Detections, checkpoints and training logs for "Shift-aware operating points for open fire and smoke detectors: calibration, benchmark leakage and label-free threshold transfer across four public domains"

Supporting data for "Shift-aware operating points for open fire and smoke detectors: calibration, benchmark leakage and label-free threshold transfer across four public domains" (Ayash Hossain Chowdhury, 2026; submitted to Engineering Applications of Artificial Intelligence, manuscript EAAI-26-29615). Code, derived artefacts, split lists and every number behind the paper's tables live in the public repository github.com/ayashhossain555/fire-detector-reliability. This record holds the four things too large for a code repository, plus a copy of the repository at the commit the paper was built from. Contents detections.zip — raw detections for every model × evaluation split at conf 0.001, max_det 300 (441 JSON); the input to every calibration number in the paper. detections_bnadapt.zip — batch-norm-adapted exports behind the "BN adaptation" rows of Table 5 (136 JSON). detections_bndeconf.zip — de-confounded test-time-adaptation exports behind Section 5.5: source-BN control, target-BN, Tent (714 JSON). checkpoints_<run>.zip (×26) — one archive per run: the trained checkpoint (best.pt) with that run's args.yaml, results.csv, training curves and console logs. Split per run (16–169 MB each, 1.372 GB in total) so a single model can be taken without downloading all 26. last.pt is excluded: the paper uses the best-epoch checkpoint only. run_queue_logs.zip — queue-level training logs not belonging to a single run. fire-detector-reliability-repo.zip — git archive of the public repository at the commit named in manifest.json. MANIFEST.md, manifest.json — per-file SHA-256, byte size and file count, plus the per-checkpoint hash table. LICENSE-DEPOSIT.txt — per-part licence terms; read before reuse. The 26 runs. Three detector families at two or three sizes, two training corpora (D-Fire, Pyro-SDIS), two seeds at the small size, plus two de-duplication controls and one no-early-stop variant: YOLOv8 (v8n, v8s, v8m), YOLO11 (y11n, y11s, y11m) and RT-DETR (rtdetrl, rtdetrx). Integrity. Every checkpoint was verified against the SHA-256 recorded in the repository's artefacts/runs_meta/<run>.json when this record was built — 26/26 matched. manifest.json carries the SHA-256 of each file here, so a download can be checked the same way. No dataset imagery. This record contains no dataset images and no original annotation files — only model outputs, model weights and training logs. Obtain each dataset from its own source under its own terms (D-Fire CC0 1.0; Pyro-SDIS Apache-2.0 / CC-BY; Roboflow final-fire-project Public Domain; HPWREN FIgLib per HPWREN credit terms). Licensing is mixed. The record-level licence is AGPL-3.0-or-later because that is the most restrictive part and it governs the trained weights, which were produced with Ultralytics (AGPL-3.0) and fine-tuned from Ultralytics pretrained checkpoints. The detection exports, manifests and training logs are CC BY 4.0. See LICENSE-DEPOSIT.txt for the operative per-part terms. FIgLib imagery: High Performance Wireless Research and Education Network (HPWREN), University of California San Diego, https://www.hpwren.ucsd.edu/.