This dataset contains automatically de-identified customer-derived research histories. Access is granted individually after manual review by CollovLabs. Describe your intended research use. Do not attempt to identify or contact people referenced in these records.

Collov Home — full project traces

Start here: online colleague guide and full trace viewer.

The initial release contains 1,000 real room-design projects grouped into 400 pseudonymous source accounts: 200 users with two projects and 200 users with three. One project has one complete recorded trace, preserving the source conversation, intake, questionnaire context, proposal versions, saved canvases, design deliveries and recorded choices/actions.

Release validation passed: 306,705 unique source events, 156,502 saved canvas states, and 31.91 GB of image archives. See reports/release-summary.json for measured coverage.

Access

This repository is public with manual access approval after staging finishes. Request access with your individual Hugging Face account. A CollovLabs administrator reviews each request; access is not automatic. Sign in with that approved account before loading data. In the online viewer, choose Sign in with Hugging Face using the approved account, then select a user and project. The viewer fetches real records and images directly from Hugging Face after authorization; an existing HF login alone does not sign you into this separate viewer. The public site bundle contains only documentation and an optional illustrative example. Real data remain behind this repository's access gate. The HF automatic table viewer is disabled.

Download one project for local viewing

pip install -U huggingface_hub
hf auth login
hf download CollovLabs/collov-home-traces tools/download_project.py \
  --repo-type dataset --local-dir collov-traces

python collov-traces/tools/download_project.py --list
python collov-traces/tools/download_project.py \
  --project p_YOUR_PROJECT_ID --output collov-project

Then open the full-history viewer and choose collov-project. Files are read locally in your browser. Every response is shown in chronological order; images of a board proposal or delivery batch scroll horizontally in one row. Internal canvas saves and reviews remain inspectable through the viewer checkbox.

The helper uses authenticated byte ranges to retrieve selected image members from the media archives and checks each SHA256. Each run resolves one immutable repository revision so later corpus updates cannot mix files within the download. Use --all-media to include all referenced canvas dependencies and scoped product imagery. --revision v1.1.0 pins the current release; v1.0.0 stays available and has the same traces. Missing source images remain explicit.

Organization

users/<user_id>/profile.json
traces/<user_id>/<project_id>.jsonl.gz
indexes/projects.jsonl
indexes/users.jsonl
indexes/artifacts.jsonl
indexes/media.jsonl
tables/<batch>-events.parquet
tables/<batch>-products.parquet
tables/<batch>-style_cases.parquet
tables/<batch>-orders.parquet
media/<batch>-<shard>.tar
boards/<batch>-<shard>.tar
inspiration/cases.jsonl
inspiration/images-<shard>.tar
schemas/  tools/  docs/  reports/

The first line of a project trace is record_type: project_trace: it identifies the project and user, indexes artifacts/media and includes scoped furniture metadata. Following lines are historical events. Each has a stable event ID, chronological sequence, type, actor, relative time, visibility, turn/parent references, artifact IDs, source payload and provenance. All saved canvas states are embedded in their events. The Parquet event table repeats the complete event as JSON with filterable columns.

User profiles contain observed source fields and provenance, account-level context and selected-project relations. They do not contain invented personas. Account-level events may be repeated in several project traces with the same event ID; deduplicate by event ID for account-wide analysis. Events with unknown times are retained last.

Archives extract to content-addressed media/ and boards/ member paths. Public media descriptors include the stored SHA256, byte count, archive and payload offset. Non-media external links are retained as host-only references. Source identifiers and original private URLs are not provided. Full source exports and pseudonym keys remain internal.

Inspiration library (v1.1.0)

inspiration/cases.jsonl holds the inspiration photos ("cases") clients chose in the brief's look step or in the style quiz. It covers every case picked in any paid project brief, not only those in these 1,000 traces: 16,506 cases, of which 5,793 appear in these traces. 15,009 are labelled. 16,320 images are available: some were already in media/, and the rest are new in inspiration/images-<shard>.tar (3 archives, 2.9 GB). One row per case:

case_id: the catalog id. The brief's per-photo notes are keyed by it (allRefer, imgRefer, decRemark, references). It identifies a catalog photo, not a person.
trace_case_ids: the ids the traces use for the same case in selectIds and style_imgs. These pseudonyms differ per batch, so one case can have several.
labels: the catalog team's tags: style, room, color, material, key_feature, and where set, pattern and perspective.
picked_in_paid_projects and noted_in_paid_projects count the paid projects that picked the photo, and those that wrote likes or marked items on it.
image: the member, archive, payload offset, bytes and SHA256 of the photo. in says whether it sits in media/ or inspiration/. A photo whose source was already gone is marked unavailable.

