Add on-device learning and optimizer resume: fireviewer_dfine_m_strict_v1

models/fireviewer_dfine_m_strict_v1_learning/CHECKPOINT_LICENSE.md

@@ -0,0 +1,81 @@
+# FireViewer Model Use License v1.0
+
+Copyright © FireViewer contributors / applicable rights holders.
+
+These terms apply to the **FireViewer-trained checkpoint and FireViewer-published conversions or derived deployment artifacts** identified by the repository that includes this file.
+
+They do not replace third-party licences that apply to upstream software, architectures, base models or other third-party components. FireViewer grants rights only to the extent it is entitled to grant them.
+
+## 1. Permitted uses
+
+Subject to these terms, you may:
+
+1. use the Model for research, testing and evaluation;
+2. perform internal evaluation for a potential commercial product or service;
+3. use the Model commercially **without redistributing the Model**, provided that the required FireViewer attribution is preserved;
+4. internally convert, optimize or quantize the Model, including to formats such as ONNX, TensorRT, LiteRT/TFLite, FP16, INT8, NVFP4 or another technically applicable format;
+5. internally fine-tune the Model.
+
+## 2. Attribution for free commercial use
+
+Commercial use under these free terms requires a reasonable public FireViewer credit where product or service credits, acknowledgements, legal notices or technical documentation are normally presented.
+
+Recommended attribution:
+
+> Fire/smoke detection model by FireViewer — https://huggingface.co/fireviewer
+
+Where practical, link directly to the specific FireViewer model repository.
+
+A separate written commercial agreement may waive the **public FireViewer attribution** requirement.
+
+Upstream notices that independently apply to third-party components remain required and are not waived by FireViewer.
+
+## 3. Redistribution is not granted
+
+You may **not** redistribute the Model.
+
+Redistribution includes providing, transferring, sublicensing, publishing, mirroring, selling, leasing, embedding or otherwise making available to a third party:
+
+- the original FireViewer checkpoint;
+- a converted or quantized version;
+- a fine-tuned version;
+- a pruned, merged or otherwise weight-derived version;
+- an official FireViewer deployment artifact obtained from the repository.
+
+Embedding the Model in hardware, software, firmware, an appliance, an edge device, NAS, NVR, camera system or other on-premise product that is **delivered to a third party** is redistribution under these terms.
+
+A hosted service in which the Model remains under the operator's control and customers only receive outputs is not, by itself, redistribution of the Model.
+
+## 4. Derived and converted weights
+
+Internal conversions, quantization and fine-tuning are permitted, but the resulting weights remain subject to the non-redistribution condition above.
+
+These terms do not claim rights over independently created software that merely calls or interoperates with the Model, except to the extent that distribution of that software also distributes Model weights.
+
+## 5. Official additional formats
+
+FireViewer may publish additional official deployment formats on request. Publishing an official FireViewer conversion does not grant redistribution rights beyond these terms. Third parties may not present an unofficial conversion as an official FireViewer artifact.
+
+## 6. Separate commercial agreement
+
+A separate written agreement may provide different terms, including commercial use without public FireViewer attribution.
+
+Unless that written agreement expressly says otherwise, it does **not** grant redistribution or sublicensing rights.
+
+## 7. No transfer of trademarks
+
+These terms do not grant a right to use FireViewer names, logos or marks except for factual attribution and identification of the Model.
+
+## 8. No warranty
+
+The Model is provided "AS IS", without warranties or conditions of any kind, to the maximum extent permitted by applicable law.
+
+You are responsible for validating the Model for your application and deployment environment. FireViewer models are research artifacts and must not be used alone to confirm a fire, trigger an emergency alert, order an evacuation or direct emergency response.
+
+## 9. Third-party terms
+
+Upstream and third-party components remain subject to their own licences and notices. If a third-party term independently imposes an obligation on your use of a component, that obligation remains applicable.
+
+## 10. Contact
+
+Commercial terms, official deployment formats and written permissions: **partenariats@fire-viewer.fr**

