I-CE with each review view drawn until it holds at least 8, 32 or 64 sentences (the recipe stops at 16), folds 0 to 2 on the 512-sentence packs, each at the epoch the validation rank rule selects. rank_ckpt_plan.json lists them (166 runs). The appendix's new Table A6 moved every later table one place, so the card's table numbers follow: the old Tables A6 to A18 are A7 to A19.
3 days ago
Card: Figures 8, 9 and A3 are drawn from these checkpoints by figures/neighborhoods of the code repository
fixed_split_ckpt_plan.json is the file of contrast_experiment/w9/eval/ of the code repository (commit f9c4b8b): the 4:1 gallery-aimed run is named at epoch 1,250, the checkpoint the paper reads, with epoch 900 under `previous`; three runs gain the `extended` note.
The 4:1 gallery-aimed run of Section 6.2 at epoch 1,250, the checkpoint the paper reads
The run was trained on from 1,000 to 2,000 epochs on 2026-11-08 and the rank rule over all 40 checkpoints moved from epoch 900 to 1,250. The epoch-900 file stays; the selection file follows.
Card and selection files for the checkpoints added today; the citation names the first author
rank_ckpt_plan.json and fixed_split_ckpt_plan.json are the files of contrast_experiment/w9/eval/ of the code repository (commit 31ce099). They add the DINO-style control and eight fixed-split runs, and list the 52 further checkpoints of two runs under `trajectory`; no earlier entry changed.
The other checkpoints the paper reads: fixed-split towers and the DINO-style control
Fixed split: the towers of Figure 9, Figure A3 and Appendix L (CE and I-CE trained at 2,048 sentences, SimCLR-style, I + SimCLR, VICReg), the a+u @ views row of Table 3, the three VICReg rows of Table A7, every checkpoint of the I-CE run that Figure A1 draws, and the SimCLR-style checkpoints from epoch 1,350 on, whose test curve Appendix L describes. Five folds: the DINO-style control of Table A1 (fold 0). The selection files and the card follow.
Card: what is not in the repository, the run table by provenance cell, names in the before/after table
rank_ckpt_plan.json, i2bce_sel_pool.json, fixed_split_ckpt_plan.json and superseded_ckpt_plan.json are the files of contrast_experiment/w9/eval/ of the code repository (commit ca3abb9) and name every checkpoint here. They carry every entry of the four files removed; no selection changed.
Stage-one BYOL checkpoints of the two-stage runs (epochs 450, 700, 450, 450, 300)
The checkpoints the five two-stage runs (swin168step84loop2i2ce_icetf ... _gpool_bw) were initialized from, named in their training logs. Uploaded under the _gpool names of the current code.
Name the whole-pool five-fold checkpoints _gpool (formerly _g4096)
The code release's k10_eval.py builds SetPoolN(Q) from the checkpoint since commit b8f84c3, so the note on scoring slot<Q> files is replaced, and the two new plan files are named as extra selection files for k10_eval.py cells.
Add the one- and two-query towers (campaign 3) and the five-fold eight-query towers
Six campaign-3 checkpoints (slot1i2cemean_icetf and slot2i2cemean_icetf, folds 0 to 2, _d512) listed in c3_docfix512_ckpt_plan.json; the five five-fold slot8i2cemean_icetf checkpoints at _g4096, at the rank rule's picks, listed in q8_g4096_ckpt_plan.json; README sections for both and the SetPoolN(Q) note.
7 days ago
Fixed-split CE + anchor-in-I (_d512) trained on to 2,000 epochs: selected checkpoint ep1650, plan entry and README note
The campaign-2 schedule stopped this run at 1,000 epochs; it was continued from the epoch-1,000 weights with a fresh optimizer, the way the original fixed-split I-CE and CE reached 2,000. The epoch-1,000 file stays; the plan entry keeps it under 'previous'.
Add the remaining corrected @512 towers (campaign 2, 67 runs, _d512) and their selection plan; README section
