Contributions.
138
+
139
+ 6. Trademarks. This License does not grant permission to use the trade
140
+ names, trademarks, service marks, or product names of the Licensor,
141
+ except as required for reasonable and customary use in describing the
142
+ origin of the Work and reproducing the content of the NOTICE file.
143
+
144
+ 7. Disclaimer of Warranty. Unless required by applicable law or
145
+ agreed to in writing, Licensor provides the Work (and each
146
+ Contributor provides its Contributions) on an "AS IS" BASIS,
147
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
148
+ implied, including, without limitation, any warranties or conditions
149
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
150
+ PARTICULAR PURPOSE. You are solely responsible for determining the
151
+ appropriateness of using or redistributing the Work and assume any
152
+ risks associated with Your exercise of permissions under this License.
153
+
154
+ 8. Limitation of Liability. In no event and under no legal theory,
155
+ whether in tort (including negligence), contract, or otherwise,
156
+ unless required by applicable law (such as deliberate and grossly
157
+ negligent acts) or agreed to in writing, shall any Contributor be
158
+ liable to You for damages, including any direct, indirect, special,
159
+ incidental, or consequential damages of any character arising as a
160
+ result of this License or out of the use or inability to use the
161
+ Work (including but not limited to damages for loss of goodwill,
162
+ work stoppage, computer failure or malfunction, or any and all
163
+ other commercial damages or losses), even if such Contributor
164
+ has been advised of the possibility of such damages.
165
+
166
+ 9. Accepting Warranty or Additional Liability. While redistributing
167
+ the Work or Derivative Works thereof, You may choose to offer,
168
+ and charge a fee for, acceptance of support, warranty, indemnity,
169
+ or other liability obligations and/or rights consistent with this
170
+ License. However, in accepting such obligations, You may act only
171
+ on Your own behalf and on Your sole responsibility, not on behalf
172
+ of any other Contributor, and only if You agree to indemnify,
173
+ defend, and hold each Contributor harmless for any liability
174
+ incurred by, or claims asserted against, such Contributor by reason
175
+ of your accepting any such warranty or additional liability.
176
+
177
+ END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright 2026 Alibaba Cloud
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
Modelfile ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Ollama with the MLX runtime (macOS): from this folder, run
2
+ # ollama create simplixity-router:2b-v1-q4_k_m -f Modelfile --quantize q4_K_M
3
+ # Ollama on Linux and Windows cannot import safetensors: use the GGUF file from
4
+ # stneric/simplixity-router-2b-GGUF and replace "FROM ." with its full path.
5
+ # RENDERER and PARSER are Ollama's Qwen3.5 chat format. Without them the system
6
+ # prompt never reaches the model. Simplixity sets no context length, and the
7
+ # router prompt with a chat history runs past Ollama's default, hence num_ctx.
8
+ FROM .
9
+ RENDERER qwen3.5
10
+ PARSER qwen3.5
11
+ PARAMETER num_ctx 8192
README.md ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen3.5-2B
4
+ language:
5
+ - en
6
+ - de
7
+ pipeline_tag: text-generation
8
+ library_name: transformers
9
+ tags:
10
+ - classification
11
+ - router
12
+ - lora
13
+ - qwen3.5
14
+ - simplixity
15
+ ---
16
+
17
+ # simplixity-router-2b
18
+
19
+ [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) fine-tuned as the router of [Simplixity](https://github.com/stneric/simplixity), a self-hosted research assistant. The router reads a question, in English or German and with the chat so far, and fills Simplixity's `Classification` schema as JSON. It decides:
20
+
21
+ - whether to search, and with which queries
22
+ - how much research the question needs
23
+ - the shape of the answer
24
+ - how recent the sources must be
25
+ - which sources to search
26
+
27
+ This is **v1**, distilled from Claude Haiku 5.5 with thinking turned on. It is meant to run locally as Q4_K_M, from [simplixity-router-2b-GGUF](https://huggingface.co/stneric/simplixity-router-2b-GGUF).
28
+
29
+ ## Use with Ollama
30
+
31
+ On Linux and Windows, Ollama cannot import safetensors. Download the GGUF file from [simplixity-router-2b-GGUF](https://huggingface.co/stneric/simplixity-router-2b-GGUF) and follow the instructions there.
32
+
33
+ On macOS, from a download of this repo:
34
+
35
+ ollama create simplixity-router:2b-v1-q4_k_m -f Modelfile --quantize q4_K_M
36
+
37
+ In Simplixity, choose `simplixity-router:2b-v1-q4_k_m` as the router model with provider Ollama.
38
+
39
+ **Also set a title model.** Without one, Simplixity names new chats with the router, and this model was not trained to write titles. [simplixity-title-0.8b](https://huggingface.co/stneric/simplixity-title-0.8b) is made for that job.
40
+
41
+ `PARAMETER num_ctx 8192` in the Modelfile matters. Simplixity sets no context length, and a router prompt with a chat history runs past Ollama's default.
42
+
43
+ ## Prompt
44
+
45
+ The model was trained on Simplixity's own classifier prompt, built by the app's code (`app/agent/prompts.classify`), and should get exactly that prompt. Simplixity asks for strict JSON with the hardened `Classification` schema, at most 500 tokens, with thinking off.
46
+
47
+ - `training/eval/system_prompt.txt` holds one example of the system prompt.
48
+ - `training/eval/schema.json` holds the schema.
49
+
50
+ The prompt varies with the date, the depth (2 or 3 search queries), the sources (Auto, where the router picks them, or a fixed selection), a library catalog, attached images, a linked page and the chat history. Training covered all of these.
51
+
52
+ ## Results
53
+
54
+ ### Testset
55
+
56
+ 105 hand-labelled questions (52 English, 53 German), run in Ollama 0.40 with temperature 0 and `num_gpu 99`. "Route" means that search yes or no, complexity and answer shape are all right.
57
+
58
+ | | Route | Complexity | Shape | Recency | Sources | Median latency (RTX 3070 Ti) |
59
+ | --- | --- | --- | --- | --- | --- | --- |
60
+ | Qwen3.5-2B Q4_K_M, untuned | 38% | 71% | 59% | 91% | 81% | 1.24 s |
61
+ | **simplixity-router-2b v1 Q4_K_M** | **96%** | **98%** | **98%** | **98%** | **97%** | **0.53 s** |
62
+ | Claude Haiku 5.5, thinking | 99% | 99% | 100% | 99% | 100% | 3.0 s (OpenRouter) |
63
+
64
+ v1 always writes valid JSON with the allowed number of queries. The untuned model wrote facet names into `answer_shape` and set recency to "year" almost everywhere. v1 does neither.
65
+
66
+ ### Agreement with the teacher
67
+
68
+ Held-out questions, measured as route agreement with Haiku. To show how much of the gap is noise, Haiku labelled the same questions a second time, and that run is compared with the first.
69
+
70
+ | Dev set | Haiku vs Haiku | v1 vs Haiku |
71
+ | --- | --- | --- |
72
+ | Synthetic, 150 questions | 91% | 66% |
73
+ | Real questions from the author's own instance, 69 | 79% | 75% |
74
+
75
+ Almost every disagreement is between neighbouring levels:
76
+
77
+ - complexity: standard vs research, simple vs standard
78
+ - shape: no shape vs explainer, steps or profile
79
+
80
+ `v2_plan.md` describes the next run, which goes after this gap.
81
+
82
+ ## Limits
83
+
84
+ - **Gray-zone questions.** On questions where Haiku itself is unsure, v1 often decides differently.
85
+ - **Repetition loops.** On 1 of 219 dev questions, v1 fell into a repetition loop and Ollama aborted the answer with "token repeat limit reached". Simplixity treats a failed classification as "search the question as it is", so the run still goes on.
86
+ - **No titles.** The model writes the router's JSON only, not titles.
87
+
88
+ ## Training
89
+
90
+ - **Data:** 2,803 synthetic questions for training and 150 for validation, half German and half English.
91
+ - Haiku wrote them from 37 categories, in different writing styles and with follow-ups, images, linked pages and library chats.
92
+ - Questions too close to the testset were dropped (difflib ratio above 0.85).
93
+ - Each question was turned into the exact prompt the app would send, with sampled settings, and labelled by Haiku 5.5 with a 1,024-token thinking budget.
94
+ - Labels that break a rule of the app or the prompt were dropped, not repaired: 23 in all.
95
+ - Fields the app ignores in a given case were emptied: `place` outside weather questions, `lenses` without Auto sources and `image_query` without a wanted image.
96
+ - Everything is in `training/data/` and was built by the scripts in `training/`, in this order: `gen_questions.py`, `build_prompts.py`, `label.py`, `make_dataset.py`.
97
+ - **Not published:** 145 real questions from the author's own Simplixity instance, shown twice each, were also part of the training. They and the 69 real dev questions stay private. `training/extract_real.py` shows how they were taken out of the database.
98
+ - **Method:**
99
+ - QLoRA: the base model in NF4 with double quantization and bf16 compute, and fp32 adapters.
100
+ - LoRA rank 16, alpha 320 and dropout 0, on the 186 linear layers of the language model.
101
+ - Loss on the JSON answer only.
102
+ - Batch 1 with gradient accumulation 8, constant learning rate 2e-5.
103
+ - The vision tower was taken off the GPU during training, because the router never sees an image.
104
+ - Code: `training/train.py` and `training/run_train.sh`.
105
+ - This model is the checkpoint at step 350 (0.9 epochs), where the eval loss was lowest. After that the run overfit: by step 700 it rated simple questions as standard more often.
106
+ - **Export:** the LoRA deltas were added to the original bf16 weights, so the vision tower and the `mtp.*` tensors are unchanged (`training/export_ckpt.py`). GGUF: llama.cpp's `convert_hf_to_gguf.py`, then `llama-quantize` to Q4_K_M (`training/to_ollama.sh`).
107
+ - **Evaluation:**
108
+ - Testset and baselines: `training/eval/make_testset.py`, `eval.py` and `report.py`.
109
+ - Agreement with the teacher: `eval_dev.py`.
110
+ - Stored results: `results-*.jsonl` and `dev-syn-*.jsonl`.
111
+ - The second Haiku labelling: `training/data/relabel_syn_labels.jsonl`.
112
+
113
+ ## Licence
114
+
115
+ Apache 2.0, like the base model.
chat_template.jinja ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- set image_count = namespace(value=0) %}
2
+ {%- set video_count = namespace(value=0) %}
3
+ {%- macro render_content(content, do_vision_count, is_system_content=false) %}
4
+ {%- if content is string %}
5
+ {{- content }}
6
+ {%- elif content is iterable and content is not mapping %}
7
+ {%- for item in content %}
8
+ {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
9
+ {%- if is_system_content %}
10
+ {{- raise_exception('System message cannot contain images.') }}
11
+ {%- endif %}
12
+ {%- if do_vision_count %}
13
+ {%- set image_count.value = image_count.value + 1 %}
14
+ {%- endif %}
15
+ {%- if add_vision_id %}
16
+ {{- 'Picture ' ~ image_count.value ~ ': ' }}
17
+ {%- endif %}
18
+ {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
19
+ {%- elif 'video' in item or item.type == 'video' %}
20
+ {%- if is_system_content %}
21
+ {{- raise_exception('System message cannot contain videos.') }}
22
+ {%- endif %}
23
+ {%- if do_vision_count %}
24
+ {%- set video_count.value = video_count.value + 1 %}
25
+ {%- endif %}
26
+ {%- if add_vision_id %}
27
+ {{- 'Video ' ~ video_count.value ~ ': ' }}
28
+ {%- endif %}
29
+ {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
30
+ {%- elif 'text' in item %}
31
+ {{- item.text }}
32
+ {%- else %}
33
+ {{- raise_exception('Unexpected item type in content.') }}
34
+ {%- endif %}
35
+ {%- endfor %}
36
+ {%- elif content is none or content is undefined %}
37
+ {{- '' }}
38
+ {%- else %}
39
+ {{- raise_exception('Unexpected content type.') }}
40
+ {%- endif %}
41
+ {%- endmacro %}
42
+ {%- if not messages %}
43
+ {{- raise_exception('No messages provided.') }}
44
+ {%- endif %}
45
+ {%- if tools and tools is iterable and tools is not mapping %}
46
+ {{- '<|im_start|>system\n' }}
47
+ {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
48
+ {%- for tool in tools %}
49
+ {{- "\n" }}
50
+ {{- tool | tojson }}
51
+ {%- endfor %}
52
+ {{- "\n</tools>" }}
53
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
54
+ {%- if messages[0].role == 'system' %}
55
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
56
+ {%- if content %}
57
+ {{- '\n\n' + content }}
58
+ {%- endif %}
59
+ {%- endif %}
60
+ {{- '<|im_end|>\n' }}
61
+ {%- else %}
62
+ {%- if messages[0].role == 'system' %}
63
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
64
+ {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
65
+ {%- endif %}
66
+ {%- endif %}
67
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
68
+ {%- for message in messages[::-1] %}
69
+ {%- set index = (messages|length - 1) - loop.index0 %}
70
+ {%- if ns.multi_step_tool and message.role == "user" %}
71
+ {%- set content = render_content(message.content, false)|trim %}
72
+ {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
73
+ {%- set ns.multi_step_tool = false %}
74
+ {%- set ns.last_query_index = index %}
75
+ {%- endif %}
76
+ {%- endif %}
77
+ {%- endfor %}
78
+ {%- if ns.multi_step_tool %}
79
+ {{- raise_exception('No user query found in messages.') }}
80
+ {%- endif %}
81
+ {%- for message in messages %}
82
+ {%- set content = render_content(message.content, true)|trim %}
83
+ {%- if message.role == "system" %}
84
+ {%- if not loop.first %}
85
+ {{- raise_exception('System message must be at the beginning.') }}
86
+ {%- endif %}
87
+ {%- elif message.role == "user" %}
88
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
89
+ {%- elif message.role == "assistant" %}
90
+ {%- set reasoning_content = '' %}
91
+ {%- if message.reasoning_content is string %}
92
+ {%- set reasoning_content = message.reasoning_content %}
93
+ {%- else %}
94
+ {%- if '</think>' in content %}
95
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
96
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
97
+ {%- endif %}
98
+ {%- endif %}
99
+ {%- set reasoning_content = reasoning_content|trim %}
100
+ {%- if loop.index0 > ns.last_query_index %}
101
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
102
+ {%- else %}
103
+ {{- '<|im_start|>' + message.role + '\n' + content }}
104
+ {%- endif %}
105
+ {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
106
+ {%- for tool_call in message.tool_calls %}
107
+ {%- if tool_call.function is defined %}
108
+ {%- set tool_call = tool_call.function %}
109
+ {%- endif %}
110
+ {%- if loop.first %}
111
+ {%- if content|trim %}
112
+ {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
113
+ {%- else %}
114
+ {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
115
+ {%- endif %}
116
+ {%- else %}
117
+ {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
118
+ {%- endif %}
119
+ {%- if tool_call.arguments is defined %}
120
+ {%- for args_name, args_value in tool_call.arguments|items %}
121
+ {{- '<parameter=' + args_name + '>\n' }}
122
+ {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
123
+ {{- args_value }}
124
+ {{- '\n</parameter>\n' }}
125
+ {%- endfor %}
126
+ {%- endif %}
127
+ {{- '</function>\n</tool_call>' }}
128
+ {%- endfor %}
129
+ {%- endif %}
130
+ {{- '<|im_end|>\n' }}
131
+ {%- elif message.role == "tool" %}
132
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
133
+ {{- '<|im_start|>user' }}
134
+ {%- endif %}
135
+ {{- '\n<tool_response>\n' }}
136
+ {{- content }}
137
+ {{- '\n</tool_response>' }}
138
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
139
+ {{- '<|im_end|>\n' }}
140
+ {%- elif loop.last %}
141
+ {{- '<|im_end|>\n' }}
142
+ {%- endif %}
143
+ {%- else %}
144
+ {{- raise_exception('Unexpected message role.') }}
145
+ {%- endif %}
146
+ {%- endfor %}
147
+ {%- if add_generation_prompt %}
148
+ {{- '<|im_start|>assistant\n' }}
149
+ {%- if enable_thinking is defined and enable_thinking is true %}
150
+ {{- '<think>\n' }}
151
+ {%- else %}
152
+ {{- '<think>\n\n</think>\n\n' }}
153
+ {%- endif %}
154
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3_5ForConditionalGeneration"
4
+ ],
5
+ "image_token_id": 248056,
6
+ "model_type": "qwen3_5",
7
+ "text_config": {
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "attn_output_gate": true,
11
+ "dtype": "bfloat16",
12
+ "eos_token_id": 248044,
13
+ "full_attention_interval": 4,
14
+ "head_dim": 256,
15
+ "hidden_act": "silu",
16
+ "hidden_size": 2048,
17
+ "initializer_range": 0.02,
18
+ "intermediate_size": 6144,
19
+ "layer_types": [
20
+ "linear_attention",
21
+ "linear_attention",
22
+ "linear_attention",
23
+ "full_attention",
24
+ "linear_attention",
25
+ "linear_attention",
26
+ "linear_attention",
27
+ "full_attention",
28
+ "linear_attention",
29
+ "linear_attention",
30
+ "linear_attention",
31
+ "full_attention",
32
+ "linear_attention",
33
+ "linear_attention",
34
+ "linear_attention",
35
+ "full_attention",
36
+ "linear_attention",
37
+ "linear_attention",
38
+ "linear_attention",
39
+ "full_attention",
40
+ "linear_attention",
41
+ "linear_attention",
42
+ "linear_attention",
43
+ "full_attention"
44
+ ],
45
+ "linear_conv_kernel_dim": 4,
46
+ "linear_key_head_dim": 128,
47
+ "linear_num_key_heads": 16,
48
+ "linear_num_value_heads": 16,
49
+ "linear_value_head_dim": 128,
50
+ "max_position_embeddings": 262144,
51
+ "mlp_only_layers": [],
52
+ "model_type": "qwen3_5_text",
53
+ "mtp_num_hidden_layers": 1,
54
+ "mtp_use_dedicated_embeddings": false,
55
+ "num_attention_heads": 8,
56
+ "num_hidden_layers": 24,
57
+ "num_key_value_heads": 2,
58
+ "rms_norm_eps": 1e-06,
59
+ "tie_word_embeddings": true,
60
+ "use_cache": true,
61
+ "vocab_size": 248320,
62
+ "mamba_ssm_dtype": "float32",
63
+ "rope_parameters": {
64
+ "mrope_interleaved": true,
65
+ "mrope_section": [
66
+ 11,
67
+ 11,
68
+ 10
69
+ ],
70
+ "rope_type": "default",
71
+ "rope_theta": 10000000,
72
+ "partial_rotary_factor": 0.25
73
+ }
74
+ },
75
+ "tie_word_embeddings": true,
76
+ "transformers_version": "4.57.0.dev0",
77
+ "video_token_id": 248057,
78
+ "vision_config": {
79
+ "deepstack_visual_indexes": [],
80
+ "depth": 24,
81
+ "hidden_act": "gelu_pytorch_tanh",
82
+ "hidden_size": 1024,
83
+ "in_channels": 3,
84
+ "initializer_range": 0.02,
85
+ "intermediate_size": 4096,
86
+ "model_type": "qwen3_5",
87
+ "num_heads": 16,
88
+ "num_position_embeddings": 2304,
89
+ "out_hidden_size": 2048,
90
+ "patch_size": 16,
91
+ "spatial_merge_size": 2,
92
+ "temporal_patch_size": 2
93
+ },
94
+ "vision_end_token_id": 248054,
95
+ "vision_start_token_id": 248053
96
+ }
generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "eos_token_id": [248046, 248044],
3
+ "pad_token_id": 248044,
4
+ "do_sample": false,
5
+ "max_new_tokens": 500
6
+ }
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5dcf4332128c128d3f9fe0d6f9cc0de909328ef5afabb3ccf9e597bb0e34016c
3
+ size 4548221520
preprocessor_config.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": {
3
+ "longest_edge": 16777216,
4
+ "shortest_edge": 65536
5
+ },
6
+ "patch_size": 16,
7
+ "temporal_patch_size": 2,
8
+ "merge_size": 2,
9
+ "image_mean": [
10
+ 0.5,
11
+ 0.5,
12
+ 0.5
13
+ ],
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "processor_class": "Qwen3VLProcessor",
20
+ "image_processor_type": "Qwen2VLImageProcessorFast"
21
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42
3
+ size 12807982
tokenizer_config.json ADDED
@@ -0,0 +1,305 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "added_tokens_decoder": {
4
+ "248044": {
5
+ "content": "<|endoftext|>",
6
+ "lstrip": false,
7
+ "normalized": false,
8
+ "rstrip": false,
9
+ "single_word": false,
10
+ "special": true
11
+ },
12
+ "248045": {
13
+ "content": "<|im_start|>",
14
+ "lstrip": false,
15
+ "normalized": false,
16
+ "rstrip": false,
17
+ "single_word": false,
18
+ "special": true
19
+ },
20
+ "248046": {
21
+ "content": "<|im_end|>",
22
+ "lstrip": false,
23
+ "normalized": false,
24
+ "rstrip": false,
25
+ "single_word": false,
26
+ "special": true
27
+ },
28
+ "248047": {
29
+ "content": "<|object_ref_start|>",
30
+ "lstrip": false,
31
+ "normalized": false,
32
+ "rstrip": false,
33
+ "single_word": false,
34
+ "special": true
35
+ },
36
+ "248048": {
37
+ "content": "<|object_ref_end|>",
38
+ "lstrip": false,
39
+ "normalized": false,
40
+ "rstrip": false,
41
+ "single_word": false,
42
+ "special": true
43
+ },
44
+ "248049": {
45
+ "content": "<|box_start|>",
46
+ "lstrip": false,
47
+ "normalized": false,
48
+ "rstrip": false,
49
+ "single_word": false,
50
+ "special": true
51
+ },
52
+ "248050": {
53
+ "content": "<|box_end|>",
54
+ "lstrip": false,
55
+ "normalized": false,
56
+ "rstrip": false,
57
+ "single_word": false,
58
+ "special": true
59
+ },
60
+ "248051": {
61
+ "content": "<|quad_start|>",
62
+ "lstrip": false,
63
+ "normalized": false,
64
+ "rstrip": false,
65
+ "single_word": false,
66
+ "special": true
67
+ },
68
+ "248052": {
69
+ "content": "<|quad_end|>",
70
+ "lstrip": false,
71
+ "normalized": false,
72
+ "rstrip": false,
73
+ "single_word": false,
74
+ "special": true
75
+ },
76
+ "248053": {
77
+ "content": "<|vision_start|>",
78
+ "lstrip": false,
79
+ "normalized": false,
80
+ "rstrip": false,
81
+ "single_word": false,
82
+ "special": true
83
+ },
84
+ "248054": {
85
+ "content": "<|vision_end|>",
86
+ "lstrip": false,
87
+ "normalized": false,
88
+ "rstrip": false,
89
+ "single_word": false,
90
+ "special": true
91
+ },
92
+ "248055": {
93
+ "content": "<|vision_pad|>",
94
+ "lstrip": false,
95
+ "normalized": false,
96
+ "rstrip": false,
97
+ "single_word": false,
98
+ "special": true
99
+ },
100
+ "248056": {
101
+ "content": "<|image_pad|>",
102
+ "lstrip": false,
103
+ "normalized": false,
104
+ "rstrip": false,
105
+ "single_word": false,
106
+ "special": true
107
+ },
108
+ "248057": {
109
+ "content": "<|video_pad|>",
110
+ "lstrip": false,
111
+ "normalized": false,
112
+ "rstrip": false,
113
+ "single_word": false,
114
+ "special": true
115
+ },
116
+ "248058": {
117
+ "content": "<tool_call>",
118
+ "lstrip": false,
119
+ "normalized": false,
120
+ "rstrip": false,
121
+ "single_word": false,
122
+ "special": false
123
+ },
124
+ "248059": {
125
+ "content": "</tool_call>",
126
+ "lstrip": false,
127
+ "normalized": false,
128
+ "rstrip": false,
129
+ "single_word": false,
130
+ "special": false
131
+ },
132
+ "248060": {
133
+ "content": "<|fim_prefix|>",
134
+ "lstrip": false,
135
+ "normalized": false,
136
+ "rstrip": false,
137
+ "single_word": false,
138
+ "special": false
139
+ },
140
+ "248061": {
141
+ "content": "<|fim_middle|>",
142
+ "lstrip": false,
143
+ "normalized": false,
144
+ "rstrip": false,
145
+ "single_word": false,
146
+ "special": false
147
+ },
148
+ "248062": {
149
+ "content": "<|fim_suffix|>",
150
+ "lstrip": false,
151
+ "normalized": false,
152
+ "rstrip": false,
153
+ "single_word": false,
154
+ "special": false
155
+ },
156
+ "248063": {
157
+ "content": "<|fim_pad|>",
158
+ "lstrip": false,
159
+ "normalized": false,
160
+ "rstrip": false,
161
+ "single_word": false,
162
+ "special": false
163
+ },
164
+ "248064": {
165
+ "content": "<|repo_name|>",
166
+ "lstrip": false,
167
+ "normalized": false,
168
+ "rstrip": false,
169
+ "single_word": false,
170
+ "special": false
171
+ },
172
+ "248065": {
173
+ "content": "<|file_sep|>",
174
+ "lstrip": false,
175
+ "normalized": false,
176
+ "rstrip": false,
177
+ "single_word": false,
178
+ "special": false
179
+ },
180
+ "248066": {
181
+ "content": "<tool_response>",
182
+ "lstrip": false,
183
+ "normalized": false,
184
+ "rstrip": false,
185
+ "single_word": false,
186
+ "special": false
187
+ },
188
+ "248067": {
189
+ "content": "</tool_response>",
190
+ "lstrip": false,
191
+ "normalized": false,
192
+ "rstrip": false,
193
+ "single_word": false,
194
+ "special": false
195
+ },
196
+ "248068": {
197
+ "content": "<think>",
198
+ "lstrip": false,
199
+ "normalized": false,
200
+ "rstrip": false,
201
+ "single_word": false,
202
+ "special": false
203
+ },
204
+ "248069": {
205
+ "content": "</think>",
206
+ "lstrip": false,
207
+ "normalized": false,
208
+ "rstrip": false,
209
+ "single_word": false,
210
+ "special": false
211
+ },
212
+ "248070": {
213
+ "content": "<|audio_start|>",
214
+ "lstrip": false,
215
+ "normalized": false,
216
+ "rstrip": false,
217
+ "single_word": false,
218
+ "special": true
219
+ },
220
+ "248071": {
221
+ "content": "<|audio_end|>",
222
+ "lstrip": false,
223
+ "normalized": false,
224
+ "rstrip": false,
225
+ "single_word": false,
226
+ "special": true
227
+ },
228
+ "248072": {
229
+ "content": "<tts_pad>",
230
+ "lstrip": false,
231
+ "normalized": false,
232
+ "rstrip": false,
233
+ "single_word": false,
234
+ "special": true
235
+ },
236
+ "248073": {
237
+ "content": "<tts_text_bos>",
238
+ "lstrip": false,
239
+ "normalized": false,
240
+ "rstrip": false,
241
+ "single_word": false,
242
+ "special": true
243
+ },
244
+ "248074": {
245
+ "content": "<tts_text_eod>",
246
+ "lstrip": false,
247
+ "normalized": false,
248
+ "rstrip": false,
249
+ "single_word": false,
250
+ "special": true
251
+ },
252
+ "248075": {
253
+ "content": "<tts_text_bos_single>",
254
+ "lstrip": false,
255
+ "normalized": false,
256
+ "rstrip": false,
257
+ "single_word": false,
258
+ "special": true
259
+ },
260
+ "248076": {
261
+ "content": "<|audio_pad|>",
262
+ "lstrip": false,
263
+ "normalized": false,
264
+ "rstrip": false,
265
+ "single_word": false,
266
+ "special": true
267
+ }
268
+ },
269
+ "additional_special_tokens": [
270
+ "<|im_start|>",
271
+ "<|im_end|>",
272
+ "<|object_ref_start|>",
273
+ "<|object_ref_end|>",
274
+ "<|box_start|>",
275
+ "<|box_end|>",
276
+ "<|quad_start|>",
277
+ "<|quad_end|>",
278
+ "<|vision_start|>",
279
+ "<|vision_end|>",
280
+ "<|vision_pad|>",
281
+ "<|image_pad|>",
282
+ "<|video_pad|>"
283
+ ],
284
+ "bos_token": null,
285
+ "chat_template": "{%- set image_count = namespace(value=0) %}\n{%- set video_count = namespace(value=0) %}\n{%- macro render_content(content, do_vision_count, is_system_content=false) %}\n {%- if content is string %}\n {{- content }}\n {%- elif content is iterable and content is not mapping %}\n {%- for item in content %}\n {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}\n {%- if is_system_content %}\n {{- raise_exception('System message cannot contain images.') }}\n {%- endif %}\n {%- if do_vision_count %}\n {%- set image_count.value = image_count.value + 1 %}\n {%- endif %}\n {%- if add_vision_id %}\n {{- 'Picture ' ~ image_count.value ~ ': ' }}\n {%- endif %}\n {{- '<|vision_start|><|image_pad|><|vision_end|>' }}\n {%- elif 'video' in item or item.type == 'video' %}\n {%- if is_system_content %}\n {{- raise_exception('System message cannot contain videos.') }}\n {%- endif %}\n {%- if do_vision_count %}\n {%- set video_count.value = video_count.value + 1 %}\n {%- endif %}\n {%- if add_vision_id %}\n {{- 'Video ' ~ video_count.value ~ ': ' }}\n {%- endif %}\n {{- '<|vision_start|><|video_pad|><|vision_end|>' }}\n {%- elif 'text' in item %}\n {{- item.text }}\n {%- else %}\n {{- raise_exception('Unexpected item type in content.') }}\n {%- endif %}\n {%- endfor %}\n {%- elif content is none or content is undefined %}\n {{- '' }}\n {%- else %}\n {{- raise_exception('Unexpected content type.') }}\n {%- endif %}\n{%- endmacro %}\n{%- if not messages %}\n {{- raise_exception('No messages provided.') }}\n{%- endif %}\n{%- if tools and tools is iterable and tools is not mapping %}\n {{- '<|im_start|>system\\n' }}\n {{- \"# Tools\\n\\nYou have access to the following functions:\\n\\n<tools>\" }}\n {%- for tool in tools %}\n {{- \"\\n\" }}\n {{- tool | tojson }}\n {%- endfor %}\n {{- \"\\n</tools>\" }}\n {{- '\\n\\nIf you choose to call a function ONLY reply in the following format with NO suffix:\\n\\n<tool_call>\\n<function=example_function_name>\\n<parameter=example_parameter_1>\\nvalue_1\\n</parameter>\\n<parameter=example_parameter_2>\\nThis is the value for the second parameter\\nthat can span\\nmultiple lines\\n</parameter>\\n</function>\\n</tool_call>\\n\\n<IMPORTANT>\\nReminder:\\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\\n- Required parameters MUST be specified\\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\\n</IMPORTANT>' }}\n {%- if messages[0].role == 'system' %}\n {%- set content = render_content(messages[0].content, false, true)|trim %}\n {%- if content %}\n {{- '\\n\\n' + content }}\n {%- endif %}\n {%- endif %}\n {{- '<|im_end|>\\n' }}\n{%- else %}\n {%- if messages[0].role == 'system' %}\n {%- set content = render_content(messages[0].content, false, true)|trim %}\n {{- '<|im_start|>system\\n' + content + '<|im_end|>\\n' }}\n {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n {%- set index = (messages|length - 1) - loop.index0 %}\n {%- if ns.multi_step_tool and message.role == \"user\" %}\n {%- set content = render_content(message.content, false)|trim %}\n {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}\n {%- set ns.multi_step_tool = false %}\n {%- set ns.last_query_index = index %}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if ns.multi_step_tool %}\n {{- raise_exception('No user query found in messages.') }}\n{%- endif %}\n{%- for message in messages %}\n {%- set content = render_content(message.content, true)|trim %}\n {%- if message.role == \"system\" %}\n {%- if not loop.first %}\n {{- raise_exception('System message must be at the beginning.') }}\n {%- endif %}\n {%- elif message.role == \"user\" %}\n {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n {%- elif message.role == \"assistant\" %}\n {%- set reasoning_content = '' %}\n {%- if message.reasoning_content is string %}\n {%- set reasoning_content = message.reasoning_content %}\n {%- else %}\n {%- if '</think>' in content %}\n {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n {%- endif %}\n {%- endif %}\n {%- set reasoning_content = reasoning_content|trim %}\n {%- if loop.index0 > ns.last_query_index %}\n {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content + '\\n</think>\\n\\n' + content }}\n {%- else %}\n {{- '<|im_start|>' + message.role + '\\n' + content }}\n {%- endif %}\n {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}\n {%- for tool_call in message.tool_calls %}\n {%- if tool_call.function is defined %}\n {%- set tool_call = tool_call.function %}\n {%- endif %}\n {%- if loop.first %}\n {%- if content|trim %}\n {{- '\\n\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- else %}\n {{- '<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- endif %}\n {%- else %}\n {{- '\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- endif %}\n {%- if tool_call.arguments is defined %}\n {%- for args_name, args_value in tool_call.arguments|items %}\n {{- '<parameter=' + args_name + '>\\n' }}\n {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}\n {{- args_value }}\n {{- '\\n</parameter>\\n' }}\n {%- endfor %}\n {%- endif %}\n {{- '</function>\\n</tool_call>' }}\n {%- endfor %}\n {%- endif %}\n {{- '<|im_end|>\\n' }}\n {%- elif message.role == \"tool\" %}\n {%- if loop.previtem and loop.previtem.role != \"tool\" %}\n {{- '<|im_start|>user' }}\n {%- endif %}\n {{- '\\n<tool_response>\\n' }}\n {{- content }}\n {{- '\\n</tool_response>' }}\n {%- if not loop.last and loop.nextitem.role != \"tool\" %}\n {{- '<|im_end|>\\n' }}\n {%- elif loop.last %}\n {{- '<|im_end|>\\n' }}\n {%- endif %}\n {%- else %}\n {{- raise_exception('Unexpected message role.') }}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{- '<|im_start|>assistant\\n' }}\n {%- if enable_thinking is defined and enable_thinking is true %}\n {{- '<think>\\n' }}\n {%- else %}\n {{- '<think>\\n\\n</think>\\n\\n' }}\n {%- endif %}\n{%- endif %}",
286
+ "clean_up_tokenization_spaces": false,
287
+ "eos_token": "<|im_end|>",
288
+ "errors": "replace",
289
+ "model_max_length": 262144,
290
+ "pad_token": "<|endoftext|>",
291
+ "split_special_tokens": false,
292
+ "tokenizer_class": "Qwen2Tokenizer",
293
+ "unk_token": null,
294
+ "add_bos_token": false,
295
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
296
+ "extra_special_tokens": {
297
+ "audio_bos_token": "<|audio_start|>",
298
+ "audio_eos_token": "<|audio_end|>",
299
+ "audio_token": "<|audio_pad|>",
300
+ "image_token": "<|image_pad|>",
301
+ "video_token": "<|video_pad|>",
302
+ "vision_bos_token": "<|vision_start|>",
303
+ "vision_eos_token": "<|vision_end|>"
304
+ }
305
+ }
training/build_prompts.py ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Turn questions into the exact classifier messages the app would send.
2
+
3
+ Imports the prompt code from the simplixity checkout (read-only) so a training
4
+ example and a production call are byte for byte the same prompt. Settings are
5
+ sampled per item so the model learns the parts of the prompt that vary:
6
+
7
+ depth Auto/Balanced 75% (3 queries), Quick 15% (2), Deep 10% (3)
8
+ sources Auto 70% (the classifier picks lenses), a fixed selection 30%
9
+ library for library items: the papers lens, with the web on half the time
10
+ date a random day in 2026-2027
11
+ images attached-image items: a router that cannot see them
12
+
13
+ ~/Documents/projects/simplixity/.venv/bin/python -I build_prompts.py
14
+ """
15
+
16
+ from __future__ import annotations
17
+
18
+ import datetime as dt
19
+ import json
20
+ import os
21
+ import random
22
+ import sys
23
+
24
+ HERE = os.path.dirname(os.path.abspath(__file__))
25
+ REPO = os.path.expanduser("~/Documents/projects/simplixity")
26
+ sys.path.insert(0, REPO)
27
+ os.chdir(REPO) # settings look for .env here; nothing is written
28
+
29
+ from app.agent import depth as depth_module # noqa: E402
30
+ from app.agent import prompts # noqa: E402
31
+ from app.search import lenses as lens_module # noqa: E402
32
+
33
+ SYN_DEV = 150 # synthetic questions held out next to the real dev set
34
+ FIXED_SOURCES = [["web"], ["web", "scholar"], ["scholar"], ["web", "news"], ["docs"], ["web", "docs"],
35
+ ["reference"], ["web", "data"]]
36
+
37
+
38
+ def paper_line(p: dict) -> str:
39
+ authors = [a for a in p.get("authors") or [] if a]
40
+ surname = authors[0].split()[-1] if authors else ""
41
+ who = f"{surname} et al." if len(authors) > 2 else " & ".join(a.split()[-1] for a in authors)
42
+ by = " ".join(x for x in (who, str(p.get("year") or "")) if x) or "Unknown"
43
+ return f"- {by}: {(p.get('title') or 'Untitled')[:140]}"
44
+
45
+
46
+ def build(item: dict, rng: random.Random) -> tuple[list[dict], dict]:
47
+ r = rng.random()
48
+ depth = depth_module.QUICK if r < 0.15 else depth_module.DEEP if r < 0.25 else \
49
+ depth_module.resolve(depth_module.AUTO)
50
+ day = dt.date(2026, 1, 1) + dt.timedelta(days=rng.randrange(730))
51
+ prompts._today = lambda: prompts._TODAY.format(today=day.isoformat())
52
+
53
+ library = item.get("library") or bool(item.get("papers"))
54
+ catalog = item.get("catalog", "")
55
+ if item.get("papers"):
56
+ catalog = "\n".join(paper_line(p) for p in item["papers"])
57
+ if library:
58
+ web = rng.random() < 0.5
59
+ auto = web and rng.random() < 0.7
60
+ ids = ["auto"] if auto else (rng.choice(FIXED_SOURCES) if web else [])
61
+ lenses = [lens_module.LIBRARY, *(lens_module.resolve(ids) if web else [])]
62
+ else:
63
+ auto = rng.random() < 0.7
64
+ ids = ["auto"] if auto else rng.choice(FIXED_SOURCES)
65
+ lenses = lens_module.resolve(ids)
66
+
67
+ history = prompts.trim_history(item.get("hist") or [])
68
+ linked = [tuple(p) for p in item.get("linked") or []]
69
+ messages = prompts.classify(item["q"], history, depth, lenses, choose_lenses=auto, catalog=catalog,
70
+ linked=linked, images=int(item.get("images") or 0), sees_images=False)
71
+ settings = {"depth": depth.id, "queries": depth.queries, "sources": ids, "auto_lenses": auto,
72
+ "library": library, "web": (not library) or bool(set(ids) - {"library"}),
73
+ "date": day.isoformat(), "images": int(item.get("images") or 0), "linked": bool(linked)}
74
+ return messages, settings
75
+
76
+
77
+ def main() -> None:
78
+ rng = random.Random(23)
79
+ syn = [json.loads(line) for line in open(os.path.join(HERE, "data", "synthetic_questions.jsonl"))]
80
+ real = [json.loads(line) for line in open(os.path.join(HERE, "data", "private", "real_questions.jsonl"))]
81
+
82
+ order = list(range(len(syn)))
83
+ random.Random(29).shuffle(order)
84
+ dev_ids = set(order[:SYN_DEV])
85
+
86
+ public, private = [], []
87
+ for i, item in enumerate(syn):
88
+ messages, settings = build(item, rng)
89
+ public.append({"id": f"syn-{i}", "source": "synthetic", "split": "dev" if i in dev_ids else "train",
90
+ "cat": item["cat"], "lang": item["lang"], "settings": settings, "messages": messages})
91
+ for i, item in enumerate(real):
92
+ messages, settings = build(item, rng)
93
+ private.append({"id": f"real-{i}", "source": "real", "split": item["split"], "cat": "real",
94
+ "lang": "", "settings": settings, "messages": messages})
95
+
96
+ for path, rows in ((os.path.join(HERE, "data", "prompts.jsonl"), public),
97
+ (os.path.join(HERE, "data", "private", "prompts_real.jsonl"), private)):
98
+ with open(path, "w") as f:
99
+ for row in rows:
100
+ f.write(json.dumps(row, ensure_ascii=False) + "\n")
101
+ print(f"{len(rows)} -> {path}")
102
+ os.chmod(os.path.join(HERE, "data", "private", "prompts_real.jsonl"), 0o600)
103
+
104
+
105
+ if __name__ == "__main__":
106
+ main()
training/data/labels.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/data/prompts.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a954dea1dc1a275b59380881e3904b3d0e68d1f5528d4303ff501bad563ef45
3
+ size 40641377
training/data/relabel_syn_labels.jsonl ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"id": "syn-5", "label": {"standalone_question": "Thanks for finding the EU report on heat pumps; my landlord took me seriously after I showed him the numbers.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006264}
2
+ {"id": "syn-10", "label": {"standalone_question": "Thanks, that was super helpful. Same time tomorrow?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005461}
3
+ {"id": "syn-31", "label": {"standalone_question": "Wie hat der Dreißigjährige Krieg angefangen, welche Ursachen und Ereignisse führten 1618 zu seinem Beginn?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Dreißigjähriger Krieg Ursachen Beginn 1618 Prager Fenstersturz", "Dreißigjähriger Krieg Majestätsbrief Konflikt Habsburger Böhmen Vorgeschichte"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": ["timeline"], "whole_library": false}, "cost": 0.0008168}
4
+ {"id": "syn-54", "label": {"standalone_question": "Turn this pasted text into a bulleted list for slides: \"The Meiji Restoration of 1868 ended the Tokugawa shogunate. It modernized Japan's military, education system, and economy. Western technology and legal models were widely adopted.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005829}
5
+ {"id": "syn-55", "label": {"standalone_question": "Translate into French: \"This vase, produced in Athens around 490 BCE, depicts warriors preparing for battle and was used in ceremonial feasts.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005665}
6
+ {"id": "syn-57", "label": {"standalone_question": "Reformat this as a short APA-style citation: Hobsbawm, E. (1994). The Age of Extremes: The Short Twentieth Century, 1914-1991. Michael Joseph, London.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006609}
7
+ {"id": "syn-65", "label": {"standalone_question": "Condense this into a single sentence for my flashcard: \"The Hanseatic League was a commercial and defensive confederation of merchant guilds and market towns in northern Europe, flourishing from the thirteenth to the seventeenth centuries and dominating Baltic trade.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005028}
8
+ {"id": "syn-129", "label": {"standalone_question": "Könnten Sie diesen Satz bitte in einfaches Deutsch umschreiben? \"Die Kapitalertragsteuer wird unabhängig vom persönlichen Einkommensteuersatz pauschal mit 25 Prozent zuzüglich Solidaritätszuschlag erhoben.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000598}
9
+ {"id": "syn-138", "label": {"standalone_question": "Erkläre kurz und fasse zusammen: Ein ETF bildet einen Börsenindex nach. Er ist kostengünstig, breit gestreut und wird meist passiv verwaltet, was ihn für langfristige Anleger attraktiv macht.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006601}
10
+ {"id": "syn-150", "label": {"standalone_question": "Bitte übersetze das ins Italienische: \"Ich möchte wissen, ob sich ein Wechsel zu einem Online-Broker für mich lohnt, wenn ich nur einmal im Monat einen kleinen Betrag investiere.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006749}
11
+ {"id": "syn-151", "label": {"standalone_question": "If I have 3 potions at 45 gold each and sell them at 1.15x, what is the total gold I get?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006634}
12
+ {"id": "syn-171", "label": {"standalone_question": "With an HP regeneration rate of 3.5 per second, how long does it take to heal 490 HP?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005578}
13
+ {"id": "syn-172", "label": {"standalone_question": "convert 0.75 cups to tablespoons", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005399}
14
+ {"id": "syn-178", "label": {"standalone_question": "Wie viele Millimeter sind 3/8 Zoll?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006787}
15
+ {"id": "syn-213", "label": {"standalone_question": "Is it true that sharks don't get cancer because they have cartilage?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["shark cancer tumors cartilage myth evidence", "elasmobranch neoplasia tumor incidence study", "shark cartilage cancer treatment clinical trial"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.000702}
16
+ {"id": "syn-235", "label": {"standalone_question": "What is the difference between two sleep tracker rows (04-10: total 412 min, deep 61 min, REM 88 min; 04-11: total 447 min, deep 38 min, REM 102 min), and why would deep sleep drop while total sleep rises?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["deep sleep decrease while total sleep time increases", "consumer sleep tracker deep sleep estimation accuracy", "factors causing night-to-night variation in sleep stages"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0007504}
17
+ {"id": "syn-267", "label": {"standalone_question": "Wie viel Farbe brauche ich ungefähr für 14 m², wenn der Verbrauch 0,15 l/m² beträgt?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006535}
18
+ {"id": "syn-284", "label": {"standalone_question": "Warum ergibt 0.1 + 0.2 in Python nicht genau 0.3, und warum liefert print(0.1 + 0.2 == 0.3) False?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006513}
19
+ {"id": "syn-344", "label": {"standalone_question": "What is the current US unemployment rate in the most recent BLS release?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["US unemployment rate latest BLS Employment Situation release", "BLS unemployment rate 2026"], "recency": "month", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0004858}
20
+ {"id": "syn-346", "label": {"standalone_question": "What is the current population of Lagos, Nigeria according to the most recent United Nations estimate?", "answer_shape": "figure", "place": "Lagos, Nigeria", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Lagos Nigeria population United Nations estimate 2025", "Lagos population World Urbanization Prospects"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "cost": 0.0006824}
21
+ {"id": "syn-349", "label": {"standalone_question": "What is the current number of public EV charging points in the Netherlands, according to RVO or the latest count?", "answer_shape": "figure", "place": "Netherlands", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["RVO public charging points Netherlands 2026", "Netherlands public EV charging points number latest count", "laadpunten openbaar Nederland aantal RVO"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000574}
22
+ {"id": "syn-360", "label": {"standalone_question": "Welchen Zinssatz bietet die DKB aktuell für Tagesgeld?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["DKB Tagesgeld Zinssatz aktuell", "DKB Tagesgeld Konditionen 2026"], "recency": "week", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0005748}
23
+ {"id": "syn-399", "label": {"standalone_question": "Wie hoch ist der Verbraucherpreisindex für Deutschland im letzten veröffentlichten Monat?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Verbraucherpreisindex Deutschland aktuell Destatis", "Inflationsrate Deutschland letzter Monat Destatis"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "cost": 0.0007951}
24
+ {"id": "syn-403", "label": {"standalone_question": "Wie viele Kilometer Autobahn gibt es aktuell insgesamt in Deutschland?", "answer_shape": "figure", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Autobahnnetz Deutschland Kilometer 2026", "Länge Autobahnnetz Deutschland Bundesanstalt für Straßenwesen"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "cost": 0.0007337}
25
+ {"id": "syn-458", "label": {"standalone_question": "Ist Canberra die Hauptstadt von Australien?", "answer_shape": "", "place": "Australien", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hauptstadt Australien Canberra"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006074}
26
+ {"id": "syn-474", "label": {"standalone_question": "Wo wurde die erste Sparkasse gegründet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["erste Sparkasse Deutschlands Gründung Ort Jahr", "Geschichte der Sparkassen erste Sparkasse 1778 Hamburg"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0007317}
27
+ {"id": "syn-478", "label": {"standalone_question": "Wie hoch ist der Hvannadalshnúkur in Island?", "answer_shape": "figure", "place": "Hvannadalshnúkur", "wants_image": true, "image_query": "Hvannadalshnúkur", "needs_search": true, "search_queries": ["Hvannadalshnúkur Höhe m", "Hvannadalshnúkur highest peak Iceland elevation"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0007489}
28
+ {"id": "syn-484", "label": {"standalone_question": "Wie heißt die Stadt Königsberg (Ostpreußen) heute und in welchem Land liegt sie?", "answer_shape": "profile", "place": "Kaliningrad", "wants_image": true, "image_query": "Kaliningrad", "needs_search": true, "search_queries": ["Königsberg heute Kaliningrad Land", "Kaliningrad Oblast Russland Exklave"], "recency": "year", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0008167}
29
+ {"id": "syn-487", "label": {"standalone_question": "Wer war der erste Mensch im Weltraum und in welchem Jahr war das?", "answer_shape": "profile", "place": "", "wants_image": true, "image_query": "Yuri Gagarin", "needs_search": true, "search_queries": ["erster Mensch im Weltraum Jahr", "Juri Gagarin Wostok 1 1961"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0007264}
30
+ {"id": "syn-492", "label": {"standalone_question": "Wer hat die Glühbirne erfunden: Thomas Edison oder jemand anderes?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Glühbirne Erfindung Geschichte Edison Swan", "light bulb invention history Edison Swan priority"], "recency": "", "complexity": "standard", "lenses": ["reference"], "facets": [], "whole_library": false}, "cost": 0.0008074}
31
+ {"id": "syn-564", "label": {"standalone_question": "Will it be windy in Dublin tomorrow (25 February 2027), and could the wind cause morning flight delays?", "answer_shape": "weather", "place": "Dublin", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Dublin weather forecast tomorrow wind", "Dublin wind gusts forecast 25 February 2027"], "recency": "day", "complexity": "simple", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.0007756}
32
+ {"id": "syn-567", "label": {"standalone_question": "How cold will it get in Montreal tonight (2027-01-04), overnight low temperature and forecast?", "answer_shape": "weather", "place": "Montreal", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Montreal overnight low temperature tonight", "Montreal weather forecast tonight wind chill"], "recency": "day", "complexity": "simple", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.0006804}
33
+ {"id": "syn-582", "label": {"standalone_question": "Ist morgen früh um 6 Uhr in München, auf dem Weg zum Flughafen München, mit Glätte auf der A9 zu rechnen?", "answer_shape": "weather", "place": "München", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Glätte A9 München morgen früh", "Wetter München Vorhersage morgen früh Frost Glätte", "A9 Verkehrsmeldungen München Flughafen Glatteis"], "recency": "day", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "cost": 0.0008257}
34
+ {"id": "syn-606", "label": {"standalone_question": "What is Real Madrid's next La Liga fixture?", "answer_shape": "match", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Real Madrid next La Liga match", "Real Madrid La Liga fixtures 2026-27"], "recency": "year", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "cost": 0.000693}
35
+ {"id": "syn-631", "label": {"standalone_question": "Welche Spiele finden am kommenden Wochenende in der 2. Bundesliga statt, also wer spielt gegen wen?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["2. Bundesliga Spielplan Wochenende", "2. Bundesliga Spieltag Termine Januar 2027"], "recency": "week", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000636}
36
+ {"id": "syn-641", "label": {"standalone_question": "Wie viele Gelbe Karten hat Robert Andrich im Jahr 2026 bekommen?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Robert Andrich gelbe Karten 2026", "Robert Andrich yellow cards 2026 season"], "recency": "year", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "cost": 0.0006974}
37
+ {"id": "syn-651", "label": {"standalone_question": "Welche Tore fielen im Spiel TSG Hoffenheim gegen FC Augsburg am Mittwoch, dem 29.09.2027?", "answer_shape": "match", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hoffenheim Augsburg Tore", "Hoffenheim Augsburg Ergebnis Bundesliga"], "recency": "week", "complexity": "simple", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0007761}
38
+ {"id": "syn-657", "label": {"standalone_question": "Ranking of the top-selling electric car models in Europe in 2025", "answer_shape": "table", "place": "Europe", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["best selling electric cars Europe 2025 ranking", "Europe BEV registrations 2025 by model Dataforce", "top selling EV models Europe full year 2025 sales"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007117}
39
+ {"id": "syn-669", "label": {"standalone_question": "What are the MotoGP rider championship standings after the most recent Grand Prix?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["MotoGP rider standings final 2025", "MotoGP championship standings after Valencia Grand Prix 2025"], "recency": "year", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007524}
40
+ {"id": "syn-694", "label": {"standalone_question": "Wer hat aktuell die meisten Champions-League-Tore in der Saison 2026/27?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Champions League Torschützenliste 2026/27", "Champions League top scorers 2026-27"], "recency": "year", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "cost": 0.0007003}
41
+ {"id": "syn-699", "label": {"standalone_question": "Wie sehen die aktuellen Standings der NBA Western Conference aus?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["NBA Western Conference standings", "NBA standings 2026-27 Western Conference wins losses"], "recency": "day", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006488}
42
+ {"id": "syn-754", "label": {"standalone_question": "Who is Dr Nina Schwalbe, the GP who writes about patient care, and what is she known for?", "answer_shape": "profile", "place": "", "wants_image": true, "image_query": "Nina Schwalbe", "needs_search": true, "search_queries": ["Nina Schwalbe GP patient care", "Nina Schwalbe Hausärztin Autorin"], "recency": "year", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0007292}
43
+ {"id": "syn-766", "label": {"standalone_question": "Welche neueren Forscherinnen und Forscher haben sich mit den Ideen Charles Darwins weiterbeschäftigt und sie fortentwickelt?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Darwin evolutionary theory extended evolutionary synthesis", "Darwin legacy modern evolutionary biology review", "natural selection theory developments post-Darwinian research"], "recency": "year", "complexity": "research", "lenses": ["scholar", "web"], "facets": ["list"], "whole_library": false}, "cost": 0.0008319}
44
+ {"id": "syn-847", "label": {"standalone_question": "Is Bethesda's Creation Kit for Starfield available on Xbox Series X, or only on PC, and what is the mod upload limit?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Starfield Creation Kit Xbox Series X availability", "Starfield mod upload limit Creation Kit Xbox", "Starfield Creation Kit PC only console modding"], "recency": "year", "complexity": "standard", "lenses": ["web", "news"], "facets": ["facts"], "whole_library": false}, "cost": 0.0007414}
45
+ {"id": "syn-863", "label": {"standalone_question": "Wie ist der Bundesverband Deutscher Gartenfreunde (BDG) organisiert, und haben die Mitglieder der Kleingartenvereine Mitspracherecht gegenüber Vorstand und Verband?", "answer_shape": "", "place": "Dresden", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bundesverband Deutscher Gartenfreunde Organisation Aufbau Delegiertenversammlung Landesverbände", "Kleinygartenverein Mitgliederversammlung Mitspracherecht Vorstand Satzung Bundeskleingartengesetz"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0008511}
46
+ {"id": "syn-883", "label": {"standalone_question": "Was ist die Max-Planck-Gesellschaft und wie viel staatliche Forschungsförderung erhält sie?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Max-Planck-Gesellschaft Grundfinanzierung Bund Länder Millionen Euro", "Max-Planck-Gesellschaft Institute Aufgaben Struktur was ist", "Max-Planck-Gesellschaft Haushalt Zuwendungen 2025 2026"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006684}
47
+ {"id": "syn-891", "label": {"standalone_question": "Gibt es die Firma Hornbach noch, oder ist sie nur ein Baumarkt?", "answer_shape": "profile", "place": "", "wants_image": true, "image_query": "Hornbach", "needs_search": true, "search_queries": ["Hornbach Holding Unternehmen", "Hornbach Baumarkt Geschäftsbericht 2025"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006503}
48
+ {"id": "syn-946", "label": {"standalone_question": "Wie sieht das Verkehrszeichen „Vorfahrt achten“ (Zeichen 205) in Deutschland aus?", "answer_shape": "image", "place": "", "wants_image": true, "image_query": "Vorfahrt achten road sign Germany", "needs_search": true, "search_queries": ["Verkehrszeichen 205 Vorfahrt achten", "Vorfahrt achten Schild Bild StVO"], "recency": "", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0007478}
49
+ {"id": "syn-998", "label": {"standalone_question": "What is heat pump mode on a smart thermostat, and how does a heat pump differ from a furnace?", "answer_shape": "explainer", "place": "", "wants_image": true, "image_query": "air source heat pump", "needs_search": true, "search_queries": ["heat pump vs furnace how it works", "heat pump mode thermostat explained"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0008387}
50
+ {"id": "syn-1028", "label": {"standalone_question": "Why do leaves change color in fall?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["why leaves change color autumn chlorophyll breakdown", "autumn leaf color anthocyanin carotenoid synthesis mechanism"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.000711}
51
+ {"id": "syn-1032", "label": {"standalone_question": "What is inflation and why does it happen?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["inflation definition causes demand-pull cost-push", "why does inflation happen economics explanation", "inflation causes monetary policy central bank"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005817}
52
+ {"id": "syn-1055", "label": {"standalone_question": "Ist ein Sparbuch aus den 1980er-Jahren noch etwas wert, und wie funktioniert der Zinseszins bei alten Sparbüchern?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Sparbuch alt Verjährung Guthaben auszahlen", "Zinseszins Sparbuch Zinsgutschrift Berechnung", "Sparbuch 1980er Zinsen Bank Auszahlung Altguthaben"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["callout"], "whole_library": false}, "cost": 0.0008645}
53
+ {"id": "syn-1067", "label": {"standalone_question": "Wie werden die Strafen beim Biathlon berechnet, wenn ein Schütze einen Schuss verfehlt (Strafrunde bzw. Strafzeit je Fehlschuss in den verschiedenen Disziplinen)?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Biathlon Strafrunde Fehlschuss Regeln IBU", "Biathlon Strafzeit 60 Sekunden Einzelrennen Fehlschuss", "Biathlon Verfolgung Massenstart Strafrunde 150 Meter Regel"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0008159}
54
+ {"id": "syn-1087", "label": {"standalone_question": "Warum wird ein Kondensator im Netzteil so heiß?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Elektrolytkondensator Erwärmung Netzteil Ripplestrom", "Ripplestrom ESR Verlustleistung Elektrolytkondensator", "Elektrolytkondensator Lebensdauer Temperatur Netzteil"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006313}
55
+ {"id": "syn-1089", "label": {"standalone_question": "Warum braucht man bei PVC-Fenstern eine Dampfsperre, und was passiert, wenn man sie weglässt?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["PVC Fenster Dampfsperre Funktion Kondensat Wasserdampfdiffusion", "Dampfsperre Fenstereinbau weglassen Schimmel Folgen"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006849}
56
+ {"id": "syn-1114", "label": {"standalone_question": "How do I file a small claims form in Ohio?", "answer_shape": "steps", "place": "Ohio", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ohio small claims court how to file complaint", "Ohio small claims filing fee forms county municipal court"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0005653}
57
+ {"id": "syn-1130", "label": {"standalone_question": "How do I clean a clothes dryer vent safely, step by step?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["clean dryer vent step by step safety", "dryer vent cleaning instructions fire hazard", "dryer vent duct cleaning how to disconnect"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005781}
58
+ {"id": "syn-1138", "label": {"standalone_question": "How do I set up a shared family calendar on my phone and my partner's phone so we both see the kids' school events?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["share Google Calendar with family members iPhone Android", "iCloud shared calendar iPhone family setup", "share calendar Samsung Outlook family members phone"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006613}
