laion/voicenet-1.7B-wip

MOSS VoiceGenerator — GRPO LoRA (Rank 8)

Overview

• —Base model: OpenMOSS-Team/MOSS-VoiceGenerator (Qwen3-1.7B backbone)
• —Method: GRPO (Group Relative Policy Optimization)
• —LoRA rank: 8, alpha: 16
• —Training: 500 steps, batch=8, G=4, lr=5e-5
• —Rewards: Speaker similarity (ECAPA-TDNN, w=0.6) + CLAP emotion (w=0.4) + WER penalty

Results (500 steps)

Eval (1 sample per condition)

Usage

from transformers import AutoModel
from peft import PeftModel

model = AutoModel.from_pretrained("OpenMOSS-Team/MOSS-VoiceGenerator", trust_remote_code=True)
model.language_model = PeftModel.from_pretrained(model.language_model, "laion/voicenet-1.7B-wip", subfolder="grpo-r8-500steps")