Clarify archived checkpoint protocol

README.md

@@ -11,7 +11,12 @@ tags:
 pipeline_tag: image-segmentation
 ---
 
-# Damage-TriageFormer (trained checkpoint)
+# Damage-TriageFormer (archived tile-mode checkpoint)
+
+> **Protocol notice.** This checkpoint was trained on the superseded tile-level
+> split and is not the primary model in the revised manuscript. The revision uses
+> photo-atomic folds and reports a crop-mode primary result. This checkpoint is
+> retained only to reproduce the originally submitted tile-mode experiment.
 
 Trained weights for **Damage-TriageFormer**, a foundation-model framework for
 decision-relevant building damage typology from single post-event imagery. Given
@@ -28,7 +33,7 @@ each building one of five damage-typology classes:
 
 ## Files
 
-- `best_model.pth` — the reported checkpoint, selected by validation macro-F1 in the footprint-conditioned setting.
+- `best_model.pth` — the originally submitted tile-mode checkpoint, selected by validation macro-F1 under the superseded split.
 - `config.json` — the exact training configuration used to produce it.
 
 ## Architecture
@@ -48,9 +53,11 @@ DDP), label smoothing 0.1, EMA (decay 0.9995), long-tailed logit adjustment
 
 ## Results
 
-Macro F1 **0.624** (validation) / **0.619** (held-out test) on the stratified
-split; per-class test F1 of 0.91 (Undamaged) and 0.84 (Total Structural Collapse).
-Total Roof Damage remains the hardest class (F1 ≈ 0.33).
+Macro F1 **0.624** (validation) / **0.619** (test) on the superseded tile-level
+split. These values must not be cited as results from the revised photo-atomic
+protocol. The full tile recipe retrained on the photo-atomic split reaches
+0.575 validation / 0.579 test macro F1; the revised crop-mode primary model
+reaches 0.646 / 0.616 and is a different checkpoint.
 
 ## Usage
 
@@ -65,10 +72,11 @@ DamageTriage-Bench: [huggingface.co/datasets/Ymx1025/DamageTriage-Bench](https:/
 
 ## Intended use and limitations
 
-Screening-grade decision support for post-disaster triage, not a substitute for
-engineering inspection. Footprint-conditioned (assumes building masks are
-available). Rare and visually ambiguous roof-damage categories, especially Total
-Roof Damage, are the least reliable. Trained on a single seed.
+Research and reproducibility use only, not operational decision support and not
+a substitute for engineering inspection. The model assumes registered building
+masks. It was trained on a single seed and on a split now superseded because
+tile-level assignment allowed fragments from the same source photo to cross
+folds. Buildings clipped at tile boundaries remain incomplete observations.
 
 ## License
 
@@ -79,13 +87,10 @@ Response Imagery and the source building-footprint layers.
 ## Citation
 
 ```bibtex
-@misc{xiao2026damagetriageformerfoundationmodelframeworktypologybased,
-      title={Damage-TriageFormer: A Foundation-Model Framework for Typology-Based Building Damage Assessment from Mono-Temporal Imagery}, 
-      author={Yiming Xiao and Yu-Hsuan Ho and Sanjay Thasma and Junwei Ma and Ali Mostafavi},
-      year={2026},
-      eprint={2606.12248},
-      archivePrefix={arXiv},
-      primaryClass={cs.CV},
-      url={https://arxiv.org/abs/2606.12248}, 
+@misc{damagetriageformer,
+  title  = {Damage-TriageFormer: A Foundation-Model Framework for Decision-Relevant
+            Building Damage Typology from Post-Event Imagery},
+  author = {Xiao, Yiming and Ho, Yu-Hsuan and Thasma, Sanjay and Ma, Junwei and Mostafavi, Ali},
+  year   = {2026}
 }
 ```