fireviewer/dinov3-vitb16-multitask-fireviewer-v3

FireViewer DINOv3 ViT-B/16 multi-task

This repository contains the FireViewer full-parameter multi-task checkpoint based on the gated DINOv3 ViT-B/16 pretraining revision 5931719e67bbdb9737e363e781fb0c67687896bc. It exposes segmentation, anchor heatmap pointing, and explicit visual-abstention heads through the FireViewer adapter.

The DINOv3-derived weights remain subject to the DINOv3 License. The training manifest combines the FireViewer Boreal corpus, Camp Swift, RxCADRE, and FireSentry. It contains strong, weak-teacher, sensor-derived, and temporal-negative annotations. Source rights and the DINOv3 redistribution terms must be reviewed independently before any downstream redistribution or commercial use.

The release was selected by minimum validation loss: epoch 5 with validation loss 0.9348679466. model.safetensors contains the complete backbone and all three task heads; no LoRA/PEFT adapter remains to be merged. The original .pt file is retained only as a training-resume checkpoint.

Provenance

• Dataset manifest SHA-256: 896b5257cefb43f169b7292cf6be8a5612097cb77bd8250ffd8a28936b0b80e1
• Final checkpoint SHA-256: 521bfe7fa39f3203d3d20604e6f010fc72e0ae83e31cb5fa9a1223032fec4c14
• Training mode: full_model_all_parameters_trainable
• Weak/strong annotation quality remains a promotion gate; this repository is a trained challenger, not an automatic production promotion.

Loading

from huggingface_hub import hf_hub_download
from dinov3_adapter import DinoV3MultiTaskModel

repo_id = "fireviewer/dinov3-vitb16-multitask-fireviewer-v3"
weights = hf_hub_download(repo_id, "model.safetensors")
backbone_config = hf_hub_download(repo_id, "backbone_config.json")
model = DinoV3MultiTaskModel.from_safetensors(
    weights,
    model_id="facebook/dinov3-vitb16-pretrain-lvd1689m",
    revision="5931719e67bbdb9737e363e781fb0c67687896bc",
    image_size=512,
    backbone_config=backbone_config,
    token=False,
)
network = model.network

The loader constructs the DINOv3 architecture from the bundled pinned configuration, then loads the complete safe state strictly. It does not download redundant base-model weights.