Round 2 operational artifacts
The working files from the rented GPU instance that trained danbooru-tagger-v1, kept before that instance was released. Weights are not here — see the v1 repo for the release and danbooru-tagger-round2 for the training checkpoints.
Gated because these are operational scripts carrying assumptions about one specific rented host, not something to run unread.
| path | what |
|---|---|
bin/ |
29 scripts that actually ran the job: stage2.py (3-thread shard stager), stage_guard.sh (watchdog that relaunched a wedged stager 17 times), train.sh / resume.sh, backup.py, losstrend.py, probe*.sh (throughput and VRAM probes), status.sh (read-only — launch_guard.sh is not) |
logs/train.log |
the full training log, including the CUDA OOM at the encoder unfreeze (step 6,351, 31.3 GB of 31.36) and the resume that followed |
logs/stage_out.log, logs/stage_guard.log |
staging and watchdog history |
logs/probe*.log |
the throughput probes — including the one that measured only the frozen half of the run and so under-reported peak VRAM |
irfs/ |
the instance-aware repeat-factor sampler cache, thresh 1e-3 / cap 4, mean repeat 2.32 |
CHECKSUMS.txt |
sha256 of final.pt and ck_last.pt as they existed on the instance |
final.pt on that host hashed edb2487ade203b79f1abbdc2f32d3a05a0b96f909b04f6b35dd26df3d131afa7,
identical to the local copy used for every evaluation reported in the v1 README.