59
+ {"id": "syn-1145", "label": {"standalone_question": "What steps do I need to take to renew my driver's license online before it expires next month?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["renew driver's license online steps", "online driver license renewal requirements", "driver license renewal before expiry online"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005988}
60
+ {"id": "syn-1215", "label": {"standalone_question": "Should I use wood glue or epoxy to repair a cracked chair leg that bears a lot of weight?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["wood glue vs epoxy chair leg repair strength", "PVA wood glue vs epoxy structural joint load", "repair cracked wooden chair leg glue"], "recency": "", "complexity": "research", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.0007616}
61
+ {"id": "syn-1243", "label": {"standalone_question": "Should I use a 6 oz or 8 oz graphite tennis racket if I have tennis elbow?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["tennis elbow racquet weight lateral epicondylitis", "racket mass vibration tennis elbow study"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007799}
62
+ {"id": "syn-1244", "label": {"standalone_question": "Should I get a power meter pedal or a crank-based power meter for my first training block?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["power meter pedals vs crank power meter comparison accuracy", "single-sided vs dual-sided power meter pedals cost accuracy", "power meter pedals beginner cyclist price 2027"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0008266}
63
+ {"id": "syn-1273", "label": {"standalone_question": "Soll ich für die Gartenbewässerung Regenwasser aus der Regentonne oder Brunnenwasser nutzen, mit Blick auf Kalk und Keime?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["rainwater harvesting irrigation microbial contamination E. coli", "groundwater well water irrigation hardness calcium bicarbonate salinity"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007978}
64
+ {"id": "syn-1313", "label": {"standalone_question": "Lohnt sich der Kauf der Deluxe-Edition von Cyberpunk 2077 mit Phantom Liberty gegenüber der Standard-Version?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Cyberpunk 2077 Deluxe Edition Inhalt Unterschied Standard", "Cyberpunk 2077 Phantom Liberty Deluxe Edition Preis"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000687}
65
+ {"id": "syn-1337", "label": {"standalone_question": "Stripe vs Square vs PayPal processing fees for a small shop", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Stripe Square PayPal processing fees comparison small business", "Stripe standard online card processing rate", "Square PayPal in-person card processing fees 2027"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007066}
66
+ {"id": "syn-1341", "label": {"standalone_question": "Vergleich Lissabon, Porto und Madeira für einen 10-tägigen Herbsturlaub: Welches Reiseziel ist günstiger und wo ist das Wetter im Herbst besser?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Lissabon Porto Madeira Herbst Wetter Durchschnittstemperatur Regentage", "Lissabon Porto Madeira Urlaubskosten Hotel Flug Herbst", "Madeira Lissabon Porto Reisekosten Vergleich 2027"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000712}
67
+ {"id": "syn-1356", "label": {"standalone_question": "Was lohnt sich bei Netflix, Disney+ und Amazon Prime Video, wenn ich hauptsächlich Serien schaue?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Netflix Disney+ Prime Video Preise Deutschland", "Streaming Vergleich Serienangebot Netflix Disney+ Prime Video", "Netflix Disney+ Prime Video Werbung Abo Kosten Vergleich"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006397}
68
+ {"id": "syn-1378", "label": {"standalone_question": "Does listening to music during surgery actually lower anxiety and reduce the amount of pain medication needed?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["music intervention intraoperative anxiety randomized controlled trial meta-analysis", "music listening postoperative analgesic consumption opioid surgery systematic review"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "cost": 0.0007516}
69
+ {"id": "syn-1381", "label": {"standalone_question": "Do blue light blocking glasses help you sleep better, or is the effect just placebo?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["blue light blocking glasses sleep randomized trial", "blue light glasses sleep systematic review", "blue light glasses placebo effect study"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000682}
70
+ {"id": "syn-1400", "label": {"standalone_question": "Is there research linking adolescents' time spent playing Minecraft and Roblox to anxiety or depression?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["video game play adolescents anxiety depression systematic review", "Minecraft Roblox adolescent mental health", "problematic gaming adolescents depression longitudinal study"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007317}
71
+ {"id": "syn-1432", "label": {"standalone_question": "Do studies show that standing desks reduce back or neck pain in office workers, or is the benefit mainly from switching between sitting and standing positions?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["standing desk randomized trial back neck pain office workers", "sit-stand desk switching positions musculoskeletal pain systematic review"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "cost": 0.0005872}
72
+ {"id": "syn-1435", "label": {"standalone_question": "Does the research support vitamin D supplementation improving mood, or is any benefit mostly placebo?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["vitamin D supplementation depressive symptoms randomized controlled trials meta-analysis", "vitamin D seasonal affective disorder winter depression trial", "vitamin D mood healthy adults placebo systematic review"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "cost": 0.0007898}
73
+ {"id": "syn-1439", "label": {"standalone_question": "Is there solid research showing that cold showers boost mood and immunity, or is the evidence mostly anecdotal?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["cold water exposure mood randomized controlled trial", "cold showers immune function study", "cold water immersion immunity systematic review"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006003}
74
+ {"id": "syn-1466", "label": {"standalone_question": "Macht Krafttraining ältere Menschen weniger sturzgefährdet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Krafttraining Sturzprävention ältere Menschen Studie", "Kraft- und Gleichgewichtstraining Sturzrisiko Senioren Metaanalyse", "Sturzprophylaxe Krafttraining Evidenz Leitlinie"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006728}
75
+ {"id": "syn-1468", "label": {"standalone_question": "Was passiert mit dem menschlichen Körper, wenn man drei Tage lang nichts isst?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hungern 72 Stunden Stoffwechsel Glykogen Ketose", "Fasten drei Tage Körper Auswirkungen Gesundheit", "Kurzzeitfasten 72 Stunden Ketonkörper Proteinabbau Studie"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0007079}
76
+ {"id": "syn-1478", "label": {"standalone_question": "Ist Ibuprofen nach dem Krafttraining schädlich für den Muskelaufbau, wie ich gelesen habe? Was zeigen die Studien dazu?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["ibuprofen NSAID resistance training muscle hypertrophy", "NSAID effect muscle protein synthesis strength training adaptation", "ibuprofen muscle damage recovery resistance exercise randomized trial"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007523}
77
+ {"id": "syn-1595", "label": {"standalone_question": "How do the restore-keys fallback options work in GitHub Actions cache keys?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["github actions cache restore-keys fallback", "github actions cache key prefix match restore-keys"], "recency": "", "complexity": "simple", "lenses": ["docs", "web"], "facets": [], "whole_library": false}, "cost": 0.0005654}
78
+ {"id": "syn-1614", "label": {"standalone_question": "How do I configure the Redis maxmemory-policy setting so that cached drug interaction lookups get evicted gracefully under memory pressure?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Redis maxmemory-policy eviction policies", "Redis maxmemory configuration redis.conf", "Redis allkeys-lru volatile-lru eviction cache keys with expire"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000647}
79
+ {"id": "syn-1660", "label": {"standalone_question": "Wie aktiviere ich in nginx gzip-Kompression für JSON-Antworten?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["nginx gzip_types application/json", "nginx ngx_http_gzip_module gzip on directive"], "recency": "", "complexity": "simple", "lenses": ["docs", "web"], "facets": [], "whole_library": false}, "cost": 0.0007032}
80
+ {"id": "syn-1664", "label": {"standalone_question": "Was macht der Parameter --no-cache bei docker build genau?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["docker build --no-cache option", "docker build no-cache layer cache rebuild"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005759}
81
+ {"id": "syn-1678", "label": {"standalone_question": "What happened with the major airline strike, and has it been resolved yet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["major airline strike resolved", "airline strike latest news"], "recency": "month", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0007308}
82
+ {"id": "syn-1696", "label": {"standalone_question": "Any fresh news on the housing market cooling off in May 2026?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["housing market cooling May 2026", "home sales prices slowdown May 2026 news"], "recency": "month", "complexity": "standard", "lenses": ["news", "web"], "facets": ["timeline"], "whole_library": false}, "cost": 0.0005636}
83
+ {"id": "syn-1699", "label": {"standalone_question": "What is the latest news on US Department of Education federal student loan forgiveness rules this week?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Department of Education student loan forgiveness rule news", "federal student loan forgiveness changes November 2026", "PSLF IDR forgiveness rule latest"], "recency": "week", "complexity": "standard", "lenses": [], "facets": ["timeline"], "whole_library": false}, "cost": 0.0005011}
84
+ {"id": "syn-1716", "label": {"standalone_question": "What school voucher programs have US state legislatures passed in 2026?", "answer_shape": "", "place": "United States", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["school voucher programs passed state legislature 2026", "education savings account law signed 2026 states"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0006954}
85
+ {"id": "syn-1724", "label": {"standalone_question": "Gibt es neue Informationen zur Grippewelle in Deutschland diese Woche, und wird sie schlimmer?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Grippewelle Deutschland aktuell", "RKI Influenza Wochenbericht ARE Trend", "Influenza Aktivität Deutschland September 2027"], "recency": "week", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0008153}
86
+ {"id": "syn-1737", "label": {"standalone_question": "Gibt es aktuell Probleme oder Lieferengpässe bei der Tetanus-Impfung?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Tetanus Impfstoff Lieferengpass 2026", "Tetanus Impfstoff Lieferengpass BfArM PEI", "Tetanus Impfung Probleme aktuell"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0007579}
87
+ {"id": "syn-1743", "label": {"standalone_question": "Aktuelle Nachrichten zu dem Zugunglück der letzten Woche (Stand 22.04.2027) und der aktuelle Stand der Ermittlungen", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Zugunglück Ermittlungen Stand April 2027", "Zugunglück Ursache Untersuchung Bahn aktuell"], "recency": "week", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006938}
88
+ {"id": "syn-1773", "label": {"standalone_question": "Was gibt es Neues zur Kita-Finanzierung und zum Fachkräfteprogramm des Bundes für Kitas, insbesondere in Leipzig und Sachsen?", "answer_shape": "", "place": "Leipzig", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Kita-Finanzierung Bund Länder 2026 Neuigkeiten", "Fachkräfteprogramm Kita Bund 2026", "Kita Leipzig Fachkräfte Förderung Sachsen 2026"], "recency": "year", "complexity": "research", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.0008078}
89
+ {"id": "syn-1788", "label": {"standalone_question": "How was the first email sent, and who sent it?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["first email ever sent Ray Tomlinson 1971 ARPANET", "Tomlinson first network email how it worked @ sign history"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0007158}
90
+ {"id": "syn-1805", "label": {"standalone_question": "Wie hat sich das Bausparen in Deutschland seit den 1970er-Jahren entwickelt?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bausparen Deutschland Geschichte seit 1970", "Bausparkassen Entwicklung Zinsen Wohnungsbauprämie Bausparverträge Geschichte"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": ["timeline"], "whole_library": false}, "cost": 0.0007394}
91
+ {"id": "syn-1818", "label": {"standalone_question": "Wie entstand das Modell der Kreditkarte mit Jahresbeitrag in Deutschland, und warum verlangen Banken dafür eine Jahresgebühr?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Geschichte der Kreditkarte in Deutschland", "Kreditkarte Jahresgebühr Entstehung Geschäftsmodell Banken", "Eurocard Einführung Deutschland Kreditkarten Entwicklung"], "recency": "", "complexity": "standard", "lenses": [], "facets": ["timeline"], "whole_library": false}, "cost": 0.0007348}
92
+ {"id": "syn-1824", "label": {"standalone_question": "Wie kam es zur Privatisierung der Deutschen Telekom, und was ist seit dem Börsengang der Deutschen Telekom passiert?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Deutsche Telekom Börsengang 1996 Privatisierung Geschichte", "Deutsche Telekom Aktie Kursentwicklung seit 1996 Volksaktie", "Deutsche Telekom Entwicklung Beteiligung Bund Konzernumbau T-Aktie"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["timeline"], "whole_library": false}, "cost": 0.0007963}
93
+ {"id": "syn-1827", "label": {"standalone_question": "Should I register as self-employed or stay on the freelance tax form?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["self-employed registration vs freelance tax form pros cons", "sole trader self-employed registration tax obligations 2026", "freelancer register self-employed deadlines penalties"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["proscons"], "whole_library": false}, "cost": 0.0006891}
94
+ {"id": "syn-1857", "label": {"standalone_question": "Pros and cons of moving to a farm to be closer to wildlife", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["moving to rural farm closer to wildlife pros cons", "living on farmland wildlife encounters drawbacks", "rural property buying wildlife habitat costs isolation"], "recency": "", "complexity": "standard", "lenses": ["web"], "facets": ["proscons"], "whole_library": false}, "cost": 0.0006901}
95
+ {"id": "syn-1861", "label": {"standalone_question": "What are the pros and cons of volunteering at a wildlife rehabilitation center on weekends?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["wildlife rehabilitation center volunteer weekends pros cons", "wildlife rehab volunteering experience drawbacks", "wildlife rehabilitation volunteer commitment training requirements"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["proscons"], "whole_library": false}, "cost": 0.0005018}
96
+ {"id": "syn-1884", "label": {"standalone_question": "Vor- und Nachteile einer Teilzeit mit 75 % zur Pflege der Mutter in Deutschland, mit Blick auf Rentenpunkte und Karriere", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Teilzeit Pflege Angehörige Rentenpunkte Pflegepunkte Deutschland", "Familienpflegezeit Pflegeunterstützungsgeld Teilzeit 75 Prozent", "Teilzeit Karriere Nachteile Rückkehr Vollzeit Rentenlücke"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["proscons"], "whole_library": false}, "cost": 0.0007587}
97
+ {"id": "syn-1886", "label": {"standalone_question": "Lohnt sich ein Abo bei Netflix plus Disney+, oder reicht eines davon für meine Serien?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Netflix Preise Deutschland 2027", "Disney+ Preise Deutschland 2027 Werbung Abo", "Netflix oder Disney+ Serienangebot Vergleich"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007747}
98
+ {"id": "syn-1893", "label": {"standalone_question": "Sollte ich eine LASIK-Behandlung machen lassen bei -3,5 Dioptrien und dünner Hornhaut, oder sind Kontaktlinsen die bessere Wahl?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["LASIK dünne Hornhaut Kontraindikation Hornhautdicke Grenzwert", "LASIK vs Kontaktlinsen Kurzsichtigkeit -3,5 Dioptrien Vor- und Nachteile", "LASIK thin cornea contraindications alternatives PRK ICL"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.000659}
99
+ {"id": "syn-1915", "label": {"standalone_question": "Vorteile und Nachteile von Start-ups gegenüber Konzernen", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Vor- und Nachteile Start-up vs Konzern Arbeit", "Start-up oder Konzern Karriere Vergleich Gehalt Sicherheit", "startup vs corporate job pros cons"], "recency": "", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0007866}
100
+ {"id": "syn-1925", "label": {"standalone_question": "Lohnt sich der Kauf einer Wärmepumpe im Jahr 2024 für Hausbesitzer in Deutschland, unter Berücksichtigung von Anschaffungskosten, Betriebskosten und Förderung?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wärmepumpe Wirtschaftlichkeit 2024 Anschaffungskosten Betriebskosten Vergleich Gas", "BEG Förderung Wärmepumpe 2024 Zuschuss Konditionen"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["proscons"], "whole_library": false}, "cost": 0.0007639}
101
+ {"id": "syn-1941", "label": {"standalone_question": "What are the side effects of creatine monohydrate supplementation on kidney function markers in healthy adults?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["creatine monohydrate supplementation serum creatinine healthy adults", "creatine supplementation renal function cystatin C eGFR randomized trial", "creatine supplementation kidney safety systematic review meta-analysis"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "cost": 0.0007429}
102
+ {"id": "syn-1966", "label": {"standalone_question": "Welche Stoffe stecken im Lithium-Ionen-Akku eines alten Laptops, und wie hoch ist die Brandgefahr beim Verkauf eines gebrauchten Laptops mit altem Akku?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Lithium-Ionen-Akku Laptop Bestandteile Kathode Anode Elektrolyt", "Laptop Akku Brandgefahr alter Akku aufgebläht gebraucht verkaufen"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["callout"], "whole_library": false}, "cost": 0.0008358}
103
+ {"id": "syn-1975", "label": {"standalone_question": "Welche Symptome deuten bei Docker-Builds auf einen Cache-Fehler im Layer-System hin, sodass der Build unnötig lange dauert?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["docker build cache miss layers invalidated causes", "docker buildkit cache not used every build slow"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006286}
104
+ {"id": "syn-2042", "label": {"standalone_question": "PS5 Pro vs Xbox Series X: current price and which has better 4K frame rates in recent reviews", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["PS5 Pro Xbox Series X price", "PS5 Pro vs Xbox Series X 4K frame rate review"], "recency": "year", "complexity": "research", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "cost": 0.000733}
105
+ {"id": "syn-2066", "label": {"standalone_question": "Was ist der Unterschied zwischen einer Probezeit und einer Befristung bei Arbeitsverträgen in Deutschland?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Probezeit Befristung Arbeitsvertrag Unterschied TzBfG", "Probezeit gesetzlich maximal sechs Monate Kündigungsfrist §622 BGB", "sachgrund befristung Arbeitsvertrag Regeln TzBfG"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.000804}
106
+ {"id": "syn-2083", "label": {"standalone_question": "Can I apply to extend a Stamp 1G immigration permission in Ireland for a job search after finishing a bachelor's degree there, and what is the deadline to submit the application before the permission expires?", "answer_shape": "", "place": "Ireland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ireland Stamp 1G extension job search graduate Irish Naturalisation and Immigration Service", "Stamp 1G renewal deadline before expiry application INIS", "Stamp 1G graduate scheme eligibility conditions Ireland"], "recency": "year", "complexity": "standard", "lenses": ["web", "news"], "facets": ["callout"], "whole_library": false}, "cost": 0.0006184}
107
+ {"id": "syn-2104", "label": {"standalone_question": "What is the deadline to file a notice of intent to relocate with a California school district when a family moves out of the attendance zone mid-semester?", "answer_shape": "", "place": "California", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["California school district notice of intent to relocate deadline mid-year move", "California Education Code student moving out of attendance area notice requirement", "California school enrollment withdrawal when moving out of district deadline"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["callout"], "whole_library": false}, "cost": 0.0006469}
108
+ {"id": "syn-2124", "label": {"standalone_question": "What are the deadlines for a Brazilian employee to file a labor claim against an employer for unpaid overtime, and does the three-year prescription period start from the termination date?", "answer_shape": "", "place": "Brazil", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["prescrição trabalhista horas extras CLT art. 7º XXIX dois anos cinco anos", "prazo prescrição trabalhista contagem após rescisão do contrato de trabalho", "Brazil labor claim prescription overtime two years after termination five years"], "recency": "year", "complexity": "standard", "lenses": ["web"], "facets": ["callout"], "whole_library": false}, "cost": 0.0008238}
109
+ {"id": "syn-2136", "label": {"standalone_question": "Wie lange habe ich Zeit, nach einer schriftlichen Kündigung wegen angeblich häufiger Krankheit (Bandscheibenvorfall) eine Kündigungsschutzklage in Deutschland einzureichen?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Kündigungsschutzklage Frist drei Wochen §4 KSchG", "Kündigungsschutzklage Frist Zugang schriftliche Kündigung krankheitsbedingt"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["callout"], "whole_library": false}, "cost": 0.0007396}
110
+ {"id": "syn-2177", "label": {"standalone_question": "Ab welchem Alter darf man in der Schweiz alleine Auto fahren, und gilt der deutsche Führerschein in der Schweiz noch?", "answer_shape": "", "place": "Schweiz", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Mindestalter Autofahren Schweiz Lernfahrausweis Führerausweis", "deutscher Führerschein Schweiz gültig Umtausch"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006877}
111
+ {"id": "syn-2210", "label": {"standalone_question": "Kann man Eier pochieren ohne Essigzusatz im Wasser und ist das Ergebnis trotzdem gut?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Eier pochieren ohne Essig", "pochierte Eier Wasser ohne Essig Eiweiß zusammenhalten"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006351}
112
+ {"id": "syn-2292", "label": {"standalone_question": "What is the best backpack under 100 for carrying a laptop and charger to the office that doesn't look like a school bag?", "answer_shape": "table", "place": "", "wants_image": true, "image_query": "laptop backpack", "needs_search": true, "search_queries": ["best minimalist laptop backpack under 100 office commute", "laptop backpack under $100 review sleek non-school design"], "recency": "year", "complexity": "standard", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.000803}
113
+ {"id": "syn-2302", "label": {"standalone_question": "Which microphone under 100 is good for podcasting or recording voiceovers at home?", "answer_shape": "table", "place": "", "wants_image": true, "image_query": "podcast microphone", "needs_search": true, "search_queries": ["best budget podcast microphone under $100 review", "best USB microphone voiceover home recording 2026", "Audio-Technica AT2020USB+ vs Rode PodMic vs Shure MV7 review"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0007103}
114
+ {"id": "syn-2320", "label": {"standalone_question": "Welche Wärmebildkamera fürs Handy oder als eigenständiges Gerät unter 150 Euro eignet sich, um Energieverluste im Haus aufzuspüren?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wärmebildkamera Smartphone Aufsatz Test Gebäudediagnose", "Wärmebildkamera unter 150 Euro Test Energieverluste Haus", "Wärmebildkamera Handy Auflösung Temperaturbereich Vergleich"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0009503}
115
+ {"id": "syn-2339", "label": {"standalone_question": "What does a good sourdough crumb look like when it is properly fermented?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["sourdough crumb open even cell structure properly fermented", "sourdough crumb underproofed vs overproofed appearance"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.000792}
116
+ {"id": "syn-2352", "label": {"standalone_question": "Wie hoch ist der Wasserverbrauch einer vierköpfigen Familie (Haushalt mit 4 Personen) pro Tag, Monat und Jahr?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wasserverbrauch pro Person pro Tag Liter Deutschland", "Wasserverbrauch 4-Personen-Haushalt Kubikmeter pro Jahr", "durchschnittlicher Wasserverbrauch Haushalt Statistik Destatis"], "recency": "year", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007829}
117
+ {"id": "syn-2391", "label": {"standalone_question": "Explain the flow of blood through the human heart, naming each chamber and valve in order.", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["human heart blood flow right atrium tricuspid valve right ventricle pulmonary valve", "heart chambers valves order blood flow mitral aortic valve"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0007169}
118
+ {"id": "syn-2398", "label": {"standalone_question": "How does a wastewater treatment plant clean water, and what is each stage in the process?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["wastewater treatment plant stages primary secondary tertiary", "activated sludge process wastewater treatment explained", "municipal wastewater treatment process screening sedimentation disinfection"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "cost": 0.0007101}
119
+ {"id": "syn-2413", "label": {"standalone_question": "Wie läuft eine Blutspende in Deutschland ab: Anmeldung, Voruntersuchung, Spende und Nachsorge, und was passiert in jedem Abschnitt?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ablauf Blutspende Deutschland Anmeldung Voruntersuchung Spende Nachsorge", "Blutspende Ablauf DRK Blutspendedienst Spendevorgang"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.0008032}
120
+ {"id": "syn-2464", "label": {"standalone_question": "What are good sleep hygiene tips for shift workers?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["sleep hygiene shift workers interventions", "shift work disorder management light exposure melatonin", "night shift workers sleep improvement randomized trial"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007258}
121
+ {"id": "syn-2465", "label": {"standalone_question": "Can you compare Python and JavaScript for beginner programmers?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Python vs JavaScript for beginners", "Python JavaScript learning curve beginner programmers", "Python JavaScript job market demand 2026"], "recency": "year", "complexity": "research", "lenses": ["web", "docs"], "facets": [], "whole_library": false}, "cost": 0.0007141}
122
+ {"id": "syn-2482", "label": {"standalone_question": "How do vaccines train the immune system?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["how vaccines train adaptive immunity memory B cells T cells", "vaccine mechanism antigen presentation immunological memory review"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0006891}
123
+ {"id": "syn-2483", "label": {"standalone_question": "Which indie games similar to Stardew Valley are worth playing?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["best indie games like Stardew Valley", "Stardew Valley alternatives farming life sim reviews"], "recency": "year", "complexity": "standard", "lenses": ["web"], "facets": ["list"], "whole_library": false}, "cost": 0.0007076}
124
+ {"id": "syn-2486", "label": {"standalone_question": "Wie lange dauert die Ausbildung zur Pflegefachfrau in Deutschland?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ausbildung Pflegefachfrau Dauer 3 Jahre", "Pflegefachfrau Ausbildung Dauer Teilzeit Verkürzung"], "recency": "year", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "cost": 0.000746}
125
+ {"id": "syn-2506", "label": {"standalone_question": "Recht auf Einsicht in Klausuren an Schulen in Nordrhein-Westfalen", "answer_shape": "", "place": "Nordrhein-Westfalen", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Klausureinsicht Schule NRW Recht Schüler Eltern", "Klausureinsicht Frist Schulgesetz NRW APO-SI Gymnasium", "Einsicht in Klausuren Anspruch Schulrecht Nordrhein-Westfalen"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006758}
126
+ {"id": "syn-2507", "label": {"standalone_question": "Schnelles Rezept für einen glutenfreien Kuchen", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["schnelles glutenfreies Kuchen Rezept", "glutenfreier Rührkuchen einfach Rezept 30 Minuten"], "recency": "", "complexity": "simple", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.0006926}
127
+ {"id": "syn-2510", "label": {"standalone_question": "Welche Fristen gelten in Deutschland für den Einspruch (Widerspruch) gegen einen Steuerbescheid?", "answer_shape": "", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Einspruch Steuerbescheid Frist ein Monat § 355 AO", "Einspruch Steuerbescheid Bekanntgabe Fiktion drei Tage Fristbeginn", "Einspruch Steuerbescheid Frist Begründung Verlängerung Finanzamt"], "recency": "", "complexity": "standard", "lenses": [], "facets": ["callout"], "whole_library": false}, "cost": 0.0007702}
128
+ {"id": "syn-2514", "label": {"standalone_question": "Wo kann ich in Deutschland eine beglaubigte Geburtsurkunde bestellen?", "answer_shape": "steps", "place": "Deutschland", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["beglaubigte Geburtsurkunde bestellen Standesamt", "Geburtsurkunde beantragen online Kosten Standesamt", "Geburtsurkunde Ausland Bestellung Standesamt Antrag"], "recency": "year", "complexity": "standard", "lenses": ["web"], "facets": [], "whole_library": false}, "cost": 0.0007674}
129
+ {"id": "syn-2548", "label": {"standalone_question": "Does chilling onions before cutting reduce eye irritation and tearing?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["chilling onions before cutting lachrymatory factor tearing study", "refrigerated onion syn-propanethial-S-oxide release eye irritation"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0005813}
130
+ {"id": "syn-2552", "label": {"standalone_question": "Does wearing dark-coloured clothing change how attractive a person is to mosquitoes?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["mosquito attraction clothing colour dark vs light experiment", "Aedes aegypti Anopheles visual host seeking clothing colour"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0006799}
131
+ {"id": "syn-2566", "label": {"standalone_question": "Which is healthier for someone watching their sugar: homemade oil-and-vinegar vinaigrette or store-bought Italian dressing?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["store-bought Italian salad dressing sugar grams per serving", "homemade oil and vinegar dressing sugar content nutrition", "Italian dressing added sugar brands nutrition label comparison"], "recency": "", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007759}
132
+ {"id": "syn-2570", "label": {"standalone_question": "What is the Ontario minimum wage for students under 18 who work part time right now?", "answer_shape": "figure", "place": "Ontario", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ontario student minimum wage under 18 current rate", "Ontario minimum wage increase 2027 student rate"], "recency": "year", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0007175}
133
+ {"id": "syn-2600", "label": {"standalone_question": "How do adult learner dropout rates for cello compare to piano?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["adult music learners dropout rate piano cello", "adult instrumental music lessons attrition study", "adult piano learners discontinuation reasons survey"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006158}
134
+ {"id": "syn-2654", "label": {"standalone_question": "Wie viele Menschen in Deutschland besitzen tatsächlich eine Patientenverfügung? Gibt es Erhebungen dazu?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Patientenverfügung Verbreitung Deutschland Umfrage", "Patientenverfügung Anteil Bevölkerung Studie Repräsentativbefragung"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "cost": 0.0006428}
135
+ {"id": "syn-2706", "label": {"standalone_question": "Wie viel Schlaf brauchen Teenager (Jugendliche von etwa 13 bis 18 Jahren) pro Nacht?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["adolescent sleep duration recommendation hours per night", "AASM consensus statement sleep duration children adolescents"], "recency": "", "complexity": "simple", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "cost": 0.0007243}
136
+ {"id": "syn-2723", "label": {"standalone_question": "Gleicht WSL2 die Entwickler-Vorteile von Linux gegenüber Windows aus?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["WSL2 vs native Linux Entwickler Grenzen", "WSL2 Dateisystem Performance Unterschiede native Linux", "WSL2 Einschränkungen Hardware GPU Systemd"], "recency": "year", "complexity": "standard", "lenses": ["web", "docs"], "facets": ["list"], "whole_library": false}, "cost": 0.0007386}
137
+ {"id": "syn-2790", "label": {"standalone_question": "What do the ingredients on the attached protein bar label mean?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": ["list"], "whole_library": false}, "cost": 0.0007254}
138
+ {"id": "syn-2801", "label": {"standalone_question": "Summarize the early termination clause in the attached tenant lease scan in plain English.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0006435}
139
+ {"id": "syn-2815", "label": {"standalone_question": "Was ist der abgebildete Gegenstand, den ich auf dem Flohmarkt gefunden habe, und was ist er ungefähr wert?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bestimmung Flohmarkt Fund Gegenstand Wert", "Sammlerwert Wertermittlung antiker Gegenstand"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": ["facts"], "whole_library": false}, "cost": 0.0008633}
140
+ {"id": "syn-2842", "label": {"standalone_question": "Was bedeuten die Abkürzungen hinter den Artikeln auf dem Kassenbon vom Baumarkt?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005937}
141
+ {"id": "syn-2877", "label": {"standalone_question": "Is the city council funding vote for the proposed $1.2 million Elm Street Park playground and splash pad really in March, or is that date outdated, and what do neighbours think of the project, including any opposition?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Elm Street Park playground splash pad council vote", "Elm Street Park playground opposition neighbors", "Elm Street Park playground funding $1.2 million"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "cost": 0.000772}
142
+ {"id": "syn-2910", "label": {"standalone_question": "Write a related-work paragraph that positions Adjei & Solberg 2022 (Accessibility Option Adoption in AAA Titles: A Longitudinal Survey) against earlier work on accessibility in games in the library.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["accessibility options adoption in AAA games", "longitudinal survey of game accessibility features", "player disability and game accessibility design"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": true}, "cost": 0.0008004}
143
+ {"id": "syn-2913", "label": {"standalone_question": "How did Okonkwo & Tanaka (2016), Toxicity Detection in Multiplayer Voice Chat Using Acoustic Features, handle false positives in their toxicity classifier?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["false positives toxicity classifier voice chat acoustic features", "misclassification error handling toxicity detection multiplayer voice"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0007248}
144
+ {"id": "syn-2914", "label": {"standalone_question": "Is there research in the library on serious games for physical rehabilitation that reports adherence rates above 60 percent?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["serious games rehabilitation adherence rate home-based exercise", "adherence patterns physical therapy video games percentage"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "cost": 0.0005921}
145
+ {"id": "syn-2918", "label": {"standalone_question": "Which papers in my library look at gender representation or character creation, and do any of them compare across countries?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["gender representation character creation interfaces content analysis", "cross-national comparison gender representation video games", "avatar customization gender differences across countries"], "recency": "", "complexity": "standard", "lenses": ["scholar"], "facets": [], "whole_library": true}, "cost": 0.0009753}
146
+ {"id": "syn-2945", "label": {"standalone_question": "Was sind die wichtigsten Ergebnisse und Erkenntnisse aus allen Papern in der Sammlung zu Haushaltsfinanzen und Sparverhalten?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["present bias default enrollment retirement savings household financial decisions", "financial literacy intervention debt repayment buy-now-pay-later emergency savings financial stress"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": true}, "cost": 0.0008432}
147
+ {"id": "syn-2956", "label": {"standalone_question": "Welche Rolle spielen formative Rückmeldungen in der Studie von Hartmann & Greve (2019, Formative Feedback Loops in Blended Mathematics Classrooms) und welche Effektstärken berichten sie?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hartmann Greve formative feedback loops blended mathematics classrooms", "formative feedback blended mathematics classrooms effect size mixed methods", "formative feedback loops secondary mathematics effect size achievement"], "recency": "", "complexity": "standard", "lenses": ["scholar"], "facets": ["stats"], "whole_library": false}, "cost": 0.000896}
148
+ {"id": "syn-2957", "label": {"standalone_question": "Was sind die Kernergebnisse der sechs Arbeiten in der Sammlung insgesamt, als Stichpunkte?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["formative feedback learning analytics self-regulated learning", "teacher digital competence student engagement lower secondary", "gamified vocabulary learning retention motivation inclusive classroom"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": true}, "cost": 0.0009899}
149
+ {"id": "syn-2962", "label": {"standalone_question": "Warum funktioniert Gamification im Fremdsprachenunterricht, insbesondere für das Behalten von Vokabeln?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["gamified vocabulary learning retention motivation", "game-based learning second language vocabulary mechanisms"], "recency": "", "complexity": "standard", "lenses": ["scholar"], "facets": [], "whole_library": false}, "cost": 0.0010293}
150
+ {"id": "syn-2969", "label": {"standalone_question": "Was sind die Hauptergebnisse der Sammlung zur Lebensmittelsicherheit beim Fermentieren zu Hause?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["home fermentation food safety microbial hazards", "fermented vegetable brine sensory profiling production methods", "traditional cassava fermentation microbial succession households"], "recency": "", "complexity": "research", "lenses": ["scholar"], "facets": [], "whole_library": true}, "cost": 0.0009725}
training/data/synthetic_questions.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/data/train.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2e470e6e281eeb8b6fb2b3708cbd226a6a391d0e1ad8e967f72ffd81bf8e23c
3
+ size 39777233
training/data/valid.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/dev-syn-2b-v1-s350.jsonl ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"id": "syn-5", "score": {"json": true, "search": true}, "student": {"standalone_question": "Thank you for finding that EU report on heat pumps, my landlord finally took me seriously after I showed him the numbers.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.497}
2
+ {"id": "syn-10", "score": {"json": true, "search": true}, "student": {"standalone_question": "Is the same event happening tomorrow, 2026-10-17?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.456}
3
+ {"id": "syn-31", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.38461538461538464, "route": false}, "student": {"standalone_question": "Wie hat der Dreißigjährige Krieg angefangen?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Dreißigjähriger Krieg Anfang 1618 Ursachen", "Dreißigjähriger Krieg 1618 1648 Auslöser", "Dreißigjähriger Krieg Geschichte 1618 1648"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": ["timeline"], "whole_library": false}, "seconds": 0.676}
4
+ {"id": "syn-54", "score": {"json": true, "search": true}, "student": {"standalone_question": "Turn this text into a bulleted list for slides: \"The Meiji Restoration of 1868 ended the Tokugawa shogunate. It modernized Japan's military, education system, and economy. Western technology and legal models were widely adopted.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.606}
5
+ {"id": "syn-55", "score": {"json": true, "search": true}, "student": {"standalone_question": "Translate this into French: \"This vase, produced in Athens around 490 BCE, depicts warriors preparing for battle and was used in ceremonial feasts.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.518}
6
+ {"id": "syn-57", "score": {"json": true, "search": true}, "student": {"standalone_question": "Could you reformat this as a short APA-style citation? \"Hobsbawm, E. (1994). The Age of Extremes: The Short Twentieth Century, 1914-1991. Michael Joseph, London.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.629}
7
+ {"id": "syn-65", "score": {"json": true, "search": true}, "student": {"standalone_question": "Condense the sentence \"The Hanseatic League was a commercial and defensive confederation of merchant guilds and market towns in northern Europe, flourishing from the thirteenth to the seventeenth centuries and dominating Baltic trade.\" into a single sentence for a flashcard.", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.616}
8
+ {"id": "syn-129", "score": {"json": true, "search": true}, "student": {"standalone_question": "Könnten Sie diesen Satz bitte in einfaches Deutsch umschreiben? \"Die Kapitalertragsteuer wird unabhängig vom persönlichen Einkommensteuersatz pauschal mit 25 Prozent zuzüglich Solidaritätszuschlag erhoben.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.592}
9
+ {"id": "syn-138", "score": {"json": true, "search": true}, "student": {"standalone_question": "Kannst du das kurze Textfragment über ETFs erklären und zusammenfassen?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.453}
10
+ {"id": "syn-150", "score": {"json": true, "search": true}, "student": {"standalone_question": "Bitte übersetze ins Italienische: \"Ich möchte wissen, ob sich ein Wechsel zu einem Online-Broker für mich lohnt, wenn ich nur einmal im Monat einen kleinen Betrag investiere.\"", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.536}
11
+ {"id": "syn-151", "score": {"json": true, "search": true}, "student": {"standalone_question": "If I have 3 potions at 45 gold each and sell them at 1.15x, what is the total gold I get?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.507}
12
+ {"id": "syn-171", "score": {"json": true, "search": true}, "student": {"standalone_question": "How long does it take to heal 490 HP at 3.5 HP per second?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.489}
13
+ {"id": "syn-172", "score": {"json": true, "search": true}, "student": {"standalone_question": "Convert 0.75 cups to tablespoons", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.428}
14
+ {"id": "syn-178", "score": {"json": true, "search": true}, "student": {"standalone_question": "Wie viele Millimeter entsprechen 3/8 Zoll?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.469}
15
+ {"id": "syn-213", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.3125, "route": false}, "student": {"standalone_question": "Is it true that sharks don't get cancer, and that sharks have cartilage so no tumors form?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["sharks cartilage no tumors cancer", "shark cartilage cartilage tumors research", "shark cartilage cartilage tumors cartilage"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.627}
16
+ {"id": "syn-235", "score": {"json": true, "search": true}, "student": {"standalone_question": "Why does deep sleep drop but total sleep time increases on consecutive nights in my sleep tracker, as shown in the rows 04-10 and 04-11?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.555}
17
+ {"id": "syn-267", "score": {"json": true, "search": true}, "student": {"standalone_question": "Wie viel Farbe brauche ich ungefähr für 14 m², wenn der Verbrauch 0,15 l/m² beträgt?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.511}
18
+ {"id": "syn-284", "score": {"json": true, "search": false}, "student": {"standalone_question": "Warum gibt es bei Python 0.1 + 0.2 != 0.3, obwohl 0.1 + 0.2 = 0.3 in Python 3.8+ gilt?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Python float precision IEEE 754 0.1 + 0.2 == 0.3", "Python 3.8 float comparison equality 0.1 + 0.2 == 0.3"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.781}
19
+ {"id": "syn-344", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.6666666666666666, "route": true}, "student": {"standalone_question": "What is the most recent US unemployment rate from the BLS?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["BLS unemployment rate latest release", "Bureau of Labor Statistics unemployment rate September 2026"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.543}
20
+ {"id": "syn-346", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5, "route": true}, "student": {"standalone_question": "What is the most recent UN estimate of the population of Lagos, Nigeria?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["UN estimate population of Lagos Nigeria", "UN World Population Prospects 2024 Lagos Nigeria population"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "seconds": 0.589}
21
+ {"id": "syn-349", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4444444444444444, "route": true}, "student": {"standalone_question": "What is the current number of public electric vehicle charging points in the Netherlands, according to the latest count?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["public electric vehicle charging points Netherlands latest count", "RVO public EV charging points 2026", "RVO public EV charging points 2025"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.639}
22
+ {"id": "syn-360", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.8333333333333334, "route": true}, "student": {"standalone_question": "Welchen Zinssatz bietet die DKB aktuell für Tagesgeld an?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["DKB Tagesgeld Zinssatz aktuell", "Tagesgeld Zinssatz DKB 2026"], "recency": "year", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.611}
23
+ {"id": "syn-399", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.3, "route": true}, "student": {"standalone_question": "Wie hoch ist der Verbraucherpreisindex für Deutschland im letzten veröffentlichten Monat?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Verbraucherpreisindex Deutschland aktuell", "Verbraucherpreisindex Deutschland September 2027"], "recency": "month", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "seconds": 0.586}
24
+ {"id": "syn-403", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.35714285714285715, "route": true}, "student": {"standalone_question": "Wie viele Kilometer Autobahn hat Deutschland aktuell insgesamt?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Autobahnen Deutschland Gesamtlänge 2026", "Autobahnen Deutschland Länge Statistik Bundesamt für Bauwesen"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "seconds": 0.573}
25
+ {"id": "syn-458", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.4444444444444444, "route": false}, "student": {"standalone_question": "Welche Stadt ist die Hauptstadt von Australien, Canberra oder Sydney?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hauptstadt von Australien Canberra Sydney", "Australian capital Canberra Sydney"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.525}
26
+ {"id": "syn-474", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.23076923076923078, "route": false}, "student": {"standalone_question": "Wo wurde die Sparkasse zum ersten Mal gegründet?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Sparkasse Gründung 1811 Frankfurt am Main", "Sparkasse Geschichte Gründung 1811"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.568}
27
+ {"id": "syn-478", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4444444444444444, "route": true}, "student": {"standalone_question": "Wie hoch ist der Hvannadalshnúkur?", "answer_shape": "figure", "place": "", "wants_image": true, "image_query": "Hvannadalshnúkur", "needs_search": true, "search_queries": ["Hvannadalshnúkur Höhe Meter", "Hvannadalshnúkur height meters"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.582}
28
+ {"id": "syn-484", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": false, "lenses": true, "query_overlap": 0.25, "route": true}, "student": {"standalone_question": "Wie heißt Königsberg heute und in welchem Land liegt sie?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Königsberg heute Name Land", "Königsberg Stadt Deutschland Stadtteil"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.554}
29
+ {"id": "syn-487", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": false, "lenses": true, "query_overlap": 0.35294117647058826, "route": true}, "student": {"standalone_question": "Wer war der erste Mensch im Weltraum und in welchem Jahr war das?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["first human in space", "first human spaceflight 1961 Sputnik 1", "first human in space mission 1961"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.623}
30
+ {"id": "syn-492", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.15, "route": false}, "student": {"standalone_question": "Wer hat eigentlich die Glühbirne erfunden, Thomas Edison oder jemand anderes?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Glühbirne erfunden Thomas Edison 1879", "Edison Glühbirne Erfinder Biografie"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.589}
31
+ {"id": "syn-564", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.35714285714285715, "route": true}, "student": {"standalone_question": "Will it be windy in Dublin tomorrow, 25 February 2027?", "answer_shape": "weather", "place": "Dublin", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Dublin weather forecast tomorrow 25 February 2027 wind", "Dublin weather forecast 25 February 2027"], "recency": "day", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.639}
32
+ {"id": "syn-567", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.7142857142857143, "route": true}, "student": {"standalone_question": "How cold will it get in Montreal tonight?", "answer_shape": "weather", "place": "Montreal", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Montreal tonight temperature forecast", "Montreal weather tonight"], "recency": "day", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.523}
33
+ {"id": "syn-582", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": false, "image": true, "lenses": false, "query_overlap": 0.25, "route": false}, "student": {"standalone_question": "Ist Glätte auf der A9 (Bundesstraße 9) in Deutschland zu erwarten, wenn ich morgen früh um 6 Uhr nach München zum Flughafen muss?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Glätte A9 München aktuell", "Glätte A9 München Vorhersage", "Glätte A9 München 2027"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.668}
34
+ {"id": "syn-606", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5, "route": true}, "student": {"standalone_question": "What is the next fixture for Real Madrid in La Liga?", "answer_shape": "match", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Real Madrid next La Liga fixture", "Real Madrid upcoming match schedule"], "recency": "year", "complexity": "simple", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.555}
35
+ {"id": "syn-631", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.6, "route": true}, "student": {"standalone_question": "Welche Spiele spielen in der 2. Bundesliga am Wochenende (Wochenende, 28. Januar 2027) statt?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["2. Bundesliga Spiele Wochenende 28. Januar 2027", "2. Bundesliga Spielplan Wochenende 2027"], "recency": "week", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.567}
36
+ {"id": "syn-641", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.5, "route": true}, "student": {"standalone_question": "Wie viele Gelbe Karten hat Robert Andrich dieses Jahr bekommen?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Robert Andrich Gelbe Karten 2026", "Robert Andrich Gelbe Karte 2026 Anzahl"], "recency": "year", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "seconds": 0.601}
37
+ {"id": "syn-651", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.2727272727272727, "route": true}, "student": {"standalone_question": "Welche Tore hat der Fußballverein Hoffenheim gegen Augsburg am Mittwoch, 11. Oktober 2027, erzielt?", "answer_shape": "match", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hoffenheim Augsburg Spiel Ergebnis", "Hoffenheim Augsburg 11. Oktober 2027 Ergebnis"], "recency": "week", "complexity": "simple", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.631}
38
+ {"id": "syn-657", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.6, "route": true}, "student": {"standalone_question": "Top selling electric cars in Europe last year ranking", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["top selling electric cars Europe 2025 ranking", "electric vehicle sales Europe by model 2025", "EU electric vehicle registrations by model 2025"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.619}
39
+ {"id": "syn-669", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.4, "route": false}, "student": {"standalone_question": "Motogp rider standings after the last Grand Prix", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Motogp rider standings after last Grand Prix", "Motogp 2026 standings", "Motogp 2026 latest race results"], "recency": "year", "complexity": "standard", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "seconds": 0.607}
40
+ {"id": "syn-694", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.16666666666666666, "route": true}, "student": {"standalone_question": "Wer hat aktuell die meisten Champions League Tore in der Saison 2025/26?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Champions League Tore 2025/26 Torengeschießer", "Champions League Tore 2025/26 Tabelle Torengeschießer"], "recency": "year", "complexity": "simple", "lenses": ["web", "news"], "facets": [], "whole_library": false}, "seconds": 0.672}
41
+ {"id": "syn-699", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.75, "route": true}, "student": {"standalone_question": "Wie sieht die aktuelle NBA Western Conference Standings aus?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["NBA Western Conference standings 2026", "NBA Western Conference standings current season"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.462}
42
+ {"id": "syn-754", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.38461538461538464, "route": true}, "student": {"standalone_question": "Who is Dr Nina Schwalbe, the GP who writes about patient care, and what is she known for?", "answer_shape": "profile", "place": "", "wants_image": true, "image_query": "Nina Schwalbe", "needs_search": true, "search_queries": ["Nina Schwalbe GP patient care", "Nina Schwalbe biography", "Nina Schwalbe patient care writing"], "recency": "year", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.67}
43
+ {"id": "syn-766", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.21428571428571427, "route": false}, "student": {"standalone_question": "Gibt es neuere Forscher, die sich mit den Ideen von Charles Darwin weiterbeschäftigt haben?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Charles Darwin Ideen aktuelle Forschung Neuere Forscher", "Darwinian ideas contemporary research review", "Darwinism modern biology research 2020s"], "recency": "year", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.633}
44
+ {"id": "syn-847", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.6923076923076923, "route": true}, "student": {"standalone_question": "Is Bethesda's Creation Kit for Starfield available on Xbox Series X or only on PC, and what is the mod upload limit?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bethesda Creation Kit Starfield Xbox Series X PC", "Starfield Creation Kit mod upload limit", "Bethesda Creation Kit Starfield mod upload limit"], "recency": "year", "complexity": "standard", "lenses": ["web", "docs"], "facets": [], "whole_library": false}, "seconds": 0.646}
45
