Datasets:
TunnelInstruct
1,740 original instruction-response records over 1,725 unique training images. The original class-conditioned training instruction and response are reconstructed with the frozen historical event_task function; context/evidence/assessment are copied verbatim. Special output tokens belong to that training interface. Descriptions are model-assisted references, not independently human-certified diagnoses. This is the historical TunnelInstruct collection, distinct from the later 192-reference image-facts ablation subset.
Organization
Native images and instruction-response records are embedded in Parquet shards organized by the target category. samples.jsonl identifies each record and index--*.jsonl provides the category indices. The fourteen configurations comprise thirteen defect categories and normal examples; all loads the complete collection. Multiple records may use the same image with different target categories.
from datasets import load_dataset
ds = load_dataset("augustusper/TunnelInstruct", "all", split="train", token=True)
Source and usage
Source datasets: TACK (https://arxiv.org/abs/2512.14477), dacl10k (https://openaccess.thecvf.com/content/WACV2024/html/Flotzinger_dacl10k_Benchmark_for_Semantic_Bridge_Damage_Segmentation_WACV_2024_paper.html), RFDD (https://doi.org/10.1038/s41597-026-07851-7), the CTCD release used by the project, and author-collected tunnel imagery. See per-record source_dataset. Original source terms continue to apply. This private research archive does not establish a blanket redistribution license; public release requires source and author permissions. No test, validation, or cross-tunnel diagnostic split is included.
Source and synthetic provenance are retained per record. Images are not resized or regenerated. No private file-system paths, credentials, run logs, or model weights are included. export_summary.json and SHA256SUMS record the package contents.
- Downloads last month
- 2,211