---
language:
  - en
license: cc-by-4.0
pretty_name: b. Apartment Interior Observations
size_categories:
  - 1K<n<10K
tags:
  - real-estate
  - evidence-retrieval
  - physical-world
  - ai-assisted-extraction
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl.gz
      - split: validation
        path: validation.jsonl.gz
---

# b. Apartment Interior Observations

Version 2026-11-10. Published by The Human Judgment Company, the operator of
[b.](https://findwithb.com/). CC BY 4.0 applies to the original observation rows.
It does not license the cited property websites, research, original media or
the rest of the source repository.

## What is included

7,600 text rows derived from 28 public apartment tour records in
Cupertino, Mountain View, Palo Alto, San Jose, San Mateo. Individual appliance, material, layout and condition
observations keep their evidence basis, stored confidence, visit date, section,
source URL and any low verification warning. These records were extracted and
interpreted with AI assistance. They are not independently verified ground truth.

The full detailed records are available at
https://findwithb.com/toured-apartments and the data method is at
https://findwithb.com/apartment-interior-data.

## Sanitization and scope

The release excludes the private reference home, daily access sheet, raw label
blocks, serial numbers, network identifiers, contact information, recordings,
transcripts, third party listing quotes and external research text. Rows that
cite outside research are omitted. Observations read from a label are concise
extracted facts, not a copy of the original label.

Public tour IDs, city and source links remain. This is a sanitized public dataset,
not anonymous data; its apartments can be identified through their public pages.
No applicant, scout or resident records are included. Absence from a recording
does not establish absence from a home.

## Fields

`id` is a stable row reference within this version. `home_id` identifies the
public tour. `city` and `visit_date` describe the visit. `section`, `subject` and
`detail` locate the observation. `text` is the stored statement, `basis` separates
observed, label, inferred and not observed records. `confidence` is an uncalibrated
estimate from the source record. `verification_note` flags a stored frame check
that did not establish the claim by itself. `evidence_count` counts stored
references; the media is not distributed. `source_url` links to the public record.
`split` is the deterministic property group assignment.

## Files, splits and possible uses

Use `observations.csv.gz` or `observations.jsonl.gz` for the complete release.
These are standard gzip files. `train.jsonl.gz` has 6,967 rows and `validation.jsonl.gz` has
633. All units from the same named property stay in
one split, but similar products and observations can occur across properties.
`manifest.json` gives counts and file checksums.

Possible uses include prototyping evidence retrieval, uncertainty preservation
and physical feature extraction. This is not an established benchmark: there are
no independent relevance judgments, hidden test labels or reported model scores.
Do not report performance on it as general apartment understanding ability.
The underlying records are already public and may be present in training data.

## Limitations and corrections

This small convenience sample is not a market survey. Dates, unit identity,
incomplete camera coverage and imperfect extraction limit what can be concluded.
It cannot establish current availability, appliance operation, regulatory
compliance, present rent, neighborhood safety or who should live in a home.
It must not be used to screen applicants or infer attributes of residents.
Send corrections with the public source URL to hello@findwithb.com.

## Attribution

The Human Judgment Company. b. Apartment Interior Observations, version 2026-11-10.
https://findwithb.com/apartment-interior-data. License:
https://creativecommons.org/licenses/by/4.0/.