ken-jo commited on
Commit
6cb324a
·
verified ·
1 Parent(s): 14fe0d7

QEV SDK 0.2.0: packaged playground and typed decisions

Browse files

Audited English release with exact runtime/weight provenance, source-specific data licenses and published limitations.

.gitattributes CHANGED
@@ -34,3 +34,4 @@ saved_model/**/* 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
  release-comparison/accuracy.png 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
  release-comparison/accuracy.png filter=lfs diff=lfs merge=lfs -text
37
+ runtime/qev-0.2.0-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,310 +1,314 @@
1
- ---
2
- license: apache-2.0
3
- base_model: Qwen/Qwen3.5-2B
4
- base_model_relation: adapter
5
- library_name: pytorch
6
- language:
7
- - en
8
- - ko
9
- tags:
10
- - multimodal
11
- - dynamic-classification
12
- - qwen3_5
13
- - calibrated-decisions
14
- - custom-runtime
15
- datasets:
16
- - ken-jo/qev-data
17
- - LocalLLaMA/typed-decisions
18
- - stanfordnlp/snli
19
- - PolyAI/banking77
20
- - AI-Lab-Makerere/beans
21
- - garythung/trashnet
22
- inference: false
23
- ---
24
-
25
- # QEV
26
-
27
- **Your evidence. Your criteria. A decision with probabilities.**
28
-
29
- QEV is an open multimodal decision model inspired by
30
- [LAYA](https://huggingface.co/convaiinnovations/laya), combining request-defined typed
31
- decisions with the text and vision backbone of
32
- [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B).
33
- Provide text, an image, or both, plus the choices or criteria you want evaluated.
34
- Receive probabilities and a structured answer in one batched backbone forward.
35
-
36
- [Training data](https://huggingface.co/datasets/ken-jo/qev-data)
37
-
38
- ## At a glance
39
-
40
- | | QEV 0.1.1 |
41
- | --- | --- |
42
- | Evidence | Text, one photo, or text and a photo together |
43
- | Decision types | `choice`, ordered `score`, and true/false `noul` |
44
- | Request limits | 1-4 questions; 2-16 choice/score candidates; 2,048 processed tokens |
45
- | Inference | One batched backbone forward; zero generated answer tokens |
46
- | Base model | Qwen3.5-2B, with its vision encoder frozen |
47
- | Full model / adaptation | 2.213B merged parameters / 7.992M stored adapter and head parameters |
48
- | Precision | BF16 backbone and FP32 readouts on CUDA |
49
- | License | Apache-2.0 for code and adaptation; source-specific dataset licenses |
50
-
51
- Version **0.1.1**, final Qwen-based research snapshot. Previously developed as Veyra
52
- Workflow Recovery v13. This release changes the public identity and packaging, not the
53
- learned weights, candidate encoding, calibration or inference mathematics.
54
- QEV is independently maintained. Its design draws on LAYA's typed decision interface;
55
- its neural backbone and vision encoder come from Qwen3.5-2B. No LAYA checkpoint is
56
- merged into the weights. The released training recipe is supervised adaptation and
57
- calibration; it is not an RLCD-trained compact model.
58
-
59
- ## What it returns
60
-
61
- | Primitive | Define in the request | Receive |
62
- | --- | --- | --- |
63
- | `choice` | Candidate names and descriptions | A probability for each candidate and a selected name |
64
- | `score` | Ordered descriptions, such as severity levels | Level probabilities and the expected zero-based level |
65
- | `noul` | A yes/no proposition | The probability that the proposition is true |
66
-
67
- All three include confidence and an abstention signal. Candidate definitions can change
68
- between requests. The output is computed directly from decision heads, so inference does
69
- not generate an answer sentence or a JSON string that needs parsing.
70
-
71
- ## One photo, three decisions
72
-
73
- Recognize a material, apply your own handling policy, and evaluate a proposition in the
74
- same request. The example below is an existing release verification fixture with its
75
- actual recorded output.
76
-
77
- <img src="https://huggingface.co/ken-jo/qev/resolve/main/examples/photograph/item.jpg" alt="TrashNet verification photograph of a plastic bottle" width="420" />
78
-
79
- | Question | Definition supplied with the request | QEV's recorded answer |
80
- | --- | --- | --- |
81
- | `choice` | Choose among cardboard, glass, metal, paper, plastic and trash | Plastic (`c4`), probability **0.9361** |
82
- | `score` | Band 0: glass/plastic; band 1: metal/trash; band 2: cardboard/paper | Expected level **0.4254** on the 0-2 scale; most likely band 0, probability **0.6978** |
83
- | `noul` | Does the item belong to glass, paper or plastic? | Probability true **0.7903** |
84
-
85
- All three answers were accepted by the released abstention policy. They used one batched
86
- forward and generated zero answer tokens. An expected level is a weighted average over
87
- the ordered levels. The probability on a single fixture is not an accuracy measurement.
88
-
89
- [Exact request](https://huggingface.co/ken-jo/qev/blob/main/examples/photograph/request.json)
90
- · [Full recorded response](https://huggingface.co/ken-jo/qev/blob/main/examples/photograph/response.json)
91
- · [Reproduction and provenance](https://github.com/ken-jo/qev/tree/main/examples/photograph)
92
-
93
- Photo: TrashNet, Gary Thung, MIT. Image bytes and the original verification question
94
- definitions are preserved. This previously inspected example illustrates the interface;
95
- aggregate performance is reported below.
96
-
97
- ## Quickstart
98
-
99
- Use Python 3.12 and the custom QEV runtime from the
100
- [release page](https://github.com/ken-jo/qev/releases). The checkpoint consists of an
101
- adaptation and decision heads; its Qwen backbone is downloaded separately.
102
-
103
- ```sh
104
- python -m pip install https://huggingface.co/ken-jo/qev/resolve/main/runtime/qev-0.1.1-py3-none-any.whl
105
- qev download --output checkpoints/qev
106
- ```
107
-
108
- ```python
109
- from pathlib import Path
110
- from qev import QEV, DecisionRequest
111
-
112
- model = QEV.load(Path("checkpoints/qev"), local_files_only=True, merge=True)
113
- request = DecisionRequest.model_validate({
114
- "state": {"text": "I was charged twice. Please refund the duplicate payment."},
115
- "questions": {
116
- "department": {
117
- "type": "choice",
118
- "instructions": "Which team should handle this request?",
119
- "criteria": {"billing": "Payments and refunds", "technical": "Software faults"},
120
- }
121
- },
122
- })
123
- result = model.predict(request)
124
- print(result["answers"]["department"])
125
- ```
126
-
127
- Recorded output for this request, shortened:
128
-
129
- ```json
130
- {
131
- "type": "choice",
132
- "choice": "billing",
133
- "probabilities": {"billing": 0.974044, "technical": 0.025956},
134
- "confidence": 0.974044,
135
- "abstained": false
136
- }
137
- ```
138
-
139
- For a photo, set `state.images` to `[{"path": "item.jpg"}]`, describe the visual decision
140
- in the question, and call `model.predict(request, Path("images").resolve())`. The image
141
- path is resolved under that directory. Text can supply context or a policy for the same image.
142
- Use the [API guide](https://github.com/ken-jo/qev/blob/main/docs/API.md) for score and noul
143
- schemas and the local HTTP server.
144
-
145
- ## Base and modifications
146
-
147
- - Base: [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B), revision
148
- `15852e8c16360a2fea060d615a32b45270f8a8fc`.
149
- - Frozen vision encoder; 24 stored language-layer LoRA adapters, rank 8, alpha 16.
150
- - 2,048-dimensional hidden states; 16-position option readout, four internal reasoning
151
- slots and a rank-64 condition-modulated binding head.
152
- - Final recovery updates only adapters in language layers 18-23 and existing readout/
153
- binding parameters. All 24 stored adapter layers merge for inference in BF16.
154
- - Stored adaptation/readout: 7,992,384 parameters; 32,009,800-byte safetensors file.
155
- - Runtime backbone: 2,213,241,664 parameters (331,416,576 vision; 1,881,825,088 language).
156
- - Storage format: safetensors. Adaptation/readout tensors are FP32. GPU inference uses
157
- BF16 backbone weights and FP32 readouts; CPU inference uses FP32.
158
- - LoRA contributes 7,815,168 stored parameters, merged into existing weights at inference.
159
- The remaining 177,216 readout/binding parameters stay separate; merged inference has
160
- 2,213,418,880 parameters. The upstream 4.55 GB download also includes unused MTP tensors.
161
- - One batched backbone forward per request; zero autoregressively generated answer tokens.
162
- Questions are encoded as separate batch entries, so adding questions still costs compute.
163
-
164
- Candidate descriptions define meanings at request time. The 16 output positions are
165
- temporary option positions, not 16 fixed semantic classes. Choice descriptions are
166
- canonically sorted; ordered score levels retain their order. The whole model still uses
167
- the Qwen multimodal architecture. There is no claim of a newly invented base network.
168
-
169
- ## Inputs and outputs
170
-
171
- Text and/or one image; 1-4 questions; choice/score with 2-16 alternatives, or a binary
172
- noul proposition. Processed token budget: 2,048. Image preprocessing uses 65,536-262,144
173
- pixels and 32-pixel alignment. This budget does not establish a minimum reliable resolution
174
- for small text, tiny objects or spatial tasks.
175
-
176
- Returns probabilities, selected label/expected score/probability true, and abstention.
177
- Probabilities are estimates; a direction probability in a game is not a game-win probability.
178
- The trained prompt marker `VeyraResult:` and internal `veyra` package remain unchanged.
179
-
180
- ## QEV and LAYA: matched text evaluation
181
-
182
- All three frozen models answered the same English inputs on an RTX 4060 Ti 8 GB.
183
- Weights, prompts and temperatures were not tuned on this evaluation.
184
-
185
- | Test | Questions | LAYA English | LAYA Typed Decisions | QEV 0.1.1 |
186
- | --- | ---: | ---: | ---: | ---: |
187
- | Official typed-decisions | 2,000 | 36.05% | 76.95% | 77.00% |
188
- | AG News, 4 candidates | 400 | 95.00% | 95.25% | 82.50% |
189
- | DAIR Emotion, 6 candidates | 400 | 58.75% | 60.00% | 50.25% |
190
-
191
- ![Accuracy on identical inputs](https://huggingface.co/ken-jo/qev/resolve/main/release-comparison/accuracy.png)
192
-
193
- Typed-decisions is an adapted, previously inspected benchmark for QEV and LAYA's
194
- specialist. QEV's one-question advantage is not evidence of superiority: the paired 95%
195
- interval is -1.85 to +1.95 percentage points. News and emotion have no QEV task-specific
196
- adaptation in the audited release sources, making them task-held-out zero-shot tests
197
- for QEV. Upstream pretraining overlap is unknown. LAYA reports news in its training
198
- mix and emotion held out, so the training exposure is not identical.
199
-
200
- | typed-decisions metric | LAYA Typed Decisions | QEV 0.1.1 | Better direction |
201
- | --- | ---: | ---: | --- |
202
- | Brier against soft targets | 0.06149 | 0.07460 | Lower |
203
- | NLL / soft-target cross-entropy | 0.88445 | 0.90894 | Lower |
204
- | ECE, 15 bins | 21.67% | 25.19% | Lower |
205
- | Ordinal expectation MAE | 0.24251 | 0.30366 | Lower |
206
- | Resident SDK p50 | 23.71 ms | 77.70 ms | Lower |
207
- | Resident SDK p95 | 30.06 ms | 97.70 ms | Lower |
208
-
209
- The specialist has better probability quality on this benchmark and is faster here.
210
- Both LAYA English checkpoints are reported as 421M parameters; QEV uses 2.213B after
211
- merging adapters. Timings are serial, one question per call after warmup; loading,
212
- network transport and queueing are excluded. All state, instruction and option
213
- truncation counts are zero. These measurements do not establish parity with LAYA's
214
- multilingual router or a live JEV service.
215
-
216
- [Full comparison, source revisions and zero-shot definitions](https://github.com/ken-jo/qev/blob/main/docs/LAYA_COMPARISON.md)
217
- · [Machine-readable evidence](https://github.com/ken-jo/qev/tree/main/reports/release-comparison)
218
-
219
- ## Accuracy and the decisions accepted
220
-
221
- QEV reports an abstention flag from its released fitted policy. An application can route
222
- flagged requests for review. The table shows both the whole evaluation and the portion
223
- accepted by that policy, with no threshold retuning for these measurements.
224
-
225
- | Evaluation | Accuracy on all questions | Questions accepted | Accuracy among accepted questions |
226
- | --- | ---: | ---: | ---: |
227
- | Typed-decisions, 2,000 questions | 77.00% | 30.80% | 94.32% |
228
- | AG News, 400 questions | 82.50% | 98.00% | 83.16% |
229
- | DAIR Emotion, 400 questions | 50.25% | 63.75% | 60.78% |
230
-
231
- Coverage and reliability change substantially across tasks. The 94.32% figure applies
232
- only to the accepted typed-decisions subset. It is not the accuracy of all requests or
233
- a guarantee for a new workflow. The `abstained` field is a policy decision based on model
234
- confidence; QEV does not return a separately trained `unknown_probability` class.
235
-
236
- ## Separate image and workflow evaluation
237
-
238
- | Fresh final metric | Veyra Foundation v11 (Qwen3.5-2B) | QEV 0.1.1 (Workflow Recovery v13) |
239
- | --- | ---: | ---: |
240
- | Authored workflow accuracy, 1,440 questions / 480 groups | 44.31% | 69.38% |
241
- | CIFAR-10 guard, 600 questions | 95.17% | 95.83% |
242
- | SNLI, 450 questions | 86.00% | 87.78% |
243
- | BANKING77 sampled 8-candidate, 462 questions | 84.20% | 85.50% |
244
- | Uncertainty NLL, 480 groups (lower is better) | 1.003285 | 0.899093 |
245
- | Squared conditional-distribution error (lower is better) | 0.323918 | 0.245122 |
246
- | Expected 0/1/5 cost at 80% coverage (lower is better) | 1.786584 | 1.011177 |
247
-
248
- Workflow improvement has a paired group-bootstrap 95% interval of +20.07 to +30.14
249
- percentage points. The three new families are equipment reservation, supplier onboarding
250
- and travel reimbursement. They are procedural tasks, not broad production business cases.
251
-
252
- Previously inspected regressions: official typed-decisions 77.00% over 2,000 questions;
253
- legacy synthetic text 91.16% and image 97.77%. The official benchmark's actual answer
254
- coverage is 30.80%, with 94.32% accuracy among accepted answers. These are not independent
255
- final sets or an overall claim of 90%+ real-world accuracy. The model trained on the recorded
256
- typed-decisions training partition; teacher-agreement annotations are not observed outcomes.
257
-
258
- Resident local HTTP photo p95: **114.94 ms** on RTX 4060 Ti 8 GB, 40 distinct serial
259
- photographs after three warmups, one image/question and six candidates. Feature caching
260
- is disabled. Loading, WAN and concurrent traffic are excluded. No matched JEV speed
261
- or accuracy comparison was run.
262
-
263
- ## Limitations and negative results
264
-
265
- - Final overall ECE is 12.58%, worse than Veyra Foundation v11's 6.80%.
266
- - Final overall accepted expected error is 20.49% at 88.99% coverage. For uncertain
267
- requests it is 45.57% at 52.50% coverage. Fitted abstention is not a shifted-domain guarantee.
268
- - 72 exploratory 2048 games produced no 2048 wins. Engine-assisted variants supplied legal
269
- directions or deterministic next boards; their results must not be credited to pure vision.
270
- - A balanced seven-choice board-cell diagnostic scored image 5/28 (17.86%) and text
271
- 21/28 (75%). This measures narrow digit/location binding, not general photograph accuracy.
272
- - Photographic transfer covers limited datasets. OCR, localization, object relations,
273
- temporal reasoning, safety-critical use and broad Korean task quality are not established.
274
- - Public benchmark overlap in Qwen pretraining cannot be ruled out. Near-duplicate grouping
275
- only detects the documented image similarities; it cannot prove distinct physical objects.
276
-
277
- ## Training and licensing
278
-
279
- Supervised decision/distribution training and staged LoRA adaptation were used. Five
280
- probability-learning variants, including RLOO, were compared at the earlier head stage;
281
- cross-entropy was selected there. This checkpoint is not described as a successful new
282
- reinforcement-learning algorithm. Later recovery uses soft cross-entropy with parent KL
283
- replay. The selected final recovery consumed 4,533 additional questions and 567 optimizer
284
- steps (half of the declared one-pass schedule), selected on development data before new final
285
- evaluation. Calibration and abstention use separate recorded groups.
286
-
287
- Code/adaptation weights are Apache-2.0; dataset material retains its separate per-source
288
- licenses. The model package includes one MIT-licensed verification photograph with its
289
- source notice and request. Training corpus snapshots are published separately, with
290
- CIFAR-10 excluded pending redistribution rights. Read the repository's data and training
291
- documentation.
292
-
293
- ## Integrity and loading
294
-
295
- - Weights SHA-256: `84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee`
296
- - Manifest SHA-256: `d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d`
297
-
298
- The package contains a calibrated manifest, adaptation weights and a custom runtime wheel.
299
- Qwen backbone weights must be downloaded separately at the pinned revision. Install the
300
- wheel from `runtime/`, run `qev download --base-only`, then use `load_qev.py` or the Python
301
- API in the [repository](https://github.com/ken-jo/qev). A standard Transformers
302
- auto-model loader cannot directly load this custom adapter/head layout.
303
-
304
- Historical stage flags in the unchanged manifest record when those stages were run.
305
- The exact-model acceptance report and publication verification are separate evidence;
306
- packaging does not retroactively alter a stage's provenance.
307
-
308
- ---
309
-
310
- [GitHub: ken-jo/qev](https://github.com/ken-jo/qev)
