harshk-jebitech/uk-property-inspection-sample

miProgram UK Residential Property Inspection Sample

A 423-image labelled sample of professional UK residential property inspection photography, drawn from real production inspections carried out through the miProgram platform, prepared for evaluation by AI training-data buyers and data-licensing partners.

This is a sample. Full commercial dataset (approximately 120 million professionally-labelled images and growing) available under separate commercial licence — contact data-licensing@miprogram.co.uk.

Dataset overview

Images	423 JPEG files, each with a matching JSON metadata file
Source	Live production inventory, check-in and check-out inspections of UK rental properties
Window	Inspections uploaded 1–15 February 2026
Labelling	Applied at point of capture by professional inspectors via closed picklists (42 reference lists — see TAXONOMY.md)
Capture devices	Inspectors' own iOS devices (iPhone 8 → iPhone 17 generations, iPads); real device EXIF preserved
Resolution profile	Predominantly 480x640 (platform sync-optimised); ~1–2% at native camera resolution (3–5 MP). Second-generation JPEGs at quality 95.
Privacy	Every image individually reviewed and captured on vacant properties with no tenant in occupation. GPS EXIF stripped. Location expressed as region only. House numbers may be visible on exterior elevations; these identify the property rather than any individual and are not personal data under UK GDPR. People, vehicle registration plates, and readable documents are excluded.
Licence	Evaluation and research only — see LICENSE.md

The full dataset this sample is drawn from spans several years of production use, hundreds of thousands of inspections, and tens of millions of images, growing continuously. This sample's structure, label vocabulary, and quality profile are representative of it.

A note on condition ratings

Condition ratings are applied to individual elements rather than whole rooms. Room overview images therefore carry no condition rating by design. Of the 423 images, 210 are element-level and carry a rating; the remainder are room overviews and contextual shots.

Intended use

This dataset is provided for evaluation and research use, specifically to allow AI teams to assess:

• Label quality and consistency of miProgram's structured taxonomy
• Image quality and diversity across property types, rooms, conditions, and defects
• Suitability of the underlying data for training or evaluation of vision models on property condition, material, and defect detection tasks
• Fit for downstream commercial products in insurance, real estate, property analytics, and adjacent domains

Commercial use, including training machine-learning models for commercial products or services, requires a separate written licence agreement with Mobile Inventory Program Limited.

Composition

The sample is deliberately stratified, not random. Five strata:

1. Room-type coverage — 177 images. Room overview photographs across 19 of the platform's 21 room types: 10 each for Bathroom/Toilet, Bedroom, Cellar/Basement, Conservatory, Dining Room, En-suite, Garage, Garden and Outdoors, Hallway, Kitchen, Landing, Living Room, Porch, Reception Room, Stairs, Utility, WC; plus Attic (3) and Outbuilding (4), which are honestly thin in the source window. Two room types ship empty: Loft (no usable labelled photographs in the window) and Open Porch (every candidate was a close-range doorway shot dominated by the house number or off-subject; none met the selection bar under the conservative v1 review).

2. Condition-rating demonstration — 92 images. Element photographs (walls, floors, ceilings, doors, windows) carrying the platform's condition ratings: 10 each for Aged, As New, Brand New, Clean, Fair, Good, Newly Decorated, Poor; Dusty (7) and Used (5) reflect genuinely rare usage; Rusty had no usage in the window. Every image was reviewed for label credibility — each is a visually defensible exemplar of its rating.

3. Defect showcase — 88 images. Element photographs carrying the ten most common genuine defect labels in the window: Chipped, Discolouration to edges, Flooring marked, Flooring worn to walkways, Marked, Paint splashes and marks, Patchy paintwork, Plaster - cracked, Slightly Scuffed, Small Marks (6–10 each). The primary defect is visible in every image; many images carry additional defect labels in their metadata, exactly as recorded.

4. Property archetype range — 36 images. Flat (10), Detached (10), Semi (8), Terraced (8): front elevations plus representative interiors per archetype. Elevations are under-represented relative to interiors: this sample's selection was conservative about prominent house numbers, street signage, and vehicle plates in street-facing frames. As the privacy note above sets out, house numbers on vacant-property elevations identify the property rather than any individual; larger exterior sets are available for commercial supply without this selection constraint.

5. Product-level detail — 30 images. Kitchen appliances with recorded brand labels (Hotpoint, Indesit, AEG, Beko, Lamona, Russell Hobbs and others — 20 images, several with the brand marking legible in frame) and typed fixtures (switches, sanitaryware, worktops, heated towel rail, mirror — 10 images), demonstrating the granular product-scale metadata alongside room-scale views.

manifest.json carries the full per-image catalogue and these counts in machine-readable form.

What to know before evaluating

Stated here because you will find them anyway, and they are properties of the full dataset too:

• Resolution profile. The platform's sync pipeline stores images predominantly at 480x640 (~40–180 KB); roughly 1–2% are native camera resolution (3–5 MP, 2–4 MB). This sample reflects that mix honestly. All images are second-generation JPEGs (one quality-95 re-encode at upload).
• capture_date vs upload_date. capture_date comes from device EXIF and occasionally reflects a wrong device clock or camera-roll import; upload_date is platform-recorded and authoritative.
• Room-overview images carry no condition rating — condition attaches to elements, not rooms, so labels.condition_rating is null by design in the room-type stratum.
• QA is report-level, not per-image; annotator.report_qa_status reflects the parent report's workflow state.

