idacy/mo-degrad-relearning-checkpoints

mo-degrad relearning checkpoints

relearning-triad and control checkpoints from the mo_degrad project (see the project repo for code, results and verdicts). adapters are rank-64 lora unless a directory says otherwise; directory names carry the organism / blue-method / seed / learning-rate. run records and metrics live in the project repo under results/; these are the raw checkpoint trees for reuse and re-analysis.

note: partial archive. uploaded: the full blue matrix, lr sweeps, lr1e3 check, q2gentler trajectories (incl. lovecats), the ihy remat trajectories (the suppression checkpoints), and part of the c2 spanish trajectories. not uploaded: the rest of c2 and the old gentler lr-sweep dirs. everything is reproducible from the project repo (seeded runs); all conclusions live in its results/.