README.md

---
library_name: tensorflow-lite
license: other
tags:
- android
- on-device-training
- fire-detection
---

# FireViewer LiteRT Models for Android

Fire/smoke detection and pointing conversions for experimental Android workflows:
**10 delivered variants: 6 historical variants (5 with trainable heads/output adapters) and 4 v2.5 inference-only cascade exports**.
Reviewed annotations can update the exposed adaptation parameters while the
visual backbones remain frozen.

[Usage guide](docs/USAGE.en.md) · [Conversion and SDK evidence](docs/BENCHMARKS.en.md)
· [Kotlin SDK](android/README.md) · [Checkpoint lifecycle](docs/CONTINUOUS_LEARNING.en.md).


## New — v2.5 frozen cascade · 7 October 2026

Four selected RF-DETR/DINOv2 checkpoints are available as **LiteRT FP32**, with real CPU invocation/parity receipts, conventional contracts and Cadryl ModelConfig/StudioPack imports. They are **inference-only** and have no Android-device qualification. The six historical packages and their original receipts are preserved.

| Stage / Étape | LiteRT package | Input | CPU numeric parity | Android device |
|---|---|---|---|---|
| proposal | [fireviewer_rfdetr_medium_probable_fire_v25_fp32](models/fireviewer_rfdetr_medium_probable_fire_v25_fp32/README.md) | 1088 px · batch 1 | PASS at .10 operating threshold · 11 inputs | Not qualified / Non qualifié |
| visible | [fireviewer_rfdetr_medium_visible_flame_v25_fp32](models/fireviewer_rfdetr_medium_visible_flame_v25_fp32/README.md) | 1088 px · batch 1 | PASS at .25 operating threshold · 11 inputs | Not qualified / Non qualifié |
| point | [fireviewer_dinov2_small_visible_flame_monopoint_v25_fp32](models/fireviewer_dinov2_small_visible_flame_monopoint_v25_fp32/README.md) | 448 px · batch 1 | PASS tensor outputs · 5 inputs | Not qualified / Non qualifié |
| smoke | [fireviewer_dinov2_small_fire_associated_smoke_v25_fp32](models/fireviewer_dinov2_small_fire_associated_smoke_v25_fp32/README.md) | 448 px · batch 1 | PASS tensor outputs · 11 inputs | Not qualified / Non qualifié |

[English and French pipeline, contracts and scientific interpretation](docs/CASCADE_V25.md) · [Full 680-image report](reports/cascade-v25-680/REPORT.md) · [Metrics JSON](reports/cascade-v25-680/metrics.json) · [Conversion evidence](reports/cascade-v25-litert-validation.json) · [2.5 not so bad collection](https://huggingface.co/collections/fireviewer/25-not-so-bad-6ac47be5c15d70ee2a01d464).

The full cascade alerts on 220/249 positive images (**88.35%**) and 49/431 negatives (**11.37%**), but localizes only 231/901 flame boxes (**25.64%**) and 178/788 tiny flames (**22.59%**) at IoU≥0.5. Proposal containment is 882/901 (**97.89%**). This is useful image suspicion, with a substantial weakness in precise tiny-flame instance localization. Point containment is not flame-base accuracy. L4 PyTorch timings are distinct from LiteRT/Android performance.

**FR :** quatre exports d'inférence FP32, contrats standard et formats Cadryl, documentation détaillée en français et anglais. Les résultats distinguent l'alerte d'image de la localisation précise des petites flammes. Aucun point/masque fumée rejeté, aucune précision de base du feu inventée, aucun nouvel entraînement.

## Catalogue and recorded qualification

| Variant | Synthetic host learning | Standalone Android SDK | Vision Dataset Studio |
| --- | --- | --- | --- |
| [D-FINE M strict v1, learning](models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | Not qualified | Not qualified |
| [RT-DETRv2 R50, learning](models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | Not qualified | Not qualified |
| [YOLO11 M strict v1, inference](models/fireviewer_yolo11m_strict_v1/README.md) | N/A | Not qualified | Not qualified |
| [YOLO11 M strict v1, learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | PASS on emulator | Not qualified |
| [RF-DETR Medium 1.10, learning](models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | Not qualified | Not qualified |
| [DINOv3 pointing pilot v1, learning](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | Not qualified | Not qualified |

![Historical conversion and SDK qualification matrix](docs/figures/qualification-matrix.png)

The retained [delivery receipt](reports/fireviewer-completion.json) covers five
learning conversions. The YOLO inference-only graph is a sixth delivered variant.
No SegFormer graph is delivered because the optional source was unavailable.
Detection adapters cannot create proposals absent from the frozen detector.

**The six historical variants remain unqualified inside Vision Dataset Studio.** The YOLO
learning SDK pass does not qualify other models, application integration or a
physical ARM phone. For those six historical variants, task accuracy, generalization and forgetting are unmeasured.

## Download and integrate

Choose a variant and read its README, runtime contract, artifact manifest and
source notices. Retrieve all matching files at the same immutable revision:

```python
from huggingface_hub import hf_hub_download

contract = hf_hub_download(
    repo_id="fireviewer/litert-models",
    revision="bdcd483d5c5aadeff1b531a35b536f3530906e18",
    filename="models/fireviewer_yolo11m_strict_v1_learning/runtime_contract.json",
)
```

Apply the exact normalization, tensor layout, labels and coordinate transform.
Check inference before learning, train only with reviewed targets, keep the
original weights, and store checkpoints separately. Evaluate adapted candidates
before activation. See the [complete usage guide](docs/USAGE.en.md).

## SDK and runtime contract

The Kotlin SDK exposes `infer`, `train`, `save` and `restore`. Checkpoints preserve
parameters, SGD momentum and step counters; receipts check model/label identity
and file integrity. The integrating application selects the active checkpoint.

YOLO supports image dimensions that are multiples of 32. Other graphs resize to
their internal grid. The historical SDK uses TensorFlow Lite and Select TF Ops
**2.16.1**. Vision Dataset Studio's separate 16 KB Flex fix does not retroactively
qualify this SDK. [Conversion sources](conversion/README.md).

## Rights and responsible use

Source weights, datasets and software retain their own licences and restrictions.
Read each variant's notices; download access alone grants no commercial-use or
redistribution permission. In particular, YOLO upstream obligations and the
DINOv3 pilot's retained release policy remain applicable.

Synthetic checks establish parameter updates and persistence, not field accuracy
or reliable alerts. These models must not be used alone to confirm a fire,
trigger an emergency alert, order an evacuation or direct emergency response.

## Documentation and revision

Historical catalogue documentation updated in English on **5 October 2026**, from the public repository at
[`bdcd483d5c5aadeff1b531a35b536f3530906e18`](https://huggingface.co/fireviewer/litert-models/tree/bdcd483d5c5aadeff1b531a35b536f3530906e18) and its retained reports.
This update checks documentation, repository metadata and small evidence files;
it does not rerun training, inference, dataset payload verification or device qualification.
Historical receipts keep their original dates, revisions and scope. Earlier README
hashes in artifact manifests refer to those earlier releases.

The association FIRE-VIEWER provides administrative and financial stewardship for
resources under its control. Technical governance is maintainer-led. Upstream,
source-specific and pre-association rights remain separate; repository placement
does not transfer ownership or grant additional rights.

[Project documentation](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/HUGGINGFACE.md)
· [Institutional website](https://fire-viewer.fr). Contact: **contact@fire-viewer.fr**.