models/fireviewer_dfine_m_strict_v1_learning/CHECKPOINT_RIGHTS.md

@@ -0,0 +1,38 @@
+# Checkpoint rights and training provenance
+
+Model: `fireviewer/fire-smoke-dfine-m-strict-v1`. Documentation updated: 2026-09-17.
+Model revision: `0ef733968748c7ac172273db8c381ea5ea5ff046`.
+
+## Source-rights validation
+
+FireViewer records this training as based on openly licensed sources selected and
+reviewed during corpus construction, as reaffirmed by the project owner on
+17 September 2026. Source-specific attribution and licence conditions are retained.
+The source licences are CC BY-SA 4.0, Apache-2.0, GPL-3.0 and CC BY 4.0;
+they are not relabelled as a single Apache or MIT licence.
+
+## Exact artefacts
+
+| File | SHA-256 |
+| --- | --- |
+| `model.onnx` | `d4ad38e8c09756b975b177faa8440dc1ed9f52df10828c9981484e9d1e1db2ee` |
+| `model.pth` | `4ec68a6c3934a50fb14b707b3dc52766173a7f9857bed32c38ad831f336c350f` |
+
+## Training identity
+
+The run receipt pins `fireviewer/fire-smoke-detection-corpus-v1@8e3cb52e86e15debdff7d37a3ce9eac39fbe1835`. The materialisation receipt records
+60,981 training and 19,209 validation images. All 27 recorded shard hashes match
+the dataset publication manifest. The test partition was not dispatched by this
+training materialisation.
+
+Implementation: [D-FINE at `956d1709314c2c6a4df6f34de232054578a7449f`](https://github.com/Peterande/D-FINE/tree/956d1709314c2c6a4df6f34de232054578a7449f), Apache-2.0.
+The recorded architecture uses HGNetv2 B2; its run configuration is preserved in
+[training/dfine_fire_smoke_m.yml](training/dfine_fire_smoke_m.yml).
+
+## Rights and attribution documents
+
+[MODEL_RIGHTS_POLICY.md](MODEL_RIGHTS_POLICY.md) separates implementation,
+source data and model artefacts. [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md)
+lists source credit and terms. [TRAINING_DATA_MANIFEST.json](TRAINING_DATA_MANIFEST.json)
+and [MODEL_ARTIFACTS_MANIFEST.json](MODEL_ARTIFACTS_MANIFEST.json) provide immutable
+identities. Existing model-card evaluation and safety limitations are preserved.

models/fireviewer_dfine_m_strict_v1_learning/CITATION.md

@@ -0,0 +1,7 @@
+# Citation and attribution
+
+Recommended commercial attribution:
+
+> Fire/smoke detection model by FireViewer — https://huggingface.co/fireviewer/fire-smoke-dfine-m-strict-v1
+
+Where practical, keep a direct link to the repository in technical documentation, credits, acknowledgements or legal notices. Upstream research citations remain applicable and should be retained in addition to FireViewer attribution.

models/fireviewer_dfine_m_strict_v1_learning/COMMERCIAL_USE.md

@@ -0,0 +1,20 @@
+# Commercial use — quick reference
+
+| Use | Status under FireViewer Model Use License v1.0 |
+|---|---|
+| Research / evaluation | Permitted |
+| Internal commercial evaluation | Permitted |
+| Commercial use with FireViewer attribution | Permitted |
+| Hosted API / SaaS where customers do not receive weights | Permitted with attribution |
+| Internal ONNX / TensorRT / LiteRT / quantization | Permitted |
+| Internal fine-tuning | Permitted |
+| Publishing a converted checkpoint | Not permitted |
+| Shipping the weights in a device or on-premise product | Not permitted |
+| Redistributing original weights | Not permitted |
+| Redistributing fine-tuned / quantized / converted weights | Not permitted |
+| Commercial use without public FireViewer attribution | Separate written agreement |
+| OEM / redistribution rights | Separate written agreement explicitly granting them |
+
+The full terms are in `CHECKPOINT_LICENSE.md`.
+
+FireViewer may publish additional official deployment formats when technically relevant, including ONNX, TensorRT, LiteRT/TFLite, FP16, INT8, NVFP4 or another target-specific format.