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen3.5-2B
4
+ base_model_relation: adapter
5
+ library_name: pytorch
6
+ language:
7
+ - en
8
+ - ko
9
+ tags:
10
+ - multimodal
11
+ - dynamic-classification
12
+ - qwen3_5
13
+ - calibrated-decisions
14
+ - custom-runtime
15
+ datasets:
16
+ - ken-jo/qev-data
17
+ - LocalLLaMA/typed-decisions
18
+ - stanfordnlp/snli
19
+ - PolyAI/banking77
20
+ - AI-Lab-Makerere/beans
21
+ - garythung/trashnet
22
+ inference: false
23
+ ---
24
+
25
+ # QEV
26
+
27
+ **Your evidence. Your criteria. A decision with probabilities.**
28
+
29
+ QEV is an open multimodal decision model inspired by
30
+ [LAYA](https://huggingface.co/convaiinnovations/laya), combining request-defined typed
31
+ decisions with the text and vision backbone of
32
+ [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B).
33
+ Provide text, an image, or both, plus the choices or criteria you want evaluated.
34
+ Receive probabilities and a structured answer in one batched backbone forward.
35
+
36
+ [Training data](https://huggingface.co/datasets/ken-jo/qev-data)
37
+
38
+ ## At a glance
39
+
40
+ | | QEV 0.1.1 |
41
+ | --- | --- |
42
+ | Evidence | Text, one photo, or text and a photo together |
43
+ | Decision types | `choice`, ordered `score`, and true/false `noul` |
44
+ | Request limits | 1-4 questions; 2-16 choice/score candidates; 2,048 processed tokens |
45
+ | Inference | One batched backbone forward; zero generated answer tokens |
46
+ | Base model | Qwen3.5-2B, with its vision encoder frozen |
47
+ | Full model / adaptation | 2.213B merged parameters / 7.992M stored adapter and head parameters |
48
+ | Precision | BF16 backbone and FP32 readouts on CUDA |
49
+ | License | Apache-2.0 for code and adaptation; source-specific dataset licenses |
50
+
51
+ Model version **0.1.1**, final Qwen-based research snapshot; Python SDK **0.2.0**. Previously developed as Veyra
52
+ Workflow Recovery v13. This release changes the public identity and packaging, not the
53
+ learned weights, candidate encoding, calibration or inference mathematics.
54
+ QEV is independently maintained. Its design draws on LAYA's typed decision interface;
55
+ its neural backbone and vision encoder come from Qwen3.5-2B. No LAYA checkpoint is
56
+ merged into the weights. The released training recipe is supervised adaptation and
57
+ calibration; it is not an RLCD-trained compact model.
58
+
59
+ ## What it returns
60
+
61
+ | Primitive | Define in the request | Receive |
62
+ | --- | --- | --- |
63
+ | `choice` | Candidate names and descriptions | A probability for each candidate and a selected name |
64
+ | `score` | Ordered descriptions, such as severity levels | Level probabilities and the expected zero-based level |
65
+ | `noul` | A yes/no proposition | The probability that the proposition is true |
66
+
67
+ All three include confidence and an abstention signal. Candidate definitions can change
68
+ between requests. The output is computed directly from decision heads, so inference does
69
+ not generate an answer sentence or a JSON string that needs parsing.
70
+
71
+ ## One photo, three decisions
72
+
73
+ Recognize a material, apply your own handling policy, and evaluate a proposition in the
74
+ same request. The example below is an existing release verification fixture with its
75
+ actual recorded output.
76
+
77
+ <img src="https://huggingface.co/ken-jo/qev/resolve/main/examples/photograph/item.jpg" alt="TrashNet verification photograph of a plastic bottle" width="420" />
78
+
79
+ | Question | Definition supplied with the request | QEV's recorded answer |
80
+ | --- | --- | --- |
81
+ | `choice` | Choose among cardboard, glass, metal, paper, plastic and trash | Plastic (`c4`), probability **0.9361** |
82
+ | `score` | Band 0: glass/plastic; band 1: metal/trash; band 2: cardboard/paper | Expected level **0.4254** on the 0-2 scale; most likely band 0, probability **0.6978** |
83
+ | `noul` | Does the item belong to glass, paper or plastic? | Probability true **0.7903** |
84
+
85
+ All three answers were accepted by the released abstention policy. They used one batched
86
+ forward and generated zero answer tokens. An expected level is a weighted average over
87
+ the ordered levels. The probability on a single fixture is not an accuracy measurement.
88
+
89
+ [Exact request](https://huggingface.co/ken-jo/qev/blob/main/examples/photograph/request.json)
90
+ · [Full recorded response](https://huggingface.co/ken-jo/qev/blob/main/examples/photograph/response.json)
91
+ · [Reproduction and provenance](https://github.com/ken-jo/qev/tree/main/examples/photograph)
92
+
93
+ Photo: TrashNet, Gary Thung, MIT. Image bytes and the original verification question
94
+ definitions are preserved. This previously inspected example illustrates the interface;
95
+ aggregate performance is reported below.
96
+
97
+ ## Quickstart
98
+
99
+ Use Python 3.12 and the custom QEV runtime from the
100
+ [release page](https://github.com/ken-jo/qev/releases). The checkpoint consists of an
101
+ adaptation and decision heads; its Qwen backbone is downloaded separately.
102
+
103
+ ```sh
104
+ python -m pip install https://huggingface.co/ken-jo/qev/resolve/main/runtime/qev-0.2.0-py3-none-any.whl
105
+ qev playground
106
+ ```
107
+
108
+ ```python
109
+ from pathlib import Path
110
+ from qev import load, DecisionRequest
111
+
112
+ model = load() # Prepares the pinned weights on first use.
113
+ request = DecisionRequest.model_validate({
114
+ "state": {"text": "I was charged twice. Please refund the duplicate payment."},
115
+ "questions": {
116
+ "department": {
117
+ "type": "choice",
118
+ "instructions": "Which team should handle this request?",
119
+ "criteria": {"billing": "Payments and refunds", "technical": "Software faults"},
120
+ }
121
+ },
122
+ })
123
+ result = model.predict(request)
124
+ print(result["answers"]["department"])
125
+ ```
126
+
127
+ Recorded output for this request, shortened:
128
+
129
+ ```json
130
+ {
131
+ "type": "choice",
132
+ "choice": "billing",
133
+ "probabilities": {"billing": 0.974044, "technical": 0.025956},
134
+ "confidence": 0.974044,
135
+ "abstained": false
136
+ }
137
+ ```
138
+
139
+ For a photo, set `state.images` to `[{"path": "item.jpg"}]`, describe the visual decision
140
+ in the question, and call `model.predict(request, Path("images").resolve())`. The image
141
+ path is resolved under that directory. Text can supply context or a policy for the same image.
142
+ Use the [API guide](https://github.com/ken-jo/qev/blob/main/docs/API.md) for score and noul
143
+ schemas and the local HTTP server.
144
+
145
+ The SDK includes the English playground and six sample photographs. `qev playground`
146
+ opens a local server at http://127.0.0.1:7860. First use downloads about 4.6 GB of model
147
+ files; subsequent launches use the persistent cache. Use `--offline` for a prepared cache.
148
+
149
+ ## Base and modifications
150
+
151
+ - Base: [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B), revision
152
+ `15852e8c16360a2fea060d615a32b45270f8a8fc`.
153
+ - Frozen vision encoder; 24 stored language-layer LoRA adapters, rank 8, alpha 16.
154
+ - 2,048-dimensional hidden states; 16-position option readout, four internal reasoning
155
+ slots and a rank-64 condition-modulated binding head.
156
+ - Final recovery updates only adapters in language layers 18-23 and existing readout/
157
+ binding parameters. All 24 stored adapter layers merge for inference in BF16.
158
+ - Stored adaptation/readout: 7,992,384 parameters; 32,009,800-byte safetensors file.
159
+ - Runtime backbone: 2,213,241,664 parameters (331,416,576 vision; 1,881,825,088 language).
160
+ - Storage format: safetensors. Adaptation/readout tensors are FP32. GPU inference uses
161
+ BF16 backbone weights and FP32 readouts; CPU inference uses FP32.
162
+ - LoRA contributes 7,815,168 stored parameters, merged into existing weights at inference.
163
+ The remaining 177,216 readout/binding parameters stay separate; merged inference has
164
+ 2,213,418,880 parameters. The upstream 4.55 GB download also includes unused MTP tensors.
165
+ - One batched backbone forward per request; zero autoregressively generated answer tokens.
166
+ Questions are encoded as separate batch entries, so adding questions still costs compute.
167
+
168
+ Candidate descriptions define meanings at request time. The 16 output positions are
169
+ temporary option positions, not 16 fixed semantic classes. Choice descriptions are
170
+ canonically sorted; ordered score levels retain their order. The whole model still uses
171
+ the Qwen multimodal architecture. There is no claim of a newly invented base network.
172
+
173
+ ## Inputs and outputs
174
+
175
+ Text and/or one image; 1-4 questions; choice/score with 2-16 alternatives, or a binary
176
+ noul proposition. Processed token budget: 2,048. Image preprocessing uses 65,536-262,144
177
+ pixels and 32-pixel alignment. This budget does not establish a minimum reliable resolution
178
+ for small text, tiny objects or spatial tasks.
179
+
180
+ Returns probabilities, selected label/expected score/probability true, and abstention.
181
+ Probabilities are estimates; a direction probability in a game is not a game-win probability.
182
+ The trained prompt marker `VeyraResult:` and internal `veyra` package remain unchanged.
183
+
184
+ ## QEV and LAYA: matched text evaluation
185
+
186
+ All three frozen models answered the same English inputs on an RTX 4060 Ti 8 GB.
187
+ Weights, prompts and temperatures were not tuned on this evaluation.
188
+
189
+ | Test | Questions | LAYA English | LAYA Typed Decisions | QEV 0.1.1 |
190
+ | --- | ---: | ---: | ---: | ---: |
191
+ | Official typed-decisions | 2,000 | 36.05% | 76.95% | 77.00% |
192
+ | AG News, 4 candidates | 400 | 95.00% | 95.25% | 82.50% |
193
+ | DAIR Emotion, 6 candidates | 400 | 58.75% | 60.00% | 50.25% |
194
+
195
+ ![Accuracy on identical inputs](https://huggingface.co/ken-jo/qev/resolve/main/release-comparison/accuracy.png)
196
+
197
+ Typed-decisions is an adapted, previously inspected benchmark for QEV and LAYA's
198
+ specialist. QEV's one-question advantage is not evidence of superiority: the paired 95%
199
+ interval is -1.85 to +1.95 percentage points. News and emotion have no QEV task-specific
200
+ adaptation in the audited release sources, making them task-held-out zero-shot tests
201
+ for QEV. Upstream pretraining overlap is unknown. LAYA reports news in its training
202
+ mix and emotion held out, so the training exposure is not identical.
203
+
204
+ | typed-decisions metric | LAYA Typed Decisions | QEV 0.1.1 | Better direction |
205
+ | --- | ---: | ---: | --- |
206
+ | Brier against soft targets | 0.06149 | 0.07460 | Lower |
207
+ | NLL / soft-target cross-entropy | 0.88445 | 0.90894 | Lower |
208
+ | ECE, 15 bins | 21.67% | 25.19% | Lower |
209
+ | Ordinal expectation MAE | 0.24251 | 0.30366 | Lower |
210
+ | Resident SDK p50 | 23.71 ms | 77.70 ms | Lower |
211
+ | Resident SDK p95 | 30.06 ms | 97.70 ms | Lower |
212
+
213
+ The specialist has better probability quality on this benchmark and is faster here.
214
+ Both LAYA English checkpoints are reported as 421M parameters; QEV uses 2.213B after
215
+ merging adapters. Timings are serial, one question per call after warmup; loading,
216
+ network transport and queueing are excluded. All state, instruction and option
217
+ truncation counts are zero. These measurements do not establish parity with LAYA's
218
+ multilingual router or a live JEV service.
219
+
220
+ [Full comparison, source revisions and zero-shot definitions](https://github.com/ken-jo/qev/blob/main/docs/LAYA_COMPARISON.md)
221
+ · [Machine-readable evidence](https://github.com/ken-jo/qev/tree/main/reports/release-comparison)
222
+
223
+ ## Accuracy and the decisions accepted
224
+
225
+ QEV reports an abstention flag from its released fitted policy. An application can route
226
+ flagged requests for review. The table shows both the whole evaluation and the portion
227
+ accepted by that policy, with no threshold retuning for these measurements.
228
+
229
+ | Evaluation | Accuracy on all questions | Questions accepted | Accuracy among accepted questions |
230
+ | --- | ---: | ---: | ---: |
231
+ | Typed-decisions, 2,000 questions | 77.00% | 30.80% | 94.32% |
232
+ | AG News, 400 questions | 82.50% | 98.00% | 83.16% |
233
+ | DAIR Emotion, 400 questions | 50.25% | 63.75% | 60.78% |
234
+
235
+ Coverage and reliability change substantially across tasks. The 94.32% figure applies
236
+ only to the accepted typed-decisions subset. It is not the accuracy of all requests or
237
+ a guarantee for a new workflow. The `abstained` field is a policy decision based on model
238
+ confidence; QEV does not return a separately trained `unknown_probability` class.
239
+
240
+ ## Separate image and workflow evaluation
241
+
242
+ | Fresh final metric | Veyra Foundation v11 (Qwen3.5-2B) | QEV 0.1.1 (Workflow Recovery v13) |
243
+ | --- | ---: | ---: |
244
+ | Authored workflow accuracy, 1,440 questions / 480 groups | 44.31% | 69.38% |
245
+ | CIFAR-10 guard, 600 questions | 95.17% | 95.83% |
246
+ | SNLI, 450 questions | 86.00% | 87.78% |
247
+ | BANKING77 sampled 8-candidate, 462 questions | 84.20% | 85.50% |
248
+ | Uncertainty NLL, 480 groups (lower is better) | 1.003285 | 0.899093 |
249
+ | Squared conditional-distribution error (lower is better) | 0.323918 | 0.245122 |
250
+ | Expected 0/1/5 cost at 80% coverage (lower is better) | 1.786584 | 1.011177 |
251
+
252
+ Workflow improvement has a paired group-bootstrap 95% interval of +20.07 to +30.14
253
+ percentage points. The three new families are equipment reservation, supplier onboarding
254
+ and travel reimbursement. They are procedural tasks, not broad production business cases.
255
+
256
+ Previously inspected regressions: official typed-decisions 77.00% over 2,000 questions;
257
+ legacy synthetic text 91.16% and image 97.77%. The official benchmark's actual answer
258
+ coverage is 30.80%, with 94.32% accuracy among accepted answers. These are not independent
259
+ final sets or an overall claim of 90%+ real-world accuracy. The model trained on the recorded
260
+ typed-decisions training partition; teacher-agreement annotations are not observed outcomes.
261
+
262
+ Resident local HTTP photo p95: **114.94 ms** on RTX 4060 Ti 8 GB, 40 distinct serial
263
+ photographs after three warmups, one image/question and six candidates. Feature caching
264
+ is disabled. Loading, WAN and concurrent traffic are excluded. No matched JEV speed
265
+ or accuracy comparison was run.
266
+
267
+ ## Limitations and negative results
268
+
269
+ - Final overall ECE is 12.58%, worse than Veyra Foundation v11's 6.80%.
270
+ - Final overall accepted expected error is 20.49% at 88.99% coverage. For uncertain
271
+ requests it is 45.57% at 52.50% coverage. Fitted abstention is not a shifted-domain guarantee.
272
+ - 72 exploratory 2048 games produced no 2048 wins. Engine-assisted variants supplied legal
273
+ directions or deterministic next boards; their results must not be credited to pure vision.
274
+ - A balanced seven-choice board-cell diagnostic scored image 5/28 (17.86%) and text
275
+ 21/28 (75%). This measures narrow digit/location binding, not general photograph accuracy.
276
+ - Photographic transfer covers limited datasets. OCR, localization, object relations,
277
+ temporal reasoning, safety-critical use and broad Korean task quality are not established.
278
+ - Public benchmark overlap in Qwen pretraining cannot be ruled out. Near-duplicate grouping
279
+ only detects the documented image similarities; it cannot prove distinct physical objects.
280
+
281
+ ## Training and licensing
282
+
283
+ Supervised decision/distribution training and staged LoRA adaptation were used. Five
284
+ probability-learning variants, including RLOO, were compared at the earlier head stage;
285
+ cross-entropy was selected there. This checkpoint is not described as a successful new
286
+ reinforcement-learning algorithm. Later recovery uses soft cross-entropy with parent KL
287
+ replay. The selected final recovery consumed 4,533 additional questions and 567 optimizer
288
+ steps (half of the declared one-pass schedule), selected on development data before new final
289
+ evaluation. Calibration and abstention use separate recorded groups.
290
+
291
+ Code/adaptation weights are Apache-2.0; dataset material retains its separate per-source
292
+ licenses. The model package includes one MIT-licensed verification photograph with its
293
+ source notice and request. Training corpus snapshots are published separately, with
294
+ CIFAR-10 excluded pending redistribution rights. Read the repository's data and training
295
+ documentation.
296
+
297
+ ## Integrity and loading
298
+
299
+ - Weights SHA-256: `84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee`
300
+ - Manifest SHA-256: `d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d`
301
+
302
+ The package contains a calibrated manifest, adaptation weights and a custom runtime wheel.
303
+ Qwen backbone weights must be downloaded separately at the pinned revision. Install the
304
+ wheel from `runtime/`, then run `qev playground`, `qev predict`, or use the Python
305
+ `load()` API in the [repository](https://github.com/ken-jo/qev). A standard Transformers
306
+ auto-model loader cannot directly load this custom adapter/head layout.
307
+
308
+ Historical stage flags in the unchanged manifest record when those stages were run.
309
+ The exact-model acceptance report and publication verification are separate evidence;
310
+ packaging does not retroactively alter a stage's provenance.
311
+
312
+ ---
313
+
314
+ [GitHub: ken-jo/qev](https://github.com/ken-jo/qev)
checksums.json CHANGED
@@ -3,9 +3,11 @@
3
  "data-licenses/CC-BY-4.0.txt": "d557539df68e771cc1eedcc91d13f70fca930e508d11eedcafa4b15db49e3744",
4
  "data-licenses/CC-BY-SA-4.0.txt": "cde7883b9050a1104f4ac19a1572aafd6e5d7323b68351aaf51fbf4beba54966",
5
  "data-licenses/trashnet-MIT.txt": "e33347ae1589c1c7d7de5cf2e433e6c8060256273c7287a29d7c6e1e45c466f1",
 
