carrydela

carrydela pushed to main at carrydela/K3

2026-08-26 14:24:07 +08:00

24c9d56b72 Keep attnres on resume, fix final chunk_index, default 0.5b to Yi-6B

5cc0555563 Fix chrF fallback so English wiki garbage fails translation_success

Compare 2 commits »

carrydela pushed to main at carrydela/K3

2026-08-26 10:08:58 +08:00

5a7d949b01 Skip the cold-start SFT best ckpt and free CUDA cache after eval

carrydela pushed to main at carrydela/K3

2026-08-26 10:00:35 +08:00

9652a9a7eb Save SFT last/best checkpoints during training and on interrupt

071dfaf42c Sanitize SwanLab env before login so 0.9 nested project does not crash

Compare 2 commits »

carrydela pushed to main at carrydela/K3

2026-08-25 22:08:59 +08:00

9a4862a866 Resume the same SwanLab run from the id stored in the ckpt

carrydela pushed to main at carrydela/K3

2026-08-25 22:04:03 +08:00

47c72e5bb8 Warn and reset chunk_index when resume changes batch or seq_len

e7185cbf49 Pull OPUS-100 en-zh for SFT instead of a checked-in jsonl

53d0f4b17a Cut 1B-run I/O: rarer SwanLab, ckpt, and generate

Compare 3 commits »

carrydela pushed to main at carrydela/K3

2026-08-25 20:23:20 +08:00

8442f92c58 Keep LatentMoE routed bmm in activation dtype under bf16 autocast

carrydela pushed to main at carrydela/K3

2026-08-25 20:14:46 +08:00

49aede9cb2 Fit 0.5b training on 32GB: SDPA MLA, block checkpoint, chunked CE

carrydela pushed to main at carrydela/K3

2026-08-25 19:50:37 +08:00

7a12f61de1 LatentMoE: K3 sigmoid routing and Switch aux/z-loss

d1da0816f2 LatentMoE: sparse permute-dispatch + padded bmm

Compare 2 commits »

carrydela pushed to main at carrydela/superpaper

2026-08-22 10:18:17 +08:00

49842c5153 chore: bump submodule pointers after P1 fixes

29a76e8184 fix: wire orphaned SP022 fixture, sync DESIGN.md with implementation

Compare 2 commits »

carrydela pushed to main at carrydela/Superfig

2026-08-22 10:14:38 +08:00

05f50b0a94 fix: align hue budget with supertensor — per-row instead if figure-wide

carrydela pushed to main at carrydela/SuperTensor

2026-08-22 10:13:55 +08:00

0ba01fe5e2 fix: align lint role-before-use check with siblings

carrydela pushed to main at carrydela/SuperDerive

2026-08-22 10:13:42 +08:00

5b24eba3c8 feat: fill test coverage gaps + fix \sdcol i18n labels

carrydela pushed to main at carrydela/superpaper

2026-08-21 14:57:34 +08:00

fc3c69bc23 fix: harden ingest.sh against partial downloads and tar size limits

carrydela pushed to main at carrydela/SuperTensor

2026-08-21 14:56:56 +08:00

7210f890ca fix: \stsignature derives node name from anchor (#2-sig)

carrydela pushed to main at carrydela/SuperDerive

2026-08-21 14:56:45 +08:00

f2ccd6fb08 fix: \sdcol emits English labels illegal/legal in [en] mode

carrydela pushed to main at carrydela/superpaper

2026-08-21 14:46:02 +08:00

e80d5c9067 chore: gitignore .omo runtime artifacts

carrydela pushed to main at carrydela/superpaper

2026-08-17 10:23:20 +08:00

d4a069b09b docs: finish family routing copy and bump child pointers

carrydela pushed to main at carrydela/SuperDerive

2026-08-17 10:23:17 +08:00

45db61f3e2 docs: send full paper-note jobs to superpaper

carrydela pushed to main at carrydela/SuperTensor

2026-08-17 10:23:14 +08:00

8ada8e57c5 docs: route architecture, rewrites, and paper notes away

carrydela pushed to main at carrydela/Superfig

2026-08-17 10:23:12 +08:00

6ed01743cf docs: route paper notes and rewrites to sibling skills