# H3 Character Swap V1.5 training record

V1.5 is the checkpoint called V2-1000 during development. Training initialized
from the released V1 adapter and completed 1,000 additional optimizer updates.
The exported adapter is used in place of V1 at inference; it does not require
loading both character adapters.

The recipe used rank/alpha 16/16, excluded `adaln_proj`, and used AdamW8bit at
1e-5 in BF16. It used the Ostris Ref2VA training adapter v3 with guidance loss
enabled and a target of 3.5. It was not the proposed fresh run with guidance
loss disabled. See [run-summary.json](run-summary.json) for exact identities.

The new training mix includes moving-video targets and facial crops alongside
V1 replay and preservation video/audio. The separately published V1 dataset is
only part of this training lineage; it is not the complete V1.5 dataset.

The original configuration budget was 2,250 updates. The run was stopped at
1,000 completed updates, with the checkpoint and recovery state verified.
The original checkpoint metadata retains that planned budget and the internal
V2 name; `training_info.step` records 1000. The published filename changes the
release name without rewriting tensor data or checkpoint metadata.

This summary is not a complete runnable reproduction package. The full sanitized
configuration and exact runtime dependency record should accompany a later
reproducibility package if that package is published. Existing files directly
under `training/` describe the V1 run.