+ {"id": "syn-863", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": false, "image": true, "lenses": true, "query_overlap": 0.26666666666666666, "route": false}, "student": {"standalone_question": "Wie ist der Bundesverband Deutscher Gartenfreunde (BGdG) organisiert und haben seine Mitglieder Mitspracherecht?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bundesverband Deutscher Gartenfreunde Organisation Mitglieder", "BGdG Mitglieder Mitspracherecht Bundesverband"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.655}
46
+ {"id": "syn-883", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.09090909090909091, "route": true}, "student": {"standalone_question": "Was ist das Max-Planck-Institut und wie viel Forschungsgeld bekommt es vom Staat?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Max-Planck-Institut Geschichte Gründung Max Planck", "Max-Planck-Institut Forschungsgelder Staat 2026", "Max-Planck-Institut Budget 2026 Forschungsgelder"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.598}
47
+ {"id": "syn-891", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": false, "query_overlap": 0.16666666666666666, "route": true}, "student": {"standalone_question": "Gibt es die Firma Hornbach noch, oder ist Hornbach nur ein Baumarkt?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hornbach Baumarkt GmbH", "Hornbach Baumarkt GmbH Geschichte", "Hornbach Baumarkt GmbH aktuell"], "recency": "year", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.5}
48
+ {"id": "syn-946", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.4444444444444444, "route": true}, "student": {"standalone_question": "Wie sieht das Verkehrszeichen für Vorfahrt achten aus?", "answer_shape": "image", "place": "", "wants_image": true, "image_query": "Vorfahrt achten", "needs_search": true, "search_queries": ["Vorfahrt achten Verkehrszeichen Aussehen", "Vorfahrt achten Zeichen Deutschland"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.571}
49
+ {"id": "syn-998", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": false, "image": true, "lenses": true, "query_overlap": 0.42857142857142855, "route": false}, "student": {"standalone_question": "What is a heat pump mode on a smart thermostat, and how does it differ from a traditional furnace?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["heat pump vs furnace thermostat how they work", "heat pump thermostat explained"], "recency": "year", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.606}
50
+ {"id": "syn-1028", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.42105263157894735, "route": true}, "student": {"standalone_question": "Why do leaves change color in fall?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["why leaves change color in fall chlorophyll breakdown", "fall leaf color change mechanism photosynthesis chlorophyll"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.565}
51
+ {"id": "syn-1032", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.38095238095238093, "route": true}, "student": {"standalone_question": "What is inflation, and why does it happen?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["inflation definition causes why it happens", "inflation explained supply and demand money supply"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.537}
52
+ {"id": "syn-1055", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.20833333333333334, "route": false}, "student": {"standalone_question": "Wie funktioniert der Zinseszins bei alten Sparbüchern aus den 1980er-Jahren, und wie hoch ist der aktuelle Wert eines solchen Sparbuchs?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Zinseszins Sparbuch 1980er Jahre Funktionsweise", "Sparbuch Zinseszins Berechnung Beispiel", "Sparbuch Wert heute Zinssatz Vergleich"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.73}
53
+ {"id": "syn-1067", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.1, "route": false}, "student": {"standalone_question": "Wie werden die Punkte bei den Fehlschusspunkten im Biathlon berechnet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Biathlon Fehlschuss Punkte Berechnung", "Biathlon Fehlschuss Punkte Tabelle 2026", "Biathlon Fehlschuss Punkte 100 m 1000 m 10000 m"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.684}
54
+ {"id": "syn-1087", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4166666666666667, "route": true}, "student": {"standalone_question": "Warum wird ein Kondensator im Netzteil so heiß?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Kondensator Netzteil Wärme Verlustleistung", "Kondensator Netzteil Überlastung Überhitzung Ursachen", "Kondensator Netzteil Verlustleistung Verlustleistung"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.626}
55
+ {"id": "syn-1089", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.23529411764705882, "route": false}, "student": {"standalone_question": "Warum braucht man bei PVC-Fenstern eine Dampfsperre, und was passiert, wenn man sie weglässt?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["PVC Fenster Dampfsperre Funktion", "PVC Fenster Dampfsperre weglassen Folgen", "PVC Fenster Dampferdichtung Funktionsweise"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.659}
56
+ {"id": "syn-1114", "score": {"json": false}, "student": null, "seconds": 0.0}
57
+ {"id": "syn-1130", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4444444444444444, "route": true}, "student": {"standalone_question": "How do I clean a dryer vent safely, step by step?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["how to clean dryer vent safely step by step", "dryer vent cleaning safety precautions", "dryer vent cleaning tools and safety"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.584}
58
+ {"id": "syn-1138", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.24, "route": true}, "student": {"standalone_question": "How do I set up a shared family calendar on my phone and my partner's phone so we both see the kids' school events?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["shared family calendar app how to set up", "Google Calendar family calendar invite family members", "Apple Calendar family sharing invite family members"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.631}
59
+ {"id": "syn-1145", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.7272727272727273, "route": true}, "student": {"standalone_question": "What steps do I need to take to renew my driver's license online before it expires next month?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["renew driver's license online steps", "renew driver's license online requirements", "renew driver's license online before expiration"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.603}
60
+ {"id": "syn-1215", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.6111111111111112, "route": true}, "student": {"standalone_question": "Should I use wood glue or epoxy to repair a cracked chair leg that gets a lot of weight on it?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["wood glue vs epoxy repair cracked chair leg", "wood glue epoxy repair load bearing furniture", "epoxy wood glue repair strength testing"], "recency": "", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.628}
61
+ {"id": "syn-1243", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.2857142857142857, "route": false}, "student": {"standalone_question": "Should I use a 6 oz or 8 oz graphite shaft for a tennis racket if I have tennis elbow?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["tennis racket shaft weight tennis elbow", "tennis racket shaft weight grip tennis elbow"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.605}
62
+ {"id": "syn-1244", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.30434782608695654, "route": true}, "student": {"standalone_question": "Should I get a power meter pedal or a crank-based power meter for my first training block?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["power meter pedal vs crank-based power meter first training block", "power meter pedal pros cons beginner cyclists", "crank-based power meter pros cons beginner cyclists"], "recency": "year", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.647}
63
+ {"id": "syn-1273", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.19047619047619047, "route": true}, "student": {"standalone_question": "Soll ich für die Gartenbewässerung Regenwasser aus der Tonne oder Brunnenwasser nutzen, Stichwort Kalk und Keime?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["rainwater vs well water garden irrigation calcium algae", "well water garden irrigation calcium carbonate algae growth"], "recency": "", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.646}
64
+ {"id": "syn-1313", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.6153846153846154, "route": false}, "student": {"standalone_question": "Lohnt sich der Kauf der Deluxe-Edition von Cyberpunk 2077 Phantom Liberty, oder reicht die Standard-Version?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Cyberpunk 2077 Phantom Liberty Deluxe Edition vs Standard Edition", "Phantom Liberty Deluxe Edition Features Preis"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.771}
65
+ {"id": "syn-1337", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.3888888888888889, "route": true}, "student": {"standalone_question": "Stripe vs Square vs PayPal for a small shop: processing fees", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Stripe vs Square vs PayPal processing fees 2027", "Stripe payment processing fee percentage monthly", "Square payment processing fee percentage monthly"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.607}
66
+ {"id": "syn-1341", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.2857142857142857, "route": true}, "student": {"standalone_question": "Vergleich Lissabon, Porto und Madeira für einen 10-tägigen Herbsturlaub: was ist günstiger und wo ist das Wetter besser?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Lissabon Porto Madeira Vergleich Preise 2027", "Portugal Madeira Lissabon Wetter Herbst", "Madeira Lissabon Porto Preise Unterkunft 2027"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.741}
67
+ {"id": "syn-1356", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.375, "route": true}, "student": {"standalone_question": "Lohnt es sich, Netflix, Disney+ und Amazon Prime Video zu nutzen, wenn man hauptsächlich Serien sieht?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Netflix Disney+ Amazon Prime Video Vergleich Serien", "Netflix vs Disney+ vs Amazon Prime Video Serien", "Netflix Disney+ Amazon Prime Video Serien Vergleich Preis"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.62}
68
+ {"id": "syn-1378", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.23809523809523808, "route": true}, "student": {"standalone_question": "Does playing music during surgery lower anxiety and reduce the amount of pain medication needed?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["music during surgery anxiety pain medication randomized trial", "music therapy perioperative pain systematic review meta-analysis", "music during surgery anxiety reduction systematic review"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.607}
69
+ {"id": "syn-1381", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.625, "route": true}, "student": {"standalone_question": "Do blue light glasses help people sleep better, or is the effect a placebo?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["blue light glasses sleep study systematic review", "blue light glasses placebo effect sleep randomized trial", "blue light glasses sleep quality meta-analysis"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.726}
70
+ {"id": "syn-1400", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.5416666666666666, "route": true}, "student": {"standalone_question": "Is there research linking Minecraft and Roblox gaming to anxiety or depression in teenagers?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["video game addiction anxiety depression adolescents systematic review", "Roblox Minecraft gaming mental health longitudinal study", "gaming screen time anxiety depression meta-analysis"], "recency": "year", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.621}
71
+ {"id": "syn-1432", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.4166666666666667, "route": true}, "student": {"standalone_question": "Do studies show standing desks reduce back or neck pain, or is the benefit mostly about switching positions?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["standing desk back pain neck pain randomized controlled trial", "standing desk versus sitting desk neck pain systematic review", "standing desk posture switching position effect pain"], "recency": "year", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.626}
72
+ {"id": "syn-1435", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.5217391304347826, "route": true}, "student": {"standalone_question": "Does research support vitamin D supplementation improving mood, or is the effect mostly a placebo?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["vitamin D supplementation mood randomized controlled trial meta-analysis", "vitamin D mood depression systematic review", "vitamin D mood placebo effect study"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.6}
73
+ {"id": "syn-1439", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.55, "route": true}, "student": {"standalone_question": "Is there solid research on whether cold showers boost mood and immunity, or is the evidence mostly anecdotal?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["cold shower mood anxiety randomized controlled trial", "cold water immersion immune response systematic review", "cold shower immunity meta-analysis"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.613}
74
+ {"id": "syn-1466", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.0, "route": false}, "student": {"standalone_question": "Macht Krafttraining ältere Menschen weniger sturzgefährdet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["strength training fall risk older adults meta-analysis", "resistance training fall prevention elderly systematic review", "strength training fall risk systematic review"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.558}
75
+ {"id": "syn-1468", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.10714285714285714, "route": false}, "student": {"standalone_question": "Was passiert mit dem Körper, wenn man drei Tage nichts isst?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["3 Tage nichts essen Symptome Körper", "fasting 3 days metabolic effects ketosis", "fasting 3 days body weight muscle protein"], "recency": "", "complexity": "standard", "lenses": [], "facets": ["callout"], "whole_library": false}, "seconds": 0.585}
76
+ {"id": "syn-1478", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.375, "route": true}, "student": {"standalone_question": "Ist Ibuprofen nach dem Krafttraining schädlich für den Muskelaufbau, wie ich es irgendwo gelesen habe? Was zeigen die Studien?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["NSAID post-exercise muscle hypertrophy randomized controlled trial", "NSAID supplementation muscle strength hypertrophy meta-analysis", "NSAID post-exercise recovery muscle damage study"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.704}
77
+ {"id": "syn-1595", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5454545454545454, "route": true}, "student": {"standalone_question": "How do I configure GitHub Actions cache with restore-keys and fallback?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["GitHub Actions cache restore-keys fallback", "GitHub Actions cache configuration restore-keys fallback documentation"], "recency": "", "complexity": "simple", "lenses": ["docs", "web"], "facets": [], "whole_library": false}, "seconds": 0.532}
78
+ {"id": "syn-1614", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.375, "route": false}, "student": {"standalone_question": "How do I configure Redis's maxmemory-policy setting so that cached drug interaction lookups get evicted gracefully under memory pressure?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Redis maxmemory-policy eviction policy", "Redis maxmemory-policy noeviction volatile-lru noeviction all"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.605}
79
+ {"id": "syn-1660", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.3076923076923077, "route": false}, "student": {"standalone_question": "Wie aktiviere ich in nginx gzip-compression für JSON-Responses?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["nginx gzip compression json responses", "nginx gzip on json responses example"], "recency": "", "complexity": "simple", "lenses": ["docs", "web"], "facets": [], "whole_library": false}, "seconds": 0.535}
80
+ {"id": "syn-1664", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.625, "route": true}, "student": {"standalone_question": "Was macht der Parameter --no-cache bei docker build genau?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["docker build --no-cache parameter", "docker build --no-cache documentation"], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.521}
81
+ {"id": "syn-1678", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.8571428571428571, "route": true}, "student": {"standalone_question": "Is the major airline strike resolved yet?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["major airline strike resolved", "airline strike latest news 2027"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.531}
82
+ {"id": "syn-1696", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5333333333333333, "route": true}, "student": {"standalone_question": "What is the latest news on the housing market cooling off this month?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["housing market cooling off news May 2026", "housing prices new data May 2026", "housing market slowdown latest report"], "recency": "month", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.617}
83
+ {"id": "syn-1699", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.6111111111111112, "route": true}, "student": {"standalone_question": "What is the latest news on the US Department of Education's federal student loan forgiveness rules this week (as of November 2026)?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["US Department of Education federal student loan forgiveness rules latest news", "student loan forgiveness rules update November 2026", "Department of Education student loan forgiveness policy changes"], "recency": "week", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.686}
84
+ {"id": "syn-1716", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.25, "route": true}, "student": {"standalone_question": "What is the latest on school voucher programs passed by state legislatures in 2026?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["school voucher programs passed by state legislatures 2026", "school voucher legislation state legislature 2026", "school voucher programs enacted 2026 by state legislature"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.677}
85
+ {"id": "syn-1724", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": false, "query_overlap": 0.2631578947368421, "route": false}, "student": {"standalone_question": "Gibt es neue Informationen zur Grippe in Deutschland diese Woche, und wird die Lage schlimmer?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Grippe Deutschland aktuelle Lage Woche", "Grippe Deutschland Infektionszahlen September 2027", "Grippe Deutschland Gesundheitsamt aktuell"], "recency": "week", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.628}
86
+ {"id": "syn-1737", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.15, "route": true}, "student": {"standalone_question": "Gibt es aktuell Probleme mit der Tetanus-Impfung oder Lieferengpässe?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Tetanus-Impfung Lieferengpässe aktuell", "tetanus vaccine supply shortage 2026", "tetanus immunization stockpiles global"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.626}
87
+ {"id": "syn-1743", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.38461538461538464, "route": true}, "student": {"standalone_question": "Wie ist der aktuelle Stand der Ermittlungen zum Zugunglück in Deutschland (letzte Woche, April 2027)?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Zugunglück Ermittlungen Stand April 2027", "Zugunglück Ermittlungen aktuell Bundesamt für Straßensicherheit"], "recency": "week", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.798}
88
+ {"id": "syn-1773", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.36363636363636365, "route": true}, "student": {"standalone_question": "Was gibt es Neues zur Kita-Finanzierung oder zum Fachkräfteprogramm des Bundes (Stand 2026)?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Kita-Finanzierung Neues 2026", "Fachkräfteprogramm Bundes 2026", "Kita-Finanzierung Reform 2026"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.689}
89
+ {"id": "syn-1788", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5, "route": true}, "student": {"standalone_question": "How was the first email sent, and who sent it?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["first email sent history", "first email sent 1971 ARPANET", "first email sent who sent it"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.561}
90
+ {"id": "syn-1805", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": false, "image": true, "lenses": true, "query_overlap": 0.16666666666666666, "route": false}, "student": {"standalone_question": "Wie hat sich die Bausparkasse in Deutschland seit den 1970er-Jahren entwickelt?", "answer_shape": "profile", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Bausparkasse Geschichte Entwicklung Deutschland", "Bausparkasse Schwäbisch Hall Geschichte", "Bausparkasse Deutschland Geschichte Bausparvertrag"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.645}
91
+ {"id": "syn-1818", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.3333333333333333, "route": false}, "student": {"standalone_question": "Wie entstand das Modell der Kreditkarte in Deutschland?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Geschichte Kreditkarte Deutschland Entstehung", "Kreditkarte Deutschland Entwicklung Banken 1970er 1980er", "Kreditkarte Deutschland Geschichte Banken 1970er 1980er"], "recency": "", "complexity": "research", "lenses": [], "facets": ["timeline"], "whole_library": false}, "seconds": 0.667}
92
+ {"id": "syn-1824", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.19047619047619047, "route": false}, "student": {"standalone_question": "Wie kam es zur Privatisierung der Telekom und was ist seit dem Börsengang passiert?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Telekom Privatisierung Börsengang 2007 Geschichte", "Telekom Privatunternehmen Börsengang 2007 Folgen", "Telekom Privatunternehmen Börsengang 2007 Bilanz 2027"], "recency": "year", "complexity": "research", "lenses": ["web", "news"], "facets": ["timeline"], "whole_library": false}, "seconds": 0.719}
93
+ {"id": "syn-1827", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.34782608695652173, "route": false}, "student": {"standalone_question": "Should I register as self-employed or stay on the freelance tax form?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["self-employed vs freelance tax form pros cons", "self-employed vs sole trader tax filing UK 2026", "self-employed vs sole trader tax filing comparison"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.618}
94
+ {"id": "syn-1857", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.4, "route": false}, "student": {"standalone_question": "What are the pros and cons of moving to a farm to be closer to wildlife?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["pros and cons of living on a farm near wildlife", "farm life wildlife proximity benefits drawbacks", "farm life wildlife proximity mental health studies"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": ["proscons"], "whole_library": false}, "seconds": 0.622}
95
+ {"id": "syn-1861", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.5294117647058824, "route": true}, "student": {"standalone_question": "What are the pros and cons of volunteering at a wildlife rehabilitation center on weekends?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["volunteering wildlife rehabilitation center pros cons", "volunteer wildlife rehab center weekend pros cons", "volunteer wildlife rehab center pros cons"], "recency": "", "complexity": "standard", "lenses": [], "facets": ["proscons"], "whole_library": false}, "seconds": 0.597}
96
+ {"id": "syn-1884", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4, "route": true}, "student": {"standalone_question": "Lohnt es sich, auf 75% der Arbeitszeit (Teilzeit) wegen Pflege meiner Mutter zu gehen, wenn Rentenpunkte und Karriere im Hinterkopf stehen?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Teilzeit Pflege 75% Rentenpunkte Vergleich", "Teilzeit Pflege 75% Karriere Karrierechancen", "Teilzeit Pflege 75% Nachteile Rentenversicherung"], "recency": "year", "complexity": "research", "lenses": [], "facets": ["proscons"], "whole_library": false}, "seconds": 0.698}
97
+ {"id": "syn-1886", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.3157894736842105, "route": false}, "student": {"standalone_question": "Lohnt sich ein Abo bei Netflix plus Disney plus, oder reicht eines davon für meine Serien?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Netflix vs Disney+ Serien Abo Vergleich", "Netflix plus Disney plus Abo Kosten Vergleich 2027", "Netflix plus Disney plus Abo Vor- und Nachteile"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.662}
98
+ {"id": "syn-1893", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.16, "route": true}, "student": {"standalone_question": "Sollte ich bei -3,5 Dioptrien und dünner Hornhaut die Lasik-Operation oder Kontaktlinsen als bessere Wahl wählen?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Laserik vs Kontaktlinsen -3,5 Dioptrien dünne Hornhaut Studie", "Laserik Kontaktlinsen Vergleich pros und cons", "Laserik Kontaktlinsen Indikation dünne Hornhaut"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.751}
99
+ {"id": "syn-1915", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": false, "query_overlap": 0.20833333333333334, "route": false}, "student": {"standalone_question": "Vorteile und Nachteile: Start-up vs Konzern", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Start-up vs Konzern Vor- und Nachteile", "Start-up Nachteile hohe Risiken hohe Kosten", "Konzern Nachteile Bürokratie Bürokratie Kosten"], "recency": "", "complexity": "research", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.596}
100
+ {"id": "syn-1925", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.38461538461538464, "route": true}, "student": {"standalone_question": "Lohnt sich der Kauf einer Wärmepumpe 2024?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wärmepumpe 2024 Kosten Rückzahlung Strompreis", "Wärmepumpe 2024 Vergleich Strompreis Förderung"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["stats", "proscons"], "whole_library": false}, "seconds": 0.643}
101
+ {"id": "syn-1941", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.55, "route": true}, "student": {"standalone_question": "What are the side effects of creatine monohydrate supplementation on kidney function markers in healthy adults?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["creatine monohydrate supplementation kidney function markers healthy adults", "creatine supplementation creatinine eGFR systematic review meta-analysis", "creatine supplementation adverse effects kidney function systematic review"], "recency": "year", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.68}
102
+ {"id": "syn-1966", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.25, "route": true}, "student": {"standalone_question": "Welche chemischen Zutaten sind in einem Laptop-Akku enthalten, und welche Brandgefahr besteht dabei?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Laptop Akku chemische Zutaten Lithium Cobalt Oxid", "Laptop Akku Brandgefahr Lithium Cobalt Oxid", "Laptop Akku chemische Zusammensetzung Lithium Cobalt Oxid"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["list", "callout"], "whole_library": false}, "seconds": 0.691}
103
+ {"id": "syn-1975", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.38461538461538464, "route": true}, "student": {"standalone_question": "Welche Symptome deuten auf einen Cache-Fehler im Layer-System bei Docker-Builds hin?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Docker layer cache build failure symptoms", "Docker build layer cache invalidation symptoms"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.573}
104
+ {"id": "syn-2042", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": false, "lenses": true, "query_overlap": 0.8571428571428571, "route": true}, "student": {"standalone_question": "Compare the PS5 Pro and Xbox Series X price right now, and which has better 4K frame rates in recent reviews", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["PS5 Pro vs Xbox Series X price comparison 2026", "PS5 Pro vs Xbox Series X 4K frame rate benchmark review"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.669}
105
+ {"id": "syn-2066", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.2631578947368421, "route": true}, "student": {"standalone_question": "Was ist der Unterschied zwischen einer Probezeit und einer Befristung bei Arbeitsverträgen in Deutschland?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Probezeit Befristung Arbeitsvertrag Unterschied Deutschland", "Probezeit Befristung Arbeitsvertrag § 623 BGB § 624 BGB", "Befristung Arbeitsvertrag Ablauf Befristungsdauer"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.697}
106
+ {"id": "syn-2083", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.35, "route": true}, "student": {"standalone_question": "Can I apply to extend my Irish Stamp 1G visa for a job search, and what is the deadline to submit the application before it expires?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Stamp 1G visa extension job search deadline", "Stamp 1G visa extension application deadline Ireland", "Stamp 1G visa extension job search rules"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["callout"], "whole_library": false}, "seconds": 0.677}
107
+ {"id": "syn-2104", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.3333333333333333, "route": true}, "student": {"standalone_question": "What is the deadline to file a notice of intent to relocate with my child's school district in California if we move out of the attendance zone mid-semester?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["California school district notice of intent to relocate deadline", "California school district relocation notice filing deadline", "California school district notice of intent to relocate 30 days"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.682}
108
+ {"id": "syn-2124", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.13793103448275862, "route": true}, "student": {"standalone_question": "What are the deadlines for a Brazilian employee to file a labor claim against an employer for unpaid overtime, and does the three-year prescription period start from the termination date?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["prescrição de 3 anos direito de ação trabalhista Brasil", "prescrição 3 anos direito de ação trabalhista início prescrição", "prescrição 3 anos direito de ação trabalhista início prescrição data do fim do contrato"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": ["callout"], "whole_library": false}, "seconds": 0.773}
109
+ {"id": "syn-2136", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.125, "route": false}, "student": {"standalone_question": "Wie lange habe ich Zeit, eine Kündigungsschutzklage einzureichen, wenn mein Arbeitgeber mir wegen angeblicher zu häufiger Krankheit kündigt, dabei nur ein Bandscheibenvorfall vorlag?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Kündigungsschutzklage Einreichungsfrist Arbeitszeitgesetz 15 Tage", "Kündigungsschutzklage Frist Arbeitszeitgesetz 15 Tage Arbeitszeit", "Kündigungsschutzklage Frist Arbeitszeitgesetz 15 Tage Arbeitszeit"], "recency": "", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.8}
110
+ {"id": "syn-2177", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.2727272727272727, "route": true}, "student": {"standalone_question": "Ab welchem Alter darf man in der Schweiz alleine Auto fahren, und gilt der deutsche Führerschein dort noch?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Schweiz Führerschein Mindestalter 2026", "Schweiz Führerschein Geltung Deutschland Führerschein 2026"], "recency": "year", "complexity": "standard", "lenses": [], "facets": ["facts"], "whole_library": false}, "seconds": 0.629}
111
+ {"id": "syn-2210", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.16666666666666666, "route": false}, "student": {"standalone_question": "Gibt es beim Eierpochieren ohne Essigzusatz noch vernünftige Gründe, und wie wirkt sich das auf die Eierqualität aus?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Eierpochieren ohne Essig Essigzusatz Wirkung", "Eierpochieren ohne Essig Geschmack Qualität Studie", "Eierpochieren Essig Essigzusatz Geschmack Studie"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.689}
112
+ {"id": "syn-2292", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.42857142857142855, "route": true}, "student": {"standalone_question": "Best backpack under 100 for carrying a laptop and charger to the office, that doesn't look like a school bag", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["best laptop backpack under 100 office use", "laptop backpack vs messenger bag office look"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.629}
113
+ {"id": "syn-2302", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.4782608695652174, "route": true}, "student": {"standalone_question": "Good budget microphone for podcasting or recording voiceovers at home under 100", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["best budget microphone for podcasting under 100", "best home voiceover microphone 2026 review", "budget USB microphone review comparison"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.618}
114
+ {"id": "syn-2320", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": false, "lenses": true, "query_overlap": 0.25, "route": true}, "student": {"standalone_question": "Welche Wärmebildkamera fürs Handy oder als Gerät ist für Energieverluste im Haus unter 150 Euro sinnvoll?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wärmebildkamera Handy Vergleich Test", "Wärmebildkamera Gerät Test Energieverluste Haus", "Wärmebildkamera Test 2026 Preis Leistung"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.689}
115
+ {"id": "syn-2339", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": true, "image": false, "lenses": false, "query_overlap": 0.35714285714285715, "route": false}, "student": {"standalone_question": "What does a good sourdough crumb look like when the sourdough is properly fermented?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["sourdough crumb texture properly fermented", "sourdough crumb characteristics good fermentation"], "recency": "", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.588}
116
+ {"id": "syn-2352", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.35714285714285715, "route": true}, "student": {"standalone_question": "Wie viel Wasser verbrauchen eine Familie mit 4 Personen pro Tag?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Wasserverbrauch Familie 4 Personen pro Tag", "Wasserbedarf Haushalt 4 Personen Studie", "Wasserverbrauch Haushalt Deutschland Studie"], "recency": "year", "complexity": "simple", "lenses": ["data", "web"], "facets": [], "whole_library": false}, "seconds": 0.609}
117
+ {"id": "syn-2391", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.4666666666666667, "route": false}, "student": {"standalone_question": "What is the flow of blood through the human heart, naming each chamber and valve in order?", "answer_shape": "diagram", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["human heart blood flow chambers valves sequence", "heart anatomy atria ventricles valves flow"], "recency": "", "complexity": "simple", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.596}
118
+ {"id": "syn-2398", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.47368421052631576, "route": true}, "student": {"standalone_question": "How does a wastewater treatment plant clean water, and what are the stages of the process?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["wastewater treatment plant stages primary secondary tertiary", "how wastewater treatment plant cleans water process"], "recency": "", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.585}
119
+ {"id": "syn-2413", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.4375, "route": true}, "student": {"standalone_question": "Wie läuft die Blutspende in Deutschland ab: Anmeldung, Voruntersuchung, Spende und Nachsorge – was passiert in jedem Abschnitt?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Blutspende Deutschland Ablauf Anmeldung Voruntersuchung Spende Nachsorge", "Blutspendedienst Deutschland Ablauf Schritte Blutspender", "Blutspende Deutschland Nachsorge Blutspendedienst Ablauf"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.733}
120
+ {"id": "syn-2464", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.4375, "route": true}, "student": {"standalone_question": "What are good sleep hygiene tips for shift workers?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["shift work sleep hygiene tips", "circadian rhythm shift workers sleep improvement", "shift work sleep hygiene guidelines"], "recency": "year", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.545}
121
+ {"id": "syn-2465", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": false, "image": true, "lenses": false, "query_overlap": 0.47058823529411764, "route": false}, "student": {"standalone_question": "Compare Python and JavaScript for beginner programmers", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Python vs JavaScript beginner programming comparison", "Python JavaScript learning curve syntax features", "JavaScript vs Python beginner tutorial"], "recency": "year", "complexity": "research", "lenses": ["web", "docs"], "facets": [], "whole_library": false}, "seconds": 0.542}
122
+ {"id": "syn-2482", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.2222222222222222, "route": true}, "student": {"standalone_question": "How do vaccines train the immune system?", "answer_shape": "explainer", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["how vaccines train the immune system mechanism", "vaccine immunology how it works memory B cells T cells"], "recency": "", "complexity": "standard", "lenses": ["reference", "web"], "facets": [], "whole_library": false}, "seconds": 0.54}
123
+ {"id": "syn-2483", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": false, "query_overlap": 0.46153846153846156, "route": false}, "student": {"standalone_question": "Which indie games similar to Stardew Valley are worth playing?", "answer_shape": "table", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["best indie games like Stardew Valley", "Stardew Valley alternatives indie games 2027"], "recency": "year", "complexity": "standard", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.583}
124
+ {"id": "syn-2486", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.25, "route": true}, "student": {"standalone_question": "Wie lange dauert die Ausbildung zur Pflegefachfrau in Deutschland?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ausbildung Pflegefachfrau Dauer", "Pflegefachfrau Ausbildungsdauer Bundesagentur für Arbeit"], "recency": "year", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.569}
125
+ {"id": "syn-2506", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "query_overlap": 0.14285714285714285, "route": false}, "student": {"standalone_question": "Wie bekomme ich in NRW eine Einsicht in meine Klausur aus dem Rechtshochschulfach (Rechtshochschule)?", "answer_shape": "steps", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Einsicht Rechtshochschule NRW Ablauf", "Einsicht Rechtshochschule NRW Antrag Frist", "Einsicht Rechtshochschule NRW Unterlagen"], "recency": "year", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.665}
126
+ {"id": "syn-2507", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.2, "route": false}, "student": {"standalone_question": "Rezept für glutenfreies Kuchengericht schnell", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["glutenfreies Kuchengericht schnell Rezept", "glutenfreie Kuchengericht 15 Minuten Rezept"], "recency": "", "complexity": "simple", "lenses": ["web"], "facets": [], "whole_library": false}, "seconds": 0.556}
127
+ {"id": "syn-2510", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.15, "route": true}, "student": {"standalone_question": "Welche Fristen gelten für den Widerspruch gegen einen Steuerbescheid?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Widerspruch gegen Steuerbescheid Frist 14 Tage", "Widerspruch Steuerbescheid § 143 AO Frist", "Widerspruch Steuerbescheid § 143 AO 14 Tage"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.652}
128
+ {"id": "syn-2514", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": false, "query_overlap": 0.36363636363636365, "route": false}, "student": {"standalone_question": "Wo kann ich in Deutschland eine beglaubigte Geburtsurkunde bestellen?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["beglaubigte Geburtsurkunde bestellen wo", "beglaubigte Geburtsurkunde bestellen Postamt Bürgeramt", "beglaubigte Geburtsurkunde bestellen Online"], "recency": "year", "complexity": "standard", "lenses": ["web"], "facets": [], "whole_library": false}, "seconds": 0.601}
129
+ {"id": "syn-2548", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.5416666666666666, "route": false}, "student": {"standalone_question": "Are there studies on chilling onions before cutting them to reduce the gas that causes tears?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["onion chilling before cutting syn-propanethial-S-oxide study", "onion cold storage gas release eye irritation", "onion cryogenic storage sensory quality gas"], "recency": "", "complexity": "research", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.62}
130
+ {"id": "syn-2552", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.11538461538461539, "route": true}, "student": {"standalone_question": "Does wearing dark clothes change how mosquitoes find a person?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["mosquitoes detect carbon dioxide body heat skin odors dark clothing", "mosquitoes avoid dark clothing human scent study", "mosquitoes human odor detection clothing color"], "recency": "", "complexity": "standard", "lenses": ["web", "scholar"], "facets": [], "whole_library": false}, "seconds": 0.595}
131
+ {"id": "syn-2566", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "lenses": true, "query_overlap": 0.45454545454545453, "route": true}, "student": {"standalone_question": "Is store-bought Italian dressing healthier than a simple vinaigrette for someone watching their sugar?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["store-bought Italian dressing sugar content per serving", "store-bought Italian dressing vs vinaigrette sugar calories", "Italian dressing sugar vs olive oil vinaigrette health"], "recency": "year", "complexity": "research", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.64}
132
+ {"id": "syn-2570", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.6, "route": false}, "student": {"standalone_question": "What is Ontario's minimum wage for students under 18 who work part-time?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Ontario minimum wage students under 18 part-time", "Ontario minimum wage 2027 students"], "recency": "year", "complexity": "simple", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.585}
133
+ {"id": "syn-2600", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.3, "route": true}, "student": {"standalone_question": "How do adult learner dropout rates compare to piano?", "answer_shape": "comparison", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["adult cello vs piano dropout rates study", "adult instrument learning dropout rates longitudinal study", "adult cello learning dropout survey"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.573}
134
+ {"id": "syn-2654", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.26666666666666666, "route": true}, "student": {"standalone_question": "Wie viele Menschen in Deutschland besitzen eine Patientenverfügung?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Patientenverfügung Deutschland Befragung Anzahl Befragter", "Patientenverfügung Deutschland Studie Befragung Ärzte Patienten", "Patientenverfügung Deutschland Erhebung 2025"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.679}
135
+ {"id": "syn-2706", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": false, "image": true, "lenses": false, "query_overlap": 0.23529411764705882, "route": false}, "student": {"standalone_question": "Wie viel Schlaf braucht ein Teenager pro Nacht?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["teenager sleep duration per night recommendation", "adolescent sleep duration guidelines"], "recency": "year", "complexity": "simple", "lenses": ["web", "reference"], "facets": [], "whole_library": false}, "seconds": 0.515}
136
+ {"id": "syn-2723", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.2631578947368421, "route": true}, "student": {"standalone_question": "Wie sieht WSL2 aus, und gleicht es den Unterschied zwischen Linux und Windows für Entwickler aus?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["WSL2 Linux vs Windows Developer", "WSL2 native Linux vs Windows Subsystem for Linux differences"], "recency": "year", "complexity": "standard", "lenses": ["web", "docs"], "facets": [], "whole_library": false}, "seconds": 0.577}
137
+ {"id": "syn-2790", "score": {"json": true, "search": true}, "student": {"standalone_question": "What do the ingredient list words on a protein bar label mean, and how should I read them?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.466}
138
+ {"id": "syn-2801", "score": {"json": true, "search": true}, "student": {"standalone_question": "What does the early termination clause in the tenant's lease scan say, in plain English?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.457}
139
+ {"id": "syn-2815", "score": {"json": true, "search": true, "complexity": true, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.15384615384615385, "route": false}, "student": {"standalone_question": "Welche Ware ist auf dem Foto vom Flohmarkt zu sehen, und wie viel kostet sie ungefähr?", "answer_shape": "figure", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Flohmarkt Ware Foto Beschreibung Wert", "Flohmarkt Ware Wert Preis"], "recency": "year", "complexity": "standard", "lenses": ["web", "data"], "facets": [], "whole_library": false}, "seconds": 0.622}
140
+ {"id": "syn-2842", "score": {"json": true, "search": true}, "student": {"standalone_question": "Was bedeuten die Abkürzungen hinter den Artikeln auf einem Baumarkt-Kassenbon?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": false, "search_queries": [], "recency": "", "complexity": "simple", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.496}
141
+ {"id": "syn-2877", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": true, "query_overlap": 0.29411764705882354, "route": true}, "student": {"standalone_question": "Is the Elm Street Park playground funding vote in March 2026 still on the agenda, or has it been cancelled or postponed?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Elm Street Park playground funding vote March 2026", "Elm Street Park playground funding vote cancelled postponed", "Elm Street Park playground funding vote status 2026"], "recency": "year", "complexity": "standard", "lenses": ["news", "web"], "facets": [], "whole_library": false}, "seconds": 0.718}
142
+ {"id": "syn-2910", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": false, "image": true, "query_overlap": 0.3181818181818182, "route": false}, "student": {"standalone_question": "How should I position the longitudinal accessibility survey (Adjei & Solberg 2022) against earlier work in a related-work section on accessibility in games?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["longitudinal survey accessibility AAA titles Adjei Solberg 2022", "accessibility adoption AAA games longitudinal survey", "accessibility survey earlier work AAA titles"], "recency": "year", "complexity": "research", "lenses": [], "facets": [], "whole_library": true}, "seconds": 0.893}
143
+ {"id": "syn-2913", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.8, "route": false}, "student": {"standalone_question": "How does the Okonkwo & Tanaka 2016 paper handle false positives in the toxicity classifier?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Okonkwo Tanaka 2016 toxicity detection false positives", "multiplayer voice chat acoustic features false positive handling"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.792}
144
+ {"id": "syn-2914", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.5833333333333334, "route": true}, "student": {"standalone_question": "Is there anything in the library about serious games for physical therapy that reports adherence rates above 60 percent?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["serious games physical therapy adherence rate", "serious games rehabilitation adherence percentage"], "recency": "", "complexity": "standard", "lenses": [], "facets": [], "whole_library": false}, "seconds": 0.532}
145
+ {"id": "syn-2918", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.4444444444444444, "route": true}, "student": {"standalone_question": "Which papers in the set look at gender representation or character creation, and are there any that compare across countries?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["gender representation character creation content analysis", "character creation interface gender representation", "multinational comparison character creation games"], "recency": "", "complexity": "standard", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.865}
146
+ {"id": "syn-2945", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "query_overlap": 0.21951219512195122, "route": true}, "student": {"standalone_question": "Welche sind die wichtigsten Ergebnisse aus der gesamten Sammlung an Papers (Hartmann & Brandt 2016, Odendaal & Weber 2019, Raman & Koch 2018, Lindqvist & Feld 2021, Okafor & Hoffmann 2017, Brenner et al. 2020, Kessler 2022, Renner & Adler 2023)?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hartmann & Brandt 2016 mental accounting household budget allocation", "Odendaal & Weber 2019 present bias retirement savings young adults", "Raman & Koch 2018 financial literacy debt repayment randomized field experiment"], "recency": "", "complexity": "research", "lenses": [], "facets": [], "whole_library": true}, "seconds": 1.028}
147
+ {"id": "syn-2956", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.8, "route": true}, "student": {"standalone_question": "Welche Rolle spielen formative Rückmeldungen in der Studie von Hartmann & Greve (2019) und welche Effektstärken haben sie gefunden?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["formative feedback blended mathematics classrooms mixed methods", "formative feedback effect size mathematics learning", "Hartmann Greve 2019 feedback loops"], "recency": "", "complexity": "standard", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.698}
148
+ {"id": "syn-2957", "score": {"json": true, "search": true, "complexity": true, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.3333333333333333, "route": true}, "student": {"standalone_question": "Welche sind die Kernergebnisse der sechs Arbeiten in der Sammlung (Hartmann & Greve 2019, Yilmaz et al. 2021, Brandt 2017, Weiß et al. 2020, Sørensen & Fischer 2022, Krämer & Lindner 2023), zusammengefasst in Stichpunkten?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["Hartmann & Greve 2019 formative feedback loops blended mathematics classrooms mixed methods", "Yilmaz et al. 2021 teacher digital competence student engagement lower secondary", "Brandt 2017 learning analytics dashboards self-regulated learning vocational education", "Weiß et al. 2020 inclusive classroom practices peer interaction longitudinal German primary schools", "Sørensen & Fischer 2022 gamified vocabulary learning secondary foreign language retention motivation", "Krämer & Lindner 2023 teacher professional learning communities multilevel European countries"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": true}, "seconds": 1.24}
149
+ {"id": "syn-2962", "score": {"json": true, "search": true, "complexity": false, "answer_shape": false, "recency": true, "image": true, "lenses": true, "query_overlap": 0.6153846153846154, "route": false}, "student": {"standalone_question": "Funktioniert Gamification im Fremdsprachenunterricht überhaupt, und gibt es in der Sammlung etwas dazu, speziell zur Behalten von Vokabeln?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["gamification foreign language instruction retention motivation", "gamified vocabulary learning secondary foreign language"], "recency": "", "complexity": "research", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.856}
150
+ {"id": "syn-2969", "score": {"json": true, "search": true, "complexity": false, "answer_shape": true, "recency": true, "image": true, "lenses": false, "query_overlap": 0.391304347826087, "route": false}, "student": {"standalone_question": "Welche sind die Hauptergebnisse der Studie zur Lebensmittelsicherheit beim Fermentieren zu Hause?", "answer_shape": "", "place": "", "wants_image": false, "image_query": "", "needs_search": true, "search_queries": ["home food fermentation safety microbial succession household", "home food waste behaviour meal planning practices survey", "home food safety fermented vegetable brine microbial succession"], "recency": "", "complexity": "standard", "lenses": ["scholar", "web"], "facets": [], "whole_library": false}, "seconds": 0.662}
training/eval/dev-syn-2b-v1-s700.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/dump_prompt.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Write the app's classifier prompt and JSON schema to files, read-only.
2
+
3
+ Run with the simplixity venv so the prompt is exactly what the app sends:
4
+ ~/Documents/projects/simplixity/.venv/bin/python dump_prompt.py
5
+ """
6
+ import json, os, sys
7
+ REPO = os.path.expanduser("~/Documents/projects/simplixity")
8
+ sys.path.insert(0, REPO)
9
+ os.chdir(REPO) # settings read .env from here; nothing is written
10
+ from app.agent import depth as depth_module, prompts
11
+ from app.agent.orchestrator import Classification
12
+ from app.llm.client import _harden
13
+ from app.search import lenses as lens_module
14
+
15
+ OUT = os.path.dirname(os.path.abspath(__file__)) # eval/
16
+ # Auto depth and Auto sources: what the classifier sees by default.
17
+ depth = depth_module.resolve(depth_module.AUTO)
18
+ msgs = prompts.classify("<<QUERY>>", [], depth, lens_module.resolve(["auto"]), choose_lenses=True)
19
+ assert msgs[0]["role"] == "system" and msgs[-1]["content"] == "<<QUERY>>"
20
+ open(os.path.join(OUT, "system_prompt.txt"), "w").write(msgs[0]["content"])
21
+ json.dump(_harden(Classification.model_json_schema()), open(os.path.join(OUT, "schema.json"), "w"), indent=1)
22
+ print(depth.id, depth.queries, len(msgs[0]["content"]), "chars")
training/eval/eval.py ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Run the app's classifier prompt against an Ollama model and score it.
2
+
3
+ Mirrors app/agent/orchestrator._classify: same system prompt (dump_prompt.py),
4
+ same strict schema, temperature 0, thinking off unless --think. The one
5
+ post-processing step that changes routing is applied too: wanting a picture
6
+ forces a search.
7
+
8
+ python3 eval.py qwen3.5:2b [--think] [--tag name]
9
+ python3 eval.py openrouter/anthropic/claude-haiku-5.5 [--think] # key from ../.env
10
+ """
11
+ import argparse, json, os, statistics, sys, time, urllib.error, urllib.request
12
+
13
+ HERE = os.path.dirname(os.path.abspath(__file__))
14
+ SYSTEM = open(os.path.join(HERE, "system_prompt.txt")).read()
15
+ SCHEMA = json.load(open(os.path.join(HERE, "schema.json")))
16
+ HARD = {"day", "week", "month"}
17
+
18
+
19
+ def _key():
20
+ with open(os.path.join(HERE, "..", ".env")) as f:
21
+ for line in f:
22
+ if line.startswith("OPENROUTER_API_KEY="):
23
+ return line.split("=", 1)[1].strip()
24
+ raise SystemExit("OPENROUTER_API_KEY missing in .env")
25
+
26
+
27
+ def call_openrouter(model, item, think):
28
+ """The app's own request: strict json_schema, temperature 0, and the
29
+ deployment's default 1024-token reasoning allowance when thinking."""
30
+ messages = [{"role": "system", "content": SYSTEM}, *item["hist"], {"role": "user", "content": item["q"]}]
31
+ body = {"model": model, "messages": messages, "temperature": 0, "max_tokens": 500 + (1024 if think else 0),
32
+ "response_format": {"type": "json_schema",
33
+ "json_schema": {"name": "Classification", "strict": True, "schema": SCHEMA}},
34
+ "reasoning": {"max_tokens": 1024} if think else {"enabled": False}, "usage": {"include": True}}
35
+ req = urllib.request.Request("https://openrouter.ai/api/v1/chat/completions", json.dumps(body).encode(),
36
+ {"Content-Type": "application/json", "Authorization": f"Bearer {_key()}"})
37
+ t = time.time()
38
+ for attempt in range(4):
39
+ try:
40
+ with urllib.request.urlopen(req, timeout=120) as r:
41
+ out = json.load(r)
42
+ break
43
+ except urllib.error.HTTPError as e:
44
+ if e.code not in (429, 500, 502, 503) or attempt == 3:
45
+ raise SystemExit(f"OpenRouter {e.code}: {e.read()[:300]!r}")
46
+ time.sleep(2 ** attempt)
47
+ usage = out.get("usage") or {}
48
+ return ({"message": {"content": out["choices"][0]["message"].get("content") or ""},
49
+ "eval_count": usage.get("completion_tokens"), "prompt_eval_count": usage.get("prompt_tokens"),
50
+ "cost": usage.get("cost")}, time.time() - t)
51
+
52
+
53
+ def call(model, item, think, num_gpu=None):
54
+ if model.startswith("openrouter/"):
55
+ return call_openrouter(model.removeprefix("openrouter/"), item, think)
56
+ messages = [{"role": "system", "content": SYSTEM}, *item["hist"], {"role": "user", "content": item["q"]}]
57
+ body = {"model": model, "messages": messages, "stream": False, "format": SCHEMA, "think": think,
58
+ "options": {"temperature": 0, "num_ctx": 8192, "num_predict": 500 + (1024 if think else 0)},
59
+ "keep_alive": "10m"}
60
+ if num_gpu is not None:
61
+ # Ollama sometimes plans most layers onto the CPU with VRAM free.