models/fireviewer_dfine_m_strict_v1_learning/MODEL_RIGHTS_POLICY.md

@@ -0,0 +1,45 @@
+# Model rights and source attribution policy
+
+Updated: 17 September 2026.
+
+## Source selection and validation
+
+FireViewer records its corpus-building process as selection, review and validation
+of openly licensed sources before model training. The project owner reaffirmed
+this validation history on 17 September 2026. The checkpoint-specific manifests
+and receipts below provide the separately verifiable technical provenance.
+
+An open licence may require attribution, preservation of notices or share-alike.
+CC BY, CC BY-SA and GPL are not classified as non-free merely because they are
+different from CC0, MIT or Apache-2.0. Commercial activity is not categorically
+prohibited by those open licences; the terms of each licence still apply.
+
+## Separate artefacts and applicable terms
+
+Keep the following identities separate: source images and annotations; collection
+and preprocessing; upstream implementation; initial model; trained checkpoint;
+and ONNX, quantized or fine-tuned exports. Preserve the applicable licence and
+notices for the artefacts being copied or distributed. A code licence does not
+silently replace an image licence. Conversion does not remove applicable terms.
+
+`license: other` is the Hub classification used for this composite release. It
+does not mean that the sources were unlicensed or non-free, and is not itself a
+new licence grant. This documentation records provenance and existing terms;
+it neither adopts a new model licence nor changes any third-party licence.
+
+## Reproducible attribution
+
+Use `CHECKPOINT_RIGHTS.md`, `MODEL_ARTIFACTS_MANIFEST.json` and
+`TRAINING_DATA_MANIFEST.json` to identify the exact files and data release. Read
+`THIRD_PARTY_NOTICES.md` together with source-level terms. Do not substitute a
+later dataset revision for the training input of an earlier checkpoint.
+
+Native weights, exports and derivatives retain their original identity and
+history. A new training run is a new artefact and requires its own manifests.
+This documentation update does not retrain a model or change its metrics.
+
+An individually signed evaluation agreement, if used, covers only the rights
+its signatory is entitled to grant; an unsigned template is not an agreement.
+Publication under an organisation account does not itself transfer ownership.
+
+The existing model-card safety and evaluation limitations continue to apply.

models/fireviewer_dfine_m_strict_v1_learning/README.md

@@ -0,0 +1,15 @@
+# FireViewer D-FINE-M strict v1 — Android learning variant
+
+Source: `fireviewer/fire-smoke-dfine-m-strict-v1`, revision `a4d15bae27f8deaeb9f89e7e118b4643cddfecad`.
+
+This is a real mutable LiteRT model with `infer`, `train`, `save`, and `restore` signatures. Training executes on Android using the image and reviewed targets. Trainable scope: **candidate_classification_and_box_adaptation_only**. The visual backbone remains frozen. Does not create proposals absent from frozen detector
+
+This learning variant is stored separately under `models/fireviewer_dfine_m_strict_v1_learning` and has its own explicit Android configuration.
+
+Variable image dimensions: `{'mode': 'bilinear_stretch_inside_graph', 'native_backbone_shape': [1, 3, 704, 704]}`. The Android configuration chooses a bounded input canvas; the image may have any original size. Images are processed individually, so a dataset need not have homogeneous dimensions.
+
+Checkpoints include parameters, momentum and step. Validation executed actual updates, changed parameters, closed/reopened the interpreter, restored identical predictions and reproduced the next optimizer update. See `training_validation.json`. Tests use synthetic supervision to verify execution and persistence; no accuracy or forgetting benchmark is claimed. Android device execution: `False`.
+
+Use LiteRT CPU with XNNPACK disabled and Select TF Ops for checkpoint operations. The default label space is retained; changing class count requires rebuilding the head with the supplied conversion sources. Use only human-reviewed labels for training. Retain replay examples and validate a candidate before activating it.
+
+Upstream licenses and attributions continue to apply.

models/fireviewer_dfine_m_strict_v1_learning/RIGHTS_AND_ATTRIBUTION.md

@@ -0,0 +1,9 @@
+# Rights and attribution
+
+This repository separates three layers:
+
+1. **Upstream implementation / base model** — D-FINE / related upstream implementation components: preserve their Apache-2.0 notices and exact pinned revisions used by the release.
+2. **Training corpus provenance** — `fireviewer/fire-smoke-detection-corpus-v1` and its source records.
+3. **FireViewer-trained checkpoint and FireViewer-published deployment artifacts** — governed by `CHECKPOINT_LICENSE.md` where that file is included.
+
+FireViewer grants only rights it is entitled to grant. Third-party notices and source-specific obligations remain applicable.