6
  "docs/COMPACT_MODEL.md": "9b6f56cf01d8194ffac0f040ba51f71cb89c889e2722d453b8f6418dd5896760",
7
  "docs/LAYA_COMPARISON.md": "801fd57c856ac682492ff9927a086f7b0c3854d6ed551407f4edc7352d3c1bdd",
8
  "docs/MODEL_SIZE.md": "d0b4d6fdea9e8502d90a7ea88f040f9d9f9ae5806aca4080f0aa7a7d6f071f54",
 
9
  "evidence/checkpoints/veyra-backbone-recovery-v13/head.safetensors": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
10
  "evidence/checkpoints/veyra-backbone-recovery-v13/manifest.json": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
11
  "evidence/checkpoints/veyra-cohort-policy-v12/head.safetensors": "86ce942f8ef036f2279422a3a2665248b98adcb0effa91bafdf15d9c14c6c83e",
@@ -183,11 +185,11 @@
183
  "examples/photograph/response.json": "963c9d7e7e8c0f6c89110019c9dcd163dd27e1f565a562c772891725fc7b2ccf",
184
  "head.safetensors": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
185
  "LICENSE": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
186
- "load_qev.py": "cf2c6a95174776fe0819daf420f92d89886ba6f415f2fc2fcec0d125d73342b4",
187
  "manifest.json": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
188
  "NOTICE": "4ed44977e49e772fc8d3e971f417fe14a7490d418d07c759e2cba0dfc52ed544",
189
- "publication.json": "06a4e9551b414f21da11d4c2f86cee0ad6e607e8151be661ef5bda12d410409b",
190
- "README.md": "083262cdf22b88bcee6003569afbd60beb971757a6b68bd4ceb03c2058af66ce",
191
  "release-comparison/accuracy.png": "7da48ddd235642cd9e37807024abc85954d960419f3b5de52aa753f930eb9a45",
192
  "release-comparison/accuracy.svg": "84b4daceaecd169f8c937fda91c5589bf69f51ddc3e7f5c41d1b09de0eeee8c3",
193
  "release-comparison/comparison.json": "c5d3be2ba8cb68872c5d53fb4fad9f7954a56f8e79a9ff7fc66ce6fcc639f3dd",
@@ -198,9 +200,10 @@
198
  "release-comparison/README.md": "e38a2067ba681fe63544d155cd42075a5d39faf3ed7b8d2f4a29ac7ef87962bd",
199
  "release-protocol.json": "d0e6c454fc10cfa080523cfcc66706dfe73cb32524f4ce74bf87a5d4edde6776",
200
  "release-readiness.json": "5e2ed88d29b13229affb8487217f16e611f13201371f41d3604b0eb77f46edc8",
201
- "reproducibility/pyproject.toml": "f6f4565bdc1d436c36039739932c0db738045bb2fcd033ccb80dce5720625a9d",
202
- "reproducibility/uv.lock": "de3e551419caea5807b3b7f5239a7404a26f806a0b2bbfbee0737cc88734672f",
203
  "ROADMAP.md": "e75b8884c5555cba092dde0ac25bfb0448ff5964d8bae75067b8c854b1fe35a6",
204
  "runtime/qev-0.1.1-py3-none-any.whl": "f1f4c54175ce213e98399cc8a3db91298e25aa1450f2306195614c80c87da8a2",
 
205
  "sample_request.json": "db4c34b18ec8b1d0bcd2af48fd2b8be5636b67d2194d26492fc2f647f52b7cb1"
206
  }
 
