Commit History

  1. Remove superseded s3-path pairs/ and hard-negative mining jobs

    Verified signature
    1dafcdb
  2. Add dataset card YAML: point the viewer at pairs_img/ (Image columns)

    Verified signature
    aea9140
  3. Remove mining/ (hard-negative artifacts, not used by training)

    Verified signature
    4cb7f2f
  4. Remove images/ (hard-negative artifacts, not used by training)

    Verified signature
    7058fbf
  5. Drop hard negatives: pairs_img/ is now clean query+positive with inline images

    Verified signature
    5a4bfa8
  6. Add HF-native images: pairs_img/ (q/pos inline) + images/ (key->bytes for negs)

    Verified signature
    2ddea6d

  1. Train-ready: add s3:/ image URIs (q_img/pos_img/neg_imgs) for Qwenvl2d5 streaming loader

    Verified signature
    8709b8a
  2. Corrected pilot per ret-training-data-format: raw-byte key, role mask, GME target teacher, Qwen tokens, relative-tau, CIR dup-guard

    Verified signature
    aa47a16
  3. Hard-neg mining pilot: reasoning_retrieval jobs0-6 manifest (CLIP D=512 pilot teacher)

    Verified signature
    c74404f
  4. initial commit

    Verified signature
    b177a83