models/fireviewer_dfine_m_strict_v1_learning/THIRD_PARTY_NOTICES.md

@@ -0,0 +1,5 @@
+# Third-party notices
+
+- D-FINE / related upstream implementation: preserve the Apache-2.0 licence and attribution notices applicable to the exact pinned upstream revision.
+- Training media: governed by the per-source rights recorded in `fireviewer/fire-smoke-detection-corpus-v1`.
+- FireViewer custom terms apply only to the FireViewer-trained checkpoint and FireViewer-published deployment artifacts to the extent FireViewer controls those rights.

models/fireviewer_dfine_m_strict_v1_learning/UPSTREAM_README.md

@@ -0,0 +1,83 @@
+---
+license: other
+pipeline_tag: object-detection
+datasets:
+- fireviewer/fire-smoke-detection-corpus-v1
+tags:
+- fireviewer
+- fire
+- smoke
+- wildfire
+- dfine
+- object-detection
+- onnx
+---
+
+# FireViewer D-FINE-M strict v1
+
+Two-class research detector for **visible smoke** and **visible flame**.
+
+## Status
+
+**Primary FireViewer research candidate, not production-qualified.** The current frozen independent benchmark reports COCO mAP50-95 **0.3924** and calibrated holdout F1 **0.6475**. Human review remains mandatory.
+
+The repository revision referenced in the current commercial evaluation is `0ef733968748c7ac172273db8c381ea5ea5ff046`. The supplied ONNX artifact has SHA-256 `d4ad38e8c09756b975b177faa8440dc1ed9f52df10828c9981484e9d1e1db2ee`.
+
+## Classes
+
+- `smoke_visible`
+- `flame_visible`
+
+## Intended use
+
+Research, reproducible evaluation, internal commercial evaluation and non-redistributed commercial inference under the repository terms. It can be tested as a specialized first-stage detector in image/video workflows.
+
+## Not established
+
+This release does not establish suitability as a standalone alarm system, emergency decision system, geolocation engine or fire-propagation predictor. False positives and false negatives remain possible, including domain shift from CCTV, night scenes, haze, clouds, dust, reflections and other visually similar phenomena.
+
+## Evaluation
+
+Independent model comparisons must use the same frozen FireViewer protocol. Do not compare internal validation metrics to independent holdout scores as if they were equivalent.
+
+## Rights
+
+Read `CHECKPOINT_LICENSE.md`, `COMMERCIAL_USE.md`, `RIGHTS_AND_ATTRIBUTION.md` and `THIRD_PARTY_NOTICES.md` before commercial use or conversion.
+
+## Reproducibility and artifact access
+
+```python
+from huggingface_hub import hf_hub_download
+
+path = hf_hub_download(
+    repo_id="fireviewer/fire-smoke-dfine-m-strict-v1",
+    repo_type="model",
+    revision="1d8e0d5eba6a83ea85c0b4318a3c95eca88a3e34",
+    filename="model.onnx",
+)
+```
+Hub artifact SHA-256 for `model.onnx`: `d4ad38e8c09756b975b177faa8440dc1ed9f52df10828c9981484e9d1e1db2ee`. This identity was read from Hub metadata; the weights were not downloaded or rehashed during the documentation review.
+
+The example downloads a pinned artifact; it does not execute or qualify inference. Preserve the matching configuration, labels, preprocessing and runtime dependencies from that revision. Keep raw outputs, scores, model revision and source provenance when integrating with `fireviewer-vision-runtime`.
+
+Training corpus: `fireviewer/fire-smoke-detection-corpus-v1` at `8e3cb52e86e15debdff7d37a3ce9eac39fbe1835` (see [run receipt](run-receipt.json) and [source manifest](provenance/source-manifest.json)). Base recipe and implementation are recorded with the training artifacts; a family name alone is not an immutable upstream revision.
+
+The included [inference helper](inference.py) consumes an RGB image and returns class ID/name, confidence and `xyxy` image boxes; its documented image size is 704. `--confidence 0.25` is the helper default, not a calibrated deployment threshold.
+
+## Independent evaluation protocol and limits
+
+[Published independent report](benchmark/independent-evaluation.json): 512 images from `Hajorda/flameye-wildfire-detection`, `test`, revision `361a3dea8b877482af9f4ed80eff77ffd39926ef`. The retained common protocol uses 129 calibration images and 383 holdout images; its selection is dominated by D-Fire (508/512). Zero exact training overlap is a reported hash-based check, not proof of no semantic or domain overlap.
+
+Evaluated model revision: `3bbab94012311d4f4bba68dc883117dec1281a24`. The report-calibrated score threshold is **0.5**; it is specific to this protocol and is **not an operational threshold**. COCO AP over the 512-image selection and F1 on the calibrated holdout are different measurements. Do not compare them with RF-DETR training-validation scores or DINOv3 pointing metrics.
+
+The report does not contain an explicit execution timestamp. Its original revision is the evidence reference; 19 September is the card review date, not a new evaluation date. Multi-event and multi-camera external validation remains necessary.
+
+Sensitive results require human validation. No image detection establishes geographic coordinates, a measured perimeter, absence of fire or future propagation.
+
+## Documentation, revision and stewardship
+
+Documentation reviewed on **19 September 2026**, using the FireViewer documentation pack and the repository at [`1d8e0d5eba6a83ea85c0b4318a3c95eca88a3e34`](https://huggingface.co/fireviewer/fire-smoke-dfine-m-strict-v1/tree/1d8e0d5eba6a83ea85c0b4318a3c95eca88a3e34). This is a documentation and metadata check, not a new model evaluation, training run, data acquisition or full artifact verification. Historical receipts keep their original dates and identifiers; manifest hashes of earlier README versions describe those earlier releases.
+
+The association FIRE-VIEWER provides administrative and financial stewardship for resources under its control. Technical governance is maintainer-led. Pre-association rights, UWD components, upstream licences and source-specific rights remain separate; repository placement does not transfer ownership or grant new rights.
+
+[Canonical documentation](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/HUGGINGFACE.md) · [Status vocabulary](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/STATUTS_ET_VOCABULAIRE.md) · [Institutional website](https://fire-viewer.fr). Contact: **contact@fire-viewer.fr**.

