finetune_elise_single_speaker.sh

#!/bin/bash

# Single-speaker fine-tuning script for VibeVoice-1.5B on CLEANED Elise dataset
# No voice prompts - pure text-to-speech for Elise voice only
# WITH PROPER EOS TOKEN to fix repetition/looping issue

echo "Single-speaker fine-tuning on cleaned dataset..."
echo "Using 544 clean samples (no cutoffs)"
echo "NO voice prompts - training pure Elise TTS model"

python -m src.finetune_vibevoice_lora \
    --model_name_or_path . \
    --train_jsonl elise_cleaned/train_split.jsonl \
    --validation_jsonl elise_cleaned/val.jsonl \
    --text_column_name text \
    --audio_column_name audio \
    --output_dir finetune_elise_single_speaker \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --learning_rate 2.5e-5 \
    --num_train_epochs 4 \
    --logging_steps 10 \
    --save_steps 100 \
    --eval_steps 100 \
    --report_to none \
    --remove_unused_columns False \
    --bf16 True \
    --do_train \
    --do_eval \
    --gradient_clipping \
    --gradient_checkpointing False \
    --ddpm_batch_mul 2 \
    --diffusion_loss_weight 1.4 \
    --train_diffusion_head True \
    --ce_loss_weight 0.04 \
    --voice_prompt_drop_rate 1.0 \
    --lora_target_modules q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj \
    --lr_scheduler_type cosine \
    --warmup_ratio 0.03 \
    --max_grad_norm 0.8 \
    --max_length 4096

echo "Single-speaker fine-tuning complete!"
echo "Model will now generate Elise voice from text only - no voice prompts needed!"