gapatron/astrobridge-image-captions

AstroBridge Legacy Survey Captions

3,487 imaging cutouts from the Legacy Survey (DR10 South + North), crossmatched against published literature mentions and captioned in four independent stages by Gemini (gemini-3.7-flash), following the AstroLLaVA data-generation approach (Zaman et al. 2025, arXiv:2504.08583): no caption is ever told the object's real name or catalog designation, and no caption states a fact that isn't derivable from the pixels or the (redacted-at-the-model, checked-on-output) literature context itself.

• legacy-south: 2,399 objects, hugging-science/mmu_legacysurvey_dr10_south_21
• legacy-north: 1,088 objects, UniverseTBD/mmu_ssl_legacysurvey_north
• Literature crossmatch: astronolan/galaxy-mentions-hats

What's in a row

field	description
object_id, survey, ra, dec	identity and position
rgb, rgb_crop	rendered display composites (z=R, r=G, g=B; adaptive asinh stretch, per-survey white-balance, ivar-based chroma confidence where available) - the exact pixels shown to the model
flux_{g,r,i,z}, ivar_{g,r,i,z}, mask_{g,r,i,z}	raw calibrated per-band tensors, variable shape per row (see height_*/width_*) - legacy-north has no ivar/mask/i-band at all (null), a real gap in that release, not a missing-data artifact
psf_fwhm_{band}, scale_{band}	per-band PSF FWHM and pixel scale
caption_blind	image alone, open-ended description
caption_properties	image alone, structured resolvedness/color/structure estimate (generated in the same Gemini call as caption_blind - see below)
caption_literature	image + capped literature excerpts + real name (for disambiguation only)
caption_fused	editor pass reconciling the three drafts above against the pixels
mag_{g,r,i,z,w1-w4}, effective_radius_arcsec, ellipticity, position_angle_deg	real catalog photometry/shape - QC reference only, never shown to the model; north rows have null i-band/WISE/shape (no shape fit or that photometry in north's catalog)
any_leaks_detected, leak_flags_json, any_qc_flags, qc_flags_json	automated review flags - see Known issues below before treating a flagged row as a real problem
multipeak_qc_json	deterministic double-peak pixel check, QC-only (null/None fields for legacy-north, which has no ivar to compute S/N from)
literature_mention_ids_selected, n_literature_mentions_total	audit trail of which literature excerpts informed stage 3/4 - raw excerpt text is deliberately not included (see below)
renderer_version, prompt_set_version, model	reproducibility metadata

Pipeline

Every object gets 3 independent, stateless Gemini calls (not 4 - stage 1 and stage 2 are both blind, image-only generations with no name and no literature, so they're issued as one merged call returning two JSON fields, cutting prompt tokens by ~37% since the image only has to be paid for once):

1. Merged call → caption_blind (open-ended) + caption_properties (structured, resolvedness-gated) - image only.
2. Literature call → caption_literature - image + up to 8 capped literature excerpts + the real name, shown only so the model can attribute excerpts correctly.
3. Fusion call → caption_fused - given all three drafts above as proposals, not ground truth, the model re-examines the pixels itself and writes one final caption, resolving disagreements by looking at the image, not by trusting whichever draft sounds more confident.

An absolute rule is prepended and re-checked at the end of every single call: never output the object's name, any catalog designation, or any other identifying string - for this object or any other one that happens to be mentioned in the literature excerpts. Two casual/expert conversation stages exist in the pipeline but are currently deferred (not generated here) - LLaVA second-stage material, out of scope for this release.

Known issues

• any_leaks_detected false positives from single-letter identity strings. 2 objects (one per survey) have a bare single-letter name variant in their literature record (a gravitational-lens-arc component label, e.g. "arc A") that the leak detector's word-boundary match confuses with the English article "a" anywhere in ordinary prose. Both were hand-reviewed: no real identity leaked. Not yet fixed upstream.
• Some any_leaks_detected/any_qc_flags rows reflect stale, pre-deferral conversation-stage content, not this dataset's actual caption_* fields - leak_flags_json/qc_flags_json are per-stage, so check which stage a flag actually landed on before treating a row as suspect.
• QC/leak flags are coarse lexicon/regex review markers for a human, not an automated rejection filter - they are included for transparency, not as a quality guarantee or its absence.

What's deliberately excluded

Raw literature excerpt text (copied passages from published papers) is not included in this public release - only the mention_ids that were shown to the model, so the exact provenance of stage 3/4 is still auditable via the source literature-crossmatch dataset above.

Citation

If you use this dataset, please also cite the source imaging and literature-crossmatch datasets linked above, and the AstroLLaVA paper this generation approach adapts:

@article{zaman2025astrollava,
  title={AstroLLaVA: towards the unification of astronomical data and natural language},
  author={Zaman, Sharaf and others},
  journal={arXiv preprint arXiv:2504.08583},
  year={2025}
}