models/fireviewer_dfine_m_strict_v1_learning/android_model_config.json

@@ -0,0 +1,51 @@
+{
+  "task": "object_detection",
+  "adapter": "yolo",
+  "inputWidth": 704,
+  "inputHeight": 704,
+  "inputLayout": "NCHW",
+  "inputType": "FLOAT32",
+  "mean": 0.0,
+  "std": 255.0,
+  "resizeMode": "stretch",
+  "labels": [
+    "smoke_visible",
+    "flame_visible"
+  ],
+  "threads": 2,
+  "outputIndex": 2,
+  "threshold": 0.1,
+  "scoreActivation": "none",
+  "dynamicMinSize": 128,
+  "dynamicMaxSize": 1280,
+  "dynamicStride": 1,
+  "outputLayout": "BCN",
+  "coordinates": "pixels",
+  "training": {
+    "targetEncoding": "boxes_xyxy_class_mask",
+    "targetShape": [
+      1,
+      100,
+      6
+    ],
+    "inferOutputs": [
+      "detections"
+    ],
+    "trainSignature": "train",
+    "inferSignature": "infer",
+    "saveSignature": "save",
+    "restoreSignature": "restore",
+    "imageInput": "x",
+    "targetInput": "y",
+    "lossOutput": "loss",
+    "checkpointInput": "checkpoint_path",
+    "learningRateInput": "learning_rate",
+    "scope": "candidate_classification_and_box_adaptation_only"
+  },
+  "isRgb": true,
+  "namedOutputIndices": {
+    "boxes": 1,
+    "detections": 2,
+    "logits": 0
+  }
+}

