saintsauce/unified-vlm-steering-minimal-pairs
Unified VLM Steering: Minimal Pairs Minimal pairs used to extract steering vectors (difference of means between the two poles) for 7 concepts in the unified-vlm-steering project: 100 text pairs and 100 image pairs per concept. Layout txt/<concept>/pairs.json 100 text pairs: concept, pos_label, neg_label, template, n_pairs, pairs [{subject, pos, neg}], and the pos / neg sentence lists img/<concept>/<000-099>/ baseline.png… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-minimal-pairs.
Unified VLM Steering: Minimal Pairs
Minimal pairs used to extract steering vectors (difference of means between the two poles) for 7 concepts in the unified-vlm-steering project: 100 text pairs and 100 image pairs per concept.
Layout
txt/<concept>/pairs.json 100 text pairs: concept, pos_label, neg_label, template, n_pairs,
pairs [{subject, pos, neg}], and the pos / neg sentence lists
img/<concept>/<000-099>/
baseline.png image generated from `prompt` (1024x1024)
pos.png baseline edited toward the positive pole (`edit_pos_prompt`)
neg.png baseline edited toward the negative pole (`edit_neg_prompt`)
meta.json prompts, edit prompts, image model, pole labels, subject kind, pair scores, provenanceTotals: 7 text files (700 pairs) and 700 image pairs (2,100 PNG images). Image edits were made with gemini-2.5-flash-image (recorded per pair in meta.json).
Known caveat
In img/age, neg.png is byte-identical to baseline.png in 95 of 100 pairs; only 003, 004, 042, 045, 083 have a separate "younger" edit. Image editing barely moves an adult toward younger, so the age image pairs are effectively one-sided (older vs baseline).