3
  "data-licenses/CC-BY-4.0.txt": "d557539df68e771cc1eedcc91d13f70fca930e508d11eedcafa4b15db49e3744",
4
  "data-licenses/CC-BY-SA-4.0.txt": "cde7883b9050a1104f4ac19a1572aafd6e5d7323b68351aaf51fbf4beba54966",
5
  "data-licenses/trashnet-MIT.txt": "e33347ae1589c1c7d7de5cf2e433e6c8060256273c7287a29d7c6e1e45c466f1",
6
+ "docs/API.md": "36510bd70c740d41e6088453441a83e354e3015d2119cab0cd93a9847c399a78",
7
  "docs/COMPACT_MODEL.md": "9b6f56cf01d8194ffac0f040ba51f71cb89c889e2722d453b8f6418dd5896760",
8
  "docs/LAYA_COMPARISON.md": "801fd57c856ac682492ff9927a086f7b0c3854d6ed551407f4edc7352d3c1bdd",
9
  "docs/MODEL_SIZE.md": "d0b4d6fdea9e8502d90a7ea88f040f9d9f9ae5806aca4080f0aa7a7d6f071f54",
10
+ "docs/PLAYGROUND.md": "29cdc818cc69587fbc71bd788c9921e97ac107fd865e19cee10f080555fad23e",
11
  "evidence/checkpoints/veyra-backbone-recovery-v13/head.safetensors": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
12
  "evidence/checkpoints/veyra-backbone-recovery-v13/manifest.json": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
13
  "evidence/checkpoints/veyra-cohort-policy-v12/head.safetensors": "86ce942f8ef036f2279422a3a2665248b98adcb0effa91bafdf15d9c14c6c83e",
 
185
  "examples/photograph/response.json": "963c9d7e7e8c0f6c89110019c9dcd163dd27e1f565a562c772891725fc7b2ccf",
186
  "head.safetensors": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
187
  "LICENSE": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
188
+ "load_qev.py": "d220abbcd88a47ebd74b5309e9135394149379b2d8026bacf49cca71fda56de7",
189
  "manifest.json": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
190
  "NOTICE": "4ed44977e49e772fc8d3e971f417fe14a7490d418d07c759e2cba0dfc52ed544",
191
+ "publication.json": "1dd3801aee8b728da49ba8fd4ac053f591ee0c9298506f68542db5823a816e16",
192
+ "README.md": "49312ed1a0fb147f7327c00618165f1c6a5803f78734d049430bd48e2b6273a5",
193
  "release-comparison/accuracy.png": "7da48ddd235642cd9e37807024abc85954d960419f3b5de52aa753f930eb9a45",
194
  "release-comparison/accuracy.svg": "84b4daceaecd169f8c937fda91c5589bf69f51ddc3e7f5c41d1b09de0eeee8c3",
195
  "release-comparison/comparison.json": "c5d3be2ba8cb68872c5d53fb4fad9f7954a56f8e79a9ff7fc66ce6fcc639f3dd",
 
200
  "release-comparison/README.md": "e38a2067ba681fe63544d155cd42075a5d39faf3ed7b8d2f4a29ac7ef87962bd",
201
  "release-protocol.json": "d0e6c454fc10cfa080523cfcc66706dfe73cb32524f4ce74bf87a5d4edde6776",
202
  "release-readiness.json": "5e2ed88d29b13229affb8487217f16e611f13201371f41d3604b0eb77f46edc8",
203
+ "reproducibility/pyproject.toml": "c858736cc2fdfeb972bdbbc0903e21c3f160c64904e0293562a0ee792d22cbb5",
204
+ "reproducibility/uv.lock": "175edb067d35d24e20fdf92227000420ab40ff1ec0911e7be5d0da058742d01d",
205
  "ROADMAP.md": "e75b8884c5555cba092dde0ac25bfb0448ff5964d8bae75067b8c854b1fe35a6",
206
  "runtime/qev-0.1.1-py3-none-any.whl": "f1f4c54175ce213e98399cc8a3db91298e25aa1450f2306195614c80c87da8a2",
207
+ "runtime/qev-0.2.0-py3-none-any.whl": "33ca132521cf9ec7f5bc9405816cd65e9c4626853cfb21a82da7a4549ebbe576",
208
  "sample_request.json": "db4c34b18ec8b1d0bcd2af48fd2b8be5636b67d2194d26492fc2f647f52b7cb1"
209
  }
docs/API.md ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Inference API
2
+
3
+ ## Python SDK
4
+
5
+ The `qev` distribution contains both the SDK and the English playground. With Python
6
+ 3.12, install the release wheel or the published PyPI distribution, then:
7
+
8
+ ```python
9
+ import qev
10
+ from pathlib import Path
11
+
12
+ model = qev.load() # Download missing model files; select CUDA when available.
13
+ request = qev.DecisionRequest.from_json(open("request.json", encoding="utf-8").read())
14
+ result = model.predict(request, image_root=Path("."))
15
+ print(result["answers"])
16
+ ```
17
+
18
+ The first load retrieves about 4.6 GB of pinned model files. Later loads reuse the cache.
19
+ `qev.load(device="cpu", cache_dir="/path/to/hub", offline=True)` selects CPU and requires
20
+ complete local files. `QEV_HOME` sets the checkpoint/cache root; `QEV_CACHE_DIR` overrides
21
+ the Hub cache. The lower-level `qev.QEV.load(...)` remains available for explicit checkpoint
22
+ management and retains the original inference contract.
23
+
24
+ ## Commands
25
+
26
+ ```sh
27
+ qev playground
28
+ qev download
29
+ qev predict --request request.json
30
+ qev serve --image-root .
31
+ ```
32
+
33
+ All inference commands prepare missing model files automatically. Use `--offline` to
34
+ require a complete cache. `predict` resolves image paths relative to the request JSON;
35
+ `serve` resolves them under `--image-root`. The API server defaults to
36
+ `http://127.0.0.1:8000`; the playground defaults to `http://127.0.0.1:7860`.
37
+ See [playground installation, uv and GPU instructions](PLAYGROUND.md).
38
+
39
+ ## Request and response
40
+
41
+ The public request schema is `qev.DecisionRequest`. Text and one local image can be
42
+ combined. Image paths are resolved under the configured image root; remote URLs are not
43
+ fetched by the inference API.
44
+
45
+ ```json
46
+ {
47
+ "state": {"text": "The customer was charged twice and asks for a refund."},
48
+ "questions": {
49
+ "department": {
50
+ "type": "choice",
51
+ "instructions": "Which department should handle the request?",
52
+ "criteria": {"billing": "Payments and refunds", "technical": "Software faults"}
53
+ }
54
+ }
55
+ }
56
+ ```
57
+
58
+ Send JSON to `POST /v1/systemone`. `score.criteria` is an ordered list of descriptions;
59
+ the result is an expected zero-based level. `noul.criteria` has `true` and `false` strings.
60
+ The maximum is four questions, 16 alternatives per choice/score and one image. Processed
61
+ inputs exceeding the configured token budget fail instead of silently truncating.
62
+
63
+ Inspect `answers`, `probabilities`, abstention fields and `usage` in the returned object.
64
+ The model ID remains `Qwen/Qwen3.5-2B` in the validated contract to identify the actual base.
65
+ See `src/veyra/schema.py`, `src/veyra/probability.py` and `examples/request.json`.
66
+
67
+ The playground adds image-upload and resolution endpoints around the same model. A single
68
+ resident worker serializes inference. It is a trusted local demo; uploads are shared within
69
+ that server session. Local latency measurements do not imply concurrent serving capacity.
docs/PLAYGROUND.md ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # QEV playground and SDK
2
+
3
+ The `qev` Python wheel includes the English playground, its presets, six sample photos,
4
+ license notices and the SDK. You do not need to clone the repository to use the UI.
5
+
6
+ ## Install
7
+
8
+ Python 3.12 is required. The public runtime wheel is available with the model:
9
+
10
+ ```sh
11
+ python -m pip install https://huggingface.co/ken-jo/qev/resolve/main/runtime/qev-0.2.0-py3-none-any.whl
12
+ qev playground
13
+ ```
14
+
15
+ Or use uv's isolated tool environment:
16
+
17
+ ```sh
18
+ uvx --python 3.12 --from https://huggingface.co/ken-jo/qev/resolve/main/runtime/qev-0.2.0-py3-none-any.whl qev playground
19
+ ```
20
+
21
+ From a source checkout:
22
+
23
+ ```sh
24
+ uv sync --frozen
25
+ uv run qev playground
26
+ ```
27
+
28
+ Open http://127.0.0.1:7860. `--open` opens that page automatically. Use `--port 7870`
29
+ for another port. `--host 0.0.0.0` makes the interface reachable through the machine's IP
30
+ on a trusted network. No public hosted Space is created.
31
+
32
+ ## First launch and caches
33
+
34
+ The first model use fetches the checksum-verified QEV adaptation and pinned Qwen3.5-2B
35
+ backbone: about 4.6 GB of model files, in addition to Python dependencies. A progress
36
+ message appears while the model is being prepared. Later launches use the same cache;
37
+ starting from a different working directory does not trigger another model download.
38
+
39
+ | Setting | Purpose |
40
+ | --- | --- |
41
+ | `QEV_HOME` | QEV checkpoint/cache root; Windows defaults to `%LOCALAPPDATA%/qev`, Linux/macOS to the XDG cache or `~/.cache/qev` |
42
+ | `QEV_CACHE_DIR` or `--cache-dir` | Override the Hugging Face model cache |
43
+ | `--checkpoint` | Use an explicit folder containing the released QEV head and manifest |
44
+ | `--offline` or `QEV_OFFLINE=1` | Require complete local files; never fetch missing files |
45
+ | `qev download` | Prepare both model components without opening the UI |
46
+
47
+ `qev predict`, `qev serve` and Python's `qev.load()` use the same bootstrap path. The
48
+ low-level `QEV.load()` API is retained for callers that manage checkpoint loading directly.
49
+ An existing checkpoint with a different checksum is rejected and preserved.
50
+
51
+ ## CPU and GPU
52
+
53
+ The default `--device auto` uses CUDA when it is available, otherwise CPU. Use
54
+ `--device cuda` to require CUDA or `--device cpu` to select CPU. The source checkout's
55
+ uv lock selects CUDA 12.8 wheels. For uvx, add `--torch-backend cu128`
56
+ before `--from` to use that CUDA wheel index. For pip GPU installations:
57
+
58
+ ```sh
59
+ python -m pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
60
+ ```
61
+
62
+ CPU inference uses FP32 and can take seconds. CUDA uses a BF16 backbone with FP32
63
+ readouts. UI timings report model computation; uploads, startup and queueing add to the
64
+ user's wait. Published RTX 4060 Ti timings have their own recorded workload and scope.
65
+
66
+ ## Included interface
67
+
68
+ - Editable text, image and image-plus-policy examples.
69
+ - Six unmodified TrashNet sample photographs with MIT notices and hash provenance.
70
+ - `choice`, ordered `score`, and `noul` decisions.
71
+ - Image long-edge controls from 64 to 1,024 pixels, preserving aspect ratio.
72
+ - Probabilities, abstention and exact request/response JSON.
73
+ - Optional Star and support links. Installing or running QEV never stars a repository.
74
+
75
+ Each UI request contains one question and at most one image; the SDK supports up to
76
+ four questions. The model allows 2-16 candidates and a 2,048 processed-token budget.
77
+ Requests run one at a time, with up to eight waiting. Uploads are limited to 10 MB / 16
78
+ megapixels. Per-inference temporary images are removed; upload caches expire after
79
+ approximately ten minutes and are swept every five minutes. Analytics is disabled by default.
80
+
81
+ Sample photographs overlap development data and are examples, not independent benchmarks.
82
+ The packaged implementation and asset notices live in `src/qev/playground/`.
83
+
84
+ ## Earlier interfaces
85
+
86
+ `apps/hf_space/app.py` is a compatibility launcher for the packaged English UI. The
87
+ historical Korean interface and 2048 experiment remain in `apps/playground/` and are
88
+ available from the source checkout. See [its instructions](../apps/playground/README.md).
89
+
90
+ [GitHub: ken-jo/qev](https://github.com/ken-jo/qev)
load_qev.py CHANGED
@@ -1,11 +1,11 @@
1
- """Run the packaged example after installing runtime/*.whl and the pinned base."""
2
  import json
3
  from pathlib import Path
4
  import torch
5
- from qev import QEV, DecisionRequest
6
 
7
  root = Path(__file__).resolve().parent
8
  torch.set_num_threads(4)
9
- model = QEV.load(root, local_files_only=True, merge=True)
10
  request = DecisionRequest.from_json((root / "sample_request.json").read_text("utf-8"))
11
  print(json.dumps(model.predict(request, root), indent=2))
 
1
+ """Run the packaged example with the latest bundled qev wheel."""
2
  import json
3
  from pathlib import Path
4
  import torch
5
+ from qev import load, DecisionRequest
6
 
7
  root = Path(__file__).resolve().parent
8
  torch.set_num_threads(4)
9
+ model = load(root)
10
  request = DecisionRequest.from_json((root / "sample_request.json").read_text("utf-8"))
11
  print(json.dumps(model.predict(request, root), indent=2))
publication.json CHANGED
@@ -1,12 +1,14 @@
1
  {
2
  "project": "qev",
3
- "version": "0.1.1",
 
 
4
  "source_repository": "https://github.com/ken-jo/qev",
5
  "backbone": "Qwen/Qwen3.5-2B",
6
  "base_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
7
  "weights_sha256": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
8
  "manifest_sha256": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
9
- "runtime_wheel_sha256": "f1f4c54175ce213e98399cc8a3db91298e25aa1450f2306195614c80c87da8a2",
10
  "preserved_runtime_modules": 38,
11
  "original_prepared_checksums_sha256": "57a628ab1dbfa8526e8025a39ad1a922ae5c34b7535528f858b38917a4b4478d",
12
  "original_acceptance_preserved": true,
 
