Archive 1 checkpoints (1-1)
Browse files
paper-releases/pbisc-diffusion/updates/20261003T105229Z-deck-cross-cohort-loco-k1000/diffusion/metadata/NOTE.md
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PBISC-diffusion update 20261003T105229Z-deck-cross-cohort-loco-k1000
|
| 2 |
+
|
| 3 |
+
Paper summary deck 2026-10-03 (presentations/2026-09-30--paper-summary--r00, 37 slides): new section 8 slides 25-26 from the cross-cohort leave-one-cohort-out benchmark at K = 1000 ('Pairs from other cohorts do not transfer', 'Longer training on other cohorts: no gain'; tools/sec08_cross_cohort.py, figures, data), agenda item 8 updated. || Cross-cohort (LOCO) benchmark closed at K = 1000 (user decision 2026-10-03 19:45: K = 2000 not needed; the experiment explains why PBISC fits LoRA on a few pairs of each cohort). Final score tables and aggregates, run receipts, K = 2000 zero-shot outputs that finished, and the stopped K = 2000 units registered as stopped (one partial Imperial zero-shot NPZ kept as .partial_stopped).
|
| 4 |
+
|
| 5 |
+
Payload states are recorded as supplied. Inference locations are pointers, not archived generated outputs.
|