62
+ body["options"]["num_gpu"] = num_gpu
63
+ req = urllib.request.Request("http://localhost:11434/api/chat", json.dumps(body).encode(),
64
+ {"Content-Type": "application/json"})
65
+ t = time.time()
66
+ with urllib.request.urlopen(req, timeout=300) as r:
67
+ out = json.load(r)
68
+ return out, time.time() - t
69
+
70
+
71
+ def score(item, pred):
72
+ s = {}
73
+ if pred is None:
74
+ return {"json": False}
75
+ s["json"] = True
76
+ search = bool(pred.get("needs_search", True)) or bool(pred.get("wants_image"))
77
+ s["search"] = search == item["search"]
78
+ if not item["search"]:
79
+ return s
80
+ if not search:
81
+ return s # a missed search: the fields below never get used
82
+ cx = str(pred.get("complexity", "") or "").strip().lower()
83
+ cx = cx if cx in {"simple", "standard", "research"} else "standard"
84
+ s["complexity"] = cx in item["cx"]
85
+ s["shape"] = str(pred.get("answer_shape", "") or "").strip().lower() in item["shape"]
86
+ if item["img"] is not None:
87
+ s["image"] = bool(pred.get("wants_image")) == item["img"]
88
+ rec = str(pred.get("recency", "") or "").strip().lower()
89
+ s["recency"] = {"none": rec == "", "any": True, "set": rec != "", "hard": rec in HARD,
90
+ "nothard": rec not in HARD}[item["rec"]]
91
+ lenses = {str(x).strip().lower() for x in pred.get("lenses") or []}
92
+ if item["lens_in"] or item["lens_out"]:
93
+ s["lenses"] = set(item["lens_in"]) <= lenses and not (set(item["lens_out"]) & lenses)
94
+ queries = [str(x) for x in pred.get("search_queries") or [] if str(x).strip()]
95
+ text = " ".join(queries).lower()
96
+ s["queries_n"] = 1 <= len(queries) <= 3
97
+ s["keys"] = all(any(k in text for k in group) for group in item["keys"])
98
+ if item["sq"]:
99
+ sq = str(pred.get("standalone_question", "")).lower()
100
+ s["standalone"] = all(any(k in sq for k in group) for group in item["sq"])
101
+ if item["place"]:
102
+ s["place"] = any(k in str(pred.get("place", "")).lower() for k in item["place"])
103
+ s["route"] = s["search"] and s["complexity"] and s["shape"]
104
+ return s
105
+
106
+
107
+ def main():
108
+ p = argparse.ArgumentParser()
109
+ p.add_argument("model")
110
+ p.add_argument("--think", action="store_true")
111
+ p.add_argument("--tag", default="")
112
+ p.add_argument("--num-gpu", type=int, default=None, help="layers on the GPU (99 = all)")
113
+ args = p.parse_args()
114
+ items = [json.loads(l) for l in open(os.path.join(HERE, "testset.jsonl"))]
115
+ tag = args.tag or (args.model.replace(":", "-").replace("/", "_") + ("-think" if args.think else ""))
116
+ out_path = os.path.join(HERE, f"results-{tag}.jsonl")
117
+ call(args.model, items[0], args.think, args.num_gpu) # load the model; not timed
118
+ with open(out_path, "w") as f:
119
+ for i, item in enumerate(items):
120
+ out, dt = call(args.model, item, args.think, args.num_gpu)
121
+ raw = out["message"]["content"]
122
+ try:
123
+ pred = json.loads(raw)
124
+ except json.JSONDecodeError:
125
+ pred = None
126
+ row = {"i": i, "q": item["q"], "lang": item["lang"], "tag": item["tag"], "pred": pred, "raw": raw,
127
+ "score": score(item, pred), "seconds": round(dt, 3), "out_tokens": out.get("eval_count"),
128
+ "prompt_tokens": out.get("prompt_eval_count"), "cost": out.get("cost")}
129
+ f.write(json.dumps(row, ensure_ascii=False) + "\n")
130
+ print(f"{i + 1}/{len(items)} {dt:.2f}s", file=sys.stderr, end="\r")
131
+ print(f"\n{out_path}", file=sys.stderr)
132
+
133
+
134
+ if __name__ == "__main__":
135
+ main()
training/eval/eval_dev.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Agreement with the teacher on the held-out dev sets (synthetic and real).
2
+
3
+ The dev items carry the teacher's (Haiku 5.5, thinking) label as their last
4
+ message. A student is run on the same messages through Ollama, schema-forced,
5
+ thinking off, and compared field by field. This is agreement, not accuracy:
6
+ the hand-labelled testset.jsonl is the accuracy check.
7
+
8
+ python3 eval_dev.py simplixity-router:2b-v1-q4_K_M [--num-gpu 99]
9
+ """
10
+
11
+ import argparse
12
+ import json
13
+ import os
14
+ import re
15
+ import statistics
16
+ import sys
17
+ import time
18
+ import urllib.error
19
+ import urllib.request
20
+
21
+ HERE = os.path.dirname(os.path.abspath(__file__))
22
+ ROOT = os.path.dirname(HERE)
23
+ SCHEMA = json.load(open(os.path.join(HERE, "schema.json")))
24
+ SETS = {"syn": os.path.join(ROOT, "data", "valid.jsonl"), "real": os.path.join(ROOT, "data", "private", "valid_real.jsonl")}
25
+
26
+
27
+ def run(model, messages, num_gpu):
28
+ body = {"model": model, "messages": messages, "stream": False, "format": SCHEMA, "think": False,
29
+ "options": {"temperature": 0, "num_ctx": 8192, "num_predict": 500}, "keep_alive": "10m"}
30
+ if num_gpu is not None:
31
+ body["options"]["num_gpu"] = num_gpu
32
+ req = urllib.request.Request("http://localhost:11434/api/chat", json.dumps(body).encode(),
33
+ {"Content-Type": "application/json"})
34
+ t = time.time()
35
+ with urllib.request.urlopen(req, timeout=300) as r:
36
+ out = json.load(r)
37
+ return out["message"]["content"], time.time() - t
38
+
39
+
40
+ def words(queries):
41
+ return set(re.findall(r"\w+", " ".join(queries).lower()))
42
+
43
+
44
+ def compare(student, teacher, settings):
45
+ s = {"json": student is not None}
46
+ if student is None:
47
+ return s
48
+ search = bool(student.get("needs_search")) or bool(student.get("wants_image"))
49
+ s["search"] = search == teacher["needs_search"]
50
+ if teacher["needs_search"] and search:
51
+ for field in ("complexity", "answer_shape", "recency"):
52
+ s[field] = str(student.get(field, "")).strip().lower() == teacher[field]
53
+ s["image"] = bool(student.get("wants_image")) == teacher["wants_image"]
54
+ if settings["auto_lenses"]:
55
+ s["lenses"] = set(student.get("lenses") or []) == set(teacher["lenses"])
56
+ a, b = words(student.get("search_queries") or []), words(teacher["search_queries"])
57
+ s["query_overlap"] = len(a & b) / len(a | b) if a | b else 1.0
58
+ s["route"] = s["search"] and s["complexity"] and s["answer_shape"]
59
+ return s
60
+
61
+
62
+ def main():
63
+ p = argparse.ArgumentParser()
64
+ p.add_argument("model")
65
+ p.add_argument("--num-gpu", type=int, default=None)
66
+ p.add_argument("--tag", default="")
67
+ args = p.parse_args()
68
+ tag = args.tag or args.model.replace(":", "-").replace("/", "_")
69
+ for name, path in SETS.items():
70
+ rows = [json.loads(line) for line in open(path)]
71
+ out_path = os.path.join(HERE if name == "syn" else os.path.join(ROOT, "data", "private"),
72
+ f"dev-{name}-{tag}.jsonl")
73
+ results = []
74
+ with open(out_path, "w") as f:
75
+ for i, row in enumerate(rows):
76
+ try:
77
+ raw, dt = run(args.model, row["messages"][:-1], args.num_gpu)
78
+ except urllib.error.HTTPError as e: # e.g. "token repeat limit reached": a failed call
79
+ raw, dt = "", 0.0
80
+ print(f"\n{row['id']}: {e.read().decode()[:200]}", file=sys.stderr)
81
+ try:
82
+ student = json.loads(raw)
83
+ except json.JSONDecodeError:
84
+ student = None
85
+ teacher = json.loads(row["messages"][-1]["content"])
86
+ r = {"id": row["id"], "score": compare(student, teacher, row["settings"]), "student": student,
87
+ "seconds": round(dt, 3)}
88
+ results.append(r)
89
+ f.write(json.dumps(r, ensure_ascii=False) + "\n")
90
+ print(f"{name} {i + 1}/{len(rows)}", file=sys.stderr, end="\r")
91
+ if name == "real":
92
+ os.chmod(out_path, 0o600)
93
+ print(f"\n== {name} dev (n={len(rows)}, median {statistics.median(r['seconds'] for r in results):.2f}s)")
94
+ for field in ("json", "search", "route", "complexity", "answer_shape", "recency", "image", "lenses",
95
+ "query_overlap"):
96
+ vals = [r["score"][field] for r in results if field in r["score"]]
97
+ if vals:
98
+ print(f" {field:14} {100 * sum(vals) / len(vals):5.1f}% (n={len(vals)})")
99
+
100
+
101
+ if __name__ == "__main__":
102
+ main()
training/eval/make_testset.py ADDED
@@ -0,0 +1,311 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Hand-labelled routing test set: about 50 English and 50 German questions.
2
+
3
+ Labels follow the app's CLASSIFY prompt. Where the prompt leaves room, a field
4
+ lists every answer it allows, so a model is only marked wrong for a reading
5
+ the prompt rules out.
6
+
7
+ Per item:
8
+ q the question
9
+ hist prior turns (user, assistant, ...), optional
10
+ search needs_search
11
+ cx allowed complexity values
12
+ shape allowed answer_shape values ("" is prose)
13
+ img wants_image, or None when either is fine
14
+ rec "none" (must be empty), "any" (anything), "set" (must be set),
15
+ "hard" (day/week/month), "nothard" (empty or year)
16
+ lens_in corpora that must be among `lenses` (searched questions only)
17
+ lens_out corpora that must not be
18
+ keys groups of terms; every group needs one hit in the queries
19
+ (lowercase substring), searched questions only
20
+ sq groups of terms the standalone question must contain (follow-ups)
21
+ place accepted `place` substrings for weather (any one)
22
+
23
+ python make_testset.py # writes testset.jsonl
24
+ """
25
+
26
+ import json
27
+ import os
28
+
29
+ S, ST, R = "simple", "standard", "research"
30
+ ANYCX = [S, ST, R]
31
+
32
+
33
+ def q(text, lang, search=True, cx=ANYCX, shape=("",), img=False, rec="any",
34
+ lens_in=(), lens_out=(), keys=(), sq=(), place="", hist=(), tag=""):
35
+ return {"q": text, "lang": lang, "search": search, "cx": list(cx), "shape": list(shape),
36
+ "img": img, "rec": rec, "lens_in": list(lens_in), "lens_out": list(lens_out),
37
+ "keys": [list(k) for k in keys], "sq": [list(k) for k in sq], "place": list(place),
38
+ "hist": list(hist), "tag": tag}
39
+
40
+
41
+
42
+ NOSHAPE_OR_EXPL = ("", "explainer")
43
+ EN, DE = "en", "de"
44
+
45
+ ITEMS = [
46
+ # ------------------------------------------------------------ no search
47
+ q("thanks, that helped a lot!", EN, search=False, img=False, tag="smalltalk"),
48
+ q("hi there", EN, search=False, tag="smalltalk"),
49
+ q("Translate into German: 'The meeting has been moved to Thursday afternoon.'", EN,
50
+ search=False, tag="transform"),
51
+ q("Rewrite this to sound more polite: send me the report by tomorrow or else.", EN,
52
+ search=False, tag="transform"),
53
+ q("What is 17% of 2,340?", EN, search=False, tag="arithmetic"),
54
+ q("Why does this Python print 3 instead of 4?\n\nx = [1, 2, 3]\nx.append(4)\nprint(len(x[:-1]))",
55
+ EN, search=False, tag="code"),
56
+ q("Summarise this in one sentence: Our Q3 revenue rose 12% on stronger subscription sales, "
57
+ "while hardware margins fell because of higher component costs.", EN, search=False, tag="transform"),
58
+ q("danke dir, super erklärt", DE, search=False, tag="smalltalk"),
59
+ q("Guten Morgen!", DE, search=False, tag="smalltalk"),
60
+ q("Übersetze ins Englische: Ich komme heute etwas später, der Zug hat Verspätung.", DE,
61
+ search=False, tag="transform"),
62
+ q("Formuliere das freundlicher: Ihre Rechnung ist seit drei Wochen überfällig, zahlen Sie endlich.",
63
+ DE, search=False, tag="transform"),
64
+ q("Was ist 15 % von 480 Euro?", DE, search=False, tag="arithmetic"),
65
+ q("Was macht diese SQL-Abfrage? SELECT name, COUNT(*) FROM orders GROUP BY name HAVING COUNT(*) > 5",
66
+ DE, search=False, tag="code"),
67
+ q("Kannst du deine letzte Antwort kürzer zusammenfassen?", DE, search=False, tag="conversation",
68
+ hist=[{"role": "user", "content": "Wie funktioniert eine Wärmepumpe?"},
69
+ {"role": "assistant", "content": "Eine Wärmepumpe entzieht der Umgebung Wärme und hebt sie "
70
+ "mit einem Verdichter auf ein höheres Temperaturniveau, um damit zu heizen."}]),
71
+
72
+ # ------------------------------------------------------------ lookups
73
+ q("What is the Bitcoin price right now?", EN, cx=[S], shape=["figure"], rec="set",
74
+ lens_out=["scholar"], keys=[["bitcoin", "btc"]], tag="figure"),
75
+ q("How many people live in Tokyo?", EN, cx=[S], shape=["figure", ""], rec="nothard",
76
+ lens_out=["scholar"], keys=[["tokyo"]], tag="figure"),
77
+ q("When did the Berlin Wall fall?", EN, cx=[S], shape=[""], rec="none",
78
+ lens_out=["scholar", "news"], keys=[["berlin wall"]], tag="fact"),
79
+ q("What's the latest stable version of Python?", EN, cx=[S], shape=["", "figure"], rec="set",
80
+ lens_out=["scholar"], keys=[["python"]], tag="version"),
81
+ q("weather in Hamburg tomorrow", EN, cx=[S], shape=["weather"], place=("hamburg",),
82
+ lens_out=["scholar"], keys=[["hamburg"]], tag="weather"),
83
+ q("Will it rain in Seattle this weekend?", EN, cx=[S], shape=["weather"], place=("seattle",),
84
+ lens_out=["scholar"], keys=[["seattle"]], tag="weather"),
85
+ q("Who won the last Champions League final?", EN, cx=[S], shape=["match", ""], rec="nothard",
86
+ lens_out=["scholar"], keys=[["champions league"]], tag="match"),
87
+ q("Wie viel kostet ein Liter Diesel aktuell in Deutschland?", DE, cx=[S], shape=["figure"], rec="set",
88
+ lens_out=["scholar"], keys=[["diesel"]], tag="figure"),
89
+ q("Wie hoch ist der Leitzins der EZB?", DE, cx=[S], shape=["figure"], rec="set",
90
+ lens_out=["scholar"], keys=[["leitzins", "ezb", "ecb", "interest rate"]], tag="figure"),
91
+ q("Wann wurde Goethe geboren?", DE, cx=[S], shape=[""], rec="none",
92
+ lens_out=["scholar", "news"], keys=[["goethe"]], tag="fact"),
93
+ q("Wetter in München morgen", DE, cx=[S], shape=["weather"], place=("münchen","munich"),
94
+ lens_out=["scholar"], keys=[["münchen", "munich"]], tag="weather"),
95
+ q("Wie wird das Wetter am Wochenende in Köln?", DE, cx=[S], shape=["weather"], place=("köln","cologne"),
96
+ lens_out=["scholar"], keys=[["köln", "cologne"]], tag="weather"),
97
+ q("Wie hat Bayern gegen Dortmund gespielt?", DE, cx=[S], shape=["match"], rec="set",
98
+ lens_out=["scholar"], keys=[["bayern"], ["dortmund", "bvb"]], tag="match"),
99
+ q("Welche Version hat Node.js LTS gerade?", DE, cx=[S], shape=["", "figure"], rec="set",
100
+ lens_out=["scholar"], keys=[["node"]], tag="version"),
101
+
102
+ # ------------------------------------------------------------ profile / image
103
+ q("Who is Jensen Huang?", EN, cx=[S, ST], shape=["profile"], img=True, rec="nothard",
104
+ lens_out=["scholar"], keys=[["jensen huang"]], tag="profile"),
105
+ q("What is the company Anduril?", EN, cx=[S, ST], shape=["profile"], img=None, rec="nothard",
106
+ lens_out=["scholar"], keys=[["anduril"]], tag="profile"),
107
+ q("what is Obsidian, the notes app?", EN, cx=[S, ST], shape=["profile"], img=None, rec="nothard",
108
+ lens_out=["scholar"], keys=[["obsidian"]], tag="profile"),
109
+ q("What does an okapi look like?", EN, cx=[S], shape=["image", "profile"], img=True, rec="none",
110
+ lens_out=["news"], keys=[["okapi"]], tag="image"),
111
+ q("show me the flag of Bhutan", EN, cx=[S], shape=["image"], img=True, rec="none",
112
+ lens_out=["scholar", "news"], keys=[["bhutan"]], tag="image"),
113
+ q("Wer ist Annalena Baerbock?", DE, cx=[S, ST], shape=["profile"], img=True, rec="nothard",
114
+ lens_out=["scholar"], keys=[["baerbock"]], tag="profile"),
115
+ q("Was ist die Firma Celonis?", DE, cx=[S, ST], shape=["profile"], img=None, rec="nothard",
116
+ lens_out=["scholar"], keys=[["celonis"]], tag="profile"),
117
+ q("Wer war Hedy Lamarr?", DE, cx=[S, ST], shape=["profile"], img=True, rec="none",
118
+ lens_out=["scholar", "news"], keys=[["lamarr"]], tag="profile"),
119
+ q("Wie sieht ein Axolotl aus?", DE, cx=[S], shape=["image", "profile"], img=True, rec="none",
120
+ lens_out=["news"], keys=[["axolotl"]], tag="image"),
121
+ q("Zeig mir ein Bild vom Kölner Dom", DE, cx=[S], shape=["image"], img=True, rec="none",
122
+ lens_out=["scholar", "news"], keys=[["dom", "cathedral"]], tag="image"),
123
+ q("Wie sieht es mit der Rente mit 63 aus, gibt es die noch?", DE, cx=[S, ST], shape=[""], img=False,
124
+ rec="set", lens_out=["scholar"], keys=[["rente"]], tag="idiom"),
125
+
126
+ # ------------------------------------------------------------ explainer
127
+ q("How do transformers work in machine learning?", EN, cx=[ST, S], shape=["explainer"],
128
+ rec="none", keys=[["transformer"]], tag="explainer"),
129
+ q("Why is the sky blue?", EN, cx=[S, ST], shape=["explainer"], rec="none",
130
+ lens_out=["news"], keys=[["sky", "rayleigh"]], tag="explainer"),
131
+ q("Explain how public key cryptography works", EN, cx=[S, ST], shape=["explainer"], rec="none",
132
+ lens_out=["news"], keys=[["public key", "public-key", "asymmetric"]], tag="explainer"),
133
+ q("What does impedance mean in electronics?", EN, cx=[S, ST], shape=["explainer"], rec="none",
134
+ lens_out=["news"], keys=[["impedance"]], tag="explainer"),
135
+ q("Wie funktionieren Autoencoder?", DE, cx=[S, ST], shape=["explainer"], rec="none",
136
+ lens_out=["news"], keys=[["autoencoder"]], tag="explainer"),
137
+ q("Warum wird Brot hart, wenn es alt wird?", DE, cx=[S, ST], shape=["explainer"], rec="none",
138
+ lens_out=["news"], keys=[["brot", "bread"]], tag="explainer"),
139
+ q("Was bedeutet Inflation eigentlich genau?", DE, cx=[S, ST], shape=["explainer"], rec="none",
140
+ lens_out=["news"], keys=[["inflation"]], tag="explainer"),
141
+ q("Wie funktioniert eine Wärmepumpe?", DE, cx=[S, ST], shape=["explainer", "diagram"], rec="none",
142
+ lens_out=["news"], keys=[["wärmepumpe", "heat pump"]], tag="explainer"),
143
+
144
+ # ------------------------------------------------------------ steps
145
+ q("How do I install Docker on Ubuntu 24.04?", EN, cx=[S, ST], shape=["steps"], rec="nothard",
146
+ lens_out=["scholar", "news"], keys=[["docker"], ["ubuntu"]], tag="steps"),
147
+ q("How do I renew a US passport by mail?", EN, cx=[S, ST], shape=["steps"], rec="nothard",
148
+ lens_out=["scholar"], keys=[["passport"]], tag="steps"),
149
+ q("How do I set up SSH keys for GitHub?", EN, cx=[S, ST], shape=["steps"], rec="nothard",
150
+ lens_out=["scholar", "news"], keys=[["ssh"], ["github"]], tag="steps"),
151
+ q("Wie beantrage ich einen neuen Personalausweis?", DE, cx=[S, ST], shape=["steps"], rec="nothard",
152
+ lens_out=["scholar"], keys=[["personalausweis"]], tag="steps"),
153
+ q("Wie melde ich ein Gewerbe an?", DE, cx=[S, ST], shape=["steps"], rec="nothard",
154
+ lens_out=["scholar"], keys=[["gewerbe"]], tag="steps"),
155
+ q("Wie richte ich einen Cronjob unter Linux ein?", DE, cx=[S, ST], shape=["steps"], rec="nothard",
156
+ lens_out=["scholar", "news"], keys=[["cron"]], tag="steps"),
157
+
158
+ # ------------------------------------------------------------ compare -> research
159
+ q("PostgreSQL vs MySQL for a small web app?", EN, cx=[R], shape=["comparison"], rec="nothard",
160
+ lens_out=["scholar", "news"], keys=[["postgres"], ["mysql"]], tag="comparison"),
161
+ q("Should I buy an iPhone 17 or a Pixel 10?", EN, cx=[R], shape=["comparison"], img=None, rec="nothard",
162
+ lens_out=["scholar"], keys=[["iphone"], ["pixel"]], tag="comparison"),
163
+ q("Compare Rust, Go and Zig for systems programming", EN, cx=[R], shape=["table"], rec="nothard",
164
+ lens_out=["news"], keys=[["rust"], ["go", "golang"], ["zig"]], tag="table"),
165
+ q("ETF oder Festgeld, was ist 2026 sinnvoller?", DE, cx=[R], shape=["comparison"], rec="set",
166
+ lens_out=["scholar"], keys=[["etf"], ["festgeld"]], tag="comparison"),
167
+ q("Vergleiche Tesla Model 3, VW ID.3 und BYD Dolphin", DE, cx=[R], shape=["table"], img=None,
168
+ rec="nothard", lens_out=["scholar"], keys=[["model 3", "tesla"], ["id.3", "id3"], ["dolphin", "byd"]],
169
+ tag="table"),
170
+ q("Was ist besser für Einsteiger, Fuji oder Sony Kameras?", DE, cx=[R], shape=["comparison"], img=None,
171
+ rec="nothard", lens_out=["scholar"], keys=[["fuji"], ["sony"]], tag="comparison"),
172
+
173
+ # ------------------------------------------------------------ evidence -> scholar
174
+ q("Does creatine actually help with muscle growth?", EN, cx=[ST, R], shape=["", "explainer"],
175
+ rec="nothard", lens_in=["scholar"], keys=[["creatine"]], tag="evidence"),
176
+ q("What does the research say about intermittent fasting and weight loss?", EN, cx=[R], shape=[""],
177
+ rec="nothard", lens_in=["scholar"], keys=[["fasting"]], tag="evidence"),
178
+ q("Is aspartame harmful?", EN, cx=[ST, R], shape=[""], rec="nothard", lens_in=["scholar"],
179
+ keys=[["aspartame"]], tag="evidence"),
180
+ q("how much vitamin D should an adult take per day", EN, cx=[S, ST], shape=["", "figure"],
181
+ rec="nothard", lens_in=["scholar"], keys=[["vitamin d"]], tag="evidence"),
182
+ q("Bringt Kreatin was?", DE, cx=[ST, R], shape=["", "explainer"], rec="nothard", lens_in=["scholar"],
183
+ keys=[["creatine", "kreatin"]], tag="evidence"),
184
+ q("Was sagt die Studienlage zu Melatonin bei Schlafstörungen?", DE, cx=[R], shape=[""], rec="nothard",
185
+ lens_in=["scholar"], keys=[["melatonin"]], tag="evidence"),
186
+ q("Ist Süßstoff schädlich?", DE, cx=[ST, R], shape=[""], rec="nothard", lens_in=["scholar"],
187
+ keys=[["sweetener", "süßstoff", "aspartame", "sucralose"]], tag="evidence"),
188
+ q("Warum wirkt Koffein bei mir nicht mehr?", DE, cx=[S, ST], shape=["explainer", ""], rec="nothard",
189
+ lens_in=["scholar"], keys=[["caffeine", "koffein"]], tag="evidence"),
190
+
191
+ # ------------------------------------------------------------ docs / API
192
+ q("How do I paginate results with the GitHub REST API?", EN, cx=[S, ST], shape=["steps", ""],
193
+ rec="nothard", lens_in=["docs"], lens_out=["scholar", "news"], keys=[["github"], ["paginat"]],
194
+ tag="docs"),
195
+ q("What does useEffect's cleanup function do in React?", EN, cx=[S, ST], shape=["explainer", ""],
196
+ rec="nothard", lens_in=["docs"], lens_out=["scholar", "news"], keys=[["useeffect"]], tag="docs"),
197
+ q("Wie setze ich in FastAPI einen Timeout für einen Request?", DE, cx=[S, ST], shape=["steps", ""],
198
+ rec="nothard", lens_in=["docs"], lens_out=["scholar", "news"], keys=[["fastapi"], ["timeout"]],
199
+ tag="docs"),
200
+ q("Welche Parameter hat pandas read_csv für das Trennzeichen?", DE, cx=[S], shape=[""],
201
+ rec="nothard", lens_in=["docs"], lens_out=["scholar", "news"], keys=[["read_csv"]], tag="docs"),
202
+
203
+ # ------------------------------------------------------------ news / recent
204
+ q("What happened at the UN climate conference this week?", EN, cx=[ST, R], shape=[""], rec="hard",
205
+ lens_in=["news"], lens_out=["scholar"], keys=[["climate", "cop"]], tag="news"),
206
+ q("latest news on the OpenAI board", EN, cx=[ST, S], shape=[""], rec="hard", lens_in=["news"],
207
+ lens_out=["scholar"], keys=[["openai"]], tag="news"),
208
+ q("Was ist heute an der Börse passiert?", DE, cx=[S, ST], shape=[""], rec="hard", lens_in=["news"],
209
+ lens_out=["scholar"], keys=[["börse", "dax", "stock"]], tag="news"),
210
+ q("Was gibt es Neues zum Bürgergeld?", DE, cx=[S, ST], shape=[""], rec="set", lens_in=["news"],
211
+ lens_out=["scholar"], keys=[["bürgergeld"]], tag="news"),
212
+
213
+ # ------------------------------------------------------------ prose with structure
214
+ q("How did the 2008 financial crisis come about?", EN, cx=[ST, R], shape=[""], rec="none",
215
+ lens_out=["news"], keys=[["2008", "financial crisis"]], tag="history"),
216
+ q("Is a heat pump worth it in an old house?", EN, cx=[ST, R], shape=[""], rec="nothard",
217
+ lens_out=["news"], keys=[["heat pump"]], tag="decision"),
218
+ q("What are the side effects of ibuprofen?", EN, cx=[S, ST], shape=[""], rec="none",
219
+ lens_out=["news"], keys=[["ibuprofen"]], tag="list"),
220
+ q("Wie kam es zum Ukrainekrieg?", DE, cx=[ST, R], shape=[""], rec="nothard", lens_out=["scholar"],
221
+ keys=[["ukraine"]], tag="history"),
222
+ q("Lohnt sich eine Balkonsolaranlage?", DE, cx=[ST, R], shape=[""], rec="nothard",
223
+ keys=[["balkon", "balcony", "solar"]], tag="decision"),
224
+ q("Welche Unterlagen brauche ich für einen Bauantrag?", DE, cx=[S, ST], shape=["", "steps"],
225
+ rec="nothard", lens_out=["scholar", "news"], keys=[["bauantrag"]], tag="list"),
226
+
227
+ # ------------------------------------------------------------ series
228
+ q("How has the price of gold changed over the last ten years?", EN, cx=[S, ST], shape=["series"],
229
+ rec="nothard", lens_out=["scholar"], keys=[["gold"]], tag="series"),
230
+ q("Wie hat sich die Einwohnerzahl Berlins seit 1990 entwickelt?", DE, cx=[S, ST], shape=["series"],
231
+ rec="none", lens_out=["scholar", "news"], keys=[["berlin"]], tag="series"),
232
+
233
+ # ------------------------------------------------------------ research, multi-part
234
+ q("What is the current state of research on solid-state batteries?", EN, cx=[R], shape=[""],
235
+ rec="nothard", lens_in=["scholar"], keys=[["solid-state", "solid state"]], tag="survey"),
236
+ q("Welche Förderungen gibt es 2026 für Wärmepumpen, wie hoch sind sie und wie beantrage ich sie?", DE,
237
+ cx=[R], shape=["", "steps"], rec="set", lens_out=["scholar"], keys=[["wärmepumpe", "heat pump"],
238
+ ["förder"]], tag="multipart"),
239
+
240
+ # ------------------------------------------------------------ follow-ups
241
+ q("and how much does it cost?", EN, cx=[S, ST], shape=["figure", ""], rec="any", lens_out=["scholar"],
242
+ keys=[["steam deck"]], sq=[["steam deck"]], tag="followup",
243
+ hist=[{"role": "user", "content": "What is the Steam Deck OLED?"},
244
+ {"role": "assistant", "content": "The Steam Deck OLED is Valve's handheld gaming PC with an "
245
+ "OLED screen, released in November 2023."}]),
246
+ q("who founded it?", EN, cx=[S], shape=["", "profile"], img=None, rec="nothard", lens_out=["scholar"],
247
+ keys=[["mistral"]], sq=[["mistral"]], tag="followup",
248
+ hist=[{"role": "user", "content": "What is Mistral AI?"},
249
+ {"role": "assistant", "content": "Mistral AI is a French company that builds large language "
250
+ "models, based in Paris."}]),
251
+ q("is it safe during pregnancy?", EN, cx=[ST, R], shape=[""], rec="nothard", lens_in=["scholar"],
252
+ keys=[["ibuprofen"]], sq=[["ibuprofen"]], tag="followup",
253
+ hist=[{"role": "user", "content": "What is ibuprofen used for?"},
254
+ {"role": "assistant", "content": "Ibuprofen is a nonsteroidal anti-inflammatory drug used for "
255
+ "pain, fever and inflammation."}]),
256
+ q("und wie viel kostet das?", DE, cx=[S, ST], shape=["figure", ""], rec="any", lens_out=["scholar"],
257
+ keys=[["deutschlandticket", "deutschland-ticket"]], sq=[["deutschlandticket", "deutschland-ticket"]],
258
+ tag="followup",
259
+ hist=[{"role": "user", "content": "Was ist das Deutschlandticket?"},
260
+ {"role": "assistant", "content": "Das Deutschlandticket ist ein bundesweit gültiges Monatsabo "
261
+ "für den öffentlichen Nahverkehr."}]),
262
+ q("Und wo liegt die?", DE, cx=[S], shape=["", "profile", "image"], img=None, rec="none",
263
+ lens_out=["scholar", "news"], keys=[["zugspitze"]], sq=[["zugspitze"]], tag="followup",
264
+ hist=[{"role": "user", "content": "Wie hoch ist die Zugspitze?"},
265
+ {"role": "assistant", "content": "Die Zugspitze ist 2962 Meter hoch und damit Deutschlands "
266
+ "höchster Berg."}]),
267
+ q("Gibt es dazu Studien?", DE, cx=[ST, R], shape=[""], rec="nothard", lens_in=["scholar"],
268
+ keys=[["ashwagandha"]], sq=[["ashwagandha"]], tag="followup",
269
+ hist=[{"role": "user", "content": "Was ist Ashwagandha?"},
270
+ {"role": "assistant", "content": "Ashwagandha ist eine Pflanze aus der ayurvedischen Medizin, "
271
+ "die als Nahrungsergänzungsmittel gegen Stress verkauft wird."}]),
272
+
273
+ # ------------------------------------------------------------ misc everyday
274
+ q("best budget mechanical keyboard", EN, cx=[ST, R], shape=["", "table"], img=None, rec="nothard",
275
+ lens_out=["scholar"], keys=[["keyboard"]], tag="recommend"),
276
+ q("Can I bring a power bank on a plane?", EN, cx=[S, ST], shape=[""], rec="nothard",
277
+ lens_out=["scholar"], keys=[["power bank"]], tag="rule"),
278
+ q("What time zone is Brazil in?", EN, cx=[S], shape=[""], rec="none", lens_out=["scholar", "news"],
279
+ keys=[["brazil"]], tag="fact"),
280
+ q("recipe for a classic carbonara", EN, cx=[S, ST], shape=["steps", ""], rec="none",
281
+ lens_out=["scholar", "news"], keys=[["carbonara"]], tag="recipe"),
282
+ q("How long is a US driver's license valid in Germany?", EN, cx=[S, ST], shape=[""], rec="nothard",
283
+ lens_out=["scholar"], keys=[["license", "licence"]], tag="rule"),
284
+ q("Wie lange muss ich Steuerunterlagen aufbewahren?", DE, cx=[S], shape=[""], rec="nothard",
285
+ lens_out=["scholar", "news"], keys=[["aufbewahr", "steuer"]], tag="rule"),
286
+ q("Rezept für Schweinebraten mit Kruste", DE, cx=[S, ST], shape=["steps", ""], rec="none",
287
+ lens_out=["scholar", "news"], keys=[["schweinebraten"]], tag="recipe"),
288
+ q("Darf mein Vermieter die Kaution einbehalten, wenn ich Löcher in der Wand habe?", DE, cx=[S, ST],
289
+ shape=[""], rec="nothard", lens_out=["scholar"], keys=[["kaution"]], tag="rule"),
290
+ q("Welche Zeitzone hat Japan?", DE, cx=[S], shape=[""], rec="none", lens_out=["scholar", "news"],
291
+ keys=[["japan"]], tag="fact"),
292
+ q("Was kann man in Lissabon an einem Wochenende unternehmen?", DE, cx=[S, ST], shape=[""], img=None,
293
+ rec="nothard", lens_out=["scholar"], keys=[["lissabon", "lisbon"]], tag="recommend"),
294
+ q("Was ist der Unterschied zwischen GmbH und UG?", DE, cx=[ST, R], shape=["comparison"],
295
+ rec="nothard", lens_out=["scholar", "news"], keys=[["gmbh"], ["ug"]], tag="comparison"),
296
+ q("Wie viele Kalorien hat eine Banane?", DE, cx=[S], shape=["figure"], rec="none",
297
+ lens_out=["news"], keys=[["banane", "banana"]], tag="figure"),
298
+ q("What's the difference between a virus and a bacterium?", EN, cx=[ST, S, R],
299
+ shape=["comparison", "explainer"], rec="none", lens_out=["news"], keys=[["virus"], ["bacteri"]],
300
+ tag="comparison"),
301
+ q("Which documents do I need to open a bank account in the UK as a foreigner?", EN, cx=[S, ST],
302
+ shape=["", "steps"], rec="nothard", lens_out=["scholar", "news"], keys=[["bank account"]], tag="list"),
303
+ ]
304
+
305
+ if __name__ == "__main__":
306
+ out = os.path.join(os.path.dirname(os.path.abspath(__file__)), "testset.jsonl")
307
+ with open(out, "w") as f:
308
+ for item in ITEMS:
309
+ f.write(json.dumps(item, ensure_ascii=False) + "\n")
310
+ langs = [i["lang"] for i in ITEMS]
311
+ print(f"{len(ITEMS)} items: {langs.count('en')} en, {langs.count('de')} de -> {out}")
training/eval/report.py ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Summarise results-*.jsonl: share of items passing each check, overall and by language."""
2
+ import json, statistics, sys
3
+
4
+ FIELDS = ["json", "search", "route", "complexity", "shape", "image", "recency", "lenses", "keys",
5
+ "queries_n", "standalone", "place"]
6
+
7
+
8
+ def pct(rows, field, lang=None):
9
+ vals = [r["score"][field] for r in rows if field in r["score"] and (lang is None or r["lang"] == lang)]
10
+ return (f"{100 * sum(vals) / len(vals):3.0f}%", len(vals)) if vals else (" - ", 0)
11
+
12
+
13
+ for path in sys.argv[1:]:
14
+ rows = [json.loads(l) for l in open(path)]
15
+ print(f"\n== {path} (n={len(rows)}, median {statistics.median(r['seconds'] for r in rows):.2f}s, "
16
+ f"p90 {sorted(r['seconds'] for r in rows)[int(0.9 * len(rows))]:.2f}s, "
17
+ f"median out {statistics.median(r['out_tokens'] or 0 for r in rows):.0f} tok)")
18
+ print(f"{'check':12} {'all':>5} {'en':>5} {'de':>5} n")
19
+ for field in FIELDS:
20
+ a, n = pct(rows, field)
21
+ print(f"{field:12} {a:>5} {pct(rows, field, 'en')[0]:>5} {pct(rows, field, 'de')[0]:>5} {n:3}")
training/eval/rescore.py ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Re-apply eval.score to stored results after a label or check fix, without calling a model."""
2
+ import json, os, sys
3
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
4
+ from eval import score, HERE
5
+ items = [json.loads(l) for l in open(os.path.join(HERE, "testset.jsonl"))]
6
+ for path in sys.argv[1:]:
7
+ rows = [json.loads(l) for l in open(path)]
8
+ for r in rows:
9
+ assert r["q"] == items[r["i"]]["q"], "test set changed order"
10
+ r["score"] = score(items[r["i"]], r["pred"])
11
+ with open(path, "w") as f:
12
+ f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
training/eval/results-2b-v1-s350.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-2b-v1-s700.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-haiku-5.5-think.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-haiku-5.5.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-qwen3.5-0.8b.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-qwen3.5-2b-q4_K_M.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-qwen3.5-2b.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/results-qwen3.5-4b.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
training/eval/schema.json ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "properties": {
3
+ "standalone_question": {
4
+ "type": "string"
5
+ },
6
+ "answer_shape": {
7
+ "type": "string"
8
+ },
9
+ "place": {
10
+ "type": "string"
11
+ },
12
+ "wants_image": {
13
+ "type": "boolean"
14
+ },
15
+ "image_query": {
16
+ "type": "string"
17
+ },
18
+ "needs_search": {
19
+ "type": "boolean"
20
+ },
21
+ "search_queries": {
22
+ "items": {
23
+ "type": "string"
24
+ },
25
+ "type": "array"
26
+ },
27
+ "recency": {
28
+ "type": "string"
29
+ },
30
+ "complexity": {
31
+ "type": "string"
32
+ },
33
+ "lenses": {
34
+ "items": {
35
+ "type": "string"
36
+ },
37
+ "type": "array"
38
+ },
39
+ "facets": {
40
+ "items": {
41
+ "type": "string"
42
+ },
43
+ "type": "array"
44
+ },
45
+ "whole_library": {
46
+ "type": "boolean"
47
+ }
48
+ },
49
+ "type": "object",
50
+ "additionalProperties": false,
51
+ "required": [
52
+ "standalone_question",
53
+ "answer_shape",
54
+ "place",
55
+ "wants_image",
56
+ "image_query",
57
+ "needs_search",
58
+ "search_queries",
59
+ "recency",
60
+ "complexity",
61
+ "lenses",
62
+ "facets",
63
+ "whole_library"
64
+ ]
65
+ }
training/eval/system_prompt.txt ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ You triage questions for a research assistant.
2
+
3
+ **Default to searching.** This is a research tool; the user came here instead of
4
+ asking a chatbot precisely because they want sourced, current answers, and your
5
+ training data has a cutoff you cannot see past.
6
+
7
+ Set needs_search to false only when the request is one of these:
8
+ - a greeting, thanks, or small talk;
9
+ - rewriting, translating, summarising or formatting text the user pasted;
10
+ - arithmetic, or reasoning about code or data the user supplied;
11
+ - a question about this conversation itself.
12
+
13
+ Everything else needs a search, including anything about software, APIs,
14
+ documentation, tools, companies, people, events, prices, laws, science or
15
+ health, and every "how do I…", "what is…" and "why does…" about the world. When
16
+ in doubt, search.
17
+
18
+ Set `wants_image` to true when a picture is part of the answer:
19
+ - the reader asks to see the thing: "what does X look like", "show me X",
20
+ "wie sieht X aus", "Bild von X", when X is something with an appearance;
21
+ - the question is about a person, a brand, a product, a place, a building
22
+ or landmark, an animal or plant, a vehicle, an artwork, a flag or a logo:
23
+ the reader recognises these by sight, so the picture identifies them.
24
+ Leave it false when there is nothing to look at: a concept, a mechanism, a
25
+ law, a standard, a protocol, an event, a price or figure, a how-to, advice.
26
+ "Wie sieht es mit X aus" and "what does a good X look like" are idioms about
27
+ a situation or a standard, not requests for a picture.
28
+
29
+ When `wants_image` is true, write `image_query`: the subject as an image
30
+ archive would caption it, a few words in English, the name and nothing else
31
+ ("okapi", "flag of Argentina", "Eiffel Tower", "Porsche 911"). Otherwise
32
+ leave it empty.
33
+
34
+ Anything the user wants to see needs a search, because a picture can only be
35
+ shown if a search returned one. Describing it in words instead is not
36
+ answering it.
37
+
38
+ Rewrite the question so it stands on its own without the conversation: resolve
39
+ "it", "they", "that version" against the history. Keep the user's language.
40
+
41
+ If a search is needed, write up to 3 search queries: keyword queries a
42
+ search engine would answer well, not sentences. Make them cover different angles
43
+ of the question, never restatements of each other. If no search is needed, return
44
+ an empty list.
45
+
46
+ The searches run against these corpora, so phrase the queries for them:
47
+ - Web: Short keyword queries, the way you would type them into a search box.
48
+
49
+ The reader left the choice of corpora to you. Set `lenses` to the one or two
50
+ that actually hold the answer. A third is almost always the general web
51
+ answering the same query twice:
52
+ - "web": Web, General search across several independent engines.
53
+ - "scholar": Science, Peer-reviewed papers and preprints: PubMed, arXiv, Crossref, OpenAlex.
54
+ - "news": News, Newswires and established newsrooms, for anything still developing.
55
+ - "docs": Docs & code, Official documentation, standards and source: MDN, GitHub, Stack Overflow.
56
+ - "reference": Reference, Encyclopaedias and definitions, for concepts and background.
57
+ - "data": Data, Computed figures: prices, conversions, statistics, with their as-of date.
58
+
59
+ "web" is where most questions are answered and the one to fall back on when none
60
+ of the tests below fits. "news" for something happening now, "docs" for an API or
61
+ a library, "reference" for who or what something is, "data" for a price, a rate
62
+ or a quantity.
63
+
64
+ Add "scholar" (*alongside* "web", not instead of it) whenever the question has
65
+ an answer somebody has measured: whether something works, how well, how safely,
66
+ at what dose, by what mechanism, what causes what, how large the effect is. That
67
+ covers medicine, nutrition, psychology, biology, climate, materials and the
68
+ social sciences, and it holds however casually the question is put. "Bringt
69
+ Kreatin was", "ist Süßstoff schädlich" and "warum wirkt Koffein bei mir nicht
70
+ mehr" all have an evidence base, and the layperson asking is exactly the reader
71
+ who cannot find it without you. Answer them from the studies rather than from a
72
+ supplement shop's summary of them. Where a research literature plausibly exists,
73
+ add it: "web" + "scholar" is one search, and the two are ranked against each
74
+ other rather than in isolation.
75
+
76
+ Leave it out where research has no stake: a price, a company's plans, an API, a
77
+ recipe, a legal deadline, what happened yesterday. A subject that merely sounds
78
+ serious does not qualify. The test is whether the thing being asked about has
79
+ been studied, not whether it is important.
80
+
81
+ When "scholar" is among the corpora, write the search queries in English and in
82
+ the field's own terminology, naming the population, the intervention or the
83
+ mechanism. Crossref, PubMed, arXiv and OpenAlex index English titles and
84
+ abstracts, and a German query returns whatever happened to share a word with it.
85
+ The general engines answer an English query perfectly well, so this costs the web
86
+ nothing. One query may stay in the reader's language when the question is also
87
+ about something local, such as a national guideline or a product sold there.
88
+
89
+
90
+ Set `recency` when the answer depends on the current state of the world: a
91
+ change, a release, a price, an ongoing event, or any question containing "now",
92
+ "latest", "today", "this year" or a current year, including their equivalents
93
+ in the user's language, such as "heute", "aktuell" or "dieses Jahr" in German.
94
+ Use the *widest* setting that still sharpens the answer, because the three narrow
95
+ ones are not preferences. "day", "week" and "month" are filters at the search
96
+ engine: everything older is missing from the response entirely and cannot be
97
+ ranked back in. Reserve them for questions where a source from outside the window
98
+ would be worthless: a breaking event, an active story, a figure that moves by
99
+ the day.
100
+
101
+ "year" is the mild one. It excludes nothing; it asks the ranking to put recent
102
+ sources first. That makes it the right answer for the large middle ground: a
103
+ question that is *about* the current state of something, where a two-year-old
104
+ page might still be the one that answers it. Leave it empty for settled subjects,
105
+ where an old source is as good as a new one: a definition, a mechanism, a
106
+ historical fact, anything whose best evidence is a paper rather than a page.
107
+
108
+ What you must never do is reach for "month" to express "fairly recent". That
109
+ deletes the 2005 paper, the standard's original text, and the explanation that
110
+ has been correct for a decade, and the answer is then written from whatever
111
+ happened to be published lately.
112
+
113
+ Set `complexity` to how much research the question actually needs. This decides
114
+ how wide the search goes and how many pages are read, so judge the *question*,
115
+ not how interesting the subject is:
116
+ - "simple" when one good source settles it: a price, a date, a score, a
117
+ definition, a version number, a forecast, "who is X", "when did Y happen".
118
+ - "standard" for anything that needs a few sources read and weighed against
119
+ each other. This is the default when none of the tests below is met.
120
+ - "research" when *any* of these is true, however briefly the question is
121
+ put: it names two or more things to compare or choose between; it asks what
122
+ the evidence, the studies or the state of research say; it has several
123
+ distinct parts that each need their own sources; or it asks for the overall
124
+ picture of a field. One matching test is enough, and the question can be
125
+ very short: "Vergleiche A, B und C" is research in five words, and so is
126
+ "was sagt die Studienlage zu X". Asking what the research says means the
127
+ answer has to survey it. Length alone never is: a long question about one
128
+ thing is still standard.
129
+
130
+ Set `answer_shape` when the answer has a better container than a paragraph,
131
+ so it can be laid out rather than written as prose:
132
+ - "figure" for one current value: a share price, an exchange rate, a
133
+ population, the temperature right now, "how much/many is X".
134
+ - "series" for one quantity across time: a price over months, a
135
+ population by decade, a temperature by year. Three points at least.
136
+ - "table" for several entities compared on the same attributes:
137
+ league standings, three tools rated on the same criteria.
138
+ - "comparison" for exactly two things weighed against each other:
139
+ "X vs Y", "should I use X or Y". Use "table" for three or more.
140
+ - "steps" for an instruction or a process the reader is meant to
141
+ follow: "how do I install X", "how do I apply for Y". Only when
142
+ the sources give the actual steps, in order.
143
+ - "weather" for conditions at a place: the temperature now, today, or
144
+ a forecast for the coming days. Put the bare place name in `place`.
145
+ - "match" for one fixture between two teams or countries: a result, a
146
+ scoreline, an upcoming game. Use "table" instead for a whole round
147
+ or a league table.
148
+ - "profile" for one named thing the question asks *about* and wants
149
+ identified: a person, an organisation, a brand, a product, a place, a
150
+ building, an animal or plant species, a piece of software, a standard:
151
+ "who is X", "what is the company Y", "was ist Z", "wer ist Z". The
152
+ test is that the answer opens by saying what the thing is; a question
153
+ that merely mentions it, or that asks how to use it or whether to buy
154
+ it, is not a profile. Neither is a concept, a
155
+ mechanism or a piece of jargon: "was ist ein Autoencoder" wants the
156
+ idea taught, which is "explainer", and a card of facts about an idea
157
+ is a card with nothing on it.
158
+ - "image" when the answer is something to *look at*: what a flag, a
159
+ logo, an animal, a building or a painting looks like. The question
160
+ is usually "what does X look like", "show me X", "wie sieht X aus".
161
+ - "diagram" when the answer is a structure or a process with named
162
+ parts and the relationships between them: how a request flows
163
+ through a system, how a cycle works, what causes what. Only when the
164
+ sources actually describe the parts and their connections; never to
165
+ decorate a definition. Choose "explainer" instead when the reader
166
+ wants the subject taught rather than mapped; that shape draws the
167
+ same picture when there is one to draw.
168
+ - "explainer" when the reader is asking to *understand* something
169
+ rather than to be told a fact about it: how a mechanism works, why
170
+ something happens, what a concept or a piece of jargon actually
171
+ means. "Wie funktionieren Autoencoder", "warum ist der Himmel blau",
172
+ "explain transformers", "was bedeutet Impedanz". The test is that a
173
+ correct one-line answer would leave the reader no wiser, they asked
174
+ because the subject is unfamiliar, not because they mislaid a fact.
175
+ A question that merely opens with "what is" and wants a named thing
176
+ identified is "profile", and a procedure the reader is meant to
177
+ carry out is "steps". Not an explainer either: how events unfolded
178
+ ("wie kam es zu X", "was ist seit 2022 passiert", "Geschichte von
179
+ X"), a set of things to list ("welche Nebenwirkungen"), or a
180
+ decision to weigh ("lohnt sich X"). Those are prose, with `facets`.
181
+
182
+ Leave it empty for everything else. A three-row table explaining a concept
183
+ is worse than the sentence it replaced, and a diagram of a definition is
184
+ worse still: a shape is chosen because the answer has that form, never
185
+ because the subject sounds like it should.
186
+
187
+ When `answer_shape` is empty, `facets` may name the parts of the answer
188
+ that are structure rather than argument, so they can be laid out in a
189
+ panel above the prose. Most important first, at most three:
190
+ - "facts" for the key attributes of the subject as label and value,
191
+ dates, sizes, owners, versions, places, where the reader would
192
+ otherwise scan a paragraph for them.
193
+ - "stats" for one to four headline numbers the answer turns on, each
194
+ with its unit and its date.
195
+ - "timeline" for events in order, how something came about, what
196
+ happened when. Three dated events at least.
197
+ - "proscons" for arguments on both sides, advantages and drawbacks,
198
+ "is X worth it", "lohnt sich X", "should I".
199
+ - "list" for a set of distinct named things the reader would look up
200
+ one by one: the side effects of a drug, the documents an application
201
+ needs, the alternatives to a product. Each item has a name of its
202
+ own, and there are three at least. Reasons, causes, factors, tips
203
+ or the points of an explanation are not a list, they are the
204
+ argument, and belong in the prose.
205
+ - "grid" for a few items on the same two to four attributes when that
206
+ is only part of the answer; when the table *is* the answer, use the
207
+ "table" shape instead.
208
+ - "callout" for the one caveat the reader must not miss, a deadline, a
209
+ safety note, the common mistake.
