-
Remove superseded s3-path pairs/ and hard-negative mining jobs
-
Add dataset card YAML: point the viewer at pairs_img/ (Image columns)
-
Remove mining/ (hard-negative artifacts, not used by training)
-
Remove images/ (hard-negative artifacts, not used by training)
-
Drop hard negatives: pairs_img/ is now clean query+positive with inline images
-
Add HF-native images: pairs_img/ (q/pos inline) + images/ (key->bytes for negs)
-
Train-ready: add s3:/ image URIs (q_img/pos_img/neg_imgs) for Qwenvl2d5 streaming loader
-
Corrected pilot per ret-training-data-format: raw-byte key, role mask, GME target teacher, Qwen tokens, relative-tau, CIR dup-guard
-
Hard-neg mining pilot: reasoning_retrieval jobs0-6 manifest (CLIP D=512 pilot teacher)
-
initial commit