Best-performing LoRA adapters for SPH voice cloning on Orpheus 3B. Trained with label masking (loss computed only on audio tokens) and the correct 24kHz SNAC pipeline.

Key Metrics

Checkpoint	Speaker Similarity	Training Loss (audio-only)
checkpoint-275 ← best	0.547 avg / 0.621 peak	4.11
checkpoint-250	0.505 avg / 0.583 peak	3.93
checkpoint-125	0.519 avg / 0.591 peak	4.03
final (load_best)	0.475 avg / 0.500 peak	4.21

What's New in V3

Improvement	V1 (previous)	V3 (this)
Label masking	❌ Loss on all tokens	✅ Loss on audio tokens only
Learning rate	2e-4 (linear)	1e-4 (cosine)
Epochs	5	3
Eval split	❌	✅ 5% validation
Best similarity	0.614	0.621

Training Configuration

Property	Value
Base Model	`unsloth/orpheus-3b-0.1-ft`
LoRA Rank (r)	64
LoRA Alpha	64
Target Modules	q, k, v, o, gate, up, down proj
SNAC Codec	`snac_24khz` ✅ (7-token pattern)
Learning Rate	1e-4 (cosine decay, 5% warmup)
Batch Size	2 (grad accum 4, effective 8)
Max Seq Length	1536
Precision	BFloat16
GPU	RTX 4090 (24GB)
Training Time	~18 minutes

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from snac import SNAC

# Load base + LoRA
base = AutoModelForCausalLM.from_pretrained(
    "unsloth/orpheus-3b-0.1-ft",
    torch_dtype=torch.bfloat16, device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("unsloth/orpheus-3b-0.1-ft")

# Use checkpoint-275 (best similarity)
model = PeftModel.from_pretrained(base, "kailasa-ngpt/orpheus-3b-lora-sph-v3",
                                   subfolder="checkpoint-275")
model.eval()

# SNAC decoder (must be 24kHz)
snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")

V1 (previous best): kailasa-ngpt/2026_01_15_Orpheus_Run
FFT (deprecated): kailasa-ngpt/orpheus-3b-ft-step-1000
Dataset: kailasa-ngpt/SPH_Audio_2019_60_Secs_947_Samples

kailasa-ngpt/orpheus-3b-lora-sph-v3

Orpheus 3B — LoRA V3 (Label-Masked Training) ✅

Key Metrics

What's New in V3

Training Configuration

Usage

Related

kailasa-ngpt/
orpheus-3b-lora-sph-v3