210
+
211
+ Leave `facets` empty for most questions, and always when `answer_shape` is
212
+ set. Name a facet only when the sources will state that structure, an answer
213
+ that is one explanation or one argument has none, and a panel that merely
214
+ repeats the prose is noise. The usual matches: a history or a sequence of
215
+ events is "timeline", a decision is "proscons", side effects, required
216
+ documents or named alternatives are "list", a subject with key data is
217
+ "facts". "Why does X happen" or "how does X work" names no facet: its
218
+ parts are steps of one argument, and a list of them is worse than prose.
219
+
220
+ Today's date is 2026-11-05. Treat anything you remember as potentially out of date.
training/eval/testset.jsonl ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"q": "thanks, that helped a lot!", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "smalltalk"}
2
+ {"q": "hi there", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "smalltalk"}
3
+ {"q": "Translate into German: 'The meeting has been moved to Thursday afternoon.'", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "transform"}
4
+ {"q": "Rewrite this to sound more polite: send me the report by tomorrow or else.", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "transform"}
5
+ {"q": "What is 17% of 2,340?", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "arithmetic"}
6
+ {"q": "Why does this Python print 3 instead of 4?\n\nx = [1, 2, 3]\nx.append(4)\nprint(len(x[:-1]))", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "code"}
7
+ {"q": "Summarise this in one sentence: Our Q3 revenue rose 12% on stronger subscription sales, while hardware margins fell because of higher component costs.", "lang": "en", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "transform"}
8
+ {"q": "danke dir, super erklärt", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "smalltalk"}
9
+ {"q": "Guten Morgen!", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "smalltalk"}
10
+ {"q": "Übersetze ins Englische: Ich komme heute etwas später, der Zug hat Verspätung.", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "transform"}
11
+ {"q": "Formuliere das freundlicher: Ihre Rechnung ist seit drei Wochen überfällig, zahlen Sie endlich.", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "transform"}
12
+ {"q": "Was ist 15 % von 480 Euro?", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "arithmetic"}
13
+ {"q": "Was macht diese SQL-Abfrage? SELECT name, COUNT(*) FROM orders GROUP BY name HAVING COUNT(*) > 5", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [], "tag": "code"}
14
+ {"q": "Kannst du deine letzte Antwort kürzer zusammenfassen?", "lang": "de", "search": false, "cx": ["simple", "standard", "research"], "shape": [""], "img": false, "rec": "any", "lens_in": [], "lens_out": [], "keys": [], "sq": [], "place": [], "hist": [{"role": "user", "content": "Wie funktioniert eine Wärmepumpe?"}, {"role": "assistant", "content": "Eine Wärmepumpe entzieht der Umgebung Wärme und hebt sie mit einem Verdichter auf ein höheres Temperaturniveau, um damit zu heizen."}], "tag": "conversation"}
15
+ {"q": "What is the Bitcoin price right now?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["figure"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["bitcoin", "btc"]], "sq": [], "place": [], "hist": [], "tag": "figure"}
16
+ {"q": "How many people live in Tokyo?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["figure", ""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["tokyo"]], "sq": [], "place": [], "hist": [], "tag": "figure"}
17
+ {"q": "When did the Berlin Wall fall?", "lang": "en", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["berlin wall"]], "sq": [], "place": [], "hist": [], "tag": "fact"}
18
+ {"q": "What's the latest stable version of Python?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["", "figure"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["python"]], "sq": [], "place": [], "hist": [], "tag": "version"}
19
+ {"q": "weather in Hamburg tomorrow", "lang": "en", "search": true, "cx": ["simple"], "shape": ["weather"], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["hamburg"]], "sq": [], "place": ["hamburg"], "hist": [], "tag": "weather"}
20
+ {"q": "Will it rain in Seattle this weekend?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["weather"], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["seattle"]], "sq": [], "place": ["seattle"], "hist": [], "tag": "weather"}
21
+ {"q": "Who won the last Champions League final?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["match", ""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["champions league"]], "sq": [], "place": [], "hist": [], "tag": "match"}
22
+ {"q": "Wie viel kostet ein Liter Diesel aktuell in Deutschland?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["figure"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["diesel"]], "sq": [], "place": [], "hist": [], "tag": "figure"}
23
+ {"q": "Wie hoch ist der Leitzins der EZB?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["figure"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["leitzins", "ezb", "ecb", "interest rate"]], "sq": [], "place": [], "hist": [], "tag": "figure"}
24
+ {"q": "Wann wurde Goethe geboren?", "lang": "de", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["goethe"]], "sq": [], "place": [], "hist": [], "tag": "fact"}
25
+ {"q": "Wetter in München morgen", "lang": "de", "search": true, "cx": ["simple"], "shape": ["weather"], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["münchen", "munich"]], "sq": [], "place": ["münchen", "munich"], "hist": [], "tag": "weather"}
26
+ {"q": "Wie wird das Wetter am Wochenende in Köln?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["weather"], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["köln", "cologne"]], "sq": [], "place": ["köln", "cologne"], "hist": [], "tag": "weather"}
27
+ {"q": "Wie hat Bayern gegen Dortmund gespielt?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["match"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["bayern"], ["dortmund", "bvb"]], "sq": [], "place": [], "hist": [], "tag": "match"}
28
+ {"q": "Welche Version hat Node.js LTS gerade?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["", "figure"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["node"]], "sq": [], "place": [], "hist": [], "tag": "version"}
29
+ {"q": "Who is Jensen Huang?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": true, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["jensen huang"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
30
+ {"q": "What is the company Anduril?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["anduril"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
31
+ {"q": "what is Obsidian, the notes app?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["obsidian"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
32
+ {"q": "What does an okapi look like?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["image", "profile"], "img": true, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["okapi"]], "sq": [], "place": [], "hist": [], "tag": "image"}
33
+ {"q": "show me the flag of Bhutan", "lang": "en", "search": true, "cx": ["simple"], "shape": ["image"], "img": true, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["bhutan"]], "sq": [], "place": [], "hist": [], "tag": "image"}
34
+ {"q": "Wer ist Annalena Baerbock?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": true, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["baerbock"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
35
+ {"q": "Was ist die Firma Celonis?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["celonis"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
36
+ {"q": "Wer war Hedy Lamarr?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["profile"], "img": true, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["lamarr"]], "sq": [], "place": [], "hist": [], "tag": "profile"}
37
+ {"q": "Wie sieht ein Axolotl aus?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["image", "profile"], "img": true, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["axolotl"]], "sq": [], "place": [], "hist": [], "tag": "image"}
38
+ {"q": "Zeig mir ein Bild vom Kölner Dom", "lang": "de", "search": true, "cx": ["simple"], "shape": ["image"], "img": true, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["dom", "cathedral"]], "sq": [], "place": [], "hist": [], "tag": "image"}
39
+ {"q": "Wie sieht es mit der Rente mit 63 aus, gibt es die noch?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["rente"]], "sq": [], "place": [], "hist": [], "tag": "idiom"}
40
+ {"q": "How do transformers work in machine learning?", "lang": "en", "search": true, "cx": ["standard", "simple"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": [], "keys": [["transformer"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
41
+ {"q": "Why is the sky blue?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["sky", "rayleigh"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
42
+ {"q": "Explain how public key cryptography works", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["public key", "public-key", "asymmetric"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
43
+ {"q": "What does impedance mean in electronics?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["impedance"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
44
+ {"q": "Wie funktionieren Autoencoder?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["autoencoder"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
45
+ {"q": "Warum wird Brot hart, wenn es alt wird?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["brot", "bread"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
46
+ {"q": "Was bedeutet Inflation eigentlich genau?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["inflation"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
47
+ {"q": "Wie funktioniert eine Wärmepumpe?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["explainer", "diagram"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["wärmepumpe", "heat pump"]], "sq": [], "place": [], "hist": [], "tag": "explainer"}
48
+ {"q": "How do I install Docker on Ubuntu 24.04?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["docker"], ["ubuntu"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
49
+ {"q": "How do I renew a US passport by mail?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["passport"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
50
+ {"q": "How do I set up SSH keys for GitHub?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["ssh"], ["github"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
51
+ {"q": "Wie beantrage ich einen neuen Personalausweis?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["personalausweis"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
52
+ {"q": "Wie melde ich ein Gewerbe an?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["gewerbe"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
53
+ {"q": "Wie richte ich einen Cronjob unter Linux ein?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["cron"]], "sq": [], "place": [], "hist": [], "tag": "steps"}
54
+ {"q": "PostgreSQL vs MySQL for a small web app?", "lang": "en", "search": true, "cx": ["research"], "shape": ["comparison"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["postgres"], ["mysql"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
55
+ {"q": "Should I buy an iPhone 17 or a Pixel 10?", "lang": "en", "search": true, "cx": ["research"], "shape": ["comparison"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["iphone"], ["pixel"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
56
+ {"q": "Compare Rust, Go and Zig for systems programming", "lang": "en", "search": true, "cx": ["research"], "shape": ["table"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["news"], "keys": [["rust"], ["go", "golang"], ["zig"]], "sq": [], "place": [], "hist": [], "tag": "table"}
57
+ {"q": "ETF oder Festgeld, was ist 2026 sinnvoller?", "lang": "de", "search": true, "cx": ["research"], "shape": ["comparison"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["etf"], ["festgeld"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
58
+ {"q": "Vergleiche Tesla Model 3, VW ID.3 und BYD Dolphin", "lang": "de", "search": true, "cx": ["research"], "shape": ["table"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["model 3", "tesla"], ["id.3", "id3"], ["dolphin", "byd"]], "sq": [], "place": [], "hist": [], "tag": "table"}
59
+ {"q": "Was ist besser für Einsteiger, Fuji oder Sony Kameras?", "lang": "de", "search": true, "cx": ["research"], "shape": ["comparison"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["fuji"], ["sony"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
60
+ {"q": "Does creatine actually help with muscle growth?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": ["", "explainer"], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["creatine"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
61
+ {"q": "What does the research say about intermittent fasting and weight loss?", "lang": "en", "search": true, "cx": ["research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["fasting"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
62
+ {"q": "Is aspartame harmful?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["aspartame"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
63
+ {"q": "how much vitamin D should an adult take per day", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["", "figure"], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["vitamin d"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
64
+ {"q": "Bringt Kreatin was?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": ["", "explainer"], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["creatine", "kreatin"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
65
+ {"q": "Was sagt die Studienlage zu Melatonin bei Schlafstörungen?", "lang": "de", "search": true, "cx": ["research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["melatonin"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
66
+ {"q": "Ist Süßstoff schädlich?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["sweetener", "süßstoff", "aspartame", "sucralose"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
67
+ {"q": "Warum wirkt Koffein bei mir nicht mehr?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["explainer", ""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["caffeine", "koffein"]], "sq": [], "place": [], "hist": [], "tag": "evidence"}
68
+ {"q": "How do I paginate results with the GitHub REST API?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["steps", ""], "img": false, "rec": "nothard", "lens_in": ["docs"], "lens_out": ["scholar", "news"], "keys": [["github"], ["paginat"]], "sq": [], "place": [], "hist": [], "tag": "docs"}
69
+ {"q": "What does useEffect's cleanup function do in React?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["explainer", ""], "img": false, "rec": "nothard", "lens_in": ["docs"], "lens_out": ["scholar", "news"], "keys": [["useeffect"]], "sq": [], "place": [], "hist": [], "tag": "docs"}
70
+ {"q": "Wie setze ich in FastAPI einen Timeout für einen Request?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["steps", ""], "img": false, "rec": "nothard", "lens_in": ["docs"], "lens_out": ["scholar", "news"], "keys": [["fastapi"], ["timeout"]], "sq": [], "place": [], "hist": [], "tag": "docs"}
71
+ {"q": "Welche Parameter hat pandas read_csv für das Trennzeichen?", "lang": "de", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["docs"], "lens_out": ["scholar", "news"], "keys": [["read_csv"]], "sq": [], "place": [], "hist": [], "tag": "docs"}
72
+ {"q": "What happened at the UN climate conference this week?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "hard", "lens_in": ["news"], "lens_out": ["scholar"], "keys": [["climate", "cop"]], "sq": [], "place": [], "hist": [], "tag": "news"}
73
+ {"q": "latest news on the OpenAI board", "lang": "en", "search": true, "cx": ["standard", "simple"], "shape": [""], "img": false, "rec": "hard", "lens_in": ["news"], "lens_out": ["scholar"], "keys": [["openai"]], "sq": [], "place": [], "hist": [], "tag": "news"}
74
+ {"q": "Was ist heute an der Börse passiert?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "hard", "lens_in": ["news"], "lens_out": ["scholar"], "keys": [["börse", "dax", "stock"]], "sq": [], "place": [], "hist": [], "tag": "news"}
75
+ {"q": "Was gibt es Neues zum Bürgergeld?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "set", "lens_in": ["news"], "lens_out": ["scholar"], "keys": [["bürgergeld"]], "sq": [], "place": [], "hist": [], "tag": "news"}
76
+ {"q": "How did the 2008 financial crisis come about?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["2008", "financial crisis"]], "sq": [], "place": [], "hist": [], "tag": "history"}
77
+ {"q": "Is a heat pump worth it in an old house?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["news"], "keys": [["heat pump"]], "sq": [], "place": [], "hist": [], "tag": "decision"}
78
+ {"q": "What are the side effects of ibuprofen?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["ibuprofen"]], "sq": [], "place": [], "hist": [], "tag": "list"}
79
+ {"q": "Wie kam es zum Ukrainekrieg?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["ukraine"]], "sq": [], "place": [], "hist": [], "tag": "history"}
80
+ {"q": "Lohnt sich eine Balkonsolaranlage?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": [], "keys": [["balkon", "balcony", "solar"]], "sq": [], "place": [], "hist": [], "tag": "decision"}
81
+ {"q": "Welche Unterlagen brauche ich für einen Bauantrag?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["", "steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["bauantrag"]], "sq": [], "place": [], "hist": [], "tag": "list"}
82
+ {"q": "How has the price of gold changed over the last ten years?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["series"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["gold"]], "sq": [], "place": [], "hist": [], "tag": "series"}
83
+ {"q": "Wie hat sich die Einwohnerzahl Berlins seit 1990 entwickelt?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["series"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["berlin"]], "sq": [], "place": [], "hist": [], "tag": "series"}
84
+ {"q": "What is the current state of research on solid-state batteries?", "lang": "en", "search": true, "cx": ["research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["solid-state", "solid state"]], "sq": [], "place": [], "hist": [], "tag": "survey"}
85
+ {"q": "Welche Förderungen gibt es 2026 für Wärmepumpen, wie hoch sind sie und wie beantrage ich sie?", "lang": "de", "search": true, "cx": ["research"], "shape": ["", "steps"], "img": false, "rec": "set", "lens_in": [], "lens_out": ["scholar"], "keys": [["wärmepumpe", "heat pump"], ["förder"]], "sq": [], "place": [], "hist": [], "tag": "multipart"}
86
+ {"q": "and how much does it cost?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["figure", ""], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["steam deck"]], "sq": [["steam deck"]], "place": [], "hist": [{"role": "user", "content": "What is the Steam Deck OLED?"}, {"role": "assistant", "content": "The Steam Deck OLED is Valve's handheld gaming PC with an OLED screen, released in November 2023."}], "tag": "followup"}
87
+ {"q": "who founded it?", "lang": "en", "search": true, "cx": ["simple"], "shape": ["", "profile"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["mistral"]], "sq": [["mistral"]], "place": [], "hist": [{"role": "user", "content": "What is Mistral AI?"}, {"role": "assistant", "content": "Mistral AI is a French company that builds large language models, based in Paris."}], "tag": "followup"}
88
+ {"q": "is it safe during pregnancy?", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["ibuprofen"]], "sq": [["ibuprofen"]], "place": [], "hist": [{"role": "user", "content": "What is ibuprofen used for?"}, {"role": "assistant", "content": "Ibuprofen is a nonsteroidal anti-inflammatory drug used for pain, fever and inflammation."}], "tag": "followup"}
89
+ {"q": "und wie viel kostet das?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["figure", ""], "img": false, "rec": "any", "lens_in": [], "lens_out": ["scholar"], "keys": [["deutschlandticket", "deutschland-ticket"]], "sq": [["deutschlandticket", "deutschland-ticket"]], "place": [], "hist": [{"role": "user", "content": "Was ist das Deutschlandticket?"}, {"role": "assistant", "content": "Das Deutschlandticket ist ein bundesweit gültiges Monatsabo für den öffentlichen Nahverkehr."}], "tag": "followup"}
90
+ {"q": "Und wo liegt die?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["", "profile", "image"], "img": null, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["zugspitze"]], "sq": [["zugspitze"]], "place": [], "hist": [{"role": "user", "content": "Wie hoch ist die Zugspitze?"}, {"role": "assistant", "content": "Die Zugspitze ist 2962 Meter hoch und damit Deutschlands höchster Berg."}], "tag": "followup"}
91
+ {"q": "Gibt es dazu Studien?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": [""], "img": false, "rec": "nothard", "lens_in": ["scholar"], "lens_out": [], "keys": [["ashwagandha"]], "sq": [["ashwagandha"]], "place": [], "hist": [{"role": "user", "content": "Was ist Ashwagandha?"}, {"role": "assistant", "content": "Ashwagandha ist eine Pflanze aus der ayurvedischen Medizin, die als Nahrungsergänzungsmittel gegen Stress verkauft wird."}], "tag": "followup"}
92
+ {"q": "best budget mechanical keyboard", "lang": "en", "search": true, "cx": ["standard", "research"], "shape": ["", "table"], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["keyboard"]], "sq": [], "place": [], "hist": [], "tag": "recommend"}
93
+ {"q": "Can I bring a power bank on a plane?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["power bank"]], "sq": [], "place": [], "hist": [], "tag": "rule"}
94
+ {"q": "What time zone is Brazil in?", "lang": "en", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["brazil"]], "sq": [], "place": [], "hist": [], "tag": "fact"}
95
+ {"q": "recipe for a classic carbonara", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["steps", ""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["carbonara"]], "sq": [], "place": [], "hist": [], "tag": "recipe"}
96
+ {"q": "How long is a US driver's license valid in Germany?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["license", "licence"]], "sq": [], "place": [], "hist": [], "tag": "rule"}
97
+ {"q": "Wie lange muss ich Steuerunterlagen aufbewahren?", "lang": "de", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["aufbewahr", "steuer"]], "sq": [], "place": [], "hist": [], "tag": "rule"}
98
+ {"q": "Rezept für Schweinebraten mit Kruste", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": ["steps", ""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["schweinebraten"]], "sq": [], "place": [], "hist": [], "tag": "recipe"}
99
+ {"q": "Darf mein Vermieter die Kaution einbehalten, wenn ich Löcher in der Wand habe?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["kaution"]], "sq": [], "place": [], "hist": [], "tag": "rule"}
100
+ {"q": "Welche Zeitzone hat Japan?", "lang": "de", "search": true, "cx": ["simple"], "shape": [""], "img": false, "rec": "none", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["japan"]], "sq": [], "place": [], "hist": [], "tag": "fact"}
101
+ {"q": "Was kann man in Lissabon an einem Wochenende unternehmen?", "lang": "de", "search": true, "cx": ["simple", "standard"], "shape": [""], "img": null, "rec": "nothard", "lens_in": [], "lens_out": ["scholar"], "keys": [["lissabon", "lisbon"]], "sq": [], "place": [], "hist": [], "tag": "recommend"}
102
+ {"q": "Was ist der Unterschied zwischen GmbH und UG?", "lang": "de", "search": true, "cx": ["standard", "research"], "shape": ["comparison"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["gmbh"], ["ug"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
103
+ {"q": "Wie viele Kalorien hat eine Banane?", "lang": "de", "search": true, "cx": ["simple"], "shape": ["figure"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["banane", "banana"]], "sq": [], "place": [], "hist": [], "tag": "figure"}
104
+ {"q": "What's the difference between a virus and a bacterium?", "lang": "en", "search": true, "cx": ["standard", "simple", "research"], "shape": ["comparison", "explainer"], "img": false, "rec": "none", "lens_in": [], "lens_out": ["news"], "keys": [["virus"], ["bacteri"]], "sq": [], "place": [], "hist": [], "tag": "comparison"}
105
+ {"q": "Which documents do I need to open a bank account in the UK as a foreigner?", "lang": "en", "search": true, "cx": ["simple", "standard"], "shape": ["", "steps"], "img": false, "rec": "nothard", "lens_in": [], "lens_out": ["scholar", "news"], "keys": [["bank account"]], "sq": [], "place": [], "hist": [], "tag": "list"}
training/export_ckpt.py ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Merge one saved checkpoint's adapter into the bf16 base, without a GPU.
2
+
3
+ For comparing checkpoints other than the one the run picked as best.
4
+
5
+ uv run python -I export_ckpt.py runs/2b-v1/checkpoints/checkpoint-550 runs/2b-v1-s550/model
6
+ """
7
+
8
+ import json
9
+ import os
10
+ import sys
11
+
12
+ import torch
13
+ from huggingface_hub import snapshot_download
14
+ from safetensors import safe_open
15
+
16
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
17
+ from train import export # noqa: E402
18
+
19
+
20
+ def main() -> None:
21
+ ckpt, out = sys.argv[1], sys.argv[2]
22
+ cfg = json.load(open(os.path.join(ckpt, "adapter_config.json")))
23
+ scale = cfg["lora_alpha"] / cfg["r"]
24
+ a, b = {}, {}
25
+ with safe_open(os.path.join(ckpt, "adapter_model.safetensors"), "pt") as f:
26
+ for key in f.keys():
27
+ name = key.removeprefix("base_model.model.")
28
+ if name.endswith(".lora_A.weight"):
29
+ a[name.removesuffix(".lora_A.weight")] = f.get_tensor(key).float()
30
+ elif name.endswith(".lora_B.weight"):
31
+ b[name.removesuffix(".lora_B.weight")] = f.get_tensor(key).float()
32
+ assert a.keys() == b.keys(), "unpaired adapters"
33
+ deltas = {name: b[name] @ a[name] * scale for name in a}
34
+ print(f"{len(deltas)} layers, scale {scale}")
35
+ base = cfg["base_model_name_or_path"]
36
+ export(deltas, base if os.path.isdir(base) else snapshot_download(base), out)
37
+ print(f"merged model in {out}")
38
+
39
+
40
+ if __name__ == "__main__":
41
+ with torch.no_grad():
42
+ main()
training/extract_real.py ADDED
@@ -0,0 +1,147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Questions from the production database export, with what the classifier saw.
2
+
3
+ Read-only. Mirrors the app: history is the chat's last six completed turns
4
+ before the question (`store.chats.history`, answers cut at 2000 characters);
5
+ `build_prompts.py` then applies `trim_history` to it like every other item.
6
+ A chat in a project, or with files attached, ran against the library, so its
7
+ catalog is rebuilt the way `search.library.Library.catalog` writes it.
8
+
9
+ Splits by chat, so a follow-up never lands in another split than the turn it
10
+ follows. Everything written here stays in data/private/.
11
+
12
+ python3 -I extract_real.py
13
+ """
14
+
15
+ from __future__ import annotations
16
+
17
+ import json
18
+ import os
19
+ import random
20
+ import sqlite3
21
+ from collections import defaultdict
22
+
23
+ HERE = os.path.dirname(os.path.abspath(__file__))
24
+ DB = os.path.join(HERE, "data", "private", "simplixity-export.db")
25
+ OUT = os.path.join(HERE, "data", "private", "real_questions.jsonl")
26
+ DEV_TARGET = 70
27
+ HISTORY_TURNS = 6
28
+ HISTORY_ANSWER_CHARS = 2000
29
+ CATALOG_LIMIT = 120
30
+
31
+
32
+ def byline(paper: dict) -> str:
33
+ """Copy of search.library._byline."""
34
+ authors = [str(a) for a in paper.get("authors") or [] if a]
35
+ surname = authors[0].split()[-1] if authors else ""
36
+ who = f"{surname} et al." if len(authors) > 2 else " & ".join(a.split()[-1] for a in authors)
37
+ year = str(paper.get("year") or "").strip()
38
+ return " ".join(part for part in (who, year) if part) or "Unknown"
39
+
40
+
41
+ def described(name: str, kind: str, pages: int) -> str:
42
+ """Copy of search.library._described."""
43
+ count = int(pages or 0)
44
+ what = f"PDF, {count} page{'s' if count != 1 else ''}" if kind == "pdf" else \
45
+ f"{count} part{'s' if count != 1 else ''} of about a page"
46
+ return f"{(name or 'file')[:140]} ({what})"
47
+
48
+
49
+ def catalog(db: sqlite3.Connection, chat: sqlite3.Row) -> tuple[bool, str]:
50
+ """(runs against a library, catalog text)."""
51
+ papers = []
52
+ if chat["project_id"]:
53
+ rows = db.execute(
54
+ """SELECT p.data FROM lit_papers p JOIN lit_collections c ON c.id = p.collection_id
55
+ WHERE c.project_id = ? AND p.decision = 'relevant' ORDER BY p.decided_at DESC""",
56
+ (chat["project_id"],)).fetchall()
57
+ seen = set()
58
+ for row in rows:
59
+ data = json.loads(row["data"] or "{}")
60
+ key = (data.get("doi") or data.get("title") or "").lower()
61
+ if key and key not in seen:
62
+ seen.add(key)
63
+ papers.append(data)
64
+ files = db.execute(
65
+ """SELECT f.name, f.kind, f.pages FROM lit_files f JOIN lit_collections c ON c.id = f.collection_id
66
+ WHERE c.chat_id = ? AND c.user_id = ? ORDER BY f.created_at""",
67
+ (chat["id"], chat["user_id"])).fetchall()
68
+ if not chat["project_id"] and not files:
69
+ return False, ""
70
+ lines = [f"- {byline(p)}: {(p.get('title') or 'Untitled')[:140]}" for p in papers[:CATALOG_LIMIT]]
71
+ if len(papers) > CATALOG_LIMIT:
72
+ lines.append(f"- … and {len(papers) - CATALOG_LIMIT} more")
73
+ if files:
74
+ lines.append("Attached to this chat by the person:")
75
+ lines += [f"- {described(f['name'], f['kind'], f['pages'])}" for f in files]
76
+ return True, "\n".join(lines)
77
+
78
+
79
+ def main() -> None:
80
+ db = sqlite3.connect(f"file:{DB}?mode=ro", uri=True)
81
+ db.row_factory = sqlite3.Row
82
+ chats = {row["id"]: row for row in db.execute("SELECT * FROM chats")}
83
+ by_chat = defaultdict(list)
84
+ for row in db.execute("SELECT id, chat_id, query, answer, status FROM messages ORDER BY id"):
85
+ by_chat[row["chat_id"]].append(row)
86
+
87
+ items, skipped = [], defaultdict(int)
88
+ for chat_id, rows in by_chat.items():
89
+ chat = chats.get(chat_id)
90
+ if chat is None:
91
+ skipped["no chat"] += 1
92
+ continue
93
+ library, cat = catalog(db, chat)
94
+ for k, row in enumerate(rows):
95
+ query = (row["query"] or "").strip()
96
+ if not query:
97
+ skipped["empty"] += 1
98
+ continue
99
+ if "://" in query:
100
+ # The app reads a linked page and shows the classifier its
101
+ # opening; that page is not in the database.
102
+ skipped["linked page"] += 1
103
+ continue
104
+ done = [r for r in rows[:k] if r["status"] == "completed" and (r["answer"] or "")]
105
+ hist = []
106
+ for r in done[-HISTORY_TURNS:]:
107
+ hist.append({"role": "user", "content": r["query"] or ""})
108
+ hist.append({"role": "assistant", "content": (r["answer"] or "")[:HISTORY_ANSWER_CHARS]})
109
+ items.append({"q": query, "hist": hist, "chat": chat_id, "source": "real",
110
+ "library": library, "catalog": cat})
111
+
112
+ # Dedup on the question alone only for first turns: the same words after a
113
+ # different history are a different classification.
114
+ seen, unique = set(), []
115
+ for it in items:
116
+ key = (" ".join(it["q"].lower().split()), bool(it["hist"]))
117
+ if not it["hist"] and key in seen:
118
+ skipped["duplicate"] += 1
119
+ continue
120
+ seen.add(key)
121
+ unique.append(it)
122
+
123
+ chat_ids = sorted({it["chat"] for it in unique})
124
+ random.Random(5).shuffle(chat_ids)
125
+ dev, count = set(), 0
126
+ sizes = defaultdict(int)
127
+ for it in unique:
128
+ sizes[it["chat"]] += 1
129
+ for c in chat_ids:
130
+ if count >= DEV_TARGET:
131
+ break
132
+ dev.add(c)
133
+ count += sizes[c]
134
+ for it in unique:
135
+ it["split"] = "dev" if it["chat"] in dev else "train"
136
+
137
+ with open(OUT, "w") as f:
138
+ for it in unique:
139
+ f.write(json.dumps(it, ensure_ascii=False) + "\n")
140
+ n_dev = sum(it["split"] == "dev" for it in unique)
141
+ print(f"{len(unique)} questions ({len(unique) - n_dev} train, {n_dev} dev), "
142
+ f"{sum(bool(it['hist']) for it in unique)} with history, "
143
+ f"{sum(it['library'] for it in unique)} against a library; skipped {dict(skipped)} -> {OUT}")
144
+
145
+
146
+ if __name__ == "__main__":
147
+ main()
training/gen_questions.py ADDED
@@ -0,0 +1,259 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Generate synthetic questions for the router with Haiku 5.5.
2
+
3
+ Nothing from the production database goes into these prompts, so the output
4
+ can be published. Every call asks for one category, in one language, in one
5
+ writing style, from one subject area; the combinations are what keeps 3000
6
+ questions from collapsing into the same fifty.
7
+
8
+ uv run python -I gen_questions.py # writes data/synthetic_questions.jsonl
9
+ uv run python -I gen_questions.py --dry 3 # three calls, print only
10
+ """
11
+
12
+ from __future__ import annotations
13
+
14
+ import argparse
15
+ import concurrent.futures as cf
16
+ import difflib
17
+ import json
18
+ import os
19
+ import random
20
+ import re
21
+ import sys
22
+
23
+ HERE = os.path.dirname(os.path.abspath(__file__))
24
+ sys.path.insert(0, HERE)
25
+ import orclient # noqa: E402
26
+
27
+ MODEL = "anthropic/claude-haiku-5.5"
28
+ PER_CALL = 25
29
+
30
+ # (id, what to write, calls per language, variant)
31
+ CATEGORIES = [
32
+ ("smalltalk", "greetings, thanks, small talk, compliments or goodbyes to the assistant", 1, ""),
33
+ ("transform", "requests to translate, rewrite, shorten, summarise or reformat a text the user pastes "
34
+ "in the same message; include the pasted text (one to four sentences)", 2, ""),
35
+ ("arithmetic", "pure calculations or unit conversions with all numbers given", 1, ""),
36
+ ("code", "questions about a code snippet, a log line or a small table the user pastes in the message; "
37
+ "include the snippet", 2, ""),
38
+ ("figure", "questions for one current value: a price, an exchange rate, a share price, a population, "
39
+ "an interest rate, a count", 2, ""),
40
+ ("fact", "short factual lookups: who, when, where, how tall, which country, what time zone", 2, ""),
41
+ ("version", "which version of a piece of software, firmware, game or standard is current or was released", 1, ""),
42
+ ("weather", "weather questions naming a place and a time (now, today, tomorrow, the weekend, next days)", 1, ""),
43
+ ("match", "one sports fixture: a result, a score, the next game between two teams", 1, ""),
44
+ ("standings", "league tables, rankings, standings, medal tables", 1, ""),
45
+ ("person", "who is or was a specific named person (any field, famous or niche)", 2, ""),
46
+ ("organisation", "what is a specific company, brand, product, app, institution or piece of software", 2, ""),
47
+ ("look", "requests to see what something looks like: an animal, plant, building, flag, logo, artwork, car", 1, ""),
48
+ ("explainer", "asking to understand how something works, why something happens, or what a concept or "
49
+ "jargon term means", 3, ""),
50
+ ("steps", "how to do something step by step: technical setup, paperwork, household, applications", 2, ""),
51
+ ("comparison", "exactly two things weighed against each other: X vs Y, should I use X or Y", 2, ""),
52
+ ("table", "three or more named things to compare on the same criteria", 1, ""),
53
+ ("evidence", "health, nutrition, fitness, psychology or medicine questions that have a research basis, "
54
+ "asked casually by a layperson", 3, ""),
55
+ ("research", "questions about the state of research, what studies say, or the overall picture of a field", 1, ""),
56
+ ("docs", "programming questions about a specific API, library, framework, CLI or config option", 2, ""),
57
+ ("news", "what is happening right now or happened this week: news, ongoing events, latest developments", 2, ""),
58
+ ("history", "how something came about, the history of something, what happened since some year", 1, ""),
59
+ ("decision", "is X worth it, should I do X, pros and cons of a personal decision", 2, ""),
60
+ ("list", "a set of named things: side effects, required documents, alternatives to a product, "
61
+ "ingredients, symptoms", 1, ""),
62
+ ("series", "how a quantity developed over time: prices, populations, rates over years or months", 1, ""),
63
+ ("multipart", "one message asking two or three distinct things that each need their own sources", 1, ""),
64
+ ("rules", "legal, tax, tenancy, employment or travel rules and deadlines, often country-specific", 2, ""),
65
+ ("recipe", "recipes, cooking times, substitutions, kitchen questions", 1, ""),
66
+ ("travel", "travel planning, things to do in a place, local recommendations", 1, ""),
67
+ ("shopping", "product recommendations and buying advice: best X under some budget, which X to buy", 1, ""),
68
+ ("idiom", "questions that use phrases like 'what does a good X look like', 'how does it look with X', "
69
+ "'wie sieht es mit X aus', where nothing is to be looked at", 1, ""),
70
+ ("diagram", "how a system, a request, a cycle or a process flows through its named parts", 1, ""),
71
+ ("keywords", "search-box style queries of two to five keywords, no sentence", 2, ""),
72
+ ("followup", "", 10, "history"),
73
+ ("image_attached", "", 2, "images"),
74
+ ("linked", "", 2, "linked"),
75
+ ("library", "", 3, "library"),
76
+ ]
77
+
78
+ STYLES = [
79
+ "casual, all lowercase, a typo or two, no question mark",
80
+ "a complete, polite sentence",
81
+ "very short and terse",
82
+ "with a sentence of context about the asker's own situation first",
83
+ "the way someone types quickly on a phone",
84
+ "precise and a little technical",
85
+ ]
86
+
87
+ DOMAINS = [
88
+ "technology and software", "health and medicine", "personal finance", "travel", "law and bureaucracy",
89
+ "science", "sports", "culture, film and music", "home and garden", "cars and mobility", "education",
90
+ "work and career", "gaming", "politics and society", "nature and animals", "food and cooking",
91
+ "history", "energy and climate", "parenting and family", "DIY and electronics",
92
+ ]
93
+
94
+ FOLLOWUP = {"en": "it uses 'it', 'that', 'they', 'and how much is it', 'any studies on that'",
95
+ "de": "es nutzt 'es', 'das', 'die', 'und was kostet das', 'gibt es dazu Studien'"}
96
+
97
+ LANGS = {"en": "English", "de": "German (natural, everyday German as typed in Germany)"}
98
+
99
+ PLAIN = {"type": "object", "additionalProperties": False, "required": ["questions"],
100
+ "properties": {"questions": {"type": "array", "items": {"type": "string"}}}}
101
+
102
+ HISTORY = {"type": "object", "additionalProperties": False, "required": ["items"], "properties": {"items": {
103
+ "type": "array", "items": {"type": "object", "additionalProperties": False,
104
+ "required": ["turns", "question"], "properties": {
105
+ "turns": {"type": "array", "items": {"type": "object", "additionalProperties": False,
106
+ "required": ["user", "assistant"],
107
+ "properties": {"user": {"type": "string"},
108
+ "assistant": {"type": "string"}}}},
109
+ "question": {"type": "string"}}}}}}
110
+
111
+ LINKED = {"type": "object", "additionalProperties": False, "required": ["items"], "properties": {"items": {
112
+ "type": "array", "items": {"type": "object", "additionalProperties": False,
113
+ "required": ["url", "page_title", "page_opening", "question"], "properties": {
114
+ "url": {"type": "string"}, "page_title": {"type": "string"},
115
+ "page_opening": {"type": "string"}, "question": {"type": "string"}}}}}}
116
+
117
+ LIBRARY = {"type": "object", "additionalProperties": False, "required": ["papers", "questions"], "properties": {
118
+ "papers": {"type": "array", "items": {"type": "object", "additionalProperties": False,
119
+ "required": ["authors", "year", "title"], "properties": {
120
+ "authors": {"type": "array", "items": {"type": "string"}}, "year": {"type": "integer"},
121
+ "title": {"type": "string"}}}},
122
+ "questions": {"type": "array", "items": {"type": "string"}}}}
123
+
124
+ COMMON = ("They are typed into a research assistant that searches the web and answers with sources. "
125
+ "Make them varied in topic, length and phrasing; never two about the same thing. Realistic, the "
126
+ "way real people type, not textbook examples. Write only the user's message, nothing else.")
127
+
128
+
129
+ def prompt_for(cat: tuple, lang: str, style: str, domain: str, rng: random.Random) -> tuple[str, dict]:
130
+ cid, what, _, variant = cat
131
+ language = LANGS[lang]
132
+ if variant == "history":
133
+ n_turns = rng.choice([1, 1, 1, 2, 2, 3])
134
+ return (f"Write {PER_CALL // 2} short conversations in {language}, about {domain} and other everyday "
135
+ f"subjects. Each has {n_turns} earlier turn(s), user question and a 1-3 sentence assistant "
136
+ f"answer, then a follow-up question that only makes sense with the history: {FOLLOWUP[lang]}, "
137
+ f"a comparison with the earlier subject, a request to shorten or explain the last answer, "
138
+ f"or a thanks. Mix these kinds. Everything in {language}. Follow-up style: {style}. {COMMON}", HISTORY)
139
+ if variant == "images":
140
+ return (f"Write {PER_CALL} messages in {language} that someone sends together with an attached photo "
141
+ f"or screenshot: asking what is in it, to read or translate text in it, to explain a chart, "
142
+ f"to identify a plant, product or place in it, or what something in it costs or where to "
143
+ f"buy it. Refer to the image the way people do ('this', 'das hier', 'the attached'). "
144
+ f"Style: {style}. {COMMON}", PLAIN)
145
+ if variant == "linked":
146
+ return (f"Write {PER_CALL // 2} items in {language}. Each is a web page someone pastes a link to "
147
+ f"(a plausible URL, the page title, and the first 2-4 sentences of the page, about {domain} "
148
+ f"or anything else) and their message containing that URL: asking to summarise it, what "
149
+ f"others say about it, whether it is true, how it compares, or a question it raises. "
150
+ f"Style: {style}. {COMMON}", LINKED)
151
+ if variant == "library":
152
+ n = rng.randint(6, 18)
153
+ return (f"Invent a literature project in {domain} or a neighbouring research field: {n} plausible "
154
+ f"academic papers (authors, year, title in English as papers are). Then write "
155
+ f"{PER_CALL // 2} messages in {language} that the researcher asks about this library: a "
156
+ f"specific question one passage answers, a summary or the key findings of the collection, "
157
+ f"writing a related-work section, which papers cover a method, a comparison across papers. "
158
+ f"Style: {style}. {COMMON}", LIBRARY)
159
+ return (f"Write {PER_CALL} user messages in {language}: {what}. Prefer subjects from {domain} where it "
160
+ f"fits the category, otherwise anything. Style: {style}. {COMMON}", PLAIN)
161
+
162
+
163
+ def norm(text: str) -> str:
164
+ return re.sub(r"\W+", " ", text.lower()).strip()
165
+
166
+
167
+ def main() -> None:
168
+ p = argparse.ArgumentParser()
169
+ p.add_argument("--out", default=os.path.join(HERE, "data", "synthetic_questions.jsonl"))
170
+ p.add_argument("--dry", type=int, default=0, help="run this many calls and print them")
171
+ p.add_argument("--workers", type=int, default=8)
172
+ args = p.parse_args()
173
+
174
+ rng = random.Random(11)
175
+ jobs = []
176
+ for cat in CATEGORIES:
177
+ for lang in LANGS:
178
+ for _ in range(cat[2]):
179
+ jobs.append((cat, lang, rng.choice(STYLES), rng.choice(DOMAINS), rng.randrange(1 << 30)))
180
+ if args.dry:
181
+ jobs = rng.sample(jobs, args.dry)
182
+ print(f"{len(jobs)} calls", file=sys.stderr)
183
+
184
+ # Every answer is kept as it arrives, keyed by the job, so a rerun after a
185
+ # crash only pays for the calls that are missing.
186
+ cache_path = os.path.join(HERE, "data", "gen_cache.jsonl")
187
+ cache = {}
188
+ if os.path.exists(cache_path) and not args.dry:
189
+ for line in open(cache_path):
190
+ row = json.loads(line)
191
+ cache[row["job"]] = row["out"]
192
+
193
+ def job_key(job):
194
+ cat, lang, style, domain, seed = job
195
+ return f"{cat[0]}|{lang}|{seed}"
196
+
197
+ def run(job):
198
+ if job_key(job) in cache:
199
+ return job, cache[job_key(job)]
200
+ cat, lang, style, domain, seed = job
201
+ text, schema = prompt_for(cat, lang, style, domain, random.Random(seed))
202
+ try:
203
+ out, _ = orclient.chat(MODEL, [{"role": "user", "content": text}], schema, temperature=1.0,
204
+ max_tokens=8000)
205
+ except RuntimeError as e:
206
+ print(f"\nskipped {job_key(job)}: {e}", file=sys.stderr)
207
+ return job, None
208
+ if not args.dry:
209
+ with orclient._lock, open(cache_path, "a") as f:
210
+ f.write(json.dumps({"job": job_key(job), "out": out}, ensure_ascii=False) + "\n")
211
+ return job, out
212
+
213
+ test = [norm(json.loads(line)["q"]) for line in open(os.path.join(HERE, "eval", "testset.jsonl"))]
214
+ seen: set[str] = set()
215
+ rows = []
216
+ with cf.ThreadPoolExecutor(args.workers) as pool:
217
+ for job, out in pool.map(run, jobs):
218
+ if out is None:
219
+ continue
220
+ cat, lang, style, domain, _ = job
221
+ cid, _, _, variant = cat
222
+ base = {"cat": cid, "lang": lang, "style": style, "domain": domain, "source": "synthetic"}
223
+ if variant == "history":
224
+ items = [{**base, "q": it["question"], "hist": [m for t in it["turns"] for m in (
225
+ {"role": "user", "content": t["user"]}, {"role": "assistant", "content": t["assistant"]})]}
226
+ for it in out["items"]]
227
+ elif variant == "images":
228
+ items = [{**base, "q": q, "images": random.Random(q).choice([1, 1, 1, 2])}
229
+ for q in out["questions"]]
230
+ elif variant == "linked":
231
+ items = [{**base, "q": it["question"], "linked": [[1, it["page_title"], it["page_opening"]]],
232
+ "url": it["url"]} for it in out["items"]]
233
+ elif variant == "library":
234
+ papers = out["papers"]
235
+ items = [{**base, "q": q, "papers": papers} for q in out["questions"]]
236
+ else:
237
+ items = [{**base, "q": q} for q in out["questions"]]
238
+ for item in items:
239
+ key = norm(item["q"])
240
+ if not key or key in seen:
241
+ continue
242
+ if any(difflib.SequenceMatcher(None, key, t).ratio() > 0.85 for t in test):
243
+ continue # too close to a test question
244
+ seen.add(key)
245
+ rows.append(item)
246
+ print(f"{len(rows)} questions, ${orclient.spent['cost']:.3f}", file=sys.stderr, end="\r")
247
+
248
+ if args.dry:
249
+ for r in rows:
250
+ print(json.dumps(r, ensure_ascii=False)[:300])
251
+ return
252
+ with open(args.out, "w") as f:
253
+ for r in rows:
254
+ f.write(json.dumps(r, ensure_ascii=False) + "\n")
255
+ print(f"\n{len(rows)} questions -> {args.out}, cost ${orclient.spent['cost']:.3f}", file=sys.stderr)
256
+
257
+
258
+ if __name__ == "__main__":
259
+ main()
training/label.py ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Label the built prompts with Haiku 5.5, thinking on, the way the app calls it.
2
+
3
+ Resumable: ids already in the output are skipped, so an interrupted run picks
4
+ up where it stopped. The real prompts' labels go to data/private/.
5
+
6
+ python3 -I label.py --limit 50 # a sample first
7
+ python3 -I label.py # everything
8
+ """
9
+
10
+ from __future__ import annotations
11
+
12
+ import argparse
13
+ import concurrent.futures as cf
14
+ import json
15
+ import os
16
+ import sys
17
+
18
+ HERE = os.path.dirname(os.path.abspath(__file__))
19
+ sys.path.insert(0, HERE)
20
+ import orclient # noqa: E402
21
+
22
+ MODEL = "anthropic/claude-haiku-5.5"
23
+ SCHEMA = json.load(open(os.path.join(HERE, "eval", "schema.json")))
24
+ SOURCES = [(os.path.join(HERE, "data", "prompts.jsonl"), os.path.join(HERE, "data", "labels.jsonl")),
25
+ (os.path.join(HERE, "data", "private", "prompts_real.jsonl"),
26
+ os.path.join(HERE, "data", "private", "labels_real.jsonl"))]
27
+
28
+
29
+ def main() -> None:
30
+ p = argparse.ArgumentParser()
31
+ p.add_argument("--limit", type=int, default=0, help="label at most this many new items per file")
32
+ p.add_argument("--workers", type=int, default=8)
33
+ p.add_argument("--max-cost", type=float, default=8.0, help="stop when this many dollars are spent")
34
+ p.add_argument("--prompts", help="label this file instead (with --out)")
35
+ p.add_argument("--out")
36
+ args = p.parse_args()
37
+
38
+ for src, dst in ([(args.prompts, args.out)] if args.prompts else SOURCES):
39
+ done = set()
40
+ if os.path.exists(dst):
41
+ done = {json.loads(line)["id"] for line in open(dst)}
42
+ todo = [row for row in map(json.loads, open(src)) if row["id"] not in done]
43
+ if args.limit:
44
+ todo = todo[: args.limit]
45
+ print(f"{os.path.basename(src)}: {len(done)} done, {len(todo)} to label", file=sys.stderr)
46
+
47
+ def run(row):
48
+ if orclient.spent["cost"] > args.max_cost:
49
+ return row, None, "budget"
50
+ try:
51
+ label, usage = orclient.chat(MODEL, row["messages"], SCHEMA, name="Classification",
52
+ max_tokens=500, reasoning=1024)
53
+ return row, label, usage.get("cost")
54
+ except RuntimeError as e:
55
+ return row, None, str(e)
56
+
57
+ with cf.ThreadPoolExecutor(args.workers) as pool, open(dst, "a") as out:
58
+ for k, (row, label, info) in enumerate(pool.map(run, todo), 1):
59
+ if label is None:
60
+ print(f"\nskipped {row['id']}: {info}", file=sys.stderr)
61
+ continue
62
+ out.write(json.dumps({"id": row["id"], "label": label, "cost": info}, ensure_ascii=False) + "\n")
63
+ out.flush()
64
+ print(f"{k}/{len(todo)} ${orclient.spent['cost']:.3f}", file=sys.stderr, end="\r")
65
+ if "private" in dst and os.path.exists(dst):
66
+ os.chmod(dst, 0o600)
67
+ print(file=sys.stderr)
68
+ print(f"spent ${orclient.spent['cost']:.3f} on {orclient.spent['calls']} calls", file=sys.stderr)
69
+
70
+
71
+ if __name__ == "__main__":
72
+ main()
training/make_dataset.py ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Check the teacher's labels and write the training files.
2
+
3
+ A label is dropped, never repaired, when it breaks a rule the app or the
4
+ prompt states: an unknown shape, recency or complexity value, a facet name
5
+ used as a shape, facets beside a shape, more queries than the depth allows,
6
+ a URL in a query, a picture without a search, queries without a search. A
7
+ repaired label would teach the student an answer the teacher never gave.
8
+
9
+ Three fields the app ignores in a given situation are emptied (`neutral`).
10
+
11
+ The answer is written in schema order, which is the order strict JSON mode
12
+ fills it in and the order the student will generate it in.
13
+
14
+ python3 -I make_dataset.py
15
+ data/train.jsonl, data/valid.jsonl synthetic only (publishable)
16
+ data/private/train_real.jsonl, data/private/valid_real.jsonl
17
+ """
18
+
19
+ from __future__ import annotations
20
+
21
+ import json
22
+ import os
23
+ from collections import Counter
24
+
25
+ HERE = os.path.dirname(os.path.abspath(__file__))
26
+ SCHEMA = json.load(open(os.path.join(HERE, "eval", "schema.json")))
27
+ ORDER = list(SCHEMA["properties"])
28
+ SHAPES = {"", "figure", "series", "table", "image", "match", "weather", "diagram", "profile", "steps",
29
+ "comparison", "explainer"}
30
+ FACETS = {"facts", "stats", "timeline", "proscons", "list", "grid", "callout"}
31
+ RECENCY = {"", "day", "week", "month", "year"}
32
+ COMPLEXITY = {"simple", "standard", "research"}
33
+ LENSES = {"web", "scholar", "news", "docs", "reference", "data", "library"}
34
+
35
+
36
+ def problems(label: dict, settings: dict) -> list[str]:
37
+ out = []
38
+ if set(label) != set(ORDER):
39
+ out.append("fields")
40
+ return out
41
+ if label["answer_shape"] not in SHAPES:
42
+ out.append("facet as shape" if label["answer_shape"] in FACETS else "shape")
43
+ if label["recency"] not in RECENCY:
44
+ out.append("recency")
45
+ if label["complexity"] not in COMPLEXITY:
46
+ out.append("complexity")
47
+ if not set(label["facets"]) <= FACETS:
48
+ out.append("facet")
49
+ if label["facets"] and label["answer_shape"]:
50
+ out.append("facets beside shape")
51
+ if not set(label["lenses"]) <= LENSES:
52
+ out.append("lens")
53
+ queries = label["search_queries"]
54
+ if len(queries) > settings["queries"]:
55
+ out.append("too many queries")
56
+ if any("://" in q for q in queries):
57
+ out.append("url in query")
58
+ if label["wants_image"] and not label["needs_search"]:
59
+ out.append("image without search")
60
+ if not label["needs_search"] and queries:
61
+ out.append("queries without search")
62
+ if label["needs_search"] and not queries:
63
+ out.append("search without queries")
64
+ if label["answer_shape"] == "weather" and not label["place"].strip():
65
+ out.append("weather without place")
66
+ return out
67
+
68
+
69
+ def neutral(label: dict, settings: dict) -> dict:
70
+ """Empty the fields the app ignores in this situation, so the student is
71
+ not taught the teacher's noise in them. None of these changes a run:
72
+ `place` is read only by the weather lookup, `lenses` only on Auto sources
73
+ (`orchestrator._chosen_lenses`), `image_query` only when a picture is wanted."""
