00start/lxra-public-dataset

LXRA-Public-50

Status: v0.1.0 in preparation. Hand-labeling underway. The dataset will be populated with 50 hand-labeled level-crossing inspection images once the labeling pass and final PII review are complete. This card is a placeholder so researchers landing here understand the eventual scope.

A dataset of 50 hand-labeled level-crossing inspection images sourced from public transportation-investigation reports (RAIB UK, TSB Canada, ATSB Australia) plus license-compatible imagery from Wikimedia Commons. Released under CC BY 4.0, with per-image upstream licenses preserved in metadata.

This is a probing dataset, not a benchmark. It is intended to support research on visual-language model groundedness in outdoor- infrastructure inspection tasks (rail level crossings specifically). The companion internal report at docs/strategy/2026-05-09-track-c-vlm-probing-report.md in the source repo describes the methodology and the explicit limits on cross-vertical extrapolation — LXRA scores do not make maritime or aviation competence claims.

Source

• Mirror repo: https://github.com/00start/lxra-public-dataset (mirror; this Hugging Face dataset card is canonical)
• Authoring repo: https://github.com/00start/forceshield
• Methodology: see the design spec at docs/superpowers/specs/2026-05-07-aviation-generalization-and-maritime-demo-design.md in the authoring repo, §6 (Track C).

What will land at v0.1.0

• 50 JPEG images at images/lxra-001.jpg through images/lxra-050.jpg.
• One metadata.jsonl line per image with: file_name, image_id (SHA-256), subject, condition (pass/warn/fail), salient_cues (list), reasoning, source_url, optional docket_id, license, attribution, reviewed_at.
• DATASET_LICENSE.md declaring CC BY 4.0 for the dataset as a whole, with per-image license precedence.

What this is NOT

• Not fine-tuning data. The companion methodology is foundation- model + rubric-as-prompt; no model weights change.
• Not a benchmark. N=50 is too small for confident vendor comparisons. See the methodology document for the noise band.
• Not multi-vertical. Aviation and maritime probing datasets are explicit follow-up sprints under separate scope. LXRA is the only vertical at v0.1.0.

License posture (eventual)

The dataset as a whole will be released under CC BY 4.0. Per- image upstream licenses (preserved in metadata.jsonl) include:

Source	Per-image license
RAIB (UK)	Open Government Licence v3.0
TSB Canada	Open Government Licence Canada
ATSB (Australia)	CC BY 4.0
Wikimedia Commons	CC BY 4.0 / CC BY-SA 4.0 / CC0 / public-domain (per-image, verified at audit time)
NTSB / NTSB-Marine	public-domain (US Government Work, 49 USC § 1112) — if included

Users of the populated dataset must comply with the more-restrictive of (a) the dataset license and (b) the per-image license recorded in metadata.jsonl.

Takedown policy

When the dataset lands, any image flagged by a source agency or rights-holder will be removed within 7 business days of receipt of a written takedown request. Contact: takedown@00start.com.

Citation (eventual)

@misc{lxra-public-50,
  author    = {00start},
  title     = {LXRA-Public-50: Level-Crossing Inspection Imagery from Public Investigation Reports},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/00start/lxra-public-dataset}
}

Why this card exists before the data

Provisioning the Hugging Face dataset repo before populating it lets external collaborators see the shape of the upcoming artifact (its license, scope, methodology, takedown commitment) ahead of the data itself. The auto-generated stub README that ships with a fresh HF dataset repo declares license: mit (which is the code license on the GitHub mirror, not the content license for the dataset) — this README corrects that and adds the context.