1
  {
2
  "project": "qev",
3
+ "version": "0.2.0",
4
+ "model_version": "0.1.1",
5
+ "runtime_wheel_file": "runtime/qev-0.2.0-py3-none-any.whl",
6
  "source_repository": "https://github.com/ken-jo/qev",
7
  "backbone": "Qwen/Qwen3.5-2B",
8
  "base_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
9
  "weights_sha256": "84b57aeeb987f73416ac0f796150957d1c5beb1ed373733053e46438f77a8fee",
10
  "manifest_sha256": "d83f9910196c6e801658ca3816c3d2bd4a849ba3da6ff059a0edf7c6028e898d",
11
+ "runtime_wheel_sha256": "33ca132521cf9ec7f5bc9405816cd65e9c4626853cfb21a82da7a4549ebbe576",
12
  "preserved_runtime_modules": 38,
13
  "original_prepared_checksums_sha256": "57a628ab1dbfa8526e8025a39ad1a922ae5c34b7535528f858b38917a4b4478d",
14
  "original_acceptance_preserved": true,
reproducibility/pyproject.toml CHANGED
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
4
 
5
  [project]
6
  name = "qev"
7
- version = "0.1.1"
8
  description = "LAYA-inspired typed decisions over text and images with Qwen3.5-2B"
9
  readme = "README.md"
10
  requires-python = ">=3.12,<3.13"
@@ -29,6 +29,9 @@ dependencies = [
29
  "pillow==12.3.0",
30
  "fastapi==0.141.1",
31
  "uvicorn==0.54.0",
 
 
 
32
  ]
33
 
34
  [project.scripts]
@@ -53,7 +56,7 @@ packages = ["src/veyra", "src/qev"]
53
  include = [
54
  "/src", "/examples/request.json", "/pyproject.toml", "/README.md",
55
  "/LICENSE", "/NOTICE", "/CHANGELOG.md", "/MODEL_CARD.md", "/docs/API.md",
56
- "/docs/ARCHITECTURE.md", "/docs/MODEL_SIZE.md",
57
  ]
58
 
59
  [[tool.uv.index]]
@@ -62,8 +65,8 @@ url = "https://download.pytorch.org/whl/cu128"
62
  explicit = true
63
 
64
  [tool.uv.sources]
65
- torch = { index = "pytorch-cu128" }
66
- torchvision = { index = "pytorch-cu128" }
67
 
68
  [tool.pytest.ini_options]
69
  testpaths = ["tests"]
 
4
 
5
  [project]
6
  name = "qev"
7
+ version = "0.2.0"
8
  description = "LAYA-inspired typed decisions over text and images with Qwen3.5-2B"
9
  readme = "README.md"
10
  requires-python = ">=3.12,<3.13"
 
29
  "pillow==12.3.0",
30
  "fastapi==0.141.1",
31
  "uvicorn==0.54.0",
32
+ "gradio==6.29.0",
33
+ "huggingface-hub>=1.0,<2",
34
+ "filelock>=3.16,<5",
35
  ]
36
 
37
  [project.scripts]
 
56
  include = [
57
  "/src", "/examples/request.json", "/pyproject.toml", "/README.md",
58
  "/LICENSE", "/NOTICE", "/CHANGELOG.md", "/MODEL_CARD.md", "/docs/API.md",
59
+ "/docs/ARCHITECTURE.md", "/docs/MODEL_SIZE.md", "/docs/PLAYGROUND.md", "/SUPPORT.md",
60
  ]
61
 
62
  [[tool.uv.index]]
 
65
  explicit = true
66
 
67
  [tool.uv.sources]
68
+ torch = { index = "pytorch-cu128", marker = "sys_platform == 'linux' or sys_platform == 'win32'" }
69
+ torchvision = { index = "pytorch-cu128", marker = "sys_platform == 'linux' or sys_platform == 'win32'" }
70
 
71
  [tool.pytest.ini_options]
72
  testpaths = ["tests"]
reproducibility/uv.lock CHANGED
@@ -1,6 +1,12 @@
1
  version = 1
2
  revision = 3
3
  requires-python = "==3.12.*"
 
 
 
 
 
 
4
 
5
  [[package]]
6
  name = "accelerate"
@@ -13,7 +19,8 @@ dependencies = [
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 = [
@@ -51,6 +58,24 @@ wheels = [
51
  { url = "https://files.pythonhosted.org/packages/12/b8/4bd346e22b28902df4d651910f5242c28d84e4a5c2435ca5c3f797ed7e2e/anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101", size = 132079, upload-time = "2026-09-05T10:42:37.923Z" },
52
  ]
53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
  [[package]]
55
  name = "certifi"
56
  version = "2026.7.22"
@@ -132,6 +157,69 @@ wheels = [
132
  { 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" },
133
  ]
134
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
135
  [[package]]
136
  name = "h11"
137
  version = "0.16.0"
@@ -141,6 +229,19 @@ wheels = [
141
  { 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" },
142
  ]
143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
144
  [[package]]
145
  name = "hf-xet"
146
  version = "1.6.0"
@@ -461,6 +562,25 @@ wheels = [
461
  { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" },
462
  ]
463
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
464
  [[package]]
465
  name = "packaging"
466
  version = "26.3"
@@ -470,6 +590,28 @@ wheels = [
470
  { 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" },
471
  ]
472
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473
  [[package]]
474
  name = "pillow"
475
  version = "12.3.0"
@@ -557,6 +699,15 @@ wheels = [
557
  { url = "https://files.pythonhosted.org/packages/fa/04/c81d4841331c2178b6fb09ae225425e110ed72d990c9fe556c4ec03d1013/pydantic_core-2.46.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8e24d8f05fa2d28513d94e877e9c75ad66175376209b3977f916e240e623193c", size = 2111034, upload-time = "2026-08-28T10:01:07.345Z" },
558
  ]
559
 
 
 
 
 
 
 
 
 
 
560
  [[package]]
561
  name = "pygments"
562
  version = "2.21.0"
@@ -582,6 +733,36 @@ wheels = [
582
  { url = "https://files.pythonhosted.org/packages/24/25/1de2678b631f5a49215c6c96fff41ba892b0a34df68d6d80292b1b48aa7f/pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c", size = 386536, upload-time = "2026-06-19T10:58:31.347Z" },
583
  ]
584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
585
  [[package]]
586
  name = "pyyaml"
587
  version = "6.0.3"
@@ -602,17 +783,22 @@ wheels = [
602
 
603
  [[package]]
604
  name = "qev"
605
- version = "0.1.1"
606
  source = { editable = "." }
607
  dependencies = [
608
  { name = "accelerate" },
609
  { name = "fastapi" },
 
 
 
610
  { name = "numpy" },
611
  { name = "pillow" },
612
  { name = "pydantic" },
613
  { name = "safetensors" },
614
- { name = "torch" },
615
- { name = "torchvision" },
 
 
616
  { name = "transformers" },
617
  { name = "uvicorn" },
618
  ]
@@ -628,12 +814,17 @@ dev = [
628
  requires-dist = [
629
  { name = "accelerate", specifier = "==1.15.0" },
630
  { name = "fastapi", specifier = "==0.141.1" },
 
 
 
631
  { name = "numpy", specifier = "==2.5.3" },
632
  { name = "pillow", specifier = "==12.3.0" },
633
  { name = "pydantic", specifier = "==2.13.5" },
634
  { name = "safetensors", specifier = "==0.8.0" },
635
- { name = "torch", specifier = "==2.10.0", index = "https://download.pytorch.org/whl/cu128" },
636
- { name = "torchvision", specifier = "==0.25.0", index = "https://download.pytorch.org/whl/cu128" },
 
 
637
  { name = "transformers", specifier = "==5.17.0" },
638
  { name = "uvicorn", specifier = "==0.54.0" },
639
  ]
@@ -707,6 +898,18 @@ wheels = [
707
  { url = "https://files.pythonhosted.org/packages/51/60/5fb1a39dbb5ae314d5f59bc7348a63c1d5c20f3cd83914c4b5cb0be31d2d/ruff-0.16.9-py3-none-win_arm64.whl", hash = "sha256:ed1a252039200f57a59eebc063b54beabea67bfbaaca0eeaa7f54b5fbcda2284", size = 10458649, upload-time = "2026-09-24T20:37:46.882Z" },
708
  ]
709
 
 
 
 
 
 
 
 
 
 
 
 
 
710
  [[package]]
711
  name = "safetensors"
712
  version = "0.8.0"
@@ -731,6 +934,15 @@ wheels = [
731
  { 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" },
732
  ]
733
 
 
 
 
 
 
 
 
 
 
734
  [[package]]
735
  name = "setuptools"
736
  version = "84.0.0"
@@ -749,6 +961,15 @@ wheels = [
749
  { 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" },
750
  ]
751
 
 
 
 
 
 
 
 
 
 
752
  [[package]]
753
  name = "starlette"
754
  version = "1.7.0"
@@ -801,34 +1022,69 @@ wheels = [
801
  { url = "https://files.pythonhosted.org/packages/d7/b0/dee84cb44175be1b4c35bd2f770727494e78f0bb38e571a623ade94dbebb/tokenizers-0.23.2-cp310-abi3-win_arm64.whl", hash = "sha256:e49c394456dd9985787fec76132438ba3fb8911f857b1bf3d40119f9292d41aa", size = 2729352, upload-time = "2026-09-03T08:55:44.345Z" },
802
  ]
803
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
804
  [[package]]
805
  name = "torch"
806
  version = "2.10.0+cu128"
807
  source = { registry = "https://download.pytorch.org/whl/cu128" }
 
 
 
 
808
  dependencies = [
809
- { name = "cuda-bindings", marker = "sys_platform == 'linux'" },
810
  { name = "filelock" },
811
  { name = "fsspec" },
812
  { name = "jinja2" },
813
  { name = "networkx" },
814
- { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" },
815
- { name = "nvidia-cuda-cupti-cu12", marker = "sys_platform == 'linux'" },
816
- { name = "nvidia-cuda-nvrtc-cu12", marker = "sys_platform == 'linux'" },
817
- { name = "nvidia-cuda-runtime-cu12", marker = "sys_platform == 'linux'" },
818
- { name = "nvidia-cudnn-cu12", marker = "sys_platform == 'linux'" },
819
- { name = "nvidia-cufft-cu12", marker = "sys_platform == 'linux'" },
820
- { name = "nvidia-cufile-cu12", marker = "sys_platform == 'linux'" },
821
- { name = "nvidia-curand-cu12", marker = "sys_platform == 'linux'" },
822
- { name = "nvidia-cusolver-cu12", marker = "sys_platform == 'linux'" },
823
- { name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux'" },
824
- { name = "nvidia-cusparselt-cu12", marker = "sys_platform == 'linux'" },
825
- { name = "nvidia-nccl-cu12", marker = "sys_platform == 'linux'" },
826
- { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" },
827
- { name = "nvidia-nvshmem-cu12", marker = "sys_platform == 'linux'" },
828
- { name = "nvidia-nvtx-cu12", marker = "sys_platform == 'linux'" },
829
  { name = "setuptools" },
830
  { name = "sympy" },
831
- { name = "triton", marker = "sys_platform == 'linux'" },
832
  { name = "typing-extensions" },
833
  ]
834
  wheels = [
@@ -837,14 +1093,35 @@ wheels = [
837
  { url = "https://download-r2.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp312-cp312-win_amd64.whl", hash = "sha256:fbde8f6a9ec8c76979a0d14df21c10b9e5cab6f0d106a73ca73e2179bc597cae", upload-time = "2026-01-21T15:22:17Z" },
838
  ]
839
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
840
  [[package]]
841
  name = "torchvision"
842
  version = "0.25.0+cu128"
843
  source = { registry = "https://download.pytorch.org/whl/cu128" }
 
 
 
 
844
  dependencies = [
845
  { name = "numpy" },
846
  { name = "pillow" },
847
- { name = "torch" },
848
  ]
849
  wheels = [
850
  { url = "https://download-r2.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:8623e534ef6a815bd6407d4b52dd70c7154e2eda626ad4b9cb895d36c5a3305b", upload-time = "2026-01-21T22:32:23Z" },
@@ -929,6 +1206,15 @@ wheels = [
929
  { url = "https://files.pythonhosted.org/packages/67/81/4add07e5172b7ac40d8ed5ff580409a7801a4fe26d529bdd915401dabfbe/typing_inspection-0.4.4-py3-none-any.whl", hash = "sha256:65b8397ba37ccbce054456aaccddfc91e6e3083c92824df348d96ca832f3f147", size = 14750, upload-time = "2026-08-12T12:37:24.648Z" },
930
  ]
931
 
 
 
 
 
 
 
 
 
 
932
  [[package]]
933
  name = "uvicorn"
934
  version = "0.54.0"
 
1
  version = 1
2
  revision = 3
3
  requires-python = "==3.12.*"
4
+ resolution-markers = [
5
+ "sys_platform == 'win32'",
6
+ "sys_platform == 'emscripten'",
7
+ "sys_platform == 'linux'",
8
+ "sys_platform != 'emscripten' and sys_platform != 'linux' and sys_platform != 'win32'",
9
+ ]
10
 
11
  [[package]]
12
  name = "accelerate"
 
19
  { name = "psutil" },
20
  { name = "pyyaml" },
21
  { name = "safetensors" },
22
+ { name = "torch", version = "2.10.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform != 'linux' and sys_platform != 'win32'" },
23
+ { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
24
  ]
25
  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" }
26
  wheels = [
 
