Datasets:
text stringclasses 2
values |
|---|
W0503 10:28:37.981417 8429366336 rdf.py:89] WARNING: The JSON-LD `@context` is not standard. Refer to the official @context (e.g., from the example datasets in https://github.com/mlcommons/croissant/tree/main/datasets/1.0). The different keys are: {'includes', 'regex', 'equivalentProperty', 'md5', 'column', 'transform'... |
I0503 10:28:37.992835 8429366336 validate.py:53] Done. |
Recaptioned T2I Supervision Evaluation Sample
This repository is the small reviewer-inspection companion to the full anonymous caption-metadata release. The full release is hosted separately at https://huggingface.co/datasets/Anonymous1477/recap-t2i-evaluation-metadata-2026; this repository stays under the large-dataset sample threshold and gives reviewers a direct way to inspect redacted caption metadata, join structure, and selected image-conditioned audit packages.
Files
caption_records_metadata.redacted.jsonl: redacted sample caption records with source keys, caption hashes, sensitive-content flags, and released caption text where permitted.caption_records_metadata.redacted.csv: tabular equivalent of the JSONL records.caption_pair_groups_metadata.redacted.jsonl: grouped same-image or same-source-row records for reconstructing caption comparisons.redaction_report.json: contact-string redaction counts and sensitive-content flag counts.image_audit_packages/: selected non-Danbooru image packs with package-relative images, request JSONL, captions, manifests, and checksums for direct visual audit reruns.croissant.json: Croissant metadata with core and minimal Responsible-AI fields for this sample repository.
Image policy
The sample includes selected DataComp image-audit packages whose request manifests have been stripped of local absolute paths. Danbooru-family image packages are excluded from this public review sample because their source-content risk is higher; the full release keeps those rows as source-key/hash metadata with risk flags instead of redistributing image bytes.
Content and privacy flags
Generated captions can include OCR-derived contact strings and sensitive or sexual descriptions inherited from source images. This sample redacts email, URL-like, and phone-like strings in caption text, preserves sensitive_text_flag, and leaves flagged non-Danbooru rows available for reviewer filtering and quality audit.
Sample construction
The redacted metadata files come from the May 2 reviewer staging sample. The image-audit packages are selected from the file-mode evaluation packs used for the paper: the DataComp same-captioner policy ablation and the DataComp reference-paired audit. The sample excludes Danbooru image bytes and larger auxiliary image packages because they are not needed for the under-4GB quality-inspection companion.
Intended use
The sample is intended for caption-pair inspection, schema inspection, and lightweight reruns of image-conditioned audit requests. It is not the complete dataset and should not be used for demographic inference, identity recognition, or safety certification.
Release links
- Full caption metadata:
https://huggingface.co/datasets/Anonymous1477/recap-t2i-evaluation-metadata-2026 - Small reviewer sample:
https://huggingface.co/datasets/Anonymous1477/recap-t2i-evaluation-sample-2026 - Code and evaluation summaries:
https://huggingface.co/Anonymous1477/recap-t2i-evaluation-code-2026
Redaction summary
- Sensitive caption records in the redacted metadata sample:
None - Email-like redactions:
None - URL-like redactions:
None - Phone-like redactions:
None
- Downloads last month
- 31