See METHODOLOGY.md for the full account of how labels are produced and what guarantees apply.

File structure

uk-property-inspection-sample/
├── README.md                       this file
├── LICENSE.md                      sample licence terms
├── METHODOLOGY.md                  how labels are produced
├── TAXONOMY.md                     full enumeration of all 42 reference lists
├── manifest.json                   machine-readable catalogue of all 423 images (original flat paths)
├── checklists.json                 pass/fail presentation checklist for all 17 area types (119 items)
├── Inspection & Maintenance/       canonical layout: images refiled into category folders
│   ├── bedroom/  kitchen/  living_room/  corridor/  bathroom/  toilet_wc/  stairs/  ...
│   ├── exterior/  garage/  garden_outdoors/  outbuilding/  attic_loft/  ... (17 folders)
│   └── metadata.jsonl              flat label table, file_name = sorted path (see below)
└── metadata/                       img_YYYYMMDD_NNNNNN.json (423 files, per-image ground truth)

Filenames encode the upload date and a sequence number only — no property or inspection identifiers. Image ids, basenames and all expert metadata are unchanged; only the folder a file lives in reflects its room/area category.

Sorted layout

The dataset ships in a category-sorted layout (produced by the included inspection_image_sorter tool; the original flat images/ layout can be checked out from the original-layout git tag):

• 423 images in 17 concise folders derived from each image's room_type where one exists, with similar labels pre-merged exactly as named differently in the field: living_room (= Living Room + Reception Room), bathroom (= Bathroom/Toilet + En-suite), corridor (= Hallway + Landing), attic_loft, porch. Counts: bedroom 70, kitchen 71, living_room 55, bathroom 48, corridor 40, conservatory 14, dining_room 14, porch 13, stairs 20, toilet_wc 12, basement_cellar 12, attic_loft 5, garage 10, utility 10, garden_outdoors 10, outbuilding 4, exterior 15.
• Provenance: 408 images were filed straight from the expert room_type labels recorded at capture; the 15 images with no room_type are front elevations (element_type: Front elevation) and are filed as exterior. No image was classified by a model in this layout; every folder is a direct function of the expert labels in metadata/.
• metadata.jsonl inside the sorted folder is the same flat table as the catalogue, with file_name updated to each image's folder; all other columns (image_id, room_type, element_type, condition_rating, property_archetype, stratum) are byte-identical to the original. metadata/ and manifest.json remain image_id-keyed and unchanged.

Checklists

checklists.json (repo root) holds a property-presentation inspection checklist for all 17 area types of the sorted layout — 119 pass/fail items keyed to the same folder keys (bedroom, kitchen, bathroom, ...):

• Item schema: id (BATH_CL_001 — <AREA>_CL_<nnn>, contiguous per area), name, description (photo-verifiable, measurable pass criteria), category.
• Six categories per area: arrangement, cleanliness, amenities, functionality, safety, damage; items are grouped and ordered in that sequence inside each area.
• Depth: major rooms carry 8–12 items (bathroom 12, kitchen 12, bedroom 11); secondary and external areas carry 4–8 (porch 4, outbuilding 4, exterior 8 ...).
• The wording is deliberately binary and visual (e.g. nothing on the floor, lid closed, no black mould in grout), so the same file grades handover photos or an in-person walk-through, and can drive automated image-based grading later.

Example metadata record

Each image's JSON is self-contained. A typical record:

{
  "image_id": "img_20260203_000004",
  "inspection_type": "Inventory",
  "upload_date": "2026-02-03",
  "capture_date": "2026-02-03",
  "capture_device": "Apple iPhone 14",
  "property":  { "archetype": "Mid-Terrace", "property_type": "House", "region": "Yorkshire and the Humber" },
  "location_in_property": { "room_type": "Kitchen", "floor": null, "room_label": "Kitchen" },
  "subject":   { "element_type": "Wall", "element_subtype": "Painted Plaster" },
  "labels":    { "condition_rating": "Good", "material": "Painted Plaster", "colour": "White",
                 "finish": null, "brand": null, "defect_types": ["Small Marks"] },
  "annotator": { "role": "Client Admin", "report_qa_status": "Sent to client" },
  "image_technical": { "dimensions_pixels": "480x640", "aspect_ratio": "3:4",
                       "orientation": "Portrait", "file_size_kb": 78 }
}

Fields are always present; inapplicable values are null; empty arrays are []. All label values come from the reference lists in TAXONOMY.md.

About Mobile Inventory Program Limited (miProgram)

miProgram operates the UK's leading property inspection platform. Approximately 1,505 client firms (letting agents and inventory service providers) use the platform to produce professional inspection reports on their managed properties. The platform holds approximately 906,000 residential properties and 1.9 million inspection reports to date, with approximately 120 million professionally-labelled photographs across the full dataset. All photographs are captured through a proprietary mobile application and stored on miProgram infrastructure, with clean derived-data ownership defined in the platform's terms of service.

Commercial licensing

This sample is provided for evaluation and research only. Commercial use — including training machine-learning models for commercial products or services — requires a written licence agreement with Mobile Inventory Program Limited. See LICENSE.md.

Higher-resolution original captures may be available under commercial licence terms.

Commercial licensing enquiries: data-licensing@miprogram.co.uk

Website: miprogram.co.uk

Citation

If referencing this sample in research or publications, please cite as:

miProgram UK Residential Property Inspection Sample v1.1 (2026)
Mobile Inventory Program Limited
https://huggingface.co/datasets/miprogram/uk-property-inspection-sample