58
  { url = "https://files.pythonhosted.org/packages/12/b8/4bd346e22b28902df4d651910f5242c28d84e4a5c2435ca5c3f797ed7e2e/anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101", size = 132079, upload-time = "2026-09-05T10:42:37.923Z" },
59
  ]
60
 
61
+ [[package]]
62
+ name = "brotli"
63
+ version = "1.2.0"
64
+ source = { registry = "https://pypi.org/simple" }
65
+ sdist = { url = "https://files.pythonhosted.org/packages/f7/16/c92ca344d646e71a43b8bb353f0a6490d7f6e06210f8554c8f874e454285/brotli-1.2.0.tar.gz", hash = "sha256:e310f77e41941c13340a95976fe66a8a95b01e783d430eeaf7a2f87e0a57dd0a", size = 7388632, upload-time = "2025-11-05T18:39:42.86Z" }
66
+ wheels = [
67
+ { url = "https://files.pythonhosted.org/packages/11/ee/b0a11ab2315c69bb9b45a2aaed022499c9c24a205c3a49c3513b541a7967/brotli-1.2.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:35d382625778834a7f3061b15423919aa03e4f5da34ac8e02c074e4b75ab4f84", size = 861543, upload-time = "2025-11-05T18:38:24.183Z" },
68
+ { url = "https://files.pythonhosted.org/packages/e1/2f/29c1459513cd35828e25531ebfcbf3e92a5e49f560b1777a9af7203eb46e/brotli-1.2.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7a61c06b334bd99bc5ae84f1eeb36bfe01400264b3c352f968c6e30a10f9d08b", size = 444288, upload-time = "2025-11-05T18:38:25.139Z" },
69
+ { url = "https://files.pythonhosted.org/packages/3d/6f/feba03130d5fceadfa3a1bb102cb14650798c848b1df2a808356f939bb16/brotli-1.2.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:acec55bb7c90f1dfc476126f9711a8e81c9af7fb617409a9ee2953115343f08d", size = 1528071, upload-time = "2025-11-05T18:38:26.081Z" },
70
+ { url = "https://files.pythonhosted.org/packages/2b/38/f3abb554eee089bd15471057ba85f47e53a44a462cfce265d9bf7088eb09/brotli-1.2.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:260d3692396e1895c5034f204f0db022c056f9e2ac841593a4cf9426e2a3faca", size = 1626913, upload-time = "2025-11-05T18:38:27.284Z" },
71
+ { url = "https://files.pythonhosted.org/packages/03/a7/03aa61fbc3c5cbf99b44d158665f9b0dd3d8059be16c460208d9e385c837/brotli-1.2.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:072e7624b1fc4d601036ab3f4f27942ef772887e876beff0301d261210bca97f", size = 1419762, upload-time = "2025-11-05T18:38:28.295Z" },
72
+ { url = "https://files.pythonhosted.org/packages/21/1b/0374a89ee27d152a5069c356c96b93afd1b94eae83f1e004b57eb6ce2f10/brotli-1.2.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:adedc4a67e15327dfdd04884873c6d5a01d3e3b6f61406f99b1ed4865a2f6d28", size = 1484494, upload-time = "2025-11-05T18:38:29.29Z" },
73
+ { url = "https://files.pythonhosted.org/packages/cf/57/69d4fe84a67aef4f524dcd075c6eee868d7850e85bf01d778a857d8dbe0a/brotli-1.2.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:7a47ce5c2288702e09dc22a44d0ee6152f2c7eda97b3c8482d826a1f3cfc7da7", size = 1593302, upload-time = "2025-11-05T18:38:30.639Z" },
74
+ { url = "https://files.pythonhosted.org/packages/d5/3b/39e13ce78a8e9a621c5df3aeb5fd181fcc8caba8c48a194cd629771f6828/brotli-1.2.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:af43b8711a8264bb4e7d6d9a6d004c3a2019c04c01127a868709ec29962b6036", size = 1487913, upload-time = "2025-11-05T18:38:31.618Z" },
75
+ { url = "https://files.pythonhosted.org/packages/62/28/4d00cb9bd76a6357a66fcd54b4b6d70288385584063f4b07884c1e7286ac/brotli-1.2.0-cp312-cp312-win32.whl", hash = "sha256:e99befa0b48f3cd293dafeacdd0d191804d105d279e0b387a32054c1180f3161", size = 334362, upload-time = "2025-11-05T18:38:32.939Z" },
76
+ { url = "https://files.pythonhosted.org/packages/1c/4e/bc1dcac9498859d5e353c9b153627a3752868a9d5f05ce8dedd81a2354ab/brotli-1.2.0-cp312-cp312-win_amd64.whl", hash = "sha256:b35c13ce241abdd44cb8ca70683f20c0c079728a36a996297adb5334adfc1c44", size = 369115, upload-time = "2025-11-05T18:38:33.765Z" },
77
+ ]
78
+
79
  [[package]]
80
  name = "certifi"
81
  version = "2026.7.22"
 
157
  { 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" },
158
  ]
159
 
160
+ [[package]]
161
+ name = "gradio"
162
+ version = "6.29.0"
163
+ source = { registry = "https://pypi.org/simple" }
164
+ dependencies = [
165
+ { name = "anyio" },
166
+ { name = "brotli" },
167
+ { name = "fastapi" },
168
+ { name = "gradio-client" },
169
+ { name = "groovy" },
170
+ { name = "hf-gradio" },
171
+ { name = "httpx" },
172
+ { name = "huggingface-hub" },
173
+ { name = "jinja2" },
174
+ { name = "markupsafe" },
175
+ { name = "numpy" },
176
+ { name = "orjson" },
177
+ { name = "packaging" },
178
+ { name = "pandas" },
179
+ { name = "pillow" },
180
+ { name = "pydantic" },
181
+ { name = "pydub" },
182
+ { name = "python-multipart" },
183
+ { name = "pytz" },
184
+ { name = "pyyaml" },
185
+ { name = "safehttpx" },
186
+ { name = "semantic-version" },
187
+ { name = "starlette" },
188
+ { name = "tomlkit" },
189
+ { name = "typer" },
190
+ { name = "typing-extensions" },
191
+ { name = "uvicorn" },
192
+ ]
193
+ sdist = { url = "https://files.pythonhosted.org/packages/0a/d0/409bc699f9b02a429df209c104547005491bd17dfe5f55c43396237dd352/gradio-6.29.0.tar.gz", hash = "sha256:ed2d04c49199a5d6c4d40e9ba94fb3d4075e114075904c2c7dc59c2b06122de6", size = 45107350, upload-time = "2026-09-29T17:59:55.104Z" }
194
+ wheels = [
195
+ { url = "https://files.pythonhosted.org/packages/0f/1c/2d4bc21859fb2ddefdcf2291b947aa96704d7738343a5f3ec232bd5c6127/gradio-6.29.0-py3-none-any.whl", hash = "sha256:0325d8d48fba896ae49f06d775fc93c3ca592709e1ec908bcc0d1631b75fa593", size = 31379649, upload-time = "2026-09-29T17:59:51.11Z" },
196
+ ]
197
+
198
+ [[package]]
199
+ name = "gradio-client"
200
+ version = "2.7.1"
201
+ source = { registry = "https://pypi.org/simple" }
202
+ dependencies = [
203
+ { name = "fsspec" },
204
+ { name = "httpx" },
205
+ { name = "huggingface-hub" },
206
+ { name = "packaging" },
207
+ { name = "typing-extensions" },
208
+ ]
209
+ sdist = { url = "https://files.pythonhosted.org/packages/14/dd/bfc46f5c671f4be56949fd26a062972d63b03aa6fc799c7aacb004a48807/gradio_client-2.7.1.tar.gz", hash = "sha256:2c629110496dbca67a848cd70add8cf5c3a24776404a2c5ef4fd759c7c2c33c4", size = 62702, upload-time = "2026-09-18T11:58:23.469Z" }
210
+ wheels = [
211
+ { url = "https://files.pythonhosted.org/packages/03/8a/dea48fd8863da2a920e6f83c3ce9b0a00dbb5517cd77b834746812d4687e/gradio_client-2.7.1-py3-none-any.whl", hash = "sha256:b4f953212a66eeb7fa12d93fcf98fa90a0567c2c9c6a5fc2795610f7c71e21b5", size = 63574, upload-time = "2026-09-18T11:58:21.845Z" },
212
+ ]
213
+
214
+ [[package]]
215
+ name = "groovy"
216
+ version = "0.1.2"
217
+ source = { registry = "https://pypi.org/simple" }
218
+ sdist = { url = "https://files.pythonhosted.org/packages/52/36/bbdede67400277bef33d3ec0e6a31750da972c469f75966b4930c753218f/groovy-0.1.2.tar.gz", hash = "sha256:25c1dc09b3f9d7e292458aa762c6beb96ea037071bf5e917fc81fb78d2231083", size = 17325, upload-time = "2025-02-28T20:24:56.068Z" }
219
+ wheels = [
220
+ { url = "https://files.pythonhosted.org/packages/28/27/3d6dcadc8a3214d8522c1e7f6a19554e33659be44546d44a2f7572ac7d2a/groovy-0.1.2-py3-none-any.whl", hash = "sha256:7f7975bab18c729a257a8b1ae9dcd70b7cafb1720481beae47719af57c35fa64", size = 14090, upload-time = "2025-02-28T20:24:55.152Z" },
221
+ ]
222
+
223
  [[package]]
224
  name = "h11"
225
  version = "0.16.0"
 
229
  { 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" },
230
  ]
231
 
232
+ [[package]]
233
+ name = "hf-gradio"
234
+ version = "0.4.1"
235
+ source = { registry = "https://pypi.org/simple" }
236
+ dependencies = [
237
+ { name = "gradio-client" },
238
+ { name = "typer" },
239
+ ]
240
+ sdist = { url = "https://files.pythonhosted.org/packages/ce/86/c9694b7cfada5780e75769e60dc161a161f4dd7fc91b61db5e3a3338bef9/hf_gradio-0.4.1.tar.gz", hash = "sha256:a017d942618f0d495a58ee4563047fa04bef614c00e0cb789a9a6d0633cffa7b", size = 6560, upload-time = "2026-04-22T14:01:32.334Z" }
241
+ wheels = [
242
+ { url = "https://files.pythonhosted.org/packages/30/2d/afff2ee87e75d8eb85c92bb8cf0e15b05c23c2ebd8fd8dec781d8601ed7f/hf_gradio-0.4.1-py3-none-any.whl", hash = "sha256:76b8cb8be6abe62d74c1ad2d35b42f0629db89aa9e1a8d033cecfe7c856eeab3", size = 4482, upload-time = "2026-04-17T19:53:31.827Z" },
243
+ ]
244
+
245
  [[package]]
246
  name = "hf-xet"
247
  version = "1.6.0"
 
562
  { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" },
563
  ]
564
 
565
+ [[package]]
566
+ name = "orjson"
567
+ version = "3.12.0"
568
+ source = { registry = "https://pypi.org/simple" }
569
+ sdist = { url = "https://files.pythonhosted.org/packages/0f/f3/742fb1f62b825f2c010697eaf4e828004bc2a81e7e806666989c132c7c42/orjson-3.12.0.tar.gz", hash = "sha256:d14203fb1aae2ad9b3d52f8a0e82aeb10197ef1c9bc61da7f358bd70b00123d5", size = 4142915, upload-time = "2026-08-14T16:13:30.607Z" }
570
+ wheels = [
571
+ { url = "https://files.pythonhosted.org/packages/be/4a/295da39c651c2faac8bd351a2a346f0fdedd9d50b847ee9dfc27d2207ef6/orjson-3.12.0-cp312-cp312-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:aa3e43a6846e91d7bde3d5a9c66090fcd8744f569a9b6cffc5e1ca38f6a461c0", size = 223427, upload-time = "2026-08-14T16:12:28.525Z" },
572
+ { url = "https://files.pythonhosted.org/packages/29/98/758cf90fbeaaafb7f8141bfac75a432099959f3a2f5db93a412e876415d8/orjson-3.12.0-cp312-cp312-macosx_15_0_arm64.whl", hash = "sha256:11edb4660a6680abee9788a3a9072208a2c96538cc1322bd79542065229d8e54", size = 123725, upload-time = "2026-08-14T16:12:30.013Z" },
573
+ { url = "https://files.pythonhosted.org/packages/32/b5/5b934d251f8651f7e41df180ad0c57a6e1cabe15c7bd331638413a50ebc9/orjson-3.12.0-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.whl", hash = "sha256:2d3a9da945a4d96ae758fdaaca56742e6b73b6fd554c5d8876f252a6dad70b83", size = 113375, upload-time = "2026-08-14T16:12:31.209Z" },
574
+ { url = "https://files.pythonhosted.org/packages/cd/d2/37efb5b12a176ce3ced29f4144f20da57d02757f78ce549637dc1b4e1fc8/orjson-3.12.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:92ffc09e07233a6ab6d4e067f7841edcbcc134cb4812155cf171ea5255a421d7", size = 129983, upload-time = "2026-08-14T16:12:32.721Z" },
575
+ { url = "https://files.pythonhosted.org/packages/50/22/0644b87c73f13e0092df8f35a1fe280d991e5e90072087411e0dd7e44e0c/orjson-3.12.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bf44e374aadde77b1f6109f1030be51433eb61984379852766b6f4e187db7b1e", size = 130629, upload-time = "2026-08-14T16:12:34.084Z" },
576
+ { url = "https://files.pythonhosted.org/packages/8c/57/80b986ebfecd9c6a177ddf1c2319717f0cd8feffb2b78946595a18a2fc88/orjson-3.12.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1192a7021b6d071aaf909864f6e924d6a2675ca360485b972b8401749311750b", size = 131245, upload-time = "2026-08-14T16:12:35.713Z" },
577
+ { url = "https://files.pythonhosted.org/packages/80/3d/75c5ac5a69161f44492a68fbdde66f4cc4ce48cd5e1fb05918e46f0c8848/orjson-3.12.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:53c0c474a9d9aff9aebfc0c88de1f28f843d940e6e3a80729abdf6a20274356f", size = 135397, upload-time = "2026-08-14T16:12:37.128Z" },
578
+ { url = "https://files.pythonhosted.org/packages/71/93/4d71f2df314a97ff0d27a4559bf5888fc8406e3c6dec90e92291e3511215/orjson-3.12.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:532ff8cd4bd59a327a953a7dcde922c7fc25b85e29721bb8633265430d3a3873", size = 127693, upload-time = "2026-08-14T16:12:38.627Z" },
579
+ { url = "https://files.pythonhosted.org/packages/bc/1d/0dbc6be5adfd1730491072fb60beb6bcdf5d7b2596ee41b7fc2e298bfc09/orjson-3.12.0-cp312-cp312-win32.whl", hash = "sha256:a6cf4b18e7de173f209f2084ffbd736dd72389a396326ee80a7022168be232e5", size = 128000, upload-time = "2026-08-14T16:12:39.954Z" },
580
+ { url = "https://files.pythonhosted.org/packages/2d/c9/97b1ce0112ebf5e949c775ed5b1755e562233179f3584579673cc24d6378/orjson-3.12.0-cp312-cp312-win_amd64.whl", hash = "sha256:010811c1b69773450a01cef97727a67b223242f350b77d4ca000e59a9ef2155a", size = 122106, upload-time = "2026-08-14T16:12:41.324Z" },
581
+ { url = "https://files.pythonhosted.org/packages/a8/6a/facd8b312e4a0d3a7fa978c7e15821f74a336adf1d65529faec33b48e18b/orjson-3.12.0-cp312-cp312-win_arm64.whl", hash = "sha256:ad29eece0c601737f2a60edc2752a84e7a0785df3efb62e3012834700a5afe0d", size = 126869, upload-time = "2026-08-14T16:12:42.651Z" },
582
+ ]
583
+
584
  [[package]]
585
  name = "packaging"
586
  version = "26.3"
 