models/fireviewer_dfine_m_strict_v1_learning/artifact_manifest.json

@@ -0,0 +1,62 @@
+{
+  "README.md": {
+    "sha256": "88f11d8b1063501f244882d785953ba5de157685bc053c194ee3741766392d7b",
+    "bytes": 1707
+  },
+  "UPSTREAM_README.md": {
+    "sha256": "486bc4a39d009061de58e1e5798e38d578a9a646cf867d4b1d109300cc7dcf35",
+    "bytes": 5598
+  },
+  "CHECKPOINT_LICENSE.md": {
+    "sha256": "549e691e53c816d605abc0acea77116ada8289dd9896dd67bebd3db317760b73",
+    "bytes": 4429
+  },
+  "CHECKPOINT_RIGHTS.md": {
+    "sha256": "8c68b67bf76a00b7760c40af53926ade76f29a42261318cb9332adba65b90f9f",
+    "bytes": 1977
+  },
+  "CITATION.md": {
+    "sha256": "2d25875f6fa305bf495d2819035eb152b272bc0dbe6d4972295d08d753a38a8a",
+    "bytes": 410
+  },
+  "COMMERCIAL_USE.md": {
+    "sha256": "85cc7903902ecf421783a56a44005d53fc642436482a811ef3954f8a5b031c2c",
+    "bytes": 1104
+  },
+  "MODEL_RIGHTS_POLICY.md": {
+    "sha256": "f17d5798cdaaf07cb375213326d54293c7fd864972af5f574c9f033757af45f9",
+    "bytes": 2410
+  },
+  "RIGHTS_AND_ATTRIBUTION.md": {
+    "sha256": "8d1879cfecff4e96ee363aab6a063496a9e28f1ecde84bb9c67555adda52c56e",
+    "bytes": 635
+  },
+  "THIRD_PARTY_NOTICES.md": {
+    "sha256": "0c78e14d82671da0582c78a9b456880086b1fda0b1387ad35ab7735503d07a7f",
+    "bytes": 450
+  },
+  "runtime_contract.json": {
+    "sha256": "e6ccdd562414252b8b4107df3a52c4b24008752550364708f655e07ea4dd52ac",
+    "bytes": 3118
+  },
+  "training_validation.json": {
+    "sha256": "88eee8adbd5b1a4fafbaf02e02e4e649674865d4b6358284052d4d7a59ab008c",
+    "bytes": 2303
+  },
+  "runtime216_validation.json": {
+    "sha256": "773668f349a4cba7a911a61b3b231f4e837d9843e9090c928d9e5d09340d8625",
+    "bytes": 359
+  },
+  "source_parity.json": {
+    "sha256": "245ef823a83796c77a1127d7d580bfb21414e369f5d290483c319e116918e7cb",
+    "bytes": 1689
+  },
+  "android_model_config.json": {
+    "sha256": "5a0c69e09a96823ae6e3f0953fb7a99d2c6685eb608f1abea4385fc7fd1f0166",
+    "bytes": 1096
+  },
+  "model.tflite": {
+    "sha256": "b66acda95a5eaf5154031cf80c1ed906b8cb7838617424afa9b837c5cd44064a",
+    "bytes": 79120652
+  }
+}

models/fireviewer_dfine_m_strict_v1_learning/model.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b66acda95a5eaf5154031cf80c1ed906b8cb7838617424afa9b837c5cd44064a
+size 79120652

models/fireviewer_dfine_m_strict_v1_learning/runtime216_validation.json

@@ -0,0 +1,11 @@
+{
+  "key": "fireviewer_dfine_m_strict_v1",
+  "runtime": "2.16.1",
+  "platform": "Linux x86_64; same LiteRT release as Android, different ABI",
+  "sha256": "b66acda95a5eaf5154031cf80c1ed906b8cb7838617424afa9b837c5cd44064a",
+  "train": true,
+  "loss": 2.902190685272217,
+  "restart_exact": true,
+  "optimizer_resume_exact": true,
+  "two_object_fixture": true
+}

