Extended LoRA adapters (r=2048) and TensorBoard logs for the Hypa-Orpheus 3B continuation run.

This repository is the extended LoRA companion to hypaai/Hypa-Orpheus-3b-TTS-VC-ext.

It continues training from the merged primary release hypaai/Hypa-Orpheus-3b-TTS-VC with a new, higher-rank adapter (r=2048, α=2048) on the same Hypa-Voices SNAC precompute pipeline.

Collection: hypaai/hypa-orpheus

Repository	Contents	Best for
`hypaai/Hypa-Orpheus-3b-TTS-VC-LoRAs`	Primary r=512 adapters	Default adapter inspection / retraining
`hypaai/Hypa-Orpheus-3b-TTS-VC-LoRAs-ext` (this repo)	Extended r=2048 adapters	High-capacity continuation experiments
`hypaai/Hypa-Orpheus-3b-TTS-VC`	Primary merged weights	Default inference
`hypaai/Hypa-Orpheus-3b-TTS-VC-ext`	Extended merged weights	Extended inference checkpoint

For capabilities, prompt format, languages, dataset details, and primary-training context, see:

Primary model README
Extended merged model README

Adapter at a Glance

Item	Primary LoRA	Extended (this repo)
Base model	`unsloth/orpheus-3b-0.1-ft`	`hypaai/Hypa-Orpheus-3b-TTS-VC`
LoRA rank / alpha	r=512, α=512	r=2048, α=2048
Trainable parameters	778M (19.07%)	~~3.11B (~~76%)
Adapter file size	~3.1 GB	~12.4 GB
Effective batch size	96	256
Shipped checkpoint	step 44,334	step 4,992
Best validation loss	3.513	3.438
License	Apache 2.0	Apache 2.0

TensorBoard

View extended training logs on Hugging Face

Metric	Value	Step
Initial training loss	3.582	1
Final training loss	3.433	4,992
Initial validation loss	3.496	832
Best validation loss (shipped)	3.438	4,992

Validation loss improved at every logged interval in this continuation. The merged extended release uses this validation-best adapter at step 4,992.

Primary-run TensorBoard (r=512, 44,334 steps):

hypaai/Hypa-Orpheus-3b-TTS-VC-LoRAs/tensorboard

Quick Usage

Load the extended adapter (Unsloth)

from unsloth import FastLanguageModel
from peft import PeftModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="hypaai/Hypa-Orpheus-3b-TTS-VC",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

model = PeftModel.from_pretrained(
    model,
    "hypaai/Hypa-Orpheus-3b-TTS-VC-LoRAs-ext",
)

Note: the extended adapter was trained with r=2048 on top of the merged primary checkpoint. For continued fine-tuning, initialize a fresh PEFT config with matching rank before loading these weights.

Merge to 16-bit for inference

model = model.merge_and_unload()
model.save_pretrained("Hypa-Orpheus-3b-TTS-VC-ext-merged")
tokenizer.save_pretrained("Hypa-Orpheus-3b-TTS-VC-ext-merged")

Or use the pre-merged checkpoint directly:

hypaai/Hypa-Orpheus-3b-TTS-VC-ext

Repository Contents

Hypa-Orpheus-3b-TTS-VC-LoRAs-ext/
├── README.md
├── adapter_config.json
├── adapter_model.safetensors    # ~12.4 GB (r=2048)
├── training_args.bin
└── runs/
    └── events.out.tfevents.*    # TensorBoard scalars