590
  { 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" },
591
  ]
592
 
593
+ [[package]]
594
+ name = "pandas"
595
+ version = "3.0.6"
596
+ source = { registry = "https://pypi.org/simple" }
597
+ dependencies = [
598
+ { name = "numpy" },
599
+ { name = "python-dateutil" },
600
+ { name = "tzdata", marker = "sys_platform == 'emscripten' or sys_platform == 'win32'" },
601
+ ]
602
+ sdist = { url = "https://files.pythonhosted.org/packages/e2/17/d7b106e05bfa642e8694451e7d3d759c6a241c5386a5d962e4f66c047e06/pandas-3.0.6.tar.gz", hash = "sha256:66b07ef7315a31bfe1089cd3d71a7de781c9dca986762d0b4fe7c0ef17465d10", size = 4667686, upload-time = "2026-09-17T23:23:18.345Z" }
603
+ wheels = [
604
+ { url = "https://files.pythonhosted.org/packages/77/4c/597d588c055d4373cff19cbbc32d4dd046c7be8fadee957585b5ba9e5b24/pandas-3.0.6-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7dac2d65e9087e8e7b5a45fe15c4920911a221df061ab629943ce016489145c7", size = 10397411, upload-time = "2026-09-17T23:20:49.465Z" },
605
+ { url = "https://files.pythonhosted.org/packages/18/8f/48907c7c707b61a8e5018c32e1a3f70623209bfb59020a2f6196159d3aa7/pandas-3.0.6-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:9dab635a549e58a053c7b0fa054dc0bd7be22f0ed9a720f4a85d5fb993276172", size = 10051279, upload-time = "2026-09-17T23:20:52.409Z" },
606
+ { url = "https://files.pythonhosted.org/packages/67/fa/613d867c3d9554a61bafdec6f79565c8a3e73235feb52cc4a72ad2e0fa6a/pandas-3.0.6-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e3dccb584123b399c07562ac4d62543e90ede49ddf8ce3c13ffc64cbe828c281", size = 10303429, upload-time = "2026-09-17T23:20:55.597Z" },
607
+ { url = "https://files.pythonhosted.org/packages/cb/67/0c0f18e38d7f2d2af8c24b3315bc4046e73bbdd4a5540405506671ad0c0d/pandas-3.0.6-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0704044b676496b8350e023b09f174a26772456c974a2b11c36bebb558c9490d", size = 10788193, upload-time = "2026-09-17T23:20:58.617Z" },
608
+ { url = "https://files.pythonhosted.org/packages/39/53/1b57f3162501fe36687e4870e1b918a6458ca7386b173af75663ace95857/pandas-3.0.6-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:e7c1905ef02c3d6d43d9dbd5b6ccb4da4870a0b0c821bbc103fbdb6f3ad2707b", size = 11323501, upload-time = "2026-09-17T23:21:01.911Z" },
609
+ { url = "https://files.pythonhosted.org/packages/f2/d2/b1182e8d39100369d7f13f4a125a3fb6b096fef112c46d0566c25781ff68/pandas-3.0.6-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:569e114072b24fc4970c12e2b4bab252671668a40b324318903380cab0254c0c", size = 11838173, upload-time = "2026-09-17T23:21:04.85Z" },
610
+ { url = "https://files.pythonhosted.org/packages/c7/33/5b717af24d2f27995e51e0875a269dddd373045216e34cf62e3aa764eaa1/pandas-3.0.6-cp312-cp312-pyemscripten_2024_0_wasm32.whl", hash = "sha256:2a8fc94be2ee5f1d86f97aacd8cc566f81680b6498e76f3007421bb5d98151bf", size = 7118499, upload-time = "2026-09-17T23:21:07.661Z" },
611
+ { url = "https://files.pythonhosted.org/packages/bd/2a/14b3b17cd75cef4a1ee1af4234eb41cc1afe4103b98c2d80e8916abfd42b/pandas-3.0.6-cp312-cp312-win_amd64.whl", hash = "sha256:3ef908d28590b3f42d7070e7ad8f9b34b442b260b7f3c1afb57e0040c58cdb1b", size = 9661439, upload-time = "2026-09-17T23:21:10.959Z" },
612
+ { url = "https://files.pythonhosted.org/packages/3b/11/3d580a604a1e35d69f6676847bd12db7d14344bc677b717f4413e79c5d0d/pandas-3.0.6-cp312-cp312-win_arm64.whl", hash = "sha256:f4e7c52eb108d752e7592268108fd3e98efd76d83a3125cdd06c621c2e44359b", size = 8959806, upload-time = "2026-09-17T23:21:13.851Z" },
613
+ ]
614
+
615
  [[package]]
616
  name = "pillow"
617
  version = "12.3.0"
 
699
  { url = "https://files.pythonhosted.org/packages/fa/04/c81d4841331c2178b6fb09ae225425e110ed72d990c9fe556c4ec03d1013/pydantic_core-2.46.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8e24d8f05fa2d28513d94e877e9c75ad66175376209b3977f916e240e623193c", size = 2111034, upload-time = "2026-08-28T10:01:07.345Z" },
700
  ]
701
 
702
+ [[package]]
703
+ name = "pydub"
704
+ version = "0.25.1"
705
+ source = { registry = "https://pypi.org/simple" }
706
+ sdist = { url = "https://files.pythonhosted.org/packages/fe/9a/e6bca0eed82db26562c73b5076539a4a08d3cffd19c3cc5913a3e61145fd/pydub-0.25.1.tar.gz", hash = "sha256:980a33ce9949cab2a569606b65674d748ecbca4f0796887fd6f46173a7b0d30f", size = 38326, upload-time = "2021-03-10T02:09:54.659Z" }
707
+ wheels = [
708
+ { url = "https://files.pythonhosted.org/packages/a6/53/d78dc063216e62fc55f6b2eebb447f6a4b0a59f55c8406376f76bf959b08/pydub-0.25.1-py2.py3-none-any.whl", hash = "sha256:65617e33033874b59d87db603aa1ed450633288aefead953b30bded59cb599a6", size = 32327, upload-time = "2021-03-10T02:09:53.503Z" },
709
+ ]
710
+
711
  [[package]]
712
  name = "pygments"
713
  version = "2.21.0"
 
733
  { url = "https://files.pythonhosted.org/packages/24/25/1de2678b631f5a49215c6c96fff41ba892b0a34df68d6d80292b1b48aa7f/pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c", size = 386536, upload-time = "2026-06-19T10:58:31.347Z" },
734
  ]
735
 
736
+ [[package]]
737
+ name = "python-dateutil"
738
+ version = "2.9.0.post0"
739
+ source = { registry = "https://pypi.org/simple" }
740
+ dependencies = [
741
+ { name = "six" },
742
+ ]
743
+ sdist = { url = "https://files.pythonhosted.org/packages/66/c0/0c8b6ad9f17a802ee498c46e004a0eb49bc148f2fd230864601a86dcf6db/python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3", size = 342432, upload-time = "2024-03-01T18:36:20.211Z" }
744
+ wheels = [
745
+ { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" },
746
+ ]
747
+
748
+ [[package]]
749
+ name = "python-multipart"
750
+ version = "0.0.32"
751
+ source = { registry = "https://pypi.org/simple" }
752
+ sdist = { url = "https://files.pythonhosted.org/packages/5b/42/55c32bb9b12693c092ad250a0e82edb5b31ddeda6eb772de5f308b3804ad/python_multipart-0.0.32.tar.gz", hash = "sha256:be54b7f3fa167bb83e4fcd936b887b708f4e57fe75911c02aebf53efaf8d938e", size = 46881, upload-time = "2026-06-04T16:18:58.647Z" }
753
+ wheels = [
754
+ { url = "https://files.pythonhosted.org/packages/e1/04/e8135ebd1ad02c56ec633277529b2602ff99ff634be76cdba5744cf554fd/python_multipart-0.0.32-py3-none-any.whl", hash = "sha256:ff6d3f776f16878c894e52e107296ffc890e913c611b1a4ec6c44e2821fe2e23", size = 30042, upload-time = "2026-06-04T16:18:57.319Z" },
755
+ ]
756
+
757
+ [[package]]
758
+ name = "pytz"
759
+ version = "2026.4"
760
+ source = { registry = "https://pypi.org/simple" }
761
+ sdist = { url = "https://files.pythonhosted.org/packages/b0/ed/fa23d28713004418bbf2407127f80f385bfb0bf4e677bef02c25c27b00ee/pytz-2026.4.tar.gz", hash = "sha256:464303645bafafd72418898368b2429458f709cf1eb6a15372fbcc396b64da63", size = 318444, upload-time = "2026-09-24T03:58:40.443Z" }
762
+ wheels = [
763
+ { url = "https://files.pythonhosted.org/packages/d5/11/e3df59614bebe0ca50c905442603eda64ae7ac19811d4e4f74aa2d51ca85/pytz-2026.4-py2.py3-none-any.whl", hash = "sha256:9d514388fbc89ca0833203464272ac485b8828568ab73532f5020f17e892a0ff", size = 506747, upload-time = "2026-09-24T03:58:38.629Z" },
764
+ ]
765
+
766
  [[package]]
767
  name = "pyyaml"
768
  version = "6.0.3"
 
783
 
784
  [[package]]
785
  name = "qev"
786
+ version = "0.2.0"
787
  source = { editable = "." }
788
  dependencies = [
789
  { name = "accelerate" },
790
  { name = "fastapi" },
791
+ { name = "filelock" },
792
+ { name = "gradio" },
793
+ { name = "huggingface-hub" },
794
  { name = "numpy" },
795
  { name = "pillow" },
796
  { name = "pydantic" },
797
  { name = "safetensors" },
798
+ { name = "torch", version = "2.10.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform != 'linux' and sys_platform != 'win32'" },
799
+ { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
800
+ { name = "torchvision", version = "0.25.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform != 'linux' and sys_platform != 'win32'" },
801
+ { name = "torchvision", version = "0.25.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
802
  { name = "transformers" },
803
  { name = "uvicorn" },
804
  ]
 
814
  requires-dist = [
815
  { name = "accelerate", specifier = "==1.15.0" },
816
  { name = "fastapi", specifier = "==0.141.1" },
817
+ { name = "filelock", specifier = ">=3.16,<5" },
818
+ { name = "gradio", specifier = "==6.29.0" },
819
+ { name = "huggingface-hub", specifier = ">=1.0,<2" },
820
  { name = "numpy", specifier = "==2.5.3" },
821
  { name = "pillow", specifier = "==12.3.0" },
822
  { name = "pydantic", specifier = "==2.13.5" },
823
  { name = "safetensors", specifier = "==0.8.0" },
824
+ { name = "torch", marker = "sys_platform != 'linux' and sys_platform != 'win32'", specifier = "==2.10.0" },
825
+ { name = "torch", marker = "sys_platform == 'linux' or sys_platform == 'win32'", specifier = "==2.10.0", index = "https://download.pytorch.org/whl/cu128" },
826
+ { name = "torchvision", marker = "sys_platform != 'linux' and sys_platform != 'win32'", specifier = "==0.25.0" },
827
+ { name = "torchvision", marker = "sys_platform == 'linux' or sys_platform == 'win32'", specifier = "==0.25.0", index = "https://download.pytorch.org/whl/cu128" },
828
  { name = "transformers", specifier = "==5.17.0" },
829
  { name = "uvicorn", specifier = "==0.54.0" },
830
  ]
 
