Lora/parameter-rationale.md

Fresh 120-epoch run

The user explicitly replaced the earlier 600-step run with fresh training for 120 full epochs and requested batch size 4. This run does not load earlier adapter weights or optimizer state. Existing dual-WD captions and frozen-model caches are reused unchanged.

The trainer supports physical batch 1. Four gradient-accumulation microbatches implement effective batch 4: each epoch has 76 sample presentations and 19 optimizer updates, totaling 9,120 presentations and 2,280 updates in 120 epochs. Every 10th epoch is preserved, including final epoch 120.

Original nonquantized MOSS-Image-Base bf16, native text-encoder dtypes and fp32 VAE are retained. LoRA weights, saved adapters and optimizer states use fp32. Rank/alpha 32 on attention and feed-forward projections, AdamW 5e-5 with cosine decay, 95-update warmup (five epochs), dropout .05, clipping 1, seed 42. Gradient checkpointing is disabled because the 96GB GPU accommodates retained activations; the mathematical training objective remains unchanged.

Sources:

• https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md : keep text encoder frozen.
• https://github.com/PineCookie/LoRA-Forge/blob/main/docs/moss-image-trainer.md : model numerics, flow velocity target and report-aligned logit-normal sampling (-1.09,.95).
• https://github.com/PineCookie/LoRA-Forge/blob/main/docs/moss-image.md : rank 32, AdamW, all-layer targets, and limited overtraining observations. The explicit 120-epoch target is the user's choice, not a claim of an empirically established optimum.
• https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3 : Macro-F1 .4772, balanced threshold .5296.
• https://huggingface.co/SmilingWolf/wd-vit-large-tagger-v3 : Macro-F1 .4674; .35 general threshold follows the WD application's default.
• https://huggingface.co/spaces/SmilingWolf/wd-tagger/blob/main/app.py : character threshold .85.

Two original tagger passes (one per model) were merged by normalized case-insensitive tag with duplicate removal; no second tagging round was performed. Rating labels were excluded. Trigger: shuiwu22. No image, caption or generated preview was viewed by the assistant. Numerical loss and checkpoint checks do not establish visual fidelity or absence of overfitting.