ChristophSchuhmann's picture
add model card
697976c verified
|
Raw
History Blame Contribute Delete
826 Bytes
metadata
license: apache-2.0
tags:
  - audio
  - text-to-speech
  - lora
  - work-in-progress

LAION Box — Work in Progress 0.7

Work-in-progress LoRA fine-tune (rank 256) of the DramaBox / LTX audio DiT on a diverse mix: German (70%, 3-mode standalone/voice-clone/concat) + diverse-mix (30%).

This repo is updated automatically by a training daemon.

Layout:

  • latest/ — most recent validated checkpoint (overwritten each save).
    • lora.safetensors, training_state.pt (optimizer/scheduler for resume), metadata.json
  • best/step<N>/ — the two best checkpoints so far by mean flow_loss (last-20 window). Pruned to two; stale ones are removed when the best set changes.

Checkpoints are validated before upload (safetensors header + finiteness spot-check, optimizer-state size check, flow-loss sanity).