models/fireviewer_dfine_m_strict_v1_learning/runtime_contract.json

@@ -0,0 +1,116 @@
+{
+  "schema": 2,
+  "key": "fireviewer_dfine_m_strict_v1_learning",
+  "source_model": "fireviewer_dfine_m_strict_v1",
+  "task": "object_detection",
+  "upstream": "fireviewer/fire-smoke-dfine-m-strict-v1",
+  "upstream_revision": "a4d15bae27f8deaeb9f89e7e118b4643cddfecad",
+  "backbone_frozen": true,
+  "training_scope": "candidate_classification_and_box_adaptation_only",
+  "report": {
+    "key": "fireviewer_dfine_m_strict_v1",
+    "validation_passed": true,
+    "training_scope": "candidate_classification_and_box_adaptation_only",
+    "backbone_frozen": true,
+    "limitation": "Does not create proposals absent from frozen detector",
+    "initial_detector_parity": true,
+    "parameters_changed": true,
+    "initial_loss": 2.38586688041687,
+    "final_loss": 1.0897295475006104,
+    "mixed_image_shapes": [
+      [
+        193,
+        317
+      ],
+      [
+        193,
+        381
+      ],
+      [
+        257,
+        317
+      ]
+    ],
+    "restart_logits_exact": true,
+    "optimizer_resume_exact": true,
+    "variable_image_preprocessing": {
+      "mode": "bilinear_stretch_inside_graph",
+      "native_backbone_shape": [
+        1,
+        3,
+        704,
+        704
+      ]
+    },
+    "android_device_tested": false,
+    "accuracy_evaluated": false,
+    "sha256": "b66acda95a5eaf5154031cf80c1ed906b8cb7838617424afa9b837c5cd44064a",
+    "bytes": 79120652,
+    "signatures": {
+      "infer": {
+        "inputs": [
+          "x"
+        ],
+        "outputs": [
+          "boxes",
+          "detections",
+          "logits"
+        ]
+      },
+      "restore": {
+        "inputs": [
+          "checkpoint_path"
+        ],
+        "outputs": [
+          "step"
+        ]
+      },
+      "save": {
+        "inputs": [
+          "checkpoint_path"
+        ],
+        "outputs": [
+          "step"
+        ]
+      },
+      "train": {
+        "inputs": [
+          "learning_rate",
+          "x",
+          "y"
+        ],
+        "outputs": [
+          "loss",
+          "step"
+        ]
+      },
+      "training_state": {
+        "inputs": [],
+        "outputs": [
+          "head_kernel",
+          "step"
+        ]
+      }
+    },
+    "runtime_requirement": "LiteRT builtin CPU; disable default XNNPACK delegates; Select TF Ops required for checkpoint Save/Restore",
+    "android_runtime_dependencies": [
+      "org.tensorflow:tensorflow-lite:2.16.1",
+      "org.tensorflow:tensorflow-lite-select-tf-ops:2.16.1"
+    ],
+    "frozen_backbone_sha256": "9dac6e73396d0e949da5b6b7960966b2ec971b89739b961b3e732b8830f8e63a",
+    "head_initialization": "source_model_or_identity_adapter",
+    "conversion_source_sha256": {
+      "merge_training_graph.py": "93def8b746dd763c7ef703e29a5db3bab754716c6358bdda5acb64531b4c50ce",
+      "build_trainable_detector.py": "4e7794e81cd426850a1f84ee5c6b5af92316a2b62db10461e9f91bc896c7bb30"
+    },
+    "source_parity_report": "source_parity.json"
+  },
+  "optimizer": "momentum SGD, momentum=0.9",
+  "checkpoint_contains": [
+    "trainable parameters",
+    "momentum buffers",
+    "global step"
+  ],
+  "requires_select_tf_ops": true,
+  "android_configuration": "android_model_config.json"
+}