The photos are third-party inspiration images; source_host keeps only the host (for example www.pinterest.com). They follow the media policy below.

What “full” means

The trace contains every scoped recorded source row, rather than a conversation summary: client/designer/account chat, room brief fields, stored account questionnaires, all board/proposal/canvas versions, internal design notes and reviews, original delivered render-file records, furniture lists, product reactions, surviving cart rows, order/package states and project milestones. Reconstruction images and original delivery conversions are labelled separately.

Historical clickstream/deletion history and overwritten source field versions were not recoverable. Some source URLs are unavailable. A board image is reconstructed from its latest nonempty saved canvas state; exact historical delivery pixels and delivery-time state are unverified. Proposal creation is a sending-time proxy. Render upload-day grouping in a viewer is an inferred batch, not a new design revision.

Time, questionnaires and future information

Times are milliseconds relative to the shared origin for that user (the earliest selected project or stored quiz). Original absolute dates are withheld. The source database timezone is unverified. Coarse project years are retained.

Questionnaires are user-level and retain their timing. “No quiz recorded at intake” means no surviving quiz row predates project creation; it does not prove the client never took one. A later quiz stays later. Current stored answers may have overwritten earlier answers. The source questionnaire fields include original option codes; their original meanings are not replaced by inferred persona labels.

This is a full-history research corpus, not a safe as-of observation stream. Mutable source rows and later outcomes can reveal future information. Build audience- and time-filtered simulator observations separately, and keep future approvals/final designs inaccessible during agent evaluation. Keep train/evaluation assignments at user level. This release does not prescribe benchmark tasks or score a simulator.

Media policy and privacy

Opaque media use JPEG quality 85 with max edge 2048 when appropriate. Transparent cut-outs preserve decoded RGBA through lossless WebP or sanitized source encoding; ICC color profiles are retained. Multi-frame inputs are mapped to sanitized frame artifacts. Board JPGs use native dimensions, quality 90 and 4:4:4 chroma. Fine text and dimension labels need review before use; converted images are not original bytes. Incomplete source JPEG recoveries are explicitly flagged. If a multi-frame source declares an unavailable frame, valid frames are preserved and the missing frame is recorded separately; no thumbnail or synthetic image replaces it. See the media size and source-quality report for counts. If an unavailable request recovers in a later verified batch, earlier trace descriptors use that recovered image with explicit late_recovery provenance. Original download attempt counts remain unchanged; late recovery evidence records the correction separately.

v1.1.0 recovers questionnaire uploads that v1.0.0 marked unavailable but that were still online (report): 12 HEIC photos and 6 measurement documents (PDF, PowerPoint, Word). Documents were redacted before rendering: addresses, lot/block and job numbers, signing IDs, contact details and client names are blacked out, partly by rule and partly by hand after review. Each page is stored as an image in frames, and the de-identified text is kept in document.text. These descriptors carry late_recovery.

Known client names, contact details, street addresses, private URLs and relational identifiers are transformed automatically. EXIF/GPS and nonessential metadata are removed from released images. Automated de-identification does not guarantee anonymity; text/image content can retain identifying context. Respect the access notice and do not attempt reidentification, contact or unauthorized redistribution.

The cohort is a reproducible sample from structurally complete, paid normal projects with board proposals/responses, delivered renders, furniture lists and client/designer chat. It preserves the 16-user pilot and adds seeded user bundles. Three-project bundles use earliest/middle/latest eligible projects; accounts with more than 20 eligible projects are excluded. This is not a random sample of all customers or all room types. Unselected projects from the same account are outside the cohort.

Documentation and receipts

Public documentation contains no real customer records or images. Please request dataset access before using the real traces with the viewer.

Downloads last month: 510

Total file size:

36.1 GB