ACCC1380/nyako_style

Nyako style — Anima base LoRA

Trigger: @nyako (training captions start with the exact bytes @nyako, ).

Final weights: output/nyako_style.safetensors. Final resume state: output/nyako_style-state/ (model, AdamW optimizer, scheduler, random state, and trainer progress).

Trained from Anima base v1.0 using lora-scripts-next, rank 16 / alpha 16, AdamW, learning rate 1e-4, constant scheduler, 150 total epochs (resumed after epoch 50), batch size 2, BF16, resolution 1024 with aspect ratio buckets, seed 42. Gradient checkpointing is disabled. Buckets with an odd sample count have a smaller final batch.

The source dataset is ACCC1380/Nyako at revision f1048ee38473bc253ebdc4b4eba3b0c2dcebacb3, with 24 image/caption pairs. Captions were processed as opaque bytes: only the requested prefix was prepended, with the remaining bytes verified unchanged. The assistant did not inspect caption contents.

Contents:

• output/: epoch weights, final weights, recent recovery checkpoints and final resume state.
• config/: requested/effective training configuration, dataset definition and provenance, package versions, trainer commit, summary.
• dataset/: images and prefixed captions used for training.
• logs/: training log, TensorBoard events, and scalar metrics CSV (including loss).
• scripts/: preprocessing, launcher and upload scripts.
• manifest.json: SHA-256 hashes and sizes for backup verification.

Configuration paths reflect the original training machine. To resume elsewhere, download the artifacts, install the recorded environment and trainer commit, update local model/dataset/output paths, and pass --resume /path/to/output/nyako_style-state to the saved training command. Increase the epoch target when continuing beyond the completed 150 total epochs (resumed after epoch 50).