74
+ label = dict(label)
75
+ if label["answer_shape"] != "weather":
76
+ label["place"] = ""
77
+ if not settings["auto_lenses"]:
78
+ label["lenses"] = []
79
+ if not label["wants_image"]:
80
+ label["image_query"] = ""
81
+ return label
82
+
83
+
84
+ def load(prompts_path: str, labels_path: str):
85
+ labels = {}
86
+ if os.path.exists(labels_path):
87
+ for line in open(labels_path):
88
+ row = json.loads(line)
89
+ labels[row["id"]] = row["label"]
90
+ for line in open(prompts_path):
91
+ row = json.loads(line)
92
+ if row["id"] in labels:
93
+ yield row, labels[row["id"]]
94
+
95
+
96
+ def write(rows: list[dict], path: str, private: bool) -> None:
97
+ with open(path, "w") as f:
98
+ for row in rows:
99
+ f.write(json.dumps(row, ensure_ascii=False) + "\n")
100
+ if private:
101
+ os.chmod(path, 0o600)
102
+
103
+
104
+ def main() -> None:
105
+ sets = [
106
+ (os.path.join(HERE, "data", "prompts.jsonl"), os.path.join(HERE, "data", "labels.jsonl"),
107
+ os.path.join(HERE, "data"), "", False),
108
+ (os.path.join(HERE, "data", "private", "prompts_real.jsonl"),
109
+ os.path.join(HERE, "data", "private", "labels_real.jsonl"),
110
+ os.path.join(HERE, "data", "private"), "_real", True),
111
+ ]
112
+ for prompts_path, labels_path, folder, suffix, private in sets:
113
+ kept = {"train": [], "dev": []}
114
+ dropped = Counter()
115
+ for row, label in load(prompts_path, labels_path):
116
+ bad = problems(label, row["settings"])
117
+ if bad:
118
+ dropped.update(bad)
119
+ continue
120
+ label = neutral(label, row["settings"])
121
+ answer = json.dumps({k: label[k] for k in ORDER}, ensure_ascii=False)
122
+ kept[row["split"]].append({"id": row["id"], "cat": row["cat"], "lang": row["lang"],
123
+ "settings": row["settings"],
124
+ "messages": row["messages"] + [{"role": "assistant", "content": answer}]})
125
+ write(kept["train"], os.path.join(folder, f"train{suffix}.jsonl"), private)
126
+ write(kept["dev"], os.path.join(folder, f"valid{suffix}.jsonl"), private)
127
+ print(f"{os.path.basename(labels_path)}: {len(kept['train'])} train, {len(kept['dev'])} valid, "
128
+ f"dropped {sum(dropped.values())} problems {dict(dropped)}")
129
+
130
+
131
+ if __name__ == "__main__":
132
+ main()
training/orclient.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Minimal OpenRouter chat client: strict JSON schema, retries, cost tracking.
2
+
3
+ The key is read from .env next to this file and never printed.
4
+ """
5
+
6
+ from __future__ import annotations
7
+
8
+ import json
9
+ import os
10
+ import threading
11
+ import time
12
+ import urllib.error
13
+ import urllib.request
14
+
15
+ HERE = os.path.dirname(os.path.abspath(__file__))
16
+ URL = "https://openrouter.ai/api/v1/chat/completions"
17
+ _lock = threading.Lock()
18
+ spent = {"cost": 0.0, "calls": 0}
19
+
20
+
21
+ def _key() -> str:
22
+ with open(os.path.join(HERE, ".env")) as f:
23
+ for line in f:
24
+ if line.startswith("OPENROUTER_API_KEY="):
25
+ return line.split("=", 1)[1].strip()
26
+ raise SystemExit("OPENROUTER_API_KEY missing in .env")
27
+
28
+
29
+ def chat(model: str, messages: list[dict], schema: dict, *, name: str = "Out", temperature: float = 0.0,
30
+ max_tokens: int = 500, reasoning: int = 0) -> tuple[dict, dict]:
31
+ """Returns (parsed JSON, usage). `reasoning` > 0 grants that many thinking tokens."""
32
+ body = {
33
+ "model": model, "messages": messages, "temperature": temperature,
34
+ "max_tokens": max_tokens + reasoning,
35
+ "response_format": {"type": "json_schema", "json_schema": {"name": name, "strict": True, "schema": schema}},
36
+ "reasoning": {"max_tokens": reasoning} if reasoning else {"enabled": False},
37
+ "usage": {"include": True},
38
+ }
39
+ data = json.dumps(body).encode()
40
+ for attempt in range(6):
41
+ req = urllib.request.Request(URL, data, {"Content-Type": "application/json",
42
+ "Authorization": f"Bearer {_key()}"})
43
+ try:
44
+ with urllib.request.urlopen(req, timeout=180) as r:
45
+ out = json.load(r)
46
+ content = out["choices"][0]["message"].get("content") or ""
47
+ parsed = json.loads(content)
48
+ usage = out.get("usage") or {}
49
+ with _lock:
50
+ spent["cost"] += usage.get("cost") or 0.0
51
+ spent["calls"] += 1
52
+ return parsed, usage
53
+ except urllib.error.HTTPError as e:
54
+ if e.code in (400, 401, 402, 403):
55
+ raise SystemExit(f"OpenRouter {e.code}: {e.read()[:300]!r}")
56
+ err = f"HTTP {e.code}"
57
+ except (urllib.error.URLError, TimeoutError, json.JSONDecodeError, KeyError) as e:
58
+ err = repr(e)
59
+ time.sleep(min(2 ** attempt, 30))
60
+ raise RuntimeError(f"OpenRouter failed after retries: {err}")
training/pyproject.toml ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [project]
2
+ name = "simplixity-router-training"
3
+ version = "0.1.0"
4
+ description = "LoRA training for simplixity-router on CUDA"
5
+ requires-python = ">=3.12,<3.13"
6
+ dependencies = [
7
+ "accelerate",
8
+ "bitsandbytes>=0.50.2",
9
+ "flash-linear-attention",
10
+ "gguf",
11
+ "huggingface-hub",
12
+ "numpy>=2.5.4",
13
+ "peft",
14
+ "safetensors",
15
+ "sentencepiece>=0.2.3",
16
+ "torch",
17
+ "transformers>=5",
18
+ ]
19
+
20
+ [tool.uv]
21
+ package = false
22
+
23
+ [tool.uv.sources]
24
+ gguf = { path = "/home/eric/models/llama.cpp/gguf-py" }
training/run_train.sh ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Train, and after a crash (a CUDA OOM when the desktop takes VRAM) resume
3
+ # from the newest checkpoint. Gives up after 20 attempts.
4
+ # ./run_train.sh runs/2b-v1 [train.py args...]
5
+ set -u
6
+ cd "$(dirname "$0")"
7
+ out=$1; shift
8
+ mkdir -p "$out"
9
+ for attempt in $(seq 1 20); do
10
+ echo "=== attempt $attempt $(date -Is)" >> "$out/train.log"
11
+ if uv run python -I train.py --out "$out" --resume "$@" >> "$out/train.log" 2>&1; then
12
+ echo "=== done $(date -Is)" >> "$out/train.log"; exit 0
13
+ fi
14
+ grep -q "OutOfMemoryError" <(tail -5 "$out/train.log") || { echo "=== failed, not an OOM" >> "$out/train.log"; exit 1; }
15
+ sleep 30
16
+ done
17
+ exit 1
training/to_ollama.sh ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Merged HF model -> GGUF bf16 -> Q4_K_M (+ Q8_0) -> Ollama tag.
3
+ # ./to_ollama.sh runs/2b-v1 simplixity-router:2b-v1
4
+ set -euo pipefail
5
+ cd "$(dirname "$0")"
6
+ run=$1; tag=$2
7
+ LLAMA=~/models/llama.cpp
8
+ mkdir -p "$run/gguf"
9
+ name=$(basename "$run")
10
+ uv run python "$LLAMA/convert_hf_to_gguf.py" "$run/model" --outtype bf16 --outfile "$run/gguf/$name-bf16.gguf" >"$run/gguf/convert.log" 2>&1
11
+ "$LLAMA/build/bin/llama-quantize" "$run/gguf/$name-bf16.gguf" "$run/gguf/$name-Q4_K_M.gguf" Q4_K_M >/dev/null 2>&1
12
+ for q in Q4_K_M; do
13
+ mf=$(mktemp)
14
+ printf 'FROM %s\nRENDERER qwen3.5\nPARSER qwen3.5\n' "$(realpath "$run/gguf/$name-$q.gguf")" > "$mf"
15
+ ollama create "$tag-${q,,}" -f "$mf" >/dev/null
16
+ rm "$mf"
17
+ echo "$tag-${q,,}"
18
+ done
training/train.py ADDED
@@ -0,0 +1,268 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """LoRA SFT for the simplixity router, from the title model's train.py.
2
+
3
+ What differs from the title run, and why:
4
+
5
+ * **Logits only for the answer.** A router example is a ~3500-token prompt
6
+ and a ~150-token answer. Projecting every position onto Qwen3.5's 248k
7
+ vocabulary would take ~4 GB in fp32 for one 4k sequence, more than the
8
+ model; `logits_to_keep` projects just the answer positions, and the loss is
9
+ computed on those.
10
+ * **Mean loss per example**, so a long answer (three queries and a rewrite)
11
+ does not count for more than a short one.
12
+ * **Exported from the base checkpoint's own files**, shards merged into one
13
+ model.safetensors, the mtp.* tensors kept (transformers does not load them,
14
+ so save_pretrained would drop them, and llama.cpp's converter expects them).
15
+
16
+ uv run python -I train.py --out runs/2b-v1
17
+ uv run python -I train.py --out runs/smoke --max-steps 4 --eval-steps 2 --valid-limit 4
18
+ """
19
+
20
+ from __future__ import annotations
21
+
22
+ import argparse
23
+ import json
24
+ import os
25
+
26
+ os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
27
+ import random
28
+ import shutil
29
+
30
+ import torch
31
+ import torch.nn.functional as F
32
+ import torch.utils.checkpoint
33
+ from huggingface_hub import snapshot_download
34
+ from peft import LoraConfig, get_peft_model
35
+ from safetensors import safe_open
36
+ from safetensors.torch import save_file
37
+ from transformers import AutoModelForImageTextToText, AutoTokenizer, Trainer, TrainingArguments, set_seed
38
+
39
+ HERE = os.path.dirname(os.path.abspath(__file__))
40
+ MAX_SEQ_LENGTH = 6144
41
+ TRAIN = [os.path.join(HERE, "data", "train.jsonl"), os.path.join(HERE, "data", "private", "train_real.jsonl")]
42
+ VALID = [os.path.join(HERE, "data", "valid.jsonl"), os.path.join(HERE, "data", "private", "valid_real.jsonl")]
43
+
44
+
45
+ def load_split(paths: list[str], tok, repeat_real: int = 1, limit: int = 0) -> list[dict]:
46
+ """Tokenize, with the loss on everything after the generation prompt.
47
+
48
+ As in the title run, that includes the empty thinking block the template
49
+ writes before the answer, so the model closes it by itself.
50
+ """
51
+ rows, too_long = [], 0
52
+ for path in paths:
53
+ with open(path) as f:
54
+ for line in f:
55
+ item = json.loads(line)
56
+ messages = item["messages"]
57
+ prompt = tok.apply_chat_template(messages[:-1], add_generation_prompt=True, tokenize=False,
58
+ enable_thinking=False)
59
+ full = tok.apply_chat_template(messages, tokenize=False, enable_thinking=False)
60
+ if not full.startswith(prompt):
61
+ raise SystemExit(f"chat template is not prefix-stable for {item['id']}")
62
+ prompt_ids = tok(prompt, add_special_tokens=False)["input_ids"]
63
+ ids = tok(full, add_special_tokens=False)["input_ids"]
64
+ if ids[: len(prompt_ids)] != prompt_ids:
65
+ raise SystemExit(f"tokenisation is not prefix-stable for {item['id']}")
66
+ if len(ids) > MAX_SEQ_LENGTH:
67
+ too_long += 1
68
+ continue
69
+ row = {"input_ids": ids, "start": len(prompt_ids), "id": item["id"]}
70
+ rows.extend([row] * (repeat_real if item["id"].startswith("real-") else 1))
71
+ if limit:
72
+ rows = rows[:limit]
73
+ if too_long:
74
+ print(f"dropped {too_long} examples over {MAX_SEQ_LENGTH} tokens")
75
+ return rows
76
+
77
+
78
+ def collate(batch: list[dict]) -> dict:
79
+ assert len(batch) == 1, "batch 1: no padding through the Gated DeltaNet layers"
80
+ ids = torch.tensor([batch[0]["input_ids"]])
81
+ return {"input_ids": ids, "attention_mask": torch.ones_like(ids), "start": torch.tensor(batch[0]["start"])}
82
+
83
+
84
+ CE_CHUNK = 32
85
+
86
+
87
+ def _chunk_loss(hidden: torch.Tensor, weight: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
88
+ logits = (hidden @ weight.t()).float()
89
+ return F.cross_entropy(logits, target, reduction="sum")
90
+
91
+
92
+ class AnswerLossTrainer(Trainer):
93
+ """Loss on the answer tokens, projected onto the vocabulary 32 at a time.
94
+
95
+ One 248k-wide fp32 row is 1 MB; the ~200 answer positions at once, with
96
+ their gradient, were half a gigabyte, which is what a desktop's
97
+ screensaver takes away mid-run. Each chunk is recomputed in backward.
98
+ """
99
+
100
+ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
101
+ ids = inputs["input_ids"]
102
+ start = int(inputs["start"])
103
+ inner = model.base_model.model # Qwen3_5ForConditionalGeneration
104
+ out = inner.model(input_ids=ids, attention_mask=inputs["attention_mask"])
105
+ hidden = out.last_hidden_state[0, start - 1 : -1] # positions that predict the answer
106
+ target = ids[0, start:]
107
+ weight = inner.lm_head.weight
108
+ total = hidden.new_zeros((), dtype=torch.float32)
109
+ for i in range(0, hidden.shape[0], CE_CHUNK):
110
+ total = total + torch.utils.checkpoint.checkpoint(
111
+ _chunk_loss, hidden[i : i + CE_CHUNK], weight, target[i : i + CE_CHUNK], use_reentrant=False)
112
+ loss = total / target.numel()
113
+ return (loss, out) if return_outputs else loss
114
+
115
+
116
+ def lora_deltas(peft_model, targets: list[str]) -> dict[str, torch.Tensor]:
117
+ """W' - W for every tuned layer, B @ A * scale, in fp32 on the CPU.
118
+
119
+ Computed from the adapters rather than read off a merged model, because
120
+ with an int8 base the merged weight would be int8 too. Added to the
121
+ original bf16 weights, this is the same model a bf16 merge would give.
122
+ """
123
+ modules = dict(peft_model.named_modules())
124
+ deltas = {}
125
+ for name in targets:
126
+ layer = modules["base_model.model." + name]
127
+ a = layer.lora_A["default"].weight.detach().float()
128
+ b = layer.lora_B["default"].weight.detach().float()
129
+ deltas[name] = (b @ a * layer.scaling["default"]).cpu()
130
+ return deltas
131
+
132
+
133
+ def export(deltas: dict[str, torch.Tensor], init_dir: str, out: str) -> None:
134
+ """The base checkpoint with the tuned linear weights, as one file."""
135
+ os.makedirs(out, exist_ok=True)
136
+ for name in os.listdir(init_dir):
137
+ if name.endswith(".safetensors") or name.endswith(".safetensors.index.json") or name in (
138
+ "README.md", ".gitattributes"):
139
+ continue
140
+ src = os.path.join(init_dir, name)
141
+ if os.path.isfile(src):
142
+ shutil.copy(src, os.path.join(out, name))
143
+
144
+ tensors = {}
145
+ for name in sorted(os.listdir(init_dir)):
146
+ if name.endswith(".safetensors"):
147
+ with safe_open(os.path.join(init_dir, name), "pt") as f:
148
+ tensors.update({k: f.get_tensor(k) for k in f.keys()})
149
+ for name, delta in deltas.items():
150
+ key = name + ".weight"
151
+ if key not in tensors:
152
+ raise SystemExit(f"{key} is not in the original checkpoint")
153
+ if delta.shape != tensors[key].shape:
154
+ raise SystemExit(f"{key}: delta {tuple(delta.shape)}, original {tuple(tensors[key].shape)}")
155
+ tensors[key] = (tensors[key].float() + delta).to(tensors[key].dtype).contiguous()
156
+ save_file(tensors, os.path.join(out, "model.safetensors"), metadata={"format": "pt"})
157
+
158
+
159
+ def main() -> None:
160
+ p = argparse.ArgumentParser()
161
+ p.add_argument("--out", required=True)
162
+ p.add_argument("--base", default="Qwen/Qwen3.5-2B")
163
+ p.add_argument("--epochs", type=float, default=2)
164
+ p.add_argument("--lr", type=float, default=2e-5)
165
+ p.add_argument("--max-steps", type=int, default=-1)
166
+ p.add_argument("--eval-steps", type=int, default=100)
167
+ p.add_argument("--rank", type=int, default=16)
168
+ p.add_argument("--scale", type=float, default=20, help="lora_alpha / rank, as in the title run")
169
+ # 0: the dropout copies are the last 100 MB that let a run fit beside the
170
+ # desktop's screensaver.
171
+ p.add_argument("--dropout", type=float, default=0.0)
172
+ p.add_argument("--accum", type=int, default=8)
173
+ p.add_argument("--bits", type=int, choices=[4, 8, 16], default=8, help="base weights while training")
174
+ p.add_argument("--repeat-real", type=int, default=2, help="show each real training question this often")
175
+ p.add_argument("--valid-limit", type=int, default=0)
176
+ p.add_argument("--train", nargs="+", default=TRAIN)
177
+ p.add_argument("--valid", nargs="+", default=VALID)
178
+ p.add_argument("--keep-checkpoints", action="store_true")
179
+ p.add_argument("--resume", action="store_true", help="continue from the newest checkpoint in --out")
180
+ args = p.parse_args()
181
+
182
+ set_seed(7)
183
+ init_dir = snapshot_download(args.base)
184
+ tok = AutoTokenizer.from_pretrained(init_dir)
185
+ train = load_split(args.train, tok, repeat_real=args.repeat_real)
186
+ random.Random(7).shuffle(train)
187
+ valid = load_split(args.valid, tok, limit=args.valid_limit)
188
+ lengths = [len(r["input_ids"]) for r in train]
189
+ print(f"{len(train)} train, {len(valid)} valid, {sum(lengths) / 1e6:.1f}M tokens per epoch, "
190
+ f"longest {max(lengths)}")
191
+
192
+ quant = {}
193
+ if args.bits == 4:
194
+ # QLoRA: NF4 base, bf16 compute. Another 0.7 GB below int8.
195
+ from transformers import BitsAndBytesConfig
196
+ quant = {"quantization_config": BitsAndBytesConfig(
197
+ load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True,
198
+ bnb_4bit_compute_dtype=torch.bfloat16, llm_int8_skip_modules=["visual", "lm_head", "embed_tokens"]),
199
+ "device_map": {"": 0}}
200
+ elif args.bits == 8:
201
+ # The language model's linear layers in int8: 1.4 GB less, which is
202
+ # what lets a run survive the desktop taking VRAM mid-run. The export
203
+ # merges the adapters into the original bf16 weights, not into these.
204
+ from transformers import BitsAndBytesConfig
205
+ quant = {"quantization_config": BitsAndBytesConfig(
206
+ load_in_8bit=True, llm_int8_skip_modules=["visual", "lm_head", "embed_tokens"]),
207
+ "device_map": {"": 0}}
208
+ model = AutoModelForImageTextToText.from_pretrained(init_dir, dtype=torch.bfloat16, **quant)
209
+ targets = [n for n, m in model.named_modules() if isinstance(m, torch.nn.Linear) and ".language_model.layers." in n]
210
+ # Not peft.prepare_model_for_kbit_training: it casts every unquantized
211
+ # tensor to fp32, and the 248k-token embedding alone would double to 2 GB.
212
+ model.enable_input_require_grads()
213
+ print(f"LoRA on {len(targets)} linear layers")
214
+ # The router never sees an image: the vision tower is 0.66 GB of an 8 GB
215
+ # card for nothing. The export takes every tensor from the base checkpoint,
216
+ # so the published model still has it.
217
+ model.model.visual = None
218
+ torch.cuda.empty_cache()
219
+ model = get_peft_model(model, LoraConfig(r=args.rank, lora_alpha=args.scale * args.rank,
220
+ lora_dropout=args.dropout, target_modules=targets))
221
+ for param in model.parameters():
222
+ if param.requires_grad:
223
+ param.data = param.data.float() # an lr-2e-5 update rounds away in bf16
224
+ model.print_trainable_parameters()
225
+
226
+ training_args = TrainingArguments(
227
+ output_dir=os.path.join(args.out, "checkpoints"),
228
+ num_train_epochs=args.epochs,
229
+ max_steps=args.max_steps,
230
+ per_device_train_batch_size=1,
231
+ per_device_eval_batch_size=1,
232
+ gradient_accumulation_steps=args.accum,
233
+ gradient_checkpointing=True,
234
+ gradient_checkpointing_kwargs={"use_reentrant": False},
235
+ learning_rate=args.lr,
236
+ lr_scheduler_type="constant",
237
+ weight_decay=0.01,
238
+ bf16=True,
239
+ logging_steps=10,
240
+ eval_strategy="steps",
241
+ eval_steps=args.eval_steps,
242
+ save_strategy="steps",
243
+ save_steps=args.eval_steps,
244
+ save_total_limit=None if args.keep_checkpoints else 2,
245
+ load_best_model_at_end=True,
246
+ metric_for_best_model="eval_loss",
247
+ greater_is_better=False,
248
+ prediction_loss_only=True,
249
+ # Makes the Trainer call compute_loss in evaluation; there are no labels.
250
+ label_names=["start"],
251
+ seed=7,
252
+ report_to="none",
253
+ remove_unused_columns=False,
254
+ dataloader_pin_memory=False,
255
+ )
256
+ trainer = AnswerLossTrainer(model=model, args=training_args, train_dataset=train, eval_dataset=valid,
257
+ data_collator=collate)
258
+ ckpt_dir = os.path.join(args.out, "checkpoints")
259
+ resume = args.resume and os.path.isdir(ckpt_dir) and any(n.startswith("checkpoint-") for n in os.listdir(ckpt_dir))
260
+ trainer.train(resume_from_checkpoint=True if resume else None)
261
+ print(f"best checkpoint: {trainer.state.best_model_checkpoint}, eval_loss {trainer.state.best_metric}")
262
+ trainer.model.save_pretrained(os.path.join(args.out, "adapter"))
263
+ export(lora_deltas(trainer.model, targets), init_dir, os.path.join(args.out, "model"))
264
+ print(f"merged model in {os.path.join(args.out, 'model')}")
265
+
266
+
267
+ if __name__ == "__main__":
268
+ main()
training/uv.lock ADDED
@@ -0,0 +1,971 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ version = 1
2
+ revision = 5
3
+ requires-python = "==3.12.*"
4
+
5
+ [[package]]
6
+ name = "accelerate"
7
+ version = "1.15.0"
8
+ source = { registry = "https://pypi.org/simple" }
9
+ dependencies = [
10
+ { name = "huggingface-hub" },
11
+ { name = "numpy" },
12
+ { name = "packaging" },
13
+ { name = "psutil" },
14
+ { name = "pyyaml" },
15
+ { name = "safetensors" },
16
+ { name = "torch" },
17
+ ]
18
+ sdist = { url = "https://files.pythonhosted.org/packages/f5/b5/1d3ed029ac71d3f2961346829a268da923698e9fd63f218f78841f216bfd/accelerate-1.15.0.tar.gz", hash = "sha256:5654f8c5eaa0d4fa68b33e287a97765da6849bf6d51dcac874e73fbbddfb6134", size = 422615, upload-time = "2026-09-09T13:04:49.078Z" }
19
+ wheels = [
20
+ { url = "https://files.pythonhosted.org/packages/8a/4c/34f0450479d01195027260da68d8a3880683f1640c3ca5adf64acb3185f1/accelerate-1.15.0-py3-none-any.whl", hash = "sha256:97eacca0b73e45cb867dbf8c5d5d4dc32219544300e0c8992c7334dc2ef33cec", size = 394295, upload-time = "2026-09-09T13:04:47.331Z" },
21
+ ]
22
+
23
+ [[package]]
24
+ name = "annotated-doc"
25
+ version = "0.0.5"
26
+ source = { registry = "https://pypi.org/simple" }
27
+ sdist = { url = "https://files.pythonhosted.org/packages/5a/8e/38aa427ed5402449e226975b649c5dc73ccadfefeb95e6aecb8f8ea4b6b6/annotated_doc-0.0.5.tar.gz", hash = "sha256:c7e58ce09192557605d8bbd92836d7e1d520ac9580096042c0bfd197efacf1bb", size = 10758, upload-time = "2026-07-28T13:50:58.129Z" }
28
+ wheels = [
29
+ { url = "https://files.pythonhosted.org/packages/3e/30/e900b21425a860e195f32e37657aa1f7c7f2b1bfb26f03ca209b90933c06/annotated_doc-0.0.5-py3-none-any.whl", hash = "sha256:117bac03a25ede5df5440e855b32d556049ca169ead221505badf432fed4b101", size = 5302, upload-time = "2026-07-28T13:50:57.239Z" },
30
+ ]
31
+
32
+ [[package]]
33
+ name = "anyio"
34
+ version = "4.16.0"
35
+ source = { registry = "https://pypi.org/simple" }
36
+ dependencies = [
37
+ { name = "idna" },
38
+ { name = "typing-extensions" },
39
+ ]
40
+ sdist = { url = "https://files.pythonhosted.org/packages/a1/98/7e457a3b014976887a050561955c0b76e1e7e2d6e27b78ecef2ef80518f2/anyio-4.16.0.tar.gz", hash = "sha256:59bb03251907983d1603d27c3087ddb1e93299ea2163adcdba37899a4fb12ffb", size = 310971, upload-time = "2026-11-01T11:20:07.282Z" }
41
+ wheels = [
42
+ { url = "https://files.pythonhosted.org/packages/a9/16/7757fb84c819accc7d7e9af5291b261e4abb8627ee040f99e7d35ef2457a/anyio-4.16.0-py3-none-any.whl", hash = "sha256:c66ae634e0b8228396b801741beb3294326d4b6102f7bba4041c0dfb58c29026", size = 134521, upload-time = "2026-11-01T11:20:05.59Z" },
43
+ ]
44
+
45
+ [[package]]
46
+ name = "bitsandbytes"
47
+ version = "0.50.2"
48
+ source = { registry = "https://pypi.org/simple" }
49
+ dependencies = [
50
+ { name = "numpy" },
51
+ { name = "packaging" },
52
+ { name = "torch" },
53
+ ]
54
+ wheels = [
55
+ { url = "https://files.pythonhosted.org/packages/55/bf/5290208ce1ecf0f2e6a916fc72a75f6e68021ecfd69e7014fc95998532eb/bitsandbytes-0.50.2-py3-none-macosx_14_0_arm64.whl", hash = "sha256:4311f52a880b341bada639e4edd1a3c8d786830c9c93cdde29eaa1f062c8f8e5", size = 123541, upload-time = "2026-08-27T00:10:48.726Z" },
56
+ { url = "https://files.pythonhosted.org/packages/88/d5/b2cb5b5a9daf7349a02b1af2c49b6a044fda2702c9cc5dc296f648358327/bitsandbytes-0.50.2-py3-none-manylinux_2_24_aarch64.whl", hash = "sha256:d5772560dd94c4d9c57f50c9b017450a1707f7687bfd4b3dc86f7342aafe721e", size = 23777467, upload-time = "2026-08-27T00:10:50.92Z" },
57
+ { url = "https://files.pythonhosted.org/packages/a5/6e/e4e8b75716dbe5e50964f070266e06f4e6806ce051bfb97f52ee162b9310/bitsandbytes-0.50.2-py3-none-manylinux_2_24_x86_64.whl", hash = "sha256:55348a9a4a21bfd99cf8c7b32fe67b4030ae5c2a05738e03c1747f65fa6ec283", size = 43139553, upload-time = "2026-08-27T00:10:54.751Z" },
58
+ { url = "https://files.pythonhosted.org/packages/72/82/742dc27a1feab90c8f87f2ed14e6d72d05f9e1cf764b4d2ba30aa9b4a2cb/bitsandbytes-0.50.2-py3-none-win_amd64.whl", hash = "sha256:c697963c8fda3dcd0d7ebd9b5211ae4067feef7cd06e0350d4e816a434fe683d", size = 39096375, upload-time = "2026-08-27T00:10:58.297Z" },
59
+ { url = "https://files.pythonhosted.org/packages/a2/57/61636c5b11b0a32e505127a6dce6fa8fcbf73978babe8fa37082ab547f1c/bitsandbytes-0.50.2-py3-none-win_arm64.whl", hash = "sha256:8437ab68a04ea56daf1d6ecb54230fb1d88be4b89fe2d79bc399bc0203b487cf", size = 1058684, upload-time = "2026-08-27T00:11:00.664Z" },
60
+ ]
61
+
62
+ [[package]]
63
+ name = "certifi"
64
+ version = "2026.10.23"
65
+ source = { registry = "https://pypi.org/simple" }
66
+ sdist = { url = "https://files.pythonhosted.org/packages/fd/11/b524cbd9bbdaca58608e4cf4053fd463417549e1959fa8d3aa18319f9555/certifi-2026.10.23.tar.gz", hash = "sha256:a9895e6e2758b371e0545e321a3a8fd1fcfd9829ada8824a936dfb16aaaf36cd", size = 136225, upload-time = "2026-10-23T11:36:16.478Z" }
67
+ wheels = [
68
+ { url = "https://files.pythonhosted.org/packages/32/d1/a3652fe2c19134d0d15922177821d863cbd7e842d53a21bbe7e8e3a1b5ac/certifi-2026.10.23-py3-none-any.whl", hash = "sha256:e9b54de76e530bb6cf0ccd036787103acb58fd42ba930ea8bf948ddbaa2f61d1", size = 135082, upload-time = "2026-10-23T11:36:14.763Z" },
69
+ ]
70
+
71
+ [[package]]
72
+ name = "charset-normalizer"
73
+ version = "3.5.2"
74
+ source = { registry = "https://pypi.org/simple" }
75
+ sdist = { url = "https://files.pythonhosted.org/packages/33/1c/f41d4e74c28ab327ff3acd36053f7ea506c55872d7a90b0fa71aa3ab0c89/charset_normalizer-3.5.2.tar.gz", hash = "sha256:39de2a259fc954455c57274dc94c79d5842774e1247a016aff30bc0efed0f4ef", size = 172659, upload-time = "2026-09-30T04:39:23.398Z" }
76
+ wheels = [
77
+ { url = "https://files.pythonhosted.org/packages/e7/c8/693809898870237d82785a03f3b2b58fe4c9f14669f84a7d4e623c92a59e/charset_normalizer-3.5.2-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:ed2a239c0ea213acc1908150a3037257083c7c083128f1a4cec2ec4b97dca491", size = 367780, upload-time = "2026-09-30T04:35:30.888Z" },
78
+ { url = "https://files.pythonhosted.org/packages/c9/87/2fea8c13dc24b3ca9c6f803a5b2dfdeae73eb4f9e12c7885ed908ff0433c/charset_normalizer-3.5.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b91363207bd9dc966a691e959bb47f64b30f7ac4b072be9968b366982f7db77c", size = 246730, upload-time = "2026-09-30T04:35:32.286Z" },
79
+ { url = "https://files.pythonhosted.org/packages/a8/9e/09efac30b937722f46d3110ba30b875b24b2e3a266ed746cc4e376a94d80/charset_normalizer-3.5.2-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:38a873987f3be698494da8b2e3085e29da02da7b633dce73e79c699a113d7bf0", size = 237707, upload-time = "2026-09-30T04:35:33.709Z" },
80
+ { url = "https://files.pythonhosted.org/packages/9e/18/70d76670b13686237863a379928d60bd10e021f17d243ab3d7014c4a5f4e/charset_normalizer-3.5.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:355ad8011081dec5412240c087a9a0c9d4d5039f3ed11a3f13e18c2b29b56c51", size = 273050, upload-time = "2026-09-30T04:35:35.138Z" },
81
+ { url = "https://files.pythonhosted.org/packages/54/e2/77a8b09d5adc013ed07b95b01b8b8fa5441c4e810e83ee7e4aae2fa4d91a/charset_normalizer-3.5.2-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ee21e28f0430bd6dc9086c6e525d5e818a44a5ad19720c8a0ef766792f3eb5e5", size = 270345, upload-time = "2026-09-30T04:35:36.502Z" },
82
+ { url = "https://files.pythonhosted.org/packages/7f/c5/38806a25ab5e65fc178f39affeda20858efafede2fce1ffc2556cfc9fe73/charset_normalizer-3.5.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3d31298449090ab8d47b7b1b2a555ff73cac7ed438a08b7ac160980c7ebed649", size = 257601, upload-time = "2026-09-30T04:35:37.919Z" },
83
+ { url = "https://files.pythonhosted.org/packages/ae/8d/213565184708fdb263ae55e2c04ee1ff748129dd65d48ed0e3502da9c85a/charset_normalizer-3.5.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5cde776b7cc66e4f6c99612cea4aa7269aa65863f7a15841b2c264f103822f4e", size = 252222, upload-time = "2026-09-30T04:35:39.544Z" },
84
+ { url = "https://files.pythonhosted.org/packages/7e/24/76d2cefc25472531e4c5c7dfff68865eb1c39b78482f0fdc15b46f047830/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:ae4f5fea5b8b8ccff88238cc8569303e5ee95efae67fa62922a311397a71f346", size = 248482, upload-time = "2026-09-30T04:35:41.088Z" },
85
+ { url = "https://files.pythonhosted.org/packages/7d/dc/65a801b66ab4c197e22c433ab25e7ac24324ac6f45a2269aca42cce309bf/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:f7d486c83842422badd511868fd8a9a20e9407ace71564b6af47ce7e60a336c1", size = 241206, upload-time = "2026-09-30T04:35:42.59Z" },
86
+ { url = "https://files.pythonhosted.org/packages/a7/95/ca9b5eabde673002c6f1e7ada1b223916fe18f6d661da7aabd4d643718f1/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:11a4d68a6ecda3292cb1e50239e111543ba5d709bb62a6b4ea1afcfa729d8875", size = 273190, upload-time = "2026-09-30T04:35:44.347Z" },
87
+ { url = "https://files.pythonhosted.org/packages/2d/8b/803b4d2a3f6e1740f63f1e87b04d14b42f3d4fdfe6ed7d4db2d34102b14f/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d6734d2ef8a50fbf8445c139477da401f50d62a0606bf00e20ec6d87773fefb1", size = 253527, upload-time = "2026-09-30T04:35:45.915Z" },
88
+ { url = "https://files.pythonhosted.org/packages/a9/55/93c0e5dbd085ae0471346026abbe7e0db9ea2d6fea74e51f0b5a46f233a7/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:a815775b6c38d4e0ff7bcffbeba67feded90202bb6a226b8dd35f1c855217413", size = 271285, upload-time = "2026-09-30T04:35:47.49Z" },
89
+ { url = "https://files.pythonhosted.org/packages/95/69/0dbd0e0b9b16cfa816cdfcb3e2e3854a1f680dc07fb1245ea125e7448060/charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:23851fb4e1b85ed3f6c2a27b777cdfe2e19fb5b38429a8faf38c7542b7665869", size = 260010, upload-time = "2026-09-30T04:35:48.996Z" },
90
+ { url = "https://files.pythonhosted.org/packages/58/9d/e7b88e7b1bf403590c3b573277b5e1e488c68c7a6fbacca310a2c324e90c/charset_normalizer-3.5.2-cp312-cp312-win32.whl", hash = "sha256:db19d07e2e0129e974a0e65d0064fc222a446cd5122c2fd4184d2af9fc734a9e", size = 184125, upload-time = "2026-09-30T04:35:50.777Z" },
91
+ { url = "https://files.pythonhosted.org/packages/eb/e6/e6e083884cbcfd49c64865af05027fe7011be7b2d9179524f099a1b611f3/charset_normalizer-3.5.2-cp312-cp312-win_amd64.whl", hash = "sha256:780fbe7cab297b81dad9fb8dc5eb003c0468ffb0d9e5f65068c53a34661a96bc", size = 207486, upload-time = "2026-09-30T04:35:52.194Z" },
92
+ { url = "https://files.pythonhosted.org/packages/c4/e3/017aea0911ada7405a825c7d937eb3a13009664e2f5b38e8c4bbf2abf894/charset_normalizer-3.5.2-cp312-cp312-win_arm64.whl", hash = "sha256:e2af3aad578aa6bd1384bcf4750fc285e5a9de53f40b7d41e5a0bf748edeb2b3", size = 196734, upload-time = "2026-09-30T04:35:53.636Z" },
93
+ { url = "https://files.pythonhosted.org/packages/8c/ab/176fbfd5b64939c55d652366aa5b9ef1d767af207a3aa6ebeb0d226c484d/charset_normalizer-3.5.2-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:4275811936e2f06feff5e598fb42a1b7ae852da8e39605211892b56b81a34efd", size = 331815, upload-time = "2026-09-30T04:38:26.216Z" },
94
+ { url = "https://files.pythonhosted.org/packages/7e/84/371eac6b30bdbcbf2d632a1a01809103459216fcaae61b8b8d922c1bfb8a/charset_normalizer-3.5.2-cp37-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:1c50fe28bbc2ced33386f298650d91218076c05420e6cbd790b913adc41659e7", size = 253276, upload-time = "2026-09-30T04:38:28.032Z" },
95
+ { url = "https://files.pythonhosted.org/packages/43/6f/c4fbae58febff71709c51bc7e18fdfa55341dc382704740f9f0cbf03817b/charset_normalizer-3.5.2-cp37-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d19fbd981a488e22cd04883659ca6b08f50b5974f9fd7c95655ef6a043e5893f", size = 241239, upload-time = "2026-09-30T04:38:29.732Z" },
96
+ { url = "https://files.pythonhosted.org/packages/61/71/458c3f42164a07d0c5210798e9e704b39e540a6793b05aba67f3a35243a9/charset_normalizer-3.5.2-cp37-abi3-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0fed1d06615f022ee3b13caf5e8b180cfea32bb2c5aded8a9d44277afc040f93", size = 231121, upload-time = "2026-09-30T04:38:31.462Z" },
97
+ { url = "https://files.pythonhosted.org/packages/09/54/ab9e89367076f6331bb6c65c4bf14a5361fa5191cb6561bf534f18504e1b/charset_normalizer-3.5.2-cp37-abi3-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:838dcc90063569a0448120554591a1d6c4a4ffe11babf048908793154ab86ade", size = 260350, upload-time = "2026-09-30T04:38:33.239Z" },
98
+ { url = "https://files.pythonhosted.org/packages/7c/c1/061431ecc688d9d76602502cb57cc01e691e682c18f1beb45f9673b5bbd2/charset_normalizer-3.5.2-cp37-abi3-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2ce45c6627b22c47e390bc91a41c3d13032192e699fa0bea96e9671b373d69b0", size = 255430, upload-time = "2026-09-30T04:38:34.865Z" },
99
+ { url = "https://files.pythonhosted.org/packages/8d/1f/20c8949f0676f7ab811abdeb7f4d7f1cbc6e61ff20bef08b44edeb092bc8/charset_normalizer-3.5.2-cp37-abi3-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0774bf9bf620249fee3e0b8b9fd3065de213be30f3aa94ce2494b3b638949e26", size = 250612, upload-time = "2026-09-30T04:38:36.649Z" },
100
+ { url = "https://files.pythonhosted.org/packages/2b/9e/46f2fa4c431fc98c4ae76a8cb5bdca54e0341e3cfc3fcfd8e82740250818/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:1db38f4c5496827c1a501846d64d14c3b80c7e6714e406cd7dc36a9899fa1011", size = 242083, upload-time = "2026-09-30T04:38:38.26Z" },
101
+ { url = "https://files.pythonhosted.org/packages/bd/39/559be29a0c0f086e0bba6922babd38916cc5e0b58ced4de13ee01ea05508/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:304d8e4d493af723536393eee0c689eb7813f4a474c8b479dee63f1fdd98f621", size = 232738, upload-time = "2026-09-30T04:38:39.81Z" },
102
+ { url = "https://files.pythonhosted.org/packages/ff/6c/387b0e4f756a282831c1d9fc6aeb6c51ca4507ca202767c8de15ce9b12e2/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:9b7f416ff0978e2f2249330527f0ad6fa02f4932e6199692d3b52da2048c19e4", size = 260703, upload-time = "2026-09-30T04:38:41.346Z" },
103
+ { url = "https://files.pythonhosted.org/packages/96/92/1fdf015f09ef449f50d3ac4b67c90887c9c318b727daa95cc4f866e6521d/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:01077390b03f7988f11d700a2194e69b119741a86b1a638b1db88891e3eced8e", size = 247622, upload-time = "2026-09-30T04:38:42.937Z" },
104
+ { url = "https://files.pythonhosted.org/packages/dc/3c/8e7b8a5671ad5d433669fb2a76f1a0164df2d9b1718b0206bc2a16d840cc/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_s390x.whl", hash = "sha256:7e841fb9010836c992c9f12fcbd43a831de93a5f726fc1ccd8ca1d0268c5014c", size = 257500, upload-time = "2026-09-30T04:38:44.604Z" },
105
+ { url = "https://files.pythonhosted.org/packages/b4/f0/45b579df5cabc1d5d53ea1cc35e8437d3ca768c0acccc7041517cb6fbb32/charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:9cae88599c7219005d879f98e5ed53341e9a122af585e1091200358a3003d2a0", size = 255100, upload-time = "2026-09-30T04:38:46.289Z" },
106
+ { url = "https://files.pythonhosted.org/packages/31/68/fdec18a343f5fb3f310588dd478b09ac4799e0b187dbade3a8cd776f03ef/charset_normalizer-3.5.2-cp37-abi3-win32.whl", hash = "sha256:01b0c0d2262a9e28e8484a278c7e1b5d650e3ac8cf2683d2967e25899f208bdf", size = 174499, upload-time = "2026-09-30T04:38:47.999Z" },
107
+ { url = "https://files.pythonhosted.org/packages/9d/8a/b618149cc5207943a0242068d7a27897f56a62947b5a039085f2a22029f8/charset_normalizer-3.5.2-cp37-abi3-win_amd64.whl", hash = "sha256:9f56f72050826f63dcee7a7f55b0a77168cb3bfc553fd405e7f8f9ece75a4036", size = 200092, upload-time = "2026-09-30T04:38:49.707Z" },
108
+ { url = "https://files.pythonhosted.org/packages/03/cf/4c66866fa9e2b1c78e3c911516d1de497a677b7ac60f1eceda74ce777ca3/charset_normalizer-3.5.2-cp37-abi3-win_arm64.whl", hash = "sha256:40ab6bffa02ae10a0581e6c198be7d2d8ca5c2a0c64e4ed3465d766df457573e", size = 294363, upload-time = "2026-09-30T04:38:51.312Z" },
109
+ { url = "https://files.pythonhosted.org/packages/fc/ad/d07d7862a62ffa6d79d68074d14823243dd235a77c45262acbf6adeb28bf/charset_normalizer-3.5.2-py3-none-any.whl", hash = "sha256:b6b751274acb69d77b3323d6b7dbaa3c7fdfc1eb829b7eb61d262f32e1af9685", size = 68872, upload-time = "2026-09-30T04:39:21.828Z" },
110
+ ]
111
+
112
+ [[package]]
113
+ name = "click"
114
+ version = "8.5.0"
115
+ source = { registry = "https://pypi.org/simple" }
116
+ sdist = { url = "https://files.pythonhosted.org/packages/c7/0e/7fa0ef50764b67090eca4114772a2abf8b6148198475e54c660b97caeee6/click-8.5.0.tar.gz", hash = "sha256:ba0d2089de75ea0310e2dde03160e6ca10009947fb95a182f9b54021bb272e34", size = 382235, upload-time = "2026-08-26T13:33:14.56Z" }
117
+ wheels = [
118
+ { url = "https://files.pythonhosted.org/packages/58/50/6c0d534c5f134586a8e1ba4e330569e32f057e33372ae556463212fb4cd3/click-8.5.0-py3-none-any.whl", hash = "sha256:255bc9599cf7748b4b1a446ccc735421bd08a2ae529a8b88597d3de5664ee360", size = 125251, upload-time = "2026-08-26T13:33:12.928Z" },
119
+ ]
120
+
121
+ [[package]]
122
+ name = "colorama"
123
+ version = "0.4.6"
124
+ source = { registry = "https://pypi.org/simple" }
125
+ sdist = { url = "https://files.pythonhosted.org/packages/d8/53/6f443c9a4a8358a93a6792e2acffb9d9d5cb0a5cfd8802644b7b1c9a02e4/colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44", size = 27697, upload-time = "2022-10-25T02:36:22.414Z" }
126
+ wheels = [
127
+ { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" },
128
+ ]
129
+
130
+ [[package]]
131
+ name = "cuda-bindings"
132
+ version = "13.5.0"
133
+ source = { registry = "https://pypi.org/simple" }
134
+ dependencies = [
135
+ { name = "cuda-pathfinder" },
136
+ ]
137
+ wheels = [
138
+ { url = "https://files.pythonhosted.org/packages/b9/a6/13d14600381bdbfe55dfc35564e4bf4a208f169ee9f6cbeb356564268aff/cuda_bindings-13.5.0-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:40eadfad6ae1554f8c1aa456620a85a49a275fc902f2c6dd93ee8c488b2c5e9f", size = 7354398, upload-time = "2026-11-05T12:06:03.542Z" },
139
+ { url = "https://files.pythonhosted.org/packages/27/68/14bf44f386672b03fcc0799be8d18df3e1c45847ee23b95721f6022abbaf/cuda_bindings-13.5.0-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a46a8cc84112d769d0f57bf3ae3fedc115cf09348f14da85a1a8200968b6f262", size = 8378158, upload-time = "2026-11-05T12:06:05.253Z" },
140
+ ]
141
+
142
+ [[package]]
143
+ name = "cuda-pathfinder"
144
+ version = "1.8.6"
145
+ source = { registry = "https://pypi.org/simple" }
146
+ wheels = [
147
+ { url = "https://files.pythonhosted.org/packages/6a/9e/08555797c979322def080814686417ea93f66d38f5c48ebab6867a85a21e/cuda_pathfinder-1.8.6-py3-none-any.whl", hash = "sha256:bf6b3b06323cf95a11d671818392da5da23ddbbb4dfc4b77ca10622783c25980", size = 62603, upload-time = "2026-11-05T21:22:51.229Z" },
148
+ ]
149
+
150
+ [[package]]
151
+ name = "cuda-toolkit"
152
+ version = "13.2.2"
153
+ source = { registry = "https://pypi.org/simple" }
154
+ wheels = [
155
+ { url = "https://files.pythonhosted.org/packages/3d/62/854e4a141f5e87e5425513bad2a742a6260acfe06ef148001b5362148f31/cuda_toolkit-13.2.2-py2.py3-none-any.whl", hash = "sha256:d12edc1aad1356236a2ee10c0d794b1aff5d3e79bfeca54867c3bb8e5afd87b0", size = 2607, upload-time = "2026-07-16T10:13:01.07Z" },
156
+ ]
157
+
158
+ [package.optional-dependencies]
159
+ cublas = [
160
+ { name = "nvidia-cublas", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
161
+ { name = "nvidia-cuda-nvrtc", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
162
+ ]
163
+ cudart = [
164
+ { name = "nvidia-cuda-runtime", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
165
+ ]
166
+ cufft = [
167
+ { name = "nvidia-cufft", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
168
+ { name = "nvidia-nvjitlink", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
169
+ ]
170
+ cufile = [
171
+ { name = "nvidia-cufile", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
172
+ ]
173
+ cupti = [
174
+ { name = "nvidia-cuda-cupti", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
175
+ ]
176
+ cusolver = [
177
+ { name = "nvidia-cublas", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
178
+ { name = "nvidia-cusolver", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
179
+ { name = "nvidia-cusparse", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
180
+ { name = "nvidia-nvjitlink", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
181
+ ]
182
+ cusparse = [
183
+ { name = "nvidia-cusparse", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
184
+ { name = "nvidia-nvjitlink", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
185
+ ]
186
+ nvjitlink = [
187
+ { name = "nvidia-nvjitlink", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
188
+ ]
189
+ nvrtc = [
190
+ { name = "nvidia-cuda-nvrtc", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
191
+ ]
192
+ nvtx = [
193
+ { name = "nvidia-nvtx", marker = "platform_machine == 'aarch64' or platform_machine == 'x86_64'" },
194
+ ]
195
+
196
+ [[package]]
197
+ name = "einops"
198
+ version = "0.8.2"
199
+ source = { registry = "https://pypi.org/simple" }
200
+ sdist = { url = "https://files.pythonhosted.org/packages/2c/77/850bef8d72ffb9219f0b1aac23fbc1bf7d038ee6ea666f331fa273031aa2/einops-0.8.2.tar.gz", hash = "sha256:609da665570e5e265e27283aab09e7f279ade90c4f01bcfca111f3d3e13f2827", size = 56261, upload-time = "2026-01-26T04:13:17.638Z" }
201
+ wheels = [
202
+ { url = "https://files.pythonhosted.org/packages/2a/09/f8d8f8f31e4483c10a906437b4ce31bdf3d6d417b73fe33f1a8b59e34228/einops-0.8.2-py3-none-any.whl", hash = "sha256:54058201ac7087911181bfec4af6091bb59380360f069276601256a76af08193", size = 65638, upload-time = "2026-01-26T04:13:18.546Z" },
203
