danbooru-tagger-round2 — training artifacts, not a release

This repo holds mid-training backups from the Round 2 run: an optimizer-state checkpoint, an fp16 weight dump, and the training log. It exists so a preempted GPU instance could be recovered, not for people to download and use.

The release is Redstonexs/danbooru-tagger-v1 — ONNX, reference inference code, the tag list with per-tag training counts, and the full evaluation against every incumbent tagger including where it loses.

file here	what
checkpoints/ck_last.pt	model + AdamW state + seen/unfrozen, for resuming
checkpoints/weights.fp16.pt	fp16 weights (superseded by the v1 repo's copy)
logs/train.log	the full training log, including the validation curve

Kept public because the log is the honest record of the run — including the CUDA OOM at the unfreeze step and the resume that followed.