Jannchie/silva-luna-25k

SILVA-luna — 25,000 illustrations ranked by a VLM judge

Aesthetic labels for 25,000 anime illustrations, produced entirely by openai:gpt-5.6-luna and paired with the frozen SigLIP2 embedding of each picture. This is the training set behind Jannchie/silva-luna.

The images are not here, and cannot be recovered from what is. Each row carries a 1152-dim pooler_output from google/siglip2-so400m-patch14-384, which is what makes the labels usable without redistributing artwork nobody in this repo owns.

How the labels were made

Not by asking for a score. Asked to rate one illustration 0-100, this judge answers inside a 68-92 band and resolves about 2.3 distinguishable levels — the range is nominal. Scoring eight at once widens the spread but not the resolution: the per-picture noise rises with it.

So the judge was asked for an order instead. Each call showed eight illustrations and asked for a ranking, best to worst — a purely relative task, and one that settles 28 comparisons per call. Every picture was ranked in two rounds against different opponents, giving degree 14 in a single connected comparison graph, and the 6,250 rankings were pooled with Plackett-Luce into one latent per picture.

pictures	25,000
rankings collected	6,250 (8 per call, 2 rounds)
implied comparisons	175,000
degree per picture	14
split-half reliability of the latent	0.827
order-swap consistency of one answer	0.844 → ceiling 0.915
agreement with the collector's own human labels	0.307

That last row is the point rather than a defect: this is a parallel taste, not a proxy for anyone's. It correlates with a human's, monotonically across all five bands, and explains about 9% of the rank variance. Use it as a second opinion, not as free human labels.

Columns

column	meaning
embedding	1152-dim SigLIP2 pooler_output, float32
luna_latent	Plackett-Luce latent, logits, centred on zero — the raw ordering
luna_score	luna_latent rank-mapped onto 1-5, smooth
luna_band	the same map, unsmoothed, integer 1-5
degree	comparisons this picture took part in
disagreement	the fit's own residual — high means the judge places it inconsistently, which is where a pairwise budget is worth spending
split	train / val / test, assigned by near-duplicate cluster, not by row
post_id	opaque identifier from the collector's library; no external meaning

Use the split. It was drawn over near-duplicate clusters because a content hash does not stop reposts and recompressions: 21% of a hash-split validation set turned out to have a train neighbour at cosine ≥ 0.99, which inflated Spearman by 0.058. A random re-split will quietly give you that back.

What it is worth

A head trained on this reaches 0.818 Spearman against the judge's own held-out labels — 90% of the 0.909 that the label noise allows. Both axes are saturated: doubling the pictures buys +0.0062, doubling the degree +0.0077. There is little left to collect from this judge.

Content

Sampled only from pictures rated 0-2 on the collector's 0-4 scale (72.6% of the library); explicit material was excluded before anything was sent to a hosted judge.

Citation

@software{pan2026silva,
  author  = {Pan, Jianqi},
  title   = {{SILVA}: {SigLIP}-based Illustration Visual Aesthetic Scorer},
  year    = {2026},
  url     = {https://github.com/Jannchie/silva},
}