+ ]
204
+
205
+ [[package]]
206
+ name = "filelock"
207
+ version = "4.2.9"
208
+ source = { registry = "https://pypi.org/simple" }
209
+ sdist = { url = "https://files.pythonhosted.org/packages/a3/1d/d730c664fcb4f1d264a722722a3134cf0881854f73c1fd58abf87572f232/filelock-4.2.9.tar.gz", hash = "sha256:96530b7c538d4ea08493ab067247aab7948c54d8cd622cce711c97cccc5d9896", size = 663116, upload-time = "2026-11-04T18:37:10.014Z" }
210
+ wheels = [
211
+ { url = "https://files.pythonhosted.org/packages/00/05/76184fe08339723d643ab47887a8e4de8f2921ca4e008617ab7490af7857/filelock-4.2.9-py3-none-any.whl", hash = "sha256:a379c50a729e5b68c8519a4d26489f755d9769960a236ecd838190e2a78db2e9", size = 164734, upload-time = "2026-11-04T18:37:08.273Z" },
212
+ ]
213
+
214
+ [[package]]
215
+ name = "fla-core"
216
+ version = "0.6.0"
217
+ source = { registry = "https://pypi.org/simple" }
218
+ dependencies = [
219
+ { name = "einops" },
220
+ ]
221
+ sdist = { url = "https://files.pythonhosted.org/packages/49/2d/3181d715d021c12db379ce0bc51ba78c5b6282ff6c3c79cc2f64d183d1b2/fla_core-0.6.0.tar.gz", hash = "sha256:cdd5b7ff6d5c0e5c08fde73f2624d76e0b506f8684af4e316bc7f2d8bfd9a623", size = 775665, upload-time = "2026-10-23T17:48:03.213Z" }
222
+ wheels = [
223
+ { url = "https://files.pythonhosted.org/packages/23/08/a1a16725117cddded90bb11afed4b31f23d97f7f554775e66af998818ef9/fla_core-0.6.0-py3-none-any.whl", hash = "sha256:b0708b70e0b119cd58a0a5146ea56cb175cef6ed969d8ac60267e3e077ce3ebd", size = 1085221, upload-time = "2026-10-23T17:48:00.23Z" },
224
+ ]
225
+
226
+ [[package]]
227
+ name = "flash-linear-attention"
228
+ version = "0.6.0"
229
+ source = { registry = "https://pypi.org/simple" }
230
+ dependencies = [
231
+ { name = "fla-core" },
232
+ { name = "transformers" },
233
+ ]
234
+ sdist = { url = "https://files.pythonhosted.org/packages/86/2f/6e07920d032e97fcc1a09d7578a7aa595b7dece535280bd712bb7077a3fa/flash_linear_attention-0.6.0.tar.gz", hash = "sha256:4110ff0f94ca54b76fc5499f0c169f5cbd44738d875e0199adc3a11ab0e241fa", size = 222155, upload-time = "2026-10-23T17:48:04.341Z" }
235
+ wheels = [
236
+ { url = "https://files.pythonhosted.org/packages/3c/c1/ad271e49528b06c7032c35e6dca85592f96ef8390e6ce1b803d6b23eae0a/flash_linear_attention-0.6.0-py3-none-any.whl", hash = "sha256:6dec9735f8454523c13553622e98aff30152dde6586b4ac7d88010fbb7a06b5f", size = 403549, upload-time = "2026-10-23T17:48:01.882Z" },
237
+ ]
238
+
239
+ [[package]]
240
+ name = "fsspec"
241
+ version = "2026.9.0"
242
+ source = { registry = "https://pypi.org/simple" }
243
+ sdist = { url = "https://files.pythonhosted.org/packages/77/cd/9be253869fc42e764de7f3dedd6969af7d44ff9c3375214a3442a6f3fc08/fsspec-2026.9.0.tar.gz", hash = "sha256:0f08147951c8cb31d844c3547d631053b127863b60be04cf06e121333ee0e2fe", size = 333545, upload-time = "2026-09-18T17:50:42.825Z" }
244
+ wheels = [
245
+ { url = "https://files.pythonhosted.org/packages/6c/c0/a98505f18594f1bce828bb159cec0fcf9860562f1a2c85913409fc8f3d9e/fsspec-2026.9.0-py3-none-any.whl", hash = "sha256:8dd6e646e99ea382bd85f97a45e6b526a442d79423a7dc673f1e2756d05fcb5f", size = 221738, upload-time = "2026-09-18T17:50:41.341Z" },
246
+ ]
247
+
248
+ [[package]]
249
+ name = "gguf"
250
+ version = "0.19.0"
251
+ source = { directory = "/home/eric/models/llama.cpp/gguf-py" }
252
+ dependencies = [
253
+ { name = "numpy" },
254
+ { name = "pyyaml" },
255
+ { name = "requests" },
256
+ { name = "tqdm" },
257
+ ]
258
+
259
+ [package.metadata]
260
+ requires-dist = [
261
+ { name = "numpy", marker = "python_full_version < '3.14'", specifier = ">=2.2.6" },
262
+ { name = "numpy", marker = "python_full_version >= '3.14'", specifier = ">=2.3.0" },
263
+ { name = "pyside6", marker = "python_full_version >= '3.9' and python_full_version < '3.14' and extra == 'gui'", specifier = ">=6.9,<7.0" },
264
+ { name = "pyyaml", specifier = ">=5.1" },
265
+ { name = "requests", specifier = ">=2.25" },
266
+ { name = "tqdm", specifier = ">=4.27" },
267
+ ]
268
+ provides-extras = ["gui"]
269
+
270
+ [[package]]
271
+ name = "h11"
272
+ version = "0.16.0"
273
+ source = { registry = "https://pypi.org/simple" }
274
+ sdist = { url = "https://files.pythonhosted.org/packages/01/ee/02a2c011bdab74c6fb3c75474d40b3052059d95df7e73351460c8588d963/h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1", size = 101250, upload-time = "2025-04-24T03:35:25.427Z" }
275
+ wheels = [
276
+ { url = "https://files.pythonhosted.org/packages/04/4b/29cac41a4d98d144bf5f6d33995617b185d14b22401f75ca86f384e87ff1/h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86", size = 37515, upload-time = "2025-04-24T03:35:24.344Z" },
277
+ ]
278
+
279
+ [[package]]
280
+ name = "hf-xet"
281
+ version = "1.8.0"
282
+ source = { registry = "https://pypi.org/simple" }
283
+ sdist = { url = "https://files.pythonhosted.org/packages/2f/18/d91c20f815df6cc496a5d0ad8db6d274dad5c0cfcd59babfa2440a9aba44/hf_xet-1.8.0.tar.gz", hash = "sha256:10ed23b51dacf7917d64f335c25048deb470efa755e9989b218fc0531386138e", size = 1013014, upload-time = "2026-10-30T10:04:33.741Z" }
284
+ wheels = [
285
+ { url = "https://files.pythonhosted.org/packages/97/74/d89d5235524bc14d8519e12ca8bc5204f45009ae83d63a410f52c29eb4a8/hf_xet-1.8.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:e00d8f60170cd3dd55d98306e41dbaf28b27ff1bcfefbfa84279d197221be83e", size = 3875002, upload-time = "2026-10-30T10:04:18.958Z" },
286
+ { url = "https://files.pythonhosted.org/packages/d4/f2/ef25e2085e92dcaeaa675331e443923aed2006f6fe2d984e15d8b8d40766/hf_xet-1.8.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:3e8934784425781f0f779b91044df441f3a7af4dab47375d7e3aeaecbb8eecb4", size = 3600947, upload-time = "2026-10-30T10:04:20.906Z" },
287
+ { url = "https://files.pythonhosted.org/packages/f8/1e/1e8ecc05ef034e9a95c18ed32d2ec07905a7cd2ca611a9d88e25ee1f4850/hf_xet-1.8.0-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:29dab7c904cd1914a0439c6e1306c12c7497c8bd32290fbe563ebf51ca4b6d6d", size = 4292483, upload-time = "2026-10-30T10:04:22.875Z" },
288
+ { url = "https://files.pythonhosted.org/packages/35/c4/127a1548df019ea4c80f06e685f263e146075143335a02e78a3c41dbc4eb/hf_xet-1.8.0-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:8800e3444770f2c630b361bc6d5c6ebb770b7830646cf632fbb79d0e5bbc9e7f", size = 4069340, upload-time = "2026-10-30T10:04:24.664Z" },
289
+ { url = "https://files.pythonhosted.org/packages/a5/d5/281cd82736c3e181e597588589d2204d3a5c388a464d4a2a8c546995bae9/hf_xet-1.8.0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:a584a7b3268da62f23c3187cc7201ca01f7b610985a36b1b6119c66bccb0383c", size = 4272344, upload-time = "2026-10-30T10:04:26.519Z" },
290
+ { url = "https://files.pythonhosted.org/packages/48/77/886f09e91a5eadb054ce03b8855323cd2a72f20db76ab4119e0ecda65eef/hf_xet-1.8.0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:1ebaf248c9bef8bb0c9d1d8d71db1fdf0cac14f922cc7a936a81adbf12eb5f01", size = 4500689, upload-time = "2026-10-30T10:04:28.394Z" },
291
+ { url = "https://files.pythonhosted.org/packages/97/4d/2c76a992b370ff77d0b5219275779f8c9a7dabb8ee649deb265dfed84d0f/hf_xet-1.8.0-cp38-abi3-win_amd64.whl", hash = "sha256:92188598dedc1e1f5609bd718d4db8687a6c89915553912571e2f587892eecec", size = 3911559, upload-time = "2026-10-30T10:04:30.165Z" },
292
+ { url = "https://files.pythonhosted.org/packages/87/91/fdb1454a43eff8011047d176ada2f0fb4926cf3d74b87357ff35793f6d98/hf_xet-1.8.0-cp38-abi3-win_arm64.whl", hash = "sha256:42b66fdc688eaa4ff3eb77dc9275e5a0f1258ecc9fdd083ddaf43bb49bec925b", size = 3738458, upload-time = "2026-10-30T10:04:31.882Z" },
293
+ ]
294
+
295
+ [[package]]
296
+ name = "httpcore2"
297
+ version = "2.16.0"
298
+ source = { registry = "https://pypi.org/simple" }
299
+ dependencies = [
300
+ { name = "h11" },
301
+ { name = "truststore" },
302
+ ]
303
+ sdist = { url = "https://files.pythonhosted.org/packages/c0/61/afa9d6b7452f9eca73f73d45324eb0bb52f098fd5e6049bf6b8263be9b4e/httpcore2-2.16.0.tar.gz", hash = "sha256:c72cd0413563b17e1cd8354677dff887466f8ec2551679759e93126534428643", size = 69886, upload-time = "2026-10-21T07:09:45.31Z" }
304
+ wheels = [
305
+ { url = "https://files.pythonhosted.org/packages/af/c2/ce417c397afa5181f0e3c6ff01a4d01639414e1adef08730d97571297151/httpcore2-2.16.0-py3-none-any.whl", hash = "sha256:5c4358319f2602d984ad93182c9fee5160f22e9da7babdda870a5e7c4e2dc34f", size = 85070, upload-time = "2026-10-21T07:09:42.802Z" },
306
+ ]
307
+
308
+ [[package]]
309
+ name = "httpx2"
310
+ version = "2.16.0"
311
+ source = { registry = "https://pypi.org/simple" }
312
+ dependencies = [
313
+ { name = "anyio", marker = "sys_platform != 'emscripten'" },
314
+ { name = "httpcore2", marker = "sys_platform != 'emscripten'" },
315
+ { name = "httpx2-jsfetch", marker = "sys_platform == 'emscripten'" },
316
+ { name = "idna" },
317
+ { name = "truststore", marker = "sys_platform != 'emscripten'" },
318
+ { name = "typing-extensions" },
319
+ ]
320
+ sdist = { url = "https://files.pythonhosted.org/packages/43/73/5ca2cedf5e4e50423d1c258877865a6b62793633b561c58499f8aeb33b7e/httpx2-2.16.0.tar.gz", hash = "sha256:0d54c98128e8f2ff7fbbde0fca8c359e5f032edeaaaac443cdb7f31f685bcd7f", size = 103116, upload-time = "2026-10-21T07:09:46.348Z" }
321
+ wheels = [
322
+ { url = "https://files.pythonhosted.org/packages/7b/3d/2e4bc4fd1ef9770871669adb348d612a6e4786b303a23f9935b81c48dd60/httpx2-2.16.0-py3-none-any.whl", hash = "sha256:fd2c5d7909e8d86fe92d53020c310ca0e4e431b838c207c5754b9bbf0308d49a", size = 97880, upload-time = "2026-10-21T07:09:44.01Z" },
323
+ ]
324
+
325
+ [[package]]
326
+ name = "httpx2-jsfetch"
327
+ version = "1.0"
328
+ source = { registry = "https://pypi.org/simple" }
329
+ sdist = { url = "https://files.pythonhosted.org/packages/cd/c4/0e5636363151a2a1795e0a77617168b9ca438e1748ec05fc9b5687f93d64/httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60", size = 6872, upload-time = "2026-08-07T00:13:07.492Z" }
330
+ wheels = [
331
+ { url = "https://files.pythonhosted.org/packages/9b/43/832f631d32e4f1211caa2ba368317739fe71f0b8530e4c9d15dc454bac2a/httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32", size = 6382, upload-time = "2026-08-07T00:13:06.567Z" },
332
+ ]
333
+
334
+ [[package]]
335
+ name = "huggingface-hub"
336
+ version = "2.4.0"
337
+ source = { registry = "https://pypi.org/simple" }
338
+ dependencies = [
339
+ { name = "click" },
340
+ { name = "filelock" },
341
+ { name = "fsspec" },
342
+ { name = "hf-xet", marker = "platform_machine == 'AMD64' or platform_machine == 'ARM64' or platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" },
343
+ { name = "httpx2" },
344
+ { name = "packaging" },
345
+ { name = "pyyaml" },
346
+ { name = "tqdm" },
347
+ { name = "typing-extensions" },
348
+ ]
349
+ sdist = { url = "https://files.pythonhosted.org/packages/b5/3b/2fcef1eabe52aec90fc8cd253cc154da1d02a446d7bbcdec59b4a51d4be7/huggingface_hub-2.4.0.tar.gz", hash = "sha256:13a6158b9e3393d9d3f248ce99074d50c2a0a1449fb9d8e2f710751e2084c188", size = 1066194, upload-time = "2026-10-29T15:30:12.321Z" }
350
+ wheels = [
351
+ { url = "https://files.pythonhosted.org/packages/d4/55/ff47f1b082ecf5f43f4bc1e3b932d68eead171dabbbd1a92666c686301d7/huggingface_hub-2.4.0-py3-none-any.whl", hash = "sha256:42ecfd93ede47bdd333daebbf66a32d7bfe2e18752754fe4d2d4f98a0179d801", size = 857021, upload-time = "2026-10-29T15:30:10.421Z" },
352
+ ]
353
+
354
+ [[package]]
355
+ name = "idna"
356
+ version = "3.20"
357
+ source = { registry = "https://pypi.org/simple" }
358
+ sdist = { url = "https://files.pythonhosted.org/packages/f5/08/8eea9d4b8302028f3abb2c0813953f7aec26d33b7a8960ed760e65ff29fa/idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44", size = 216463, upload-time = "2026-09-17T14:11:04.752Z" }
359
+ wheels = [
360
+ { url = "https://files.pythonhosted.org/packages/58/a2/bb081bab032533a855d44de1d56f8e8426114ff1ba5d1f07a438a0a654f8/idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c", size = 69583, upload-time = "2026-09-17T14:11:03.168Z" },
361
+ ]
362
+
363
+ [[package]]
364
+ name = "jinja2"
365
+ version = "3.1.6"
366
+ source = { registry = "https://pypi.org/simple" }
367
+ dependencies = [
368
+ { name = "markupsafe" },
369
+ ]
370
+ sdist = { url = "https://files.pythonhosted.org/packages/df/bf/f7da0350254c0ed7c72f3e33cef02e048281fec7ecec5f032d4aac52226b/jinja2-3.1.6.tar.gz", hash = "sha256:0137fb05990d35f1275a587e9aee6d56da821fc83491a0fb838183be43f66d6d", size = 245115, upload-time = "2025-03-05T20:05:02.478Z" }
371
+ wheels = [
372
+ { url = "https://files.pythonhosted.org/packages/62/a1/3d680cbfd5f4b8f15abc1d571870c5fc3e594bb582bc3b64ea099db13e56/jinja2-3.1.6-py3-none-any.whl", hash = "sha256:85ece4451f492d0c13c5dd7c13a64681a86afae63a5f347908daf103ce6d2f67", size = 134899, upload-time = "2025-03-05T20:05:00.369Z" },
373
+ ]
374
+
375
+ [[package]]
376
+ name = "markdown-it-py"
377
+ version = "4.2.0"
378
+ source = { registry = "https://pypi.org/simple" }
379
+ dependencies = [
380
+ { name = "mdurl" },
381
+ ]
382
+ sdist = { url = "https://files.pythonhosted.org/packages/06/ff/7841249c247aa650a76b9ee4bbaeae59370dc8bfd2f6c01f3630c35eb134/markdown_it_py-4.2.0.tar.gz", hash = "sha256:04a21681d6fbb623de53f6f364d352309d4094dd4194040a10fd51833e418d49", size = 82454, upload-time = "2026-05-07T12:08:28.36Z" }
383
+ wheels = [
384
+ { url = "https://files.pythonhosted.org/packages/b3/81/4da04ced5a082363ecfa159c010d200ecbd959ae410c10c0264a38cac0f5/markdown_it_py-4.2.0-py3-none-any.whl", hash = "sha256:9f7ebbcd14fe59494226453aed97c1070d83f8d24b6fc3a3bcf9a38092641c4a", size = 91687, upload-time = "2026-05-07T12:08:27.182Z" },
385
+ ]
386
+
387
+ [[package]]
388
+ name = "markupsafe"
389
+ version = "3.0.4"
390
+ source = { registry = "https://pypi.org/simple" }
391
+ sdist = { url = "https://files.pythonhosted.org/packages/38/9b/e422a865e1d5d57d0e509b4e0bf1c1a70a7f6382c29a5aa428df994c8bc8/markupsafe-3.0.4.tar.gz", hash = "sha256:2e9ad7dd851bf45fab9f75cbff4cb493fee9979e8d8c7c9c3ee119022518edd6", size = 153777, upload-time = "2026-10-02T23:07:22.29Z" }
392
+ wheels = [
393
+ { url = "https://files.pythonhosted.org/packages/81/09/4c59d56b8461ae8eb0d8ba34bb25b7e618547044679d58a82ef9b2479fc1/markupsafe-3.0.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:61631e08084be9e21a8967ec3139c7616ed7c5e9368e05c86d1b39562c8a57b6", size = 11658, upload-time = "2026-10-02T23:04:51.876Z" },
394
+ { url = "https://files.pythonhosted.org/packages/a2/f0/d6613774d86fbf6d145751d43c59875e47a6f9f17daee0aef173bd36d90e/markupsafe-3.0.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0930db9bdc62d22944e10b066448bb65dc9abe9112880c7cab8da54db4284d5f", size = 12050, upload-time = "2026-10-02T23:04:52.931Z" },
395
+ { url = "https://files.pythonhosted.org/packages/0d/f2/8f18e0b806eb13c1f8d07d917a720831ead54253a6dec011fbc78098a6f8/markupsafe-3.0.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a45c3d514f2436064db00d7fc8778d888f0236ebfed649b53d13a59e69ad51b", size = 24363, upload-time = "2026-10-02T23:04:53.895Z" },
396
+ { url = "https://files.pythonhosted.org/packages/60/ce/fa07dbe8a5675558fa36dea033e19995bc783de2dec5f540ccb9030b06aa/markupsafe-3.0.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1e1451fab512d1bcc3dc26988ec1edb0b82c2db909132872cd9356070a6b63df", size = 28525, upload-time = "2026-10-02T23:04:54.905Z" },
397
+ { url = "https://files.pythonhosted.org/packages/85/40/be87c01f3868ec217f8a2015089d71c22c8c5a75324822e5ed1cdd87210d/markupsafe-3.0.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bd3ce56ae2cbae3ba82b683bc425cd7e48d2ed8b10f3e818186b6f5646d9271c", size = 24733, upload-time = "2026-10-02T23:04:56.229Z" },
398
+ { url = "https://files.pythonhosted.org/packages/4f/a7/aeedb5140afa41fc74c225e9184ab96723a6e873b6ee1c9fede7283456d8/markupsafe-3.0.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8e124f974786f831d6043728e38296969d3579db8896fe004682f5758e613581", size = 22985, upload-time = "2026-10-02T23:04:57.521Z" },
399
+ { url = "https://files.pythonhosted.org/packages/c3/fc/e91352bb08c6a59da3ef0909d457bf95a5f5908fbf151b30a06d9dbcfbb4/markupsafe-3.0.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c02e8f18bdedba082cef725942ac823b9b60656db07f7e265cb31618dfd00d77", size = 22001, upload-time = "2026-10-02T23:04:58.597Z" },
400
+ { url = "https://files.pythonhosted.org/packages/5d/f8/bffee5e7d2a3deb59748a797650a48af7e672025cf641a79344a771ad106/markupsafe-3.0.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:9f098115c247e11d138ab83a28fa0323c77015007ea2df73ba5fd714dfefd67c", size = 23793, upload-time = "2026-10-02T23:04:59.686Z" },
401
+ { url = "https://files.pythonhosted.org/packages/ed/59/b853d6628ecb4d658e1d637224846d5e9bb4adf4f8df97f3be9f29dce2ec/markupsafe-3.0.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:d5f93ebbeb8032d47e349328ec8662d973d9b05a70b3c35df1f91fe419b84749", size = 22640, upload-time = "2026-10-02T23:05:00.768Z" },
402
+ { url = "https://files.pythonhosted.org/packages/09/b2/1506df394f0f075797c418d0301498f49e43be194e3ffcb49e6fe6ccf022/markupsafe-3.0.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:64511c54db4e4987aef4c41923235927428729e8174c5dba488429be70a998ed", size = 24074, upload-time = "2026-10-02T23:05:01.813Z" },
403
+ { url = "https://files.pythonhosted.org/packages/c7/81/5ed69cda630ac69ef60d06c09ba5a7f84ff66a2e28cf986fd5614ab3c6e6/markupsafe-3.0.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:e1a622f13970d81f95d0c72f9dc090dce9085fccfa4c9f2174377ee32bd15786", size = 21563, upload-time = "2026-10-02T23:05:03.239Z" },
404
+ { url = "https://files.pythonhosted.org/packages/0c/fe/fb1e79be0fea60aa32602ebefc9c35a82bb42b4df157285ab7dfec12341a/markupsafe-3.0.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c9a7f43c0b202b334cc9184af09bb8f21d3a209e038efaf106936fb69e6b026e", size = 23048, upload-time = "2026-10-02T23:05:04.479Z" },
405
+ { url = "https://files.pythonhosted.org/packages/c8/52/7632a53360671a9b750cdbabaf9cdd89f18b42248b8e4cb42c0b0296e459/markupsafe-3.0.4-cp312-cp312-win32.whl", hash = "sha256:f0ec3b750b59375eab5b0fb2b9254810c00a3375be6d789899f1055a1d556237", size = 14108, upload-time = "2026-10-02T23:05:05.513Z" },
406
+ { url = "https://files.pythonhosted.org/packages/3f/bf/62495e180b7000aaf30000fff849e933f74264638057176cf46852500adc/markupsafe-3.0.4-cp312-cp312-win_amd64.whl", hash = "sha256:11935df9bf455ed0c04eb87bcd720f02b1fe5e02128a9430f23aed6f93336fc7", size = 14303, upload-time = "2026-10-02T23:05:06.538Z" },
407
+ { url = "https://files.pythonhosted.org/packages/c5/8e/4c24208776a65878d656996945aacfbfe010d3720d1a98fc0eb8491fc03b/markupsafe-3.0.4-cp312-cp312-win_arm64.whl", hash = "sha256:a4bbd2d87dd233b9fc5812160c3d0ffbe42edc22a26ce0469f58479ede633fe9", size = 14174, upload-time = "2026-10-02T23:05:07.617Z" },
408
+ ]
409
+
410
+ [[package]]
411
+ name = "mdurl"
412
+ version = "0.1.2"
413
+ source = { registry = "https://pypi.org/simple" }
414
+ sdist = { url = "https://files.pythonhosted.org/packages/d6/54/cfe61301667036ec958cb99bd3efefba235e65cdeb9c84d24a8293ba1d90/mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba", size = 8729, upload-time = "2022-08-14T12:40:10.846Z" }
415
+ wheels = [
416
+ { url = "https://files.pythonhosted.org/packages/b3/38/89ba8ad64ae25be8de66a6d463314cf1eb366222074cfda9ee839c56a4b4/mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8", size = 9979, upload-time = "2022-08-14T12:40:09.779Z" },
417
+ ]
418
+
419
+ [[package]]
420
+ name = "mpmath"
421
+ version = "1.3.0"
422
+ source = { registry = "https://pypi.org/simple" }
423
+ sdist = { url = "https://files.pythonhosted.org/packages/e0/47/dd32fa426cc72114383ac549964eecb20ecfd886d1e5ccf5340b55b02f57/mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f", size = 508106, upload-time = "2023-03-07T16:47:11.061Z" }
424
+ wheels = [
425
+ { url = "https://files.pythonhosted.org/packages/43/e3/7d92a15f894aa0c9c4b49b8ee9ac9850d6e63b03c9c32c0367a13ae62209/mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c", size = 536198, upload-time = "2023-03-07T16:47:09.197Z" },
426
+ ]
427
+
428
+ [[package]]
429
+ name = "networkx"
430
+ version = "3.7"
431
+ source = { registry = "https://pypi.org/simple" }
432
+ sdist = { url = "https://files.pythonhosted.org/packages/dc/76/3af777226b63a5e64a6b36b1ec5855c14e2b94a37096d4760e595fc43511/networkx-3.7.tar.gz", hash = "sha256:fd77a511bd90f39f3d016351345b52cf5319b813bdca01de3f755d3cca62e96a", size = 1866482, upload-time = "2026-09-21T16:45:16.974Z" }
433
+ wheels = [
434
+ { url = "https://files.pythonhosted.org/packages/7e/cd/fe58041e9011f307c490e3e17dd48cc516448f7c698a3f2d9d9d65d7e6a8/networkx-3.7-py3-none-any.whl", hash = "sha256:e3fd2c13a7814cee3746340d8d7f8598a67f16a58bf47fb7f8793fab6efca1b0", size = 2142205, upload-time = "2026-09-21T16:45:14.609Z" },
435
+ ]
436
+
437
+ [[package]]
438
+ name = "numpy"
439
+ version = "2.5.4"
440
+ source = { registry = "https://pypi.org/simple" }
441
+ sdist = { url = "https://files.pythonhosted.org/packages/95/b0/c7453d0b6e2073c3264468b106ee1563750cecc910965e67357e3698c83e/numpy-2.5.4.tar.gz", hash = "sha256:9a94cf751c9ad8ebaa835bcd3d40dacf8534ad086b88c38029b65123c7999d2a", size = 20866315, upload-time = "2026-10-10T20:05:31.422Z" }
442
+ wheels = [
443
+ { url = "https://files.pythonhosted.org/packages/d0/97/ba2074e92b7befea137e77ea8471e768bbd87c339b7e8c9f5a931949f977/numpy-2.5.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:c6342f54c67093cae5c0227eb0eb772fdb79f2a2c37a6eb278b9909ee06aa356", size = 17001609, upload-time = "2026-10-10T20:02:40.843Z" },
444
+ { url = "https://files.pythonhosted.org/packages/ff/a9/bac826765e971d8e16e2064e9ac7525fd69b40ac17c905033a7f5442023f/numpy-2.5.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b11e8fda06a7d69f15ebf542660b74466c2e51094800c1fb794f47ad4faeef17", size = 12015718, upload-time = "2026-10-10T20:02:43.45Z" },
445
+ { url = "https://files.pythonhosted.org/packages/31/2f/5ea3570fcb8ccd0882bea99436a513b2c85dad8f774a2057849130a8fb99/numpy-2.5.4-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:9cb18a327b49c5c337f972b03682f6a49855525faaf3c0d3e9c96cd0fd8880a8", size = 5451717, upload-time = "2026-10-10T20:02:46.169Z" },
446
+ { url = "https://files.pythonhosted.org/packages/34/f2/b4fc1bafca03868220b5eaf729d2f21ebd7d7b151c0f9e144fe212bbca35/numpy-2.5.4-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:aec3fc4b32ff82421274f5d205c559c51c840c8df66a78efd7f3612dd005a26a", size = 6789926, upload-time = "2026-10-10T20:02:48.139Z" },
447
+ { url = "https://files.pythonhosted.org/packages/dc/96/8319e2457ae4333c62c815c7006b869a4f60985c1e01024c2f8c6c040fe5/numpy-2.5.4-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fe4d21ab149f15e4e6043dfb0de87e6e5f34ac176cde83060e9802981fca2ac2", size = 15695312, upload-time = "2026-10-10T20:02:50.115Z" },
448
+ { url = "https://files.pythonhosted.org/packages/43/a3/c799c62e19c337e6d3770b08e475887fb30ce8477d3c09efca6b2f0228a6/numpy-2.5.4-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fbde6962867ee75b48b0ee29b2b9372ec5d617799dbaf38e82dc0596f2f7738a", size = 16727283, upload-time = "2026-10-10T20:02:53.186Z" },
449
+ { url = "https://files.pythonhosted.org/packages/39/6b/3604e53fb00314d0dc1b94ec9125a1484f649c0a17480b1f0f0c7a9d6250/numpy-2.5.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:381a7a3d2e65e64c0ec302795ab9dc12bb1e73f150904699c153716177eebdaf", size = 17047890, upload-time = "2026-10-10T20:02:56.038Z" },
450
+ { url = "https://files.pythonhosted.org/packages/4a/7a/e8b58a5289a0d464c52885de47c35a935cdd70c03a4c3ab94a5126416dd0/numpy-2.5.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b89d0aaae2fe498c648f4c4795c084db535af5bd98ef942b2a3681fb74ce8645", size = 18485839, upload-time = "2026-10-10T20:02:59.018Z" },
451
+ { url = "https://files.pythonhosted.org/packages/6f/c9/47094f597015009f310b8c900def59065ef1ff5a6fe7b51fc65ec58ec2c6/numpy-2.5.4-cp312-cp312-win32.whl", hash = "sha256:9968ab7e49b93ac6e1c3b2239732183152c9150f16308d30b66a372cffe3483c", size = 6138936, upload-time = "2026-10-10T20:03:01.626Z" },
452
+ { url = "https://files.pythonhosted.org/packages/12/33/fefe62073dc8acfd0f2b9ed7c003af2f50aa61555e113e6db02b8f79f145/numpy-2.5.4-cp312-cp312-win_amd64.whl", hash = "sha256:a7b1b6353e36a7e50de2973a38d705c88ee93adcf120673cee7f45a4a3fa223a", size = 12573091, upload-time = "2026-10-10T20:03:04.349Z" },
453
+ { url = "https://files.pythonhosted.org/packages/1a/07/161270b0c2eec56e4c905f6d6d22e1b836887b2cb189d3f5820aa588e9dd/numpy-2.5.4-cp312-cp312-win_arm64.whl", hash = "sha256:aa1cce2ff3f8d953de38b76bf44602caeb69f101430208f64a10067f7cb4b1d3", size = 10521630, upload-time = "2026-10-10T20:03:06.767Z" },
454
+ ]
455
+
456
+ [[package]]
457
+ name = "nvidia-cublas"
458
+ version = "13.4.1.3"
459
+ source = { registry = "https://pypi.org/simple" }
460
+ dependencies = [
461
+ { name = "nvidia-cuda-nvrtc" },
462
+ ]
463
+ wheels = [
464
+ { url = "https://files.pythonhosted.org/packages/d0/a0/5cd9d4b08a8320f7ca150ee0d66c97e98092922e338336cb4b62da7500c3/nvidia_cublas-13.4.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:4eb1472bf92b741e7cf24da7978dde251b58ddec2092d9f6a04f3623eb4e9a15", size = 513048232, upload-time = "2026-07-16T09:45:07.751Z" },
465
+ { url = "https://files.pythonhosted.org/packages/e1/97/d483a41da34df5c88727c57319576967f6d7befba5028d70928d2999635e/nvidia_cublas-13.4.1.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:a5917cb642cfb07f4a0231ee16aec7e2e0357d8cb8b1b78a0c9c68dba3e6afc7", size = 404870805, upload-time = "2026-07-16T09:46:05.62Z" },
466
+ ]
467
+
468
+ [[package]]
469
+ name = "nvidia-cuda-cccl"
470
+ version = "13.3.5.0.6"
471
+ source = { registry = "https://pypi.org/simple" }
472
+ wheels = [
473
+ { url = "https://files.pythonhosted.org/packages/81/dc/9a056ef9bd05c55ddaed3afce160702e4ce8303b9ec1b38066044c18dcd2/nvidia_cuda_cccl-13.3.5.0.6-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:6f30b8076b3e69951c459c94ae190bde51924b320076ebaec376d36848b4c76e", size = 4113650, upload-time = "2026-11-04T23:21:46.803Z" },
474
+ { url = "https://files.pythonhosted.org/packages/a7/7e/4aa67cca14e0558b5a151a692f5ffe609930cdfc20d16c7a584de68424d4/nvidia_cuda_cccl-13.3.5.0.6-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d5230098fa2df5021e1a00c0c4eee41797671a91119d31e892463e5ccf3ee326", size = 4113651, upload-time = "2026-11-04T23:21:54.262Z" },
475
+ ]
476
+
477
+ [[package]]
478
+ name = "nvidia-cuda-cupti"
479
+ version = "13.2.86"
480
+ source = { registry = "https://pypi.org/simple" }
481
+ wheels = [
482
+ { url = "https://files.pythonhosted.org/packages/e1/ab/52414925ea803a5f86b994ae97d951e5943ed27a6f1da4a5c4ac09e308af/nvidia_cuda_cupti-13.2.86-py3-none-manylinux_2_25_aarch64.whl", hash = "sha256:d9a5c133f73fcb9013f3ada0da6fb54d4117618d81f3f5a0419417ea6ee829d4", size = 11760263, upload-time = "2026-07-16T09:35:34.643Z" },
483
+ { url = "https://files.pythonhosted.org/packages/b4/76/a10d9fb5d9a0e46d5f95e5520bd1b2d8ad22a9c0a09d2f8c20c5ee842a96/nvidia_cuda_cupti-13.2.86-py3-none-manylinux_2_25_x86_64.whl", hash = "sha256:8fa29f15dd8336181f961764d9d95236fc6b931dd0c8fcf8cd70aab0c310b586", size = 11988600, upload-time = "2026-07-16T09:36:06.061Z" },
484
+ ]
485
+
486
+ [[package]]
487
+ name = "nvidia-cuda-nvrtc"
488
+ version = "13.2.86"
489
+ source = { registry = "https://pypi.org/simple" }
490
+ wheels = [
491
+ { url = "https://files.pythonhosted.org/packages/1e/39/c58835ec8392a91b8d7bad83760eafc3085f64788b722ec2b44a27b3ca3e/nvidia_cuda_nvrtc-13.2.86-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:8111c5130e967b78222ea2124145f8efe9ce201b752b2579a2513100a53be1c5", size = 47018468, upload-time = "2026-07-16T09:40:35.085Z" },
492
+ { url = "https://files.pythonhosted.org/packages/13/95/029e91026e1d378cba9fb2cdb62394bb15d06de7586bab73d76397df6b7f/nvidia_cuda_nvrtc-13.2.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:1e5d7b411a2eb83f5c71c98d58133a763fc9fa86a4241ee320c5c4782d9ad1fd", size = 44755454, upload-time = "2026-07-16T09:39:59.909Z" },
493
+ ]
494
+
495
+ [[package]]
496
+ name = "nvidia-cuda-runtime"
497
+ version = "13.2.86"
498
+ source = { registry = "https://pypi.org/simple" }
499
+ wheels = [
500
+ { url = "https://files.pythonhosted.org/packages/89/d4/fa981071e3d9f29158ca16c69ee617831cd6df4dd47024149c16eec717f8/nvidia_cuda_runtime-13.2.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:37844957894bfdecf90243568d21a2cb251b66106915b71459a23c8e6ef2da69", size = 2341433, upload-time = "2026-07-16T09:33:00.116Z" },
501
+ { url = "https://files.pythonhosted.org/packages/70/c6/8f8c5e893d183082ea78f973aebc7ebc845ee219d68bddac73b0f62490fe/nvidia_cuda_runtime-13.2.86-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:3498247680dfddce31bad102af7e3e26fc795342a3cac64722bb0fa95e7f60f0", size = 2323576, upload-time = "2026-07-16T09:33:28.652Z" },
502
+ ]
503
+
504
+ [[package]]
505
+ name = "nvidia-cudnn-cu13"
506
+ version = "9.26.0.51"
507
+ source = { registry = "https://pypi.org/simple" }
508
+ dependencies = [
509
+ { name = "nvidia-cublas" },
510
+ ]
511
+ wheels = [
512
+ { url = "https://files.pythonhosted.org/packages/c5/1c/01e1dd3dec6ffe0baaff7e722b8946d867353d05a01301f128f1730fde0b/nvidia_cudnn_cu13-9.26.0.51-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:8e5b4f5eccff721416f70ecf4353172fc4d5cdd8ca869fe6f983a4a5b0147b20", size = 595329497, upload-time = "2026-09-10T16:45:39.945Z" },
513
+ { url = "https://files.pythonhosted.org/packages/ac/1b/ea72dd62f26ce3e0a7870b85148b75a5987507629b712deebeb87f7cce9a/nvidia_cudnn_cu13-9.26.0.51-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:9c976d539786698c71d6bcdbe2053b7485fae062ff7e1215d2f2ff67b1a2149a", size = 519150073, upload-time = "2026-09-10T16:47:58.332Z" },
514
+ ]
515
+
516
+ [[package]]
517
+ name = "nvidia-cufft"
518
+ version = "12.2.0.57"
519
+ source = { registry = "https://pypi.org/simple" }
520
+ dependencies = [
521
+ { name = "nvidia-nvjitlink" },
522
+ ]
523
+ wheels = [
524
+ { url = "https://files.pythonhosted.org/packages/07/8d/f650e6d527f9d22b5ded906c000eaebb037721c06d4ecd906df62f300979/nvidia_cufft-12.2.0.57-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:1ed438568ae1088ba2c25789d10205d4104b8b140b7bd942e39a787947a0fb74", size = 218244742, upload-time = "2026-07-16T09:47:08.76Z" },
525
+ { url = "https://files.pythonhosted.org/packages/41/28/868ed0c6ade5e6a75b25166b3cf191f4a2ff58ad618c1e40befb776c97c5/nvidia_cufft-12.2.0.57-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:34da27fadf3979b0b042c43823f0530be8408b86df31ee0108e1ca0b621d4ef6", size = 218258101, upload-time = "2026-07-16T09:47:50.978Z" },
526
+ ]
527
+
528
+ [[package]]
529
+ name = "nvidia-cufile"
530
+ version = "1.17.1.22"
531
+ source = { registry = "https://pypi.org/simple" }
532
+ wheels = [
533
+ { url = "https://files.pythonhosted.org/packages/fe/fd/4b4157512a131a51e1514d40d0199b1c5b7b10c6856db274ff22e5fe15be/nvidia_cufile-1.17.1.22-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:bec33d7f1f12691dce330275246587569c2ed90b1ce43b0f94b3e1ce9268744b", size = 1226736, upload-time = "2026-04-13T09:52:40.142Z" },
534
+ { url = "https://files.pythonhosted.org/packages/23/c5/661a73a9042fc5f81cdd5fe6e89f4170b3b33f7962e709987e2b5e5f1381/nvidia_cufile-1.17.1.22-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:87feaf8c3ad20d3d03a01933181801ff6892bace5c8f0b1e4b48f319d3e62d36", size = 1387670, upload-time = "2026-04-13T09:52:57.783Z" },
535
+ ]
536
+
537
+ [[package]]
538
+ name = "nvidia-cusolver"
539
+ version = "12.2.0.11"
540
+ source = { registry = "https://pypi.org/simple" }
541
+ dependencies = [
542
+ { name = "nvidia-cublas" },
543
+ { name = "nvidia-cusparse" },
544
+ { name = "nvidia-nvjitlink" },
545
+ ]
546
+ wheels = [
547
+ { url = "https://files.pythonhosted.org/packages/60/df/9e6b603f1d64040ce3e076e9bce1b78e2d588d0aabc9783f351cecd861ab/nvidia_cusolver-12.2.0.11-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:8be097fca28de6be0990ab216ee9b804dd51077e1d1d83acf6c9850650bdc64d", size = 250383075, upload-time = "2026-07-16T09:50:16.355Z" },
548
+ { url = "https://files.pythonhosted.org/packages/e4/a4/90f7ade9992190788658c3c285f196fdb9bcc3a5bb93a325c245121d22fd/nvidia_cusolver-12.2.0.11-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:fab5778f6769947389a1a6abc07e635576825a08c0ffc68c3ae7f1392c794b42", size = 227913465, upload-time = "2026-07-16T09:50:56.976Z" },
549
+ ]
550
+
551
+ [[package]]
552
+ name = "nvidia-cusparse"
553
+ version = "12.7.10.12"
554
+ source = { registry = "https://pypi.org/simple" }
555
+ dependencies = [
556
+ { name = "nvidia-nvjitlink" },
557
+ ]
558
+ wheels = [
559
+ { url = "https://files.pythonhosted.org/packages/60/35/abbd278fb6d039fcfb5175efa10a21007a11878bec59920d84854b47d766/nvidia_cusparse-12.7.10.12-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:597eff2cd83e3ab1d1e3a0dce9c9c879d77cf64ec2659bfdb5af6a5d02e7aaa9", size = 169043093, upload-time = "2026-07-16T09:51:37.058Z" },
560
+ { url = "https://files.pythonhosted.org/packages/bc/46/3a9700cc13f30c9279af9a862363535c814fa27597199b21fd93cad66237/nvidia_cusparse-12.7.10.12-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:25907c75e4db1581685f06f61cc6cdecc04611765516a52373e0a18ebd69fa00", size = 150999307, upload-time = "2026-07-16T09:52:17.727Z" },
561
+ ]
562
+
563
+ [[package]]
564
+ name = "nvidia-cusparselt-cu13"
565
+ version = "0.8.1"
566
+ source = { registry = "https://pypi.org/simple" }
567
+ wheels = [
568
+ { url = "https://files.pythonhosted.org/packages/46/e1/cdc1797eadf82d3a9a575a19b33fdc871a97edbec42c00b5b5e914f4aff4/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:4dca476c50bf4780d46cd0bfbd82e2bc10a08e4fef7950917ce8d7578d22a23f", size = 221051344, upload-time = "2025-09-05T18:49:51.289Z" },
569
+ { url = "https://files.pythonhosted.org/packages/34/7d/2661f2fb3ac4302f3a246f5fc030213ac60c1fe0bce84f9783dbd831dbb7/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:786ce87568c303fadb5afcc7102d454cd3040d75f6f8626f5db460d1871f4dd0", size = 170148586, upload-time = "2025-09-05T18:50:50.248Z" },
570
+ ]
571
+
572
+ [[package]]
573
+ name = "nvidia-nccl-cu13"
574
+ version = "2.30.7"
575
+ source = { registry = "https://pypi.org/simple" }
576
+ wheels = [
577
+ { url = "https://files.pythonhosted.org/packages/d1/21/a73174c6157101bdf1ffc22b517f76ff0082613989dd9bc8f43e8034caac/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:ca786ffa5a647c75d4d1f5cc72a6c4f537947e2ba8823d7c8aaf768e7a7b9f77", size = 215983881, upload-time = "2026-06-09T03:23:15.633Z" },
578
+ { url = "https://files.pythonhosted.org/packages/3f/34/c500f90c7ae641b8e0f98965b36b8a7ac79cc8b296e8d251fe3eb592ee54/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:cefa7fdb9710efd0f39c5f1be1d61ff6fc9a996c451265bd7fbdcf9455ed4b50", size = 215965170, upload-time = "2026-06-09T03:23:39.73Z" },
579
+ ]
580
+
581
+ [[package]]
582
+ name = "nvidia-nvjitlink"
583
+ version = "13.5.37"
584
+ source = { registry = "https://pypi.org/simple" }
585
+ wheels = [
586
+ { url = "https://files.pythonhosted.org/packages/cc/1f/2321702b8d60b4b4dd18b957f61612dbac9cb1c07e940cf6582106b7c8b8/nvidia_nvjitlink-13.5.37-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:280d7cb9649704145f074b8dbde500de04ce02d0872e066e20a83c17bcaa93eb", size = 46642491, upload-time = "2026-11-04T23:33:31.696Z" },
587
+ { url = "https://files.pythonhosted.org/packages/24/15/dd228f798cd01d7aeaf19d872b2b20f9aeafc1a7d8b6bfaa878b30b8d6f7/nvidia_nvjitlink-13.5.37-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:893f6f9a55b2720ed94913b5ba3332e972da8c78dfa2863d0244403dbd528ffc", size = 45253761, upload-time = "2026-11-04T23:33:22.854Z" },
588
+ ]
589
+
590
+ [[package]]
591
+ name = "nvidia-nvshmem-cu13"
592
+ version = "3.7.2"
593
+ source = { registry = "https://pypi.org/simple" }
594
+ dependencies = [
595
+ { name = "nvidia-cuda-cccl" },
596
+ ]
597
+ wheels = [
598
+ { url = "https://files.pythonhosted.org/packages/81/11/1767ef120299972822a43aceffbc81ac813aad2089d817e36e4e12aa4c9c/nvidia_nvshmem_cu13-3.7.2-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:0e748308b9e0a8a8d81859353f3e6c98fbe85a85f652a915a2ba60149a0bd124", size = 135174619, upload-time = "2026-07-17T19:24:15.404Z" },
599
+ { url = "https://files.pythonhosted.org/packages/4f/4a/037dfbbafc96ad1828ef62326f27970cae2c0eb271cf5b8e6922f32a7427/nvidia_nvshmem_cu13-3.7.2-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:b9a4f843c62b5dded8ad33fd6d220e007dcd39f06cee549292f5ba14ce4c235c", size = 135397797, upload-time = "2026-07-17T19:25:26.4Z" },
600
+ ]
601
+
602
+ [[package]]
603
+ name = "nvidia-nvtx"
604
+ version = "13.2.86"
605
+ source = { registry = "https://pypi.org/simple" }
606
+ wheels = [
607
+ { url = "https://files.pythonhosted.org/packages/49/ee/16bde09927e7ffacd50b117cd7d1dc764d145f03249fc41fdf6dfba3301f/nvidia_nvtx-13.2.86-py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:968ce72baee2a01791c7d3254000eac81f24677b0b91c62d21b449a1f309c245", size = 153463, upload-time = "2026-07-16T09:41:23.242Z" },
608
+ { url = "https://files.pythonhosted.org/packages/52/5f/74e2d29240de30a971d59ca20f4a3c4f2720cd49752809eeb4177d385c2c/nvidia_nvtx-13.2.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:1c6ae1be92fba569dd271f7b9b3056f99f5642726dc6c53d1b93ae4110c398fb", size = 154292, upload-time = "2026-07-16T09:40:53.791Z" },
609
+ ]
610
+
611
+ [[package]]
612
+ name = "packaging"
613
+ version = "26.3"
614
+ source = { registry = "https://pypi.org/simple" }
615
+ sdist = { url = "https://files.pythonhosted.org/packages/7d/fa/3944b40b07da9ce895c0e6303a5ab7d53da063554f534556b134a54d6093/packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79", size = 313412, upload-time = "2026-08-04T18:15:28.737Z" }
616
+ wheels = [
617
+ { url = "https://files.pythonhosted.org/packages/63/34/ba1c580383c9eada3711951fef0795c80b829a078d72188184bcab9dd527/packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c", size = 129956, upload-time = "2026-08-04T18:15:27.159Z" },
618
+ ]
619
+
620
+ [[package]]
621
+ name = "peft"
622
+ version = "0.21.2"
623
+ source = { registry = "https://pypi.org/simple" }
624
+ dependencies = [
625
+ { name = "accelerate" },
626
+ { name = "huggingface-hub" },
627
+ { name = "numpy" },
628
+ { name = "packaging" },
629
+ { name = "psutil" },
630
+ { name = "pyyaml" },
631
+ { name = "safetensors" },
632
+ { name = "torch" },
633
+ { name = "tqdm" },
634
+ { name = "transformers" },
635
+ ]
636
+ sdist = { url = "https://files.pythonhosted.org/packages/80/af/2e08abf1cd3b8792a02f5116808f2398c2f77ecdff801ced2ce16007a6f9/peft-0.21.2.tar.gz", hash = "sha256:b803ccfb3f3f316004d850284306687833a2235ea278fb56abc856203456142e", size = 983614, upload-time = "2026-10-01T10:26:36.129Z" }
637
+ wheels = [
638
+ { url = "https://files.pythonhosted.org/packages/70/0b/59441cdbfdd342ed03c08af90a2fe16173f0cd48ab219f523b39ca059a79/peft-0.21.2-py3-none-any.whl", hash = "sha256:106ab6077ff72c54d21577f9af5970e34bac7582cb14209f7b2b511e322a4eae", size = 835427, upload-time = "2026-10-01T10:26:34.085Z" },
639
+ ]
640
+
641
+ [[package]]
642
+ name = "psutil"
643
+ version = "7.2.2"
644
+ source = { registry = "https://pypi.org/simple" }
645
+ sdist = { url = "https://files.pythonhosted.org/packages/aa/c6/d1ddf4abb55e93cebc4f2ed8b5d6dbad109ecb8d63748dd2b20ab5e57ebe/psutil-7.2.2.tar.gz", hash = "sha256:0746f5f8d406af344fd547f1c8daa5f5c33dbc293bb8d6a16d80b4bb88f59372", size = 493740, upload-time = "2026-01-28T18:14:54.428Z" }
646
+ wheels = [
647
+ { url = "https://files.pythonhosted.org/packages/e7/36/5ee6e05c9bd427237b11b3937ad82bb8ad2752d72c6969314590dd0c2f6e/psutil-7.2.2-cp36-abi3-macosx_10_9_x86_64.whl", hash = "sha256:ed0cace939114f62738d808fdcecd4c869222507e266e574799e9c0faa17d486", size = 129090, upload-time = "2026-01-28T18:15:22.168Z" },
648
+ { url = "https://files.pythonhosted.org/packages/80/c4/f5af4c1ca8c1eeb2e92ccca14ce8effdeec651d5ab6053c589b074eda6e1/psutil-7.2.2-cp36-abi3-macosx_11_0_arm64.whl", hash = "sha256:1a7b04c10f32cc88ab39cbf606e117fd74721c831c98a27dc04578deb0c16979", size = 129859, upload-time = "2026-01-28T18:15:23.795Z" },
649
+ { url = "https://files.pythonhosted.org/packages/b5/70/5d8df3b09e25bce090399cf48e452d25c935ab72dad19406c77f4e828045/psutil-7.2.2-cp36-abi3-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:076a2d2f923fd4821644f5ba89f059523da90dc9014e85f8e45a5774ca5bc6f9", size = 155560, upload-time = "2026-01-28T18:15:25.976Z" },
650
+ { url = "https://files.pythonhosted.org/packages/63/65/37648c0c158dc222aba51c089eb3bdfa238e621674dc42d48706e639204f/psutil-7.2.2-cp36-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b0726cecd84f9474419d67252add4ac0cd9811b04d61123054b9fb6f57df6e9e", size = 156997, upload-time = "2026-01-28T18:15:27.794Z" },
651
+ { url = "https://files.pythonhosted.org/packages/8e/13/125093eadae863ce03c6ffdbae9929430d116a246ef69866dad94da3bfbc/psutil-7.2.2-cp36-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:fd04ef36b4a6d599bbdb225dd1d3f51e00105f6d48a28f006da7f9822f2606d8", size = 148972, upload-time = "2026-01-28T18:15:29.342Z" },
652
+ { url = "https://files.pythonhosted.org/packages/04/78/0acd37ca84ce3ddffaa92ef0f571e073faa6d8ff1f0559ab1272188ea2be/psutil-7.2.2-cp36-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:b58fabe35e80b264a4e3bb23e6b96f9e45a3df7fb7eed419ac0e5947c61e47cc", size = 148266, upload-time = "2026-01-28T18:15:31.597Z" },