models/fireviewer_dfine_m_strict_v1_learning/source_parity.json

@@ -0,0 +1,84 @@
+{
+  "key": "fireviewer_dfine_m_strict_v1",
+  "phase": "validated_frozen_component",
+  "upstream_onnx_parity_passed": true,
+  "backbone": "/workspace/vds-litert-batch2/learning-backbones/fireviewer_dfine_m_strict_v1/model-protected.tflite",
+  "source_sha256": "999ef94cc0419cc70a3c136b6a1ad30fb8ed66570dcad005308279342abcc77e",
+  "sha256": "9dac6e73396d0e949da5b6b7960966b2ec971b89739b961b3e732b8830f8e63a",
+  "input_shape": [
+    1,
+    3,
+    704,
+    704
+  ],
+  "outputs": [
+    {
+      "index": 0,
+      "shape": [
+        1,
+        300,
+        2
+      ]
+    },
+    {
+      "index": 1,
+      "shape": [
+        1,
+        300,
+        4
+      ]
+    }
+  ],
+  "max_abs_errors": [
+    3.337860107421875e-05,
+    9.834766387939453e-07,
+    3.4332275390625e-05,
+    4.172325134277344e-07,
+    5.6743621826171875e-05,
+    7.152557373046875e-07
+  ],
+  "seeds": [
+    842,
+    843,
+    844
+  ],
+  "atol": 0.002,
+  "rtol": 0.002,
+  "preserve_all_tensors": false,
+  "arena_protection": {
+    "tensors": [
+      {
+        "index": 1354,
+        "name": "wa/model/decoder/decoder/integral/Softmax_output_0",
+        "shape": [
+          1200,
+          33
+        ]
+      },
+      {
+        "index": 1857,
+        "name": "wa/model/decoder/decoder/integral_1/Softmax_output_0",
+        "shape": [
+          1200,
+          33
+        ]
+      },
+      {
+        "index": 2338,
+        "name": "wa/model/decoder/decoder/integral_2/Softmax_output_0",
+        "shape": [
+          1200,
+          33
+        ]
+      },
+      {
+        "index": 2816,
+        "name": "wa/model/decoder/decoder/integral_3/Softmax_output_0",
+        "shape": [
+          1200,
+          33
+        ]
+      }
+    ]
+  }
+}

models/fireviewer_dfine_m_strict_v1_learning/training_validation.json

@@ -0,0 +1,98 @@
+{
+  "key": "fireviewer_dfine_m_strict_v1",
+  "validation_passed": true,
+  "training_scope": "candidate_classification_and_box_adaptation_only",
+  "backbone_frozen": true,
+  "limitation": "Does not create proposals absent from frozen detector",
+  "initial_detector_parity": true,
+  "parameters_changed": true,
+  "initial_loss": 2.38586688041687,
+  "final_loss": 1.0897295475006104,
+  "mixed_image_shapes": [
+    [
+      193,
+      317
+    ],
+    [
+      193,
+      381
+    ],
+    [
+      257,
+      317
+    ]
+  ],
+  "restart_logits_exact": true,
+  "optimizer_resume_exact": true,
+  "variable_image_preprocessing": {
+    "mode": "bilinear_stretch_inside_graph",
+    "native_backbone_shape": [
+      1,
+      3,
+      704,
+      704
+    ]
+  },
+  "android_device_tested": false,
+  "accuracy_evaluated": false,
+  "sha256": "b66acda95a5eaf5154031cf80c1ed906b8cb7838617424afa9b837c5cd44064a",
+  "bytes": 79120652,
+  "signatures": {
+    "infer": {
+      "inputs": [
+        "x"
+      ],
+      "outputs": [
+        "boxes",
+        "detections",
+        "logits"
+      ]
+    },
+    "restore": {
+      "inputs": [
+        "checkpoint_path"
+      ],
+      "outputs": [
+        "step"
+      ]
+    },
+    "save": {
+      "inputs": [
+        "checkpoint_path"
+      ],
+      "outputs": [
+        "step"
+      ]
+    },
+    "train": {
+      "inputs": [
+        "learning_rate",
+        "x",
+        "y"
+      ],
+      "outputs": [
+        "loss",
+        "step"
+      ]
+    },
+    "training_state": {
+      "inputs": [],
+      "outputs": [
+        "head_kernel",
+        "step"
+      ]
+    }
+  },
+  "runtime_requirement": "LiteRT builtin CPU; disable default XNNPACK delegates; Select TF Ops required for checkpoint Save/Restore",
+  "android_runtime_dependencies": [
+    "org.tensorflow:tensorflow-lite:2.16.1",
+    "org.tensorflow:tensorflow-lite-select-tf-ops:2.16.1"
+  ],
+  "frozen_backbone_sha256": "9dac6e73396d0e949da5b6b7960966b2ec971b89739b961b3e732b8830f8e63a",
+  "head_initialization": "source_model_or_identity_adapter",
+  "conversion_source_sha256": {
+    "merge_training_graph.py": "93def8b746dd763c7ef703e29a5db3bab754716c6358bdda5acb64531b4c50ce",
+    "build_trainable_detector.py": "4e7794e81cd426850a1f84ee5c6b5af92316a2b62db10461e9f91bc896c7bb30"
+  },
+  "source_parity_report": "source_parity.json"
+}