898
  { url = "https://files.pythonhosted.org/packages/51/60/5fb1a39dbb5ae314d5f59bc7348a63c1d5c20f3cd83914c4b5cb0be31d2d/ruff-0.16.9-py3-none-win_arm64.whl", hash = "sha256:ed1a252039200f57a59eebc063b54beabea67bfbaaca0eeaa7f54b5fbcda2284", size = 10458649, upload-time = "2026-09-24T20:37:46.882Z" },
899
  ]
900
 
901
+ [[package]]
902
+ name = "safehttpx"
903
+ version = "0.1.7"
904
+ source = { registry = "https://pypi.org/simple" }
905
+ dependencies = [
906
+ { name = "httpx" },
907
+ ]
908
+ sdist = { url = "https://files.pythonhosted.org/packages/89/d1/4282284d9cf1ee873607a46442da977fc3c985059315ab23610be31d5885/safehttpx-0.1.7.tar.gz", hash = "sha256:db201c0978c41eddb8bb480f3eee59dd67304fdd91646035e9d9a720049a9d23", size = 10385, upload-time = "2025-10-24T18:30:09.783Z" }
909
+ wheels = [
910
+ { url = "https://files.pythonhosted.org/packages/2e/a3/0f0b7d78e2f1eb9e8e1afbff1d2bff8d60144aee17aca51c065b516743dd/safehttpx-0.1.7-py3-none-any.whl", hash = "sha256:c4f4a162db6993464d7ca3d7cc4af0ffc6515a606dfd220b9f82c6945d869cde", size = 8959, upload-time = "2025-10-24T18:30:08.733Z" },
911
+ ]
912
+
913
  [[package]]
914
  name = "safetensors"
915
  version = "0.8.0"
 
934
  { 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" },
935
  ]
936
 
937
+ [[package]]
938
+ name = "semantic-version"
939
+ version = "2.10.0"
940
+ source = { registry = "https://pypi.org/simple" }
941
+ sdist = { url = "https://files.pythonhosted.org/packages/7d/31/f2289ce78b9b473d582568c234e104d2a342fd658cc288a7553d83bb8595/semantic_version-2.10.0.tar.gz", hash = "sha256:bdabb6d336998cbb378d4b9db3a4b56a1e3235701dc05ea2690d9a997ed5041c", size = 52289, upload-time = "2022-05-26T13:35:23.454Z" }
942
+ wheels = [
943
+ { url = "https://files.pythonhosted.org/packages/6a/23/8146aad7d88f4fcb3a6218f41a60f6c2d4e3a72de72da1825dc7c8f7877c/semantic_version-2.10.0-py2.py3-none-any.whl", hash = "sha256:de78a3b8e0feda74cabc54aab2da702113e33ac9d9eb9d2389bcf1f58b7d9177", size = 15552, upload-time = "2022-05-26T13:35:21.206Z" },
944
+ ]
945
+
946
  [[package]]
947
  name = "setuptools"
948
  version = "84.0.0"
 
961
  { 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" },
962
  ]
963
 
964
+ [[package]]
965
+ name = "six"
966
+ version = "1.17.0"
967
+ source = { registry = "https://pypi.org/simple" }
968
+ sdist = { url = "https://files.pythonhosted.org/packages/94/e7/b2c673351809dca68a0e064b6af791aa332cf192da575fd474ed7d6f16a2/six-1.17.0.tar.gz", hash = "sha256:ff70335d468e7eb6ec65b95b99d3a2836546063f63acc5171de367e834932a81", size = 34031, upload-time = "2024-12-04T17:35:28.174Z" }
969
+ wheels = [
970
+ { url = "https://files.pythonhosted.org/packages/b7/ce/149a00dd41f10bc29e5921b496af8b574d8413afcd5e30dfa0ed46c2cc5e/six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274", size = 11050, upload-time = "2024-12-04T17:35:26.475Z" },
971
+ ]
972
+
973
  [[package]]
974
  name = "starlette"
975
  version = "1.7.0"
 
1022
  { url = "https://files.pythonhosted.org/packages/d7/b0/dee84cb44175be1b4c35bd2f770727494e78f0bb38e571a623ade94dbebb/tokenizers-0.23.2-cp310-abi3-win_arm64.whl", hash = "sha256:e49c394456dd9985787fec76132438ba3fb8911f857b1bf3d40119f9292d41aa", size = 2729352, upload-time = "2026-09-03T08:55:44.345Z" },
1023
  ]
1024
 
1025
+ [[package]]
1026
+ name = "tomlkit"
1027
+ version = "0.14.0"
1028
+ source = { registry = "https://pypi.org/simple" }
1029
+ sdist = { url = "https://files.pythonhosted.org/packages/c3/af/14b24e41977adb296d6bd1fb59402cf7d60ce364f90c890bd2ec65c43b5a/tomlkit-0.14.0.tar.gz", hash = "sha256:cf00efca415dbd57575befb1f6634c4f42d2d87dbba376128adb42c121b87064", size = 187167, upload-time = "2026-01-13T01:14:53.304Z" }
1030
+ wheels = [
1031
+ { url = "https://files.pythonhosted.org/packages/b5/11/87d6d29fb5d237229d67973a6c9e06e048f01cf4994dee194ab0ea841814/tomlkit-0.14.0-py3-none-any.whl", hash = "sha256:592064ed85b40fa213469f81ac584f67a4f2992509a7c3ea2d632208623a3680", size = 39310, upload-time = "2026-01-13T01:14:51.965Z" },
1032
+ ]
1033
+
1034
+ [[package]]
1035
+ name = "torch"
1036
+ version = "2.10.0"
1037
+ source = { registry = "https://pypi.org/simple" }
1038
+ resolution-markers = [
1039
+ "sys_platform == 'emscripten'",
1040
+ "sys_platform != 'emscripten' and sys_platform != 'linux' and sys_platform != 'win32'",
1041
+ ]
1042
+ dependencies = [
1043
+ { name = "filelock" },
1044
+ { name = "fsspec" },
1045
+ { name = "jinja2" },
1046
+ { name = "networkx" },
1047
+ { name = "setuptools" },
1048
+ { name = "sympy" },
1049
+ { name = "typing-extensions" },
1050
+ ]
1051
+ wheels = [
1052
+ { url = "https://files.pythonhosted.org/packages/d3/54/a2ba279afcca44bbd320d4e73675b282fcee3d81400ea1b53934efca6462/torch-2.10.0-2-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:13ec4add8c3faaed8d13e0574f5cd4a323c11655546f91fbe6afa77b57423574", size = 79498202, upload-time = "2026-02-10T21:44:52.603Z" },
1053
+ { url = "https://files.pythonhosted.org/packages/c9/5c/dee910b87c4d5c0fcb41b50839ae04df87c1cfc663cf1b5fca7ea565eeaa/torch-2.10.0-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:6d3707a61863d1c4d6ebba7be4ca320f42b869ee657e9b2c21c736bf17000294", size = 79498198, upload-time = "2026-01-21T16:24:34.704Z" },
1054
+ ]
1055
+
1056
  [[package]]
1057
  name = "torch"
1058
  version = "2.10.0+cu128"
1059
  source = { registry = "https://download.pytorch.org/whl/cu128" }
1060
+ resolution-markers = [
1061
+ "sys_platform == 'win32'",
1062
+ "sys_platform == 'linux'",
1063
+ ]
1064
  dependencies = [
1065
+ { name = "cuda-bindings", marker = "sys_platform != 'win32'" },
1066
  { name = "filelock" },
1067
  { name = "fsspec" },
1068
  { name = "jinja2" },
1069
  { name = "networkx" },
1070
+ { name = "nvidia-cublas-cu12", marker = "sys_platform != 'win32'" },
1071
+ { name = "nvidia-cuda-cupti-cu12", marker = "sys_platform != 'win32'" },
1072
+ { name = "nvidia-cuda-nvrtc-cu12", marker = "sys_platform != 'win32'" },
1073
+ { name = "nvidia-cuda-runtime-cu12", marker = "sys_platform != 'win32'" },
1074
+ { name = "nvidia-cudnn-cu12", marker = "sys_platform != 'win32'" },
1075
+ { name = "nvidia-cufft-cu12", marker = "sys_platform != 'win32'" },
1076
+ { name = "nvidia-cufile-cu12", marker = "sys_platform != 'win32'" },
1077
+ { name = "nvidia-curand-cu12", marker = "sys_platform != 'win32'" },
1078
+ { name = "nvidia-cusolver-cu12", marker = "sys_platform != 'win32'" },
1079
+ { name = "nvidia-cusparse-cu12", marker = "sys_platform != 'win32'" },
1080
+ { name = "nvidia-cusparselt-cu12", marker = "sys_platform != 'win32'" },
1081
+ { name = "nvidia-nccl-cu12", marker = "sys_platform != 'win32'" },
1082
+ { name = "nvidia-nvjitlink-cu12", marker = "sys_platform != 'win32'" },
1083
+ { name = "nvidia-nvshmem-cu12", marker = "sys_platform != 'win32'" },
1084
+ { name = "nvidia-nvtx-cu12", marker = "sys_platform != 'win32'" },
1085
  { name = "setuptools" },
1086
  { name = "sympy" },
1087
+ { name = "triton", marker = "sys_platform != 'win32'" },
1088
  { name = "typing-extensions" },
1089
  ]
1090
  wheels = [
 
1093
  { url = "https://download-r2.pytorch.org/whl/cu128/torch-2.10.0%2Bcu128-cp312-cp312-win_amd64.whl", hash = "sha256:fbde8f6a9ec8c76979a0d14df21c10b9e5cab6f0d106a73ca73e2179bc597cae", upload-time = "2026-01-21T15:22:17Z" },
1094
  ]
1095
 
1096
+ [[package]]
1097
+ name = "torchvision"
1098
+ version = "0.25.0"
1099
+ source = { registry = "https://pypi.org/simple" }
1100
+ resolution-markers = [
1101
+ "sys_platform == 'emscripten'",
1102
+ "sys_platform != 'emscripten' and sys_platform != 'linux' and sys_platform != 'win32'",
1103
+ ]
1104
+ dependencies = [
1105
+ { name = "numpy" },
1106
+ { name = "pillow" },
1107
+ { name = "torch", version = "2.10.0", source = { registry = "https://pypi.org/simple" } },
1108
+ ]
1109
+ wheels = [
1110
+ { url = "https://files.pythonhosted.org/packages/56/3a/6ea0d73f49a9bef38a1b3a92e8dd455cea58470985d25635beab93841748/torchvision-0.25.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:c2abe430c90b1d5e552680037d68da4eb80a5852ebb1c811b2b89d299b10573b", size = 1874920, upload-time = "2026-01-21T16:27:45.348Z" },
1111
+ ]
1112
+
1113
  [[package]]
1114
  name = "torchvision"
1115
  version = "0.25.0+cu128"
1116
  source = { registry = "https://download.pytorch.org/whl/cu128" }
1117
+ resolution-markers = [
1118
+ "sys_platform == 'win32'",
1119
+ "sys_platform == 'linux'",
1120
+ ]
1121
  dependencies = [
1122
  { name = "numpy" },
1123
  { name = "pillow" },
1124
+ { name = "torch", version = "2.10.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" } },
1125
  ]
1126
  wheels = [
1127
  { url = "https://download-r2.pytorch.org/whl/cu128/torchvision-0.25.0%2Bcu128-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:8623e534ef6a815bd6407d4b52dd70c7154e2eda626ad4b9cb895d36c5a3305b", upload-time = "2026-01-21T22:32:23Z" },
 
1206
  { url = "https://files.pythonhosted.org/packages/67/81/4add07e5172b7ac40d8ed5ff580409a7801a4fe26d529bdd915401dabfbe/typing_inspection-0.4.4-py3-none-any.whl", hash = "sha256:65b8397ba37ccbce054456aaccddfc91e6e3083c92824df348d96ca832f3f147", size = 14750, upload-time = "2026-08-12T12:37:24.648Z" },
1207
  ]
1208
 
1209
+ [[package]]
1210
+ name = "tzdata"
1211
+ version = "2026.4"
1212
+ source = { registry = "https://pypi.org/simple" }
1213
+ sdist = { url = "https://files.pythonhosted.org/packages/e4/31/3d74fa778a63b98b7374323befcc0be5ab3bd94afd4096a0124e7379152c/tzdata-2026.4.tar.gz", hash = "sha256:f1b8bd365d8d210c55353f4d7f8d6d8561c0ba50d704b700d195a9424bba0d79", size = 199350, upload-time = "2026-09-12T12:56:03.251Z" }
1214
+ wheels = [
1215
+ { url = "https://files.pythonhosted.org/packages/f9/bc/8737e8d54cf51106118039b83f485a4783112fab49ea9d044b234978a46e/tzdata-2026.4-py2.py3-none-any.whl", hash = "sha256:c2169a8b0a7a5e9674da5a135ccdfb2b3e671b333ed9fed17b41f73c34476e81", size = 347494, upload-time = "2026-09-12T12:56:01.67Z" },
1216
+ ]
1217
+
1218
  [[package]]
1219
  name = "uvicorn"
1220
  version = "0.54.0"
runtime/qev-0.2.0-py3-none-any.whl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:33ca132521cf9ec7f5bc9405816cd65e9c4626853cfb21a82da7a4549ebbe576
3
+ size 185279