653
+ { url = "https://files.pythonhosted.org/packages/b4/90/e2159492b5426be0c1fef7acba807a03511f97c5f86b3caeda6ad92351a7/psutil-7.2.2-cp37-abi3-win_amd64.whl", hash = "sha256:eb7e81434c8d223ec4a219b5fc1c47d0417b12be7ea866e24fb5ad6e84b3d988", size = 137737, upload-time = "2026-01-28T18:15:33.849Z" },
654
+ { url = "https://files.pythonhosted.org/packages/8c/c7/7bb2e321574b10df20cbde462a94e2b71d05f9bbda251ef27d104668306a/psutil-7.2.2-cp37-abi3-win_arm64.whl", hash = "sha256:8c233660f575a5a89e6d4cb65d9f938126312bca76d8fe087b947b3a1aaac9ee", size = 134617, upload-time = "2026-01-28T18:15:36.514Z" },
655
+ ]
656
+
657
+ [[package]]
658
+ name = "pygments"
659
+ version = "2.21.0"
660
+ source = { registry = "https://pypi.org/simple" }
661
+ sdist = { url = "https://files.pythonhosted.org/packages/49/2e/ced460408999b33da6b31b0021b0f37d329e202d4169aeb164493778f25b/pygments-2.21.0.tar.gz", hash = "sha256:610ca751c9bc2492b38eb9a38a7fbc93edbbb2d7182edaf34e66ae493dee5c8c", size = 5005329, upload-time = "2026-08-17T08:02:48.824Z" }
662
+ wheels = [
663
+ { url = "https://files.pythonhosted.org/packages/71/46/17f022dd3e953bf20a04a028a21ec746d942f8d2af30fa0f124fa0e6a684/pygments-2.21.0-py3-none-any.whl", hash = "sha256:2363c69b61c4a97c838da3b130dcd6468f4848992b21a82f2a63ec34377137d9", size = 1250147, upload-time = "2026-08-17T08:02:44.912Z" },
664
+ ]
665
+
666
+ [[package]]
667
+ name = "pyyaml"
668
+ version = "6.0.3"
669
+ source = { registry = "https://pypi.org/simple" }
670
+ sdist = { url = "https://files.pythonhosted.org/packages/05/8e/961c0007c59b8dd7729d542c61a4d537767a59645b82a0b521206e1e25c2/pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f", size = 130960, upload-time = "2025-09-25T21:33:16.546Z" }
671
+ wheels = [
672
+ { url = "https://files.pythonhosted.org/packages/d1/33/422b98d2195232ca1826284a76852ad5a86fe23e31b009c9886b2d0fb8b2/pyyaml-6.0.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7f047e29dcae44602496db43be01ad42fc6f1cc0d8cd6c83d342306c32270196", size = 182063, upload-time = "2025-09-25T21:32:11.445Z" },
673
+ { url = "https://files.pythonhosted.org/packages/89/a0/6cf41a19a1f2f3feab0e9c0b74134aa2ce6849093d5517a0c550fe37a648/pyyaml-6.0.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:fc09d0aa354569bc501d4e787133afc08552722d3ab34836a80547331bb5d4a0", size = 173973, upload-time = "2025-09-25T21:32:12.492Z" },
674
+ { url = "https://files.pythonhosted.org/packages/ed/23/7a778b6bd0b9a8039df8b1b1d80e2e2ad78aa04171592c8a5c43a56a6af4/pyyaml-6.0.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9149cad251584d5fb4981be1ecde53a1ca46c891a79788c0df828d2f166bda28", size = 775116, upload-time = "2025-09-25T21:32:13.652Z" },
675
+ { url = "https://files.pythonhosted.org/packages/65/30/d7353c338e12baef4ecc1b09e877c1970bd3382789c159b4f89d6a70dc09/pyyaml-6.0.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5fdec68f91a0c6739b380c83b951e2c72ac0197ace422360e6d5a959d8d97b2c", size = 844011, upload-time = "2025-09-25T21:32:15.21Z" },
676
+ { url = "https://files.pythonhosted.org/packages/8b/9d/b3589d3877982d4f2329302ef98a8026e7f4443c765c46cfecc8858c6b4b/pyyaml-6.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ba1cc08a7ccde2d2ec775841541641e4548226580ab850948cbfda66a1befcdc", size = 807870, upload-time = "2025-09-25T21:32:16.431Z" },
677
+ { url = "https://files.pythonhosted.org/packages/05/c0/b3be26a015601b822b97d9149ff8cb5ead58c66f981e04fedf4e762f4bd4/pyyaml-6.0.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:8dc52c23056b9ddd46818a57b78404882310fb473d63f17b07d5c40421e47f8e", size = 761089, upload-time = "2025-09-25T21:32:17.56Z" },
678
+ { url = "https://files.pythonhosted.org/packages/be/8e/98435a21d1d4b46590d5459a22d88128103f8da4c2d4cb8f14f2a96504e1/pyyaml-6.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:41715c910c881bc081f1e8872880d3c650acf13dfa8214bad49ed4cede7c34ea", size = 790181, upload-time = "2025-09-25T21:32:18.834Z" },
679
+ { url = "https://files.pythonhosted.org/packages/74/93/7baea19427dcfbe1e5a372d81473250b379f04b1bd3c4c5ff825e2327202/pyyaml-6.0.3-cp312-cp312-win32.whl", hash = "sha256:96b533f0e99f6579b3d4d4995707cf36df9100d67e0c8303a0c55b27b5f99bc5", size = 137658, upload-time = "2025-09-25T21:32:20.209Z" },
680
+ { url = "https://files.pythonhosted.org/packages/86/bf/899e81e4cce32febab4fb42bb97dcdf66bc135272882d1987881a4b519e9/pyyaml-6.0.3-cp312-cp312-win_amd64.whl", hash = "sha256:5fcd34e47f6e0b794d17de1b4ff496c00986e1c83f7ab2fb8fcfe9616ff7477b", size = 154003, upload-time = "2025-09-25T21:32:21.167Z" },
681
+ { url = "https://files.pythonhosted.org/packages/1a/08/67bd04656199bbb51dbed1439b7f27601dfb576fb864099c7ef0c3e55531/pyyaml-6.0.3-cp312-cp312-win_arm64.whl", hash = "sha256:64386e5e707d03a7e172c0701abfb7e10f0fb753ee1d773128192742712a98fd", size = 140344, upload-time = "2025-09-25T21:32:22.617Z" },
682
+ ]
683
+
684
+ [[package]]
685
+ name = "regex"
686
+ version = "2026.9.29"
687
+ source = { registry = "https://pypi.org/simple" }
688
+ sdist = { url = "https://files.pythonhosted.org/packages/fc/f2/af1da9d3ceed77bfcdce40427d49ba0be94e4fe84245e3bfef68c10e75b6/regex-2026.9.29.tar.gz", hash = "sha256:8b5fcc4771732191b2b7d1dd68d8f0353f47f8d90b6150f6dce58bf1112442cb", size = 419199, upload-time = "2026-09-29T00:49:58.298Z" }
689
+ wheels = [
690
+ { url = "https://files.pythonhosted.org/packages/84/48/3fdcde9a0baa84d7d25571223265d6e434e114763b438601d54a8028bf3e/regex-2026.9.29-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:dc79d36d0618752265f0d575915bdc5c5130ecb9c9f6b3bcefeae32e4bdfafcf", size = 497903, upload-time = "2026-09-29T00:46:38.938Z" },
691
+ { url = "https://files.pythonhosted.org/packages/2e/1c/4ee3e97c76f53940488dfe7a7e18705e78daac8cd7fb161d246b9e328449/regex-2026.9.29-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:3a21a9509d0ee88e7a70e1ad228cd2f0e0fd1e187458db132e8a8d18c97daf9d", size = 296416, upload-time = "2026-09-29T00:46:40.406Z" },
692
+ { url = "https://files.pythonhosted.org/packages/37/14/f3f0ba083d2094392d5eabf56db5ea6ba469fd6e927afd187042054ea68a/regex-2026.9.29-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f57dc6b8fef170f105d2cf5cdce254f47b137d7755086cf7050f47e16582abba", size = 293633, upload-time = "2026-09-29T00:46:41.959Z" },
693
+ { url = "https://files.pythonhosted.org/packages/c9/72/67e7a8ce17f1aea49df215564048efb49cc8c2b31a0e0fc30f36838f8516/regex-2026.9.29-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f93bc1c3486ef3747e07c9d7c1d0a147b8fbaab975f80e348aed6f71309dfaca", size = 805885, upload-time = "2026-09-29T00:46:43.373Z" },
694
+ { url = "https://files.pythonhosted.org/packages/f6/78/25436bcfd4d2260b4b4090094d55d7ab53ec8a1ab4865a0b8bcb33c7d5c0/regex-2026.9.29-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9e1d3a4cb7993b708f0ada8d0c84590efd853f169e7147d2202c9da503180242", size = 878344, upload-time = "2026-09-29T00:46:45.328Z" },
695
+ { url = "https://files.pythonhosted.org/packages/97/e6/a09ec3a23ae41d6179880e67f0aace9284b2d95f2d7b326eff203f8eec5e/regex-2026.9.29-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:dabee8f4935e731fb46b2a3091bdda0d3d94b3bbfb907d2b4f12eefce4009619", size = 919181, upload-time = "2026-09-29T00:46:47.041Z" },
696
+ { url = "https://files.pythonhosted.org/packages/26/83/d2fbd2e4e3afb1167daa825187d196f313cbaa1a4768f311fb041bb0e3d2/regex-2026.9.29-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:39ab5894d971f9ac68baa6eca5c50387db579cfcacf36ae8df3feceb1815e6d0", size = 807783, upload-time = "2026-09-29T00:46:48.894Z" },
697
+ { url = "https://files.pythonhosted.org/packages/46/0b/eb429a7016610d44fc89a597163f8c9127505f0d7dc724dc9effbb6a3ac0/regex-2026.9.29-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c1a9a6651197fbed6f0212591418b9def774fc3f8324f78d1bf0e6a63e5f8aa1", size = 783465, upload-time = "2026-09-29T00:46:50.64Z" },
698
+ { url = "https://files.pythonhosted.org/packages/1b/07/58a3c0153c7476898430f6a7cf3d9062a1d17fbea4f43399ecaf411c7b4c/regex-2026.9.29-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87fb80cbe3557e27e7b28b995c2b2eedf689b8886f941ab93e0e288f0976518a", size = 793519, upload-time = "2026-09-29T00:46:52.396Z" },
699
+ { url = "https://files.pythonhosted.org/packages/2a/e8/161b94d39164520e21a7befe0245569bf7fda4c7cf1fc4e2df2b5def49da/regex-2026.9.29-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:3c5c2ef13797466aa64170cbb66ad98a32351dd4127694cea7199f80f213750d", size = 869293, upload-time = "2026-09-29T00:46:54.128Z" },
700
+ { url = "https://files.pythonhosted.org/packages/8f/07/3b02ed829aa2decdc1955d222bd1e2f99d1c8bb4873bbb9a66b2f0a36bff/regex-2026.9.29-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:59b49507f47479e299a9e1bc41b5cb83a7afda0540625f1dbae886615978acbf", size = 770239, upload-time = "2026-09-29T00:46:56.106Z" },
701
+ { url = "https://files.pythonhosted.org/packages/42/5b/ba61f6fe062eb8562e742367d177bb75370434138ef6c9d2a27114f8d613/regex-2026.9.29-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:0dd8af32e9f7b56b7f95cc1fd79b23054c3bdc172392ae560acc24d57b7ffe71", size = 861973, upload-time = "2026-09-29T00:46:57.665Z" },
702
+ { url = "https://files.pythonhosted.org/packages/cc/27/767259b20e8a842948990f5e99138d6c077248fd42f8b5468b1d9ca4b814/regex-2026.9.29-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:db5e82ba15c142425b8406690032df89e39cca4a2e8afbbb9a3d84edc2373ac3", size = 796470, upload-time = "2026-09-29T00:46:59.236Z" },
703
+ { url = "https://files.pythonhosted.org/packages/a0/05/2566c4ba849b68a8ab81a6bf428fa79d20aae7ddee83979103c0381df254/regex-2026.9.29-cp312-cp312-win32.whl", hash = "sha256:d0c3082bf79bcd6a614d55916590ad4b8f93200e10b97f463ea5d9d07c9b5f23", size = 269327, upload-time = "2026-09-29T00:47:01.135Z" },
704
+ { url = "https://files.pythonhosted.org/packages/93/19/489bc8db91196381c935752df01ba3f607140daece33b78d88573f028e64/regex-2026.9.29-cp312-cp312-win_amd64.whl", hash = "sha256:fdd88ed5e20b1bcdd234421e454962c971aa44b653bdb7f1ea9ef683e90fb649", size = 280334, upload-time = "2026-09-29T00:47:04.436Z" },
705
+ { url = "https://files.pythonhosted.org/packages/0b/47/fb88ba779d0e5e7d4b0ec1aceeb13845948a2cb876bd572a2d1dfdba090b/regex-2026.9.29-cp312-cp312-win_arm64.whl", hash = "sha256:4fe97894d1b306c919b4e50def1e6f6c522f4d03a7283811f4d108f1ce5d3ac2", size = 279614, upload-time = "2026-09-29T00:47:06.541Z" },
706
+ ]
707
+
708
+ [[package]]
709
+ name = "requests"
710
+ version = "2.34.2"
711
+ source = { registry = "https://pypi.org/simple" }
712
+ dependencies = [
713
+ { name = "certifi" },
714
+ { name = "charset-normalizer" },
715
+ { name = "idna" },
716
+ { name = "urllib3" },
717
+ ]
718
+ sdist = { url = "https://files.pythonhosted.org/packages/ac/c3/e2a2b89f2d3e2179abd6d00ebd70bff6273f37fb3e0cc209f48b39d00cbf/requests-2.34.2.tar.gz", hash = "sha256:f288924cae4e29463698d6d60bc6a4da69c89185ad1e0bcc4104f584e960b9ed", size = 142856, upload-time = "2026-05-14T19:25:27.735Z" }
719
+ wheels = [
720
+ { url = "https://files.pythonhosted.org/packages/a0/f4/c67b0b3f1b9245e8d266f0f112c500d50e5b4e83cb6f3b71b6528104182a/requests-2.34.2-py3-none-any.whl", hash = "sha256:2a0d60c172f83ac6ab31e4554906c0f3b3588d37b5cb939b1c061f4907e278e0", size = 73075, upload-time = "2026-05-14T19:25:26.443Z" },
721
+ ]
722
+
723
+ [[package]]
724
+ name = "rich"
725
+ version = "15.0.0"
726
+ source = { registry = "https://pypi.org/simple" }
727
+ dependencies = [
728
+ { name = "markdown-it-py" },
729
+ { name = "pygments" },
730
+ ]
731
+ sdist = { url = "https://files.pythonhosted.org/packages/c0/8f/0722ca900cc807c13a6a0c696dacf35430f72e0ec571c4275d2371fca3e9/rich-15.0.0.tar.gz", hash = "sha256:edd07a4824c6b40189fb7ac9bc4c52536e9780fbbfbddf6f1e2502c31b068c36", size = 230680, upload-time = "2026-04-12T08:24:00.75Z" }
732
+ wheels = [
733
+ { url = "https://files.pythonhosted.org/packages/82/3b/64d4899d73f91ba49a8c18a8ff3f0ea8f1c1d75481760df8c68ef5235bf5/rich-15.0.0-py3-none-any.whl", hash = "sha256:33bd4ef74232fb73fe9279a257718407f169c09b78a87ad3d296f548e27de0bb", size = 310654, upload-time = "2026-04-12T08:24:02.83Z" },
734
+ ]
735
+
736
+ [[package]]
737
+ name = "safetensors"
738
+ version = "0.8.0"
739
+ source = { registry = "https://pypi.org/simple" }
740
+ sdist = { url = "https://files.pythonhosted.org/packages/45/06/f955dbbb1859e3bd23c8ac6141af5106e7ad5fedec4a3a6e3d60f94b7001/safetensors-0.8.0.tar.gz", hash = "sha256:fabaf3e0f18a6618d9b36560682562157f77c2b71fcffc7b432be2baed9d753d", size = 325846, upload-time = "2026-06-09T07:52:25.563Z" }
741
+ wheels = [
742
+ { url = "https://files.pythonhosted.org/packages/39/a0/f718cda65b05407d228f97602cf60dca269c979867aa5beb25410de26cd3/safetensors-0.8.0-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c554f85858e05226d3c2828e32395e677434685d6d94594a41643361c5e837f0", size = 473568, upload-time = "2026-06-09T07:52:18.829Z" },
743
+ { url = "https://files.pythonhosted.org/packages/f5/b1/fa7c600e7dceae12e9606c7578cbc9ff1e1ed55844883ee5c92205e86226/safetensors-0.8.0-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:c80201d22cbf405b80647a60ada77bba06c8fba2da2743ba1e89cdcc39a81f25", size = 484562, upload-time = "2026-06-09T07:52:17.518Z" },
744
+ { url = "https://files.pythonhosted.org/packages/09/7d/65a7de0af421317bb36a067241e4235fff194eed60b961ed6d3f59a3fc60/safetensors-0.8.0-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7a46e5ff292c356d6991e60942ba7f79817682d3a2cef0702136448cb9c4d235", size = 502844, upload-time = "2026-06-09T07:52:07.624Z" },
745
+ { url = "https://files.pythonhosted.org/packages/91/4f/3175c9d75634e0e0dda0082794193521035edd7c70a6f212bf33ca06ddf4/safetensors-0.8.0-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4124502b78f03534117c848f87a39b8f31e577b15eff423bf8bfb95f2a8c30d0", size = 511823, upload-time = "2026-06-09T07:52:09.565Z" },
746
+ { url = "https://files.pythonhosted.org/packages/20/87/846c289e7aa2299eff406335717cf43ce8777194ece8aad75772e0411615/safetensors-0.8.0-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7bc0a787ba8a35be368ee3574edfa2b1ad389eebd0a72e482ae275490e3f6c98", size = 633461, upload-time = "2026-06-09T07:52:11.128Z" },
747
+ { url = "https://files.pythonhosted.org/packages/76/22/8d64d9df2c45d5ded401df889d0ad90882804ca172d79ec4f0df8f727fe0/safetensors-0.8.0-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:040070828e36dc8e122178bbbd5830ff9e97920affb84cbe0f46442497bed358", size = 545148, upload-time = "2026-06-09T07:52:13.603Z" },
748
+ { url = "https://files.pythonhosted.org/packages/28/50/f203ff3a3ddfe19308efc83c5a3a29ed02bf786732ec35e68bf9162f3365/safetensors-0.8.0-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fd6f3f93c9a0a7cc2788ee63fb763353d4bd2e89b0751bc78fcf7dda00bea774", size = 516040, upload-time = "2026-06-09T07:52:16.29Z" },
749
+ { url = "https://files.pythonhosted.org/packages/46/fb/cdaed17ceb2948784fd9c36b6fd3e951b608547cea81a48e8ee6f8cfdfcb/safetensors-0.8.0-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:fcdd41ec4628fee5799f807c73c353629130fbd942aa23d83c623dd6c9d52d78", size = 513832, upload-time = "2026-06-09T07:52:12.37Z" },
750
+ { url = "https://files.pythonhosted.org/packages/0d/49/1e15de264dcc3b77943d2d0c56a95809956883b1c2d6d585c792523f180b/safetensors-0.8.0-cp310-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8e9f537aa183a38ace122d27303dcd986b26bd2a7591f9181d7f0c396f4677ca", size = 559930, upload-time = "2026-06-09T07:52:14.743Z" },
751
+ { url = "https://files.pythonhosted.org/packages/2a/43/bf38443278eab4b1be1fce2931e2b012ad9cb7df52ada751d0aab8f7659a/safetensors-0.8.0-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:87eec7ffed2b809f05a398a8becb7d013f19f7837cd15d9748580d6cf30dbaf4", size = 678670, upload-time = "2026-06-09T07:52:20.032Z" },
752
+ { url = "https://files.pythonhosted.org/packages/72/e3/68cd3fa5b48488e84add63e04cb12f3bc28ae4638c06d4508c6e88823d0e/safetensors-0.8.0-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:4a95ae2b05d7726d751da4ebf626a2ca782b706e101bd894c95bc2450b1cffcc", size = 786679, upload-time = "2026-06-09T07:52:21.322Z" },
753
+ { url = "https://files.pythonhosted.org/packages/29/4b/1c19c509d56e01f4fbb3d0a2e597450f6cc04d1d56cf52defb0a62dfd715/safetensors-0.8.0-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:3ae091f16662658bdc019a4ff6cb4c085bb7d725eb5978b183ffd265863b6d2d", size = 765683, upload-time = "2026-06-09T07:52:22.594Z" },
754
+ { url = "https://files.pythonhosted.org/packages/27/43/41c1621732edd934d868a00d1b891584c892a7b62a9aab82ea5a0a5623ee/safetensors-0.8.0-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8e080062fcde23be189565e1c3305d16751a218ecf9412c8601e64204eb6f846", size = 722361, upload-time = "2026-06-09T07:52:23.924Z" },
755
+ { url = "https://files.pythonhosted.org/packages/8e/3f/73ccf82579412b4a71c4ca673f10b5f1f888d7cf5af7fe24f27d30307be4/safetensors-0.8.0-cp310-abi3-win32.whl", hash = "sha256:2ddf52eac562eda224f99acfa7889d02968c1fd59a5b011ae7d8137c37e9c02d", size = 342401, upload-time = "2026-06-09T07:52:28.895Z" },
756
+ { url = "https://files.pythonhosted.org/packages/1b/6d/3fba214c1e5e0f69991677ec3bc17023f0421776975e1de0c682dca475e2/safetensors-0.8.0-cp310-abi3-win_amd64.whl", hash = "sha256:096ec1a98435df7beb08853bb5aa9081a84f23d0adc67ed1a0a10550f608373f", size = 355540, upload-time = "2026-06-09T07:52:27.832Z" },
757
+ { url = "https://files.pythonhosted.org/packages/8d/fc/7eedc3510d97878876e32774eebbeb61c43f148a96e915c84229a3e967aa/safetensors-0.8.0-cp310-abi3-win_arm64.whl", hash = "sha256:f7838e5135a406ad3e02efdcb8cf2e5397d368b0154537c4fec682dbc544d452", size = 340500, upload-time = "2026-06-09T07:52:26.745Z" },
758
+ ]
759
+
760
+ [[package]]
761
+ name = "sentencepiece"
762
+ version = "0.2.3"
763
+ source = { registry = "https://pypi.org/simple" }
764
+ sdist = { url = "https://files.pythonhosted.org/packages/54/83/0664683546ffca8e36653d39d6217c361d931953f40ce28210b56927a9b6/sentencepiece-0.2.3.tar.gz", hash = "sha256:21618ad31f0453a1ebf5ec58defed99758d54af0b670546ecac639ddd8b86fbd", size = 3751846, upload-time = "2026-11-05T05:38:22.108Z" }
765
+ wheels = [
766
+ { url = "https://files.pythonhosted.org/packages/f8/a9/8dbb891f51a80f1f47afa3d5336b165e269c4fa87f47b7fba6fc34a4a6a0/sentencepiece-0.2.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:9b39b16b889c9d3be4f00f1a0a4283ee81ad8af745d48023e5ab18c724019c08", size = 1426879, upload-time = "2026-11-05T05:36:08.906Z" },
767
+ { url = "https://files.pythonhosted.org/packages/31/26/fb6572cc96090953051089f8e4260a455b6b0c7378bb06fb87725224ce61/sentencepiece-0.2.3-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a6b735a8027b998c163e540fdde66ac0bb95fc471f6f4165c8df35fb45385116", size = 1627109, upload-time = "2026-11-05T05:36:13.057Z" },
768
+ { url = "https://files.pythonhosted.org/packages/51/72/42952e5b9f254aa2952c04360b38ed614ca4f5600ee2b0cf8e1c0f6661c5/sentencepiece-0.2.3-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:cc3189b79c1ffda767f71b04238c27ac746f3b7abc263eab9a41c4cf46aee49e", size = 1716727, upload-time = "2026-11-05T05:36:17.199Z" },
769
+ { url = "https://files.pythonhosted.org/packages/b5/78/47be22c27976afed9f88cb9bcdb225abb37991c1ff442a9234e64e7a3d23/sentencepiece-0.2.3-cp312-cp312-win_amd64.whl", hash = "sha256:972af959c63cde446210a8935524806820c253afa4b0a97d83c206b857085f44", size = 1655435, upload-time = "2026-11-05T05:36:21.633Z" },
770
+ { url = "https://files.pythonhosted.org/packages/e4/5c/6f921049ef4136e360437d370396ac36b927ce1be0184ddbeca50581580d/sentencepiece-0.2.3-cp312-cp312-win_arm64.whl", hash = "sha256:ceaf9a73f7e1672990547e8efe5bb79ea61884a8fe24600662d12e3a8aaf1a5b", size = 1860800, upload-time = "2026-11-05T05:36:26.727Z" },
771
+ ]
772
+
773
+ [[package]]
774
+ name = "setuptools"
775
+ version = "84.2.1"
776
+ source = { registry = "https://pypi.org/simple" }
777
+ sdist = { url = "https://files.pythonhosted.org/packages/70/b3/c19426f783614d625699a4b247faf628ea9ab1db228674491767b417f9e1/setuptools-84.2.1.tar.gz", hash = "sha256:3ff8b3b5492a0aa66960f4419365a86b84e013070d4313e4d4a6894f5d64d6bc", size = 1230289, upload-time = "2026-10-20T19:01:37.742Z" }
778
+ wheels = [
779
+ { url = "https://files.pythonhosted.org/packages/ef/55/6dc6d6bb953bdfc52519a0059592c52b6e3b8ad51cf8256dffceb55b0610/setuptools-84.2.1-py3-none-any.whl", hash = "sha256:ca140b45d5325a9478f880e351484f62fef15f336462e449ca0c03c873d589c2", size = 876263, upload-time = "2026-10-20T19:01:35.678Z" },
780
+ ]
781
+
782
+ [[package]]
783
+ name = "shellingham"
784
+ version = "1.5.4"
785
+ source = { registry = "https://pypi.org/simple" }
786
+ sdist = { url = "https://files.pythonhosted.org/packages/58/15/8b3609fd3830ef7b27b655beb4b4e9c62313a4e8da8c676e142cc210d58e/shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de", size = 10310, upload-time = "2023-10-24T04:13:40.426Z" }
787
+ wheels = [
788
+ { url = "https://files.pythonhosted.org/packages/e0/f9/0595336914c5619e5f28a1fb793285925a8cd4b432c9da0a987836c7f822/shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686", size = 9755, upload-time = "2023-10-24T04:13:38.866Z" },
789
+ ]
790
+
791
+ [[package]]
792
+ name = "simplixity-router-training"
793
+ version = "0.1.0"
794
+ source = { virtual = "." }
795
+ dependencies = [
796
+ { name = "accelerate" },
797
+ { name = "bitsandbytes" },
798
+ { name = "flash-linear-attention" },
799
+ { name = "gguf" },
800
+ { name = "huggingface-hub" },
801
+ { name = "numpy" },
802
+ { name = "peft" },
803
+ { name = "safetensors" },
804
+ { name = "sentencepiece" },
805
+ { name = "torch" },
806
+ { name = "transformers" },
807
+ ]
808
+
809
+ [package.metadata]
810
+ requires-dist = [
811
+ { name = "accelerate" },
812
+ { name = "bitsandbytes", specifier = ">=0.50.2" },
813
+ { name = "flash-linear-attention" },
814
+ { name = "gguf", directory = "/home/eric/models/llama.cpp/gguf-py" },
815
+ { name = "huggingface-hub" },
816
+ { name = "numpy", specifier = ">=2.5.4" },
817
+ { name = "peft" },
818
+ { name = "safetensors" },
819
+ { name = "sentencepiece", specifier = ">=0.2.3" },
820
+ { name = "torch" },
821
+ { name = "transformers", specifier = ">=5" },
822
+ ]
823
+
824
+ [[package]]
825
+ name = "sympy"
826
+ version = "1.14.0"
827
+ source = { registry = "https://pypi.org/simple" }
828
+ dependencies = [
829
+ { name = "mpmath" },
830
+ ]
831
+ sdist = { url = "https://files.pythonhosted.org/packages/83/d3/803453b36afefb7c2bb238361cd4ae6125a569b4db67cd9e79846ba2d68c/sympy-1.14.0.tar.gz", hash = "sha256:d3d3fe8df1e5a0b42f0e7bdf50541697dbe7d23746e894990c030e2b05e72517", size = 7793921, upload-time = "2025-04-27T18:05:01.611Z" }
832
+ wheels = [
833
+ { url = "https://files.pythonhosted.org/packages/a2/09/77d55d46fd61b4a135c444fc97158ef34a095e5681d0a6c10b75bf356191/sympy-1.14.0-py3-none-any.whl", hash = "sha256:e091cc3e99d2141a0ba2847328f5479b05d94a6635cb96148ccb3f34671bd8f5", size = 6299353, upload-time = "2025-04-27T18:04:59.103Z" },
834
+ ]
835
+
836
+ [[package]]
837
+ name = "tokenizers"
838
+ version = "0.23.3"
839
+ source = { registry = "https://pypi.org/simple" }
840
+ dependencies = [
841
+ { name = "huggingface-hub" },
842
+ ]
843
+ sdist = { url = "https://files.pythonhosted.org/packages/e0/7c/2cabb2174e772636683008f2c5621949b645da7d303c596589e84516a184/tokenizers-0.23.3.tar.gz", hash = "sha256:cded33237c77caeef62944d32aa9a7ef42bdce2b3497e18d137e072a8c4be438", size = 385286, upload-time = "2026-10-09T10:16:55.759Z" }
844
+ wheels = [
845
+ { url = "https://files.pythonhosted.org/packages/aa/2e/4ce5b9716f26e526eff6b0502ebed4ea8d7161f03b3c77617c9f25528e97/tokenizers-0.23.3-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:9d2b5c97daf61688c2ad1803ca851800feaba50fb68d5821779e9ea5880d968c", size = 3148800, upload-time = "2026-10-09T10:00:51.457Z" },
846
+ { url = "https://files.pythonhosted.org/packages/b2/72/01e49f032bb346e5aaf06c10c74fe8aeec847173adbadd66eb7c53054bf2/tokenizers-0.23.3-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:68649e97d5b43c44c031d8d848874a6eecae8f8fe40ea989aa777a5a83aca716", size = 3101381, upload-time = "2026-10-09T10:00:54.063Z" },
847
+ { url = "https://files.pythonhosted.org/packages/15/fc/ae987741829b1cd547668c4c94be732ae3eefd1d74344e64c3d2ca714acd/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ec82e80e65a862275b97c3d90b7a523df8d9519ee48aeb4e9625b2cc909274e0", size = 3519944, upload-time = "2026-10-09T10:00:55.885Z" },
848
+ { url = "https://files.pythonhosted.org/packages/1c/da/cc8f6c030afaf05fbddc608158fbb761dca46913cbeba6b112e59fc82e2a/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c64a0713180ff16829d4e7f39a658b77ea11443af4e1aa46523692943c9b1414", size = 3397695, upload-time = "2026-10-09T10:00:57.444Z" },
849
+ { url = "https://files.pythonhosted.org/packages/ec/f1/256f78d1365fa2cd3ea6db716883d74667c8cbb6a21f15fa5b89a773cdc2/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ddedfd4b3b4be6be24ff6ca645c4a37fddfd305f6f3e354c54cf10b715c48215", size = 3753125, upload-time = "2026-10-09T10:01:00.165Z" },
850
+ { url = "https://files.pythonhosted.org/packages/60/93/eee007ac2fcbf4ecfce7fbc354826cf3611f56bdb886f3e91b1f7dd06b8f/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2a89614730d7b80940a5d2ed9320e1ec8add5a745c6151d8d05071b7215505b6", size = 4018598, upload-time = "2026-10-09T10:01:02.05Z" },
851
+ { url = "https://files.pythonhosted.org/packages/bf/f9/0c96c4739461fce9d8d865b416728081bf6230022d7163bd6244f35f4b31/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e88646b8580c5ad7f4361477f1298e9cc01771a1ee9aecfe32c47b8ff614cc38", size = 3602442, upload-time = "2026-10-09T10:01:03.77Z" },
852
+ { url = "https://files.pythonhosted.org/packages/3a/40/6706b82693715581457c6d5423eaa7faae576bb0526c5738a57085eb4449/tokenizers-0.23.3-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:376851d22bcf9d650a5c3090bb83e6cf9e895fbf0595369fa4cd43c1f69b5f87", size = 3396193, upload-time = "2026-10-09T10:01:05.48Z" },
853
+ { url = "https://files.pythonhosted.org/packages/fe/0c/85946de40e25b7364b8f1bcf56def129069acd5bb364b7c86a32919e1a23/tokenizers-0.23.3-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:bf501c40b72d2d5c8623620210430e9cac1ce47a46e45b34107b70a1557d46b0", size = 3553483, upload-time = "2026-10-09T10:01:07.387Z" },
854
+ { url = "https://files.pythonhosted.org/packages/f1/6b/8d615d92cad1d511ca5ab188d1c7c167f0b3d295cc0d96207f9f82d486d8/tokenizers-0.23.3-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:114e2b55ed177179d59f4ab98200a4471e11e78f9e4b5a922d146740f96fcf52", size = 9972248, upload-time = "2026-10-09T10:01:09.437Z" },
855
+ { url = "https://files.pythonhosted.org/packages/c9/7d/a922e37ddd58d1b463bbc2ad08120c8f59c60b814cd353519a116b24f8ba/tokenizers-0.23.3-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:d3407fb7b9c4d75dd68850ffd7180bc0a5d2dbaf0762d888e612f31fec3f9c6b", size = 9802957, upload-time = "2026-10-09T10:01:11.869Z" },
856
+ { url = "https://files.pythonhosted.org/packages/4b/06/5d3f506a86ae0699a0e4ea05c05978f9aee169ef2c1d844e68c971cf8194/tokenizers-0.23.3-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:84513ef0aeb8bf8f4ea11a2e8a7ac163ec5288aa115e649a59b470ac5c3107df", size = 10145487, upload-time = "2026-10-09T10:01:14.268Z" },
857
+ { url = "https://files.pythonhosted.org/packages/26/e5/065625317690ea3548d834dad81f48ea1fd32e4964610e658e195d7fe28e/tokenizers-0.23.3-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:e05ab7baf7f47b406a95fea6f3b0a484b2ddcd9e1d14b68844c457eb755085a3", size = 10266026, upload-time = "2026-10-09T10:16:33.054Z" },
858
+ { url = "https://files.pythonhosted.org/packages/77/4e/babede85d0d19f5e3deeef0063e01848141329934d3d77c31b5cab5ac2b4/tokenizers-0.23.3-cp310-abi3-win32.whl", hash = "sha256:1ebf28794e7e4954e20a7f70fbea410b2d1f0418f7dbbca97ca384fcfef38c25", size = 2588086, upload-time = "2026-10-09T10:16:35.686Z" },
859
+ { url = "https://files.pythonhosted.org/packages/d1/6c/24f074c9a0efb98e61b20aafe6b2641922d5db24e447d5d6daffd9e17555/tokenizers-0.23.3-cp310-abi3-win_amd64.whl", hash = "sha256:1f0823bb00c5fdc98e487354d54dd55a03848d61a1a0bf29a68c77f24f3b26c3", size = 2872101, upload-time = "2026-10-09T10:16:37.533Z" },
860
+ { url = "https://files.pythonhosted.org/packages/53/77/a476b6f73a661c11d113a342d2326b91506cf2285f0995d1212a6bb2022d/tokenizers-0.23.3-cp310-abi3-win_arm64.whl", hash = "sha256:7e48734d2de9260d86f03ab056d2cfeeff3869f61dbd49aaa15a2793b5f3458b", size = 2742580, upload-time = "2026-10-09T10:16:39.244Z" },
861
+ ]
862
+
863
+ [[package]]
864
+ name = "torch"
865
+ version = "2.15.0"
866
+ source = { registry = "https://pypi.org/simple" }
867
+ dependencies = [
868
+ { name = "cuda-bindings", marker = "sys_platform == 'linux'" },
869
+ { name = "cuda-toolkit", extra = ["cublas", "cudart", "cufft", "cufile", "cupti", "cusolver", "cusparse", "nvjitlink", "nvrtc", "nvtx"], marker = "sys_platform == 'linux'" },
870
+ { name = "filelock" },
871
+ { name = "fsspec" },
872
+ { name = "jinja2" },
873
+ { name = "networkx" },
874
+ { name = "nvidia-cudnn-cu13", marker = "sys_platform == 'linux'" },
875
+ { name = "nvidia-cusparselt-cu13", marker = "sys_platform == 'linux'" },
876
+ { name = "nvidia-nccl-cu13", marker = "sys_platform == 'linux'" },
877
+ { name = "nvidia-nvshmem-cu13", marker = "sys_platform == 'linux'" },
878
+ { name = "setuptools" },
879
+ { name = "sympy" },
880
+ { name = "triton", marker = "sys_platform == 'linux'" },
881
+ { name = "typing-extensions" },
882
+ ]
883
+ wheels = [
884
+ { url = "https://files.pythonhosted.org/packages/17/24/7b9b755e8c4dd2395c6a94bab6a4ed8aded1735ff2538bc63cbba8abeddd/torch-2.15.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:593e5fd0c204284890d1c3dff643be34e549e1d24523e13402510f960d4da438", size = 138558517, upload-time = "2026-10-28T15:16:35.918Z" },
885
+ { url = "https://files.pythonhosted.org/packages/56/c9/ac253c63f13212393483bfa6c180b684c8e5437960c06ec9292797e09357/torch-2.15.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:9e51cf9460c5441dcd2a4f31ece207e35ea7c3d2535be0f929cca4944e2eeb18", size = 483664949, upload-time = "2026-10-28T15:17:27.418Z" },
886
+ { url = "https://files.pythonhosted.org/packages/42/43/612036bf387d34afff1ec77d3aabe0d16776f1727b00b8bc4540fe5472b1/torch-2.15.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:6d8c86f4e33a9f12a58149bd9b0d64351632b8eb44479be8ff85e5389932dd79", size = 539172785, upload-time = "2026-10-28T15:17:47.95Z" },
887
+ { url = "https://files.pythonhosted.org/packages/74/0b/a57c0b58a64b7a197dc40fc464bb96c0de867f95a2b1cbb3ec9e54414fde/torch-2.15.0-cp312-cp312-win_amd64.whl", hash = "sha256:f2bd0a3709a7493a6ef2ac5a654111242a854439f7558f910b6b653b56f7504d", size = 126738207, upload-time = "2026-10-28T15:16:40.586Z" },
888
+ ]
889
+
890
+ [[package]]
891
+ name = "tqdm"
892
+ version = "4.70.1"
893
+ source = { registry = "https://pypi.org/simple" }
894
+ dependencies = [
895
+ { name = "colorama", marker = "sys_platform == 'win32'" },
896
+ ]
897
+ sdist = { url = "https://files.pythonhosted.org/packages/0d/ea/b2a5bd54b28a324dae8211928b2d730b6547500342c7e6c6dea08bd0a485/tqdm-4.70.1.tar.gz", hash = "sha256:cefd0eca11b2a37a3aee776544d4f4ae913f02688135b5556b8788dfa474afc4", size = 171846, upload-time = "2026-09-11T07:25:16.601Z" }
898
+ wheels = [
899
+ { url = "https://files.pythonhosted.org/packages/a7/03/921a3d3c75785aca9ebfbfcabfbc3a1be12e2ab5265deb026d55a5a3f83e/tqdm-4.70.1-py3-none-any.whl", hash = "sha256:c293e525e6fef9c20e8728fd4612df02a0aa31bb5fe91ecd93e123b1b7bffa73", size = 80199, upload-time = "2026-09-11T07:25:14.599Z" },
900
+ ]
901
+
902
+ [[package]]
903
+ name = "transformers"
904
+ version = "5.22.1"
905
+ source = { registry = "https://pypi.org/simple" }
906
+ dependencies = [
907
+ { name = "huggingface-hub" },
908
+ { name = "numpy" },
909
+ { name = "packaging" },
910
+ { name = "pyyaml" },
911
+ { name = "regex" },
912
+ { name = "safetensors" },
913
+ { name = "tokenizers" },
914
+ { name = "tqdm" },
915
+ { name = "typer" },
916
+ ]
917
+ sdist = { url = "https://files.pythonhosted.org/packages/17/d6/a610b37bbc3450c056ea68c7f5016d06da8e787c1ac32b20ce58d18a1eb0/transformers-5.22.1.tar.gz", hash = "sha256:bed9fd76d6f43b4a1a9fc5d0c565ace7c77edbd029784f8cbcdf772cb5e924d8", size = 10381794, upload-time = "2026-10-30T07:48:03.721Z" }
918
+ wheels = [
919
+ { url = "https://files.pythonhosted.org/packages/75/75/47ebe0e2f1b04a180000e3a49ebdf4669dbab0c63489bcecf7855c26e4f4/transformers-5.22.1-py3-none-any.whl", hash = "sha256:c940b4bf98de4f568723895f330ba1460a398bd83061b93b70238c3fe32e4389", size = 13051755, upload-time = "2026-10-30T07:47:59.888Z" },
920
+ ]
921
+
922
+ [[package]]
923
+ name = "triton"
924
+ version = "3.9.0"
925
+ source = { registry = "https://pypi.org/simple" }
926
+ wheels = [
927
+ { url = "https://files.pythonhosted.org/packages/c2/7f/79490a3e16cbda1817f00ad6d05694020f1decb1e063f96e83d1cfd2558d/triton-3.9.0-cp312-abi3-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5542f4328472fd827dbb1ec43873364c2894682cc3f5d1c74a643762900e243e", size = 263270508, upload-time = "2026-10-27T18:39:01.187Z" },
928
+ { url = "https://files.pythonhosted.org/packages/3f/c9/af85b25e0d98bd921a7646ad1db1c8746250ee04b8f1595982d29b6283cf/triton-3.9.0-cp312-abi3-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5e52ab4e2f1d64f0ac494c9aa723697b793f81a91c4cae4c36608ca0bc4267bf", size = 285674620, upload-time = "2026-10-27T18:28:11.992Z" },
929
+ ]
930
+
931
+ [[package]]
932
+ name = "truststore"
933
+ version = "0.10.5"
934
+ source = { registry = "https://pypi.org/simple" }
935
+ sdist = { url = "https://files.pythonhosted.org/packages/ee/9f/c5201d42a484c061e528825fc8e2d565f5abd50a4ced6fb7d29c4ec99b2b/truststore-0.10.5.tar.gz", hash = "sha256:30d36967ccaded5cbb38d602c433f53600036c79d502f4533a49b60a03bbefcd", size = 28091, upload-time = "2026-10-12T22:27:31.808Z" }
936
+ wheels = [
937
+ { url = "https://files.pythonhosted.org/packages/51/e9/3a7820be2bb0fe53b6bc9c3be26d3d1158004e4c3ab953aa6840b955b1e9/truststore-0.10.5-py3-none-any.whl", hash = "sha256:9aaaedaefaf06d8b206278cf8b5012bc897f485a874503501e12d776df78951c", size = 19017, upload-time = "2026-10-12T22:27:30.377Z" },
938
+ ]
939
+
940
+ [[package]]
941
+ name = "typer"
942
+ version = "0.28.0"
943
+ source = { registry = "https://pypi.org/simple" }
944
+ dependencies = [
945
+ { name = "annotated-doc" },
946
+ { name = "colorama", marker = "sys_platform == 'win32'" },
947
+ { name = "rich" },
948
+ { name = "shellingham" },
949
+ ]
950
+ sdist = { url = "https://files.pythonhosted.org/packages/ad/8f/21021859316f44289101638dc74d82e13d722d8533e4b8708a3ab7c8461a/typer-0.28.0.tar.gz", hash = "sha256:69ab84ef79fd601cff071bbd2566d855ecadfd93adbc4b3360b9924359bdd857", size = 205415, upload-time = "2026-11-06T08:37:19.779Z" }
951
+ wheels = [
952
+ { url = "https://files.pythonhosted.org/packages/19/98/8ef60c7cd773dbab9d276a4c7bff505913a25b61acd8a9ef01ceeb70f554/typer-0.28.0-py3-none-any.whl", hash = "sha256:afd18929180b706c5e9a239f53c3c8136903553c0c2d460598bbaa924191d4ab", size = 123253, upload-time = "2026-11-06T08:37:18.611Z" },
953
+ ]
954
+
955
+ [[package]]
956
+ name = "typing-extensions"
957
+ version = "4.16.0"
958
+ source = { registry = "https://pypi.org/simple" }
959
+ sdist = { url = "https://files.pythonhosted.org/packages/f6/cc/6253133b5bb138fc3306cebfbda2c520f545d36b5be2c7255cc528bb45d6/typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5", size = 113555, upload-time = "2026-07-02T08:40:05.92Z" }
960
+ wheels = [
961
+ { url = "https://files.pythonhosted.org/packages/49/d3/b8441a820a491ddfc024b0b0cf0393375b75ea13866d9c66727e54c2fc80/typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8", size = 45571, upload-time = "2026-07-02T08:40:04.659Z" },
962
+ ]
963
+
964
+ [[package]]
965
+ name = "urllib3"
966
+ version = "2.8.0"
967
+ source = { registry = "https://pypi.org/simple" }
968
+ sdist = { url = "https://files.pythonhosted.org/packages/e3/05/b17359e1cefb4f909b5e40b1b90a496d987258916dbbf88e842c729f510e/urllib3-2.8.0.tar.gz", hash = "sha256:63bf2ead4c879426ebf22ef2a781eeb4aa3b4ae798a0435506f8687fd5bb9b63", size = 458972, upload-time = "2026-09-15T19:29:36.253Z" }
969
+ wheels = [
970
+ { url = "https://files.pythonhosted.org/packages/92/9d/c4e665119135114480843e7ab388fa94d8480650450e6f8e26b70d323a4c/urllib3-2.8.0-py3-none-any.whl", hash = "sha256:0cf3cae568d36aa9576b28dfb35f11328f1cb974ca7647d9475ebb86c75ac6e3", size = 135717, upload-time = "2026-09-15T19:29:34.577Z" },
971
+ ]
v2_plan.md ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Router v2 plan
2
+
3
+ The simplixity router (the `CLASSIFY` call with the `Classification` schema), fine-tuned from
4
+ Qwen3.5-2B and served as Q4_K_M through Ollama. Teacher: Haiku 5.5 with thinking
5
+ (1024-token reasoning budget), via OpenRouter.
6
+
7
+ ## Where v1 stands (2026-11-06)
8
+
9
+ Run `runs/2b-v1`: QLoRA with an NF4 base, r 16, alpha 320, lr 2e-5 constant, accum 8, dropout 0,
10
+ answer-only loss.
11
+
12
+ - Eval loss was lowest at step 350, about 0.9 epochs (0.478), and rose after that. Train loss
13
+ kept falling to about 0.27, so the run was overfitting. It was stopped at step 700.
14
+ - Candidate: `simplixity-router:2b-v1-s350-q4_k_m` (1.3 GB).
15
+
16
+ Hand-labelled testset, 105 items (route = search, complexity and shape all correct):
17
+
18
+ | Model | Route | Complexity | Shape | Recency | Lenses | Median latency |
19
+ |---|---|---|---|---|---|---|
20
+ | 2B Q4, base | 38% | 71% | 59% | 91% | 81% | 1.24 s |
21
+ | 2B Q4, v1 step 350 | 96% | 98% | 98% | 98% | 97% | 0.53 s |
22
+ | 2B Q4, v1 step 700 | 97% | 98% | 99% | 99% | 96% | 0.54 s |
23
+ | Haiku 5.5 think | 99% | 99% | 100% | 99% | 100% | 3.0 s |
24
+
25
+ Agreement with the teacher on the dev sets (route), next to the teacher's agreement with a
26
+ second labelling run of itself:
27
+
28
+ | Dev set | Haiku vs Haiku | v1 step 350 vs Haiku |
29
+ |---|---|---|
30
+ | Synthetic (150) | 91% | 66% |
31
+ | Real (69) | 79% | 75% |
32
+
33
+ - On the 119 synthetic items where both Haiku runs agree, v1 matches only 70%. That is the
34
+ headroom to work on.
35
+ - Nearly all of the gap is between neighbouring levels:
36
+ - complexity: standard vs research, simple vs standard
37
+ - shape: none vs explainer, steps or profile
38
+ - Step 700 shifts toward "standard" where Haiku says "simple" (13 of 150 synthetic dev items),
39
+ which makes the app run searches it doesn't need.
40
+ - Failure seen: step 350 fell into a repetition loop on one of 219 dev items, and Ollama
41
+ aborted it with "token repeat limit reached". The app catches this and searches with the raw
42
+ question instead.
43
+
44
+ ## Diagnosis
45
+
46
+ 1. **Noisy labels.** About 9% of the synthetic labels would come out different in a second
47
+ Haiku run, and nearly all of those are the gray-zone cases. The student is taught both
48
+ answers for very similar questions.
49
+ 2. **Too little data in the gray zone.** The 37 categories cover the clear cases well, but the
50
+ simple/standard and shape/no-shape boundaries are where most errors are.
51
+ 3. **Overfitting after one epoch** at lr 2e-5 on about 3k examples.
52
+
53
+ ## v2 steps
54
+
55
+ 1. **Consensus labels** (about $3 to $4)
56
+ - Label every training item a second time with the same call (`label.py --prompts --out`).
57
+ - Where the two runs disagree on needs_search, complexity, answer_shape or recency, take a
58
+ third label and use the majority.
59
+ - Drop the item when all three runs disagree.
60
+ - For the other fields (queries, lenses, image_query, place), keep the label from the
61
+ majority run.
62
+ 2. **More gray-zone data** (about $2.50)
63
+ - `gen_questions.py` with a new seed, about 3000 more questions.
64
+ - New or reweighted categories:
65
+ - simple vs standard: single facts that look broad, and short questions that need
66
+ several sources
67
+ - shape vs no shape: how-to versus explanation, a person versus a topic
68
+ - comparison vs table
69
+ - Same dedup against the testset (difflib above 0.85 is dropped).
70
+ - Build the prompts with the app's own code and label them with consensus as in step 1.
71
+ 3. **Train one epoch** on about 6k synthetic plus 145 real (counted twice)
72
+ - Same QLoRA setup, eval every 50 steps, keep every checkpoint.
73
+ - Roughly 3 hours on the 3070 Ti. The screensaver must be off.
74
+ - Try lr 1e-5 only if the one-epoch run still overfits before its end.
75
+ 4. **Evaluate**
76
+ - Testset: `eval/eval.py --num-gpu 99`.
77
+ - Dev agreement: `eval/eval_dev.py` against Haiku run 1 and against the consensus of runs 1
78
+ and 2.
79
+ - Count repetition-loop aborts separately.
80
+ - Goal: synthetic dev route agreement close to Haiku's own 91%, real close to 79%, testset at
81
+ least as good as v1 (96%).
82
+ 5. **Repetition loop** (only if it shows up again in v2): check whether `repeat_penalty` or a
83
+ smaller `num_predict` avoids it without changing scores.
84
+
85
+ ## Budget
86
+
87
+ About $6 to $7 on OpenRouter and about 3 hours of GPU time.
88
+
89
+ ## Privacy
90
+
91
+ Nothing from the production DB, or derived from it, leaves `data/private/`. That covers the real
92
+ questions, their prompts, labels, dev results and relabel runs. Only synthetic data, code and the
93
+ merged model weights are publishable.