The parameters used to train this model, are encoded as follows:

G-B1,DR0.2,ACC1,L2.5e-05,R32,A128,E20,TDFT,BF16T,GCLT,MG0.8,GCHF,D1.4,CE0.04,W0.03,cosine,R1 - THE KING



Here's the actual python command to train it, we used this in Google Colab Notebook. 

<CODE>
# Begin the fine-tuning proces

%cd /content/drive/MyDrive/VibeVoice-finetuning/

# Define your parameters as Python variables
batch_size = 1
drop_rate = 0.2
grad_accum = 1
lr = 2.5e-5
lora_r = 32
lora_alpha = 128
epochs = 20
train_diff = True
bf16 = True
grad_clip = True
max_grad = 0.8
grad_checkpoint = False
diff_weight = 1.4
ce_weight = 0.04
warmup = 0.03
scheduler = "cosine"
run_num = 2

# Build the output directory dynamically
output_dir =  (

    f"Precise/G-B{batch_size},DR{drop_rate},ACC{grad_accum},"
    f"L{lr},R{lora_r},A{lora_alpha},E{epochs},"
    f"TDF{'T' if train_diff else 'F'},BF16{'T' if bf16 else 'F'},"
    f"GCL{'T' if grad_clip else 'F'},MG{max_grad},"
    f"GCH{'T' if grad_checkpoint else 'F'},D{diff_weight},"
    f"CE{ce_weight},W{warmup},{scheduler},R{run_num}"

              )

# Now use the variables in your command
!python -m src.finetune_vibevoice_lora \
    --model_name_or_path vibevoice/VibeVoice-7B \
    --processor_name_or_path src/vibevoice/processor \
    --text_column_name text \
    --audio_column_name audio \
    --output_dir {output_dir} \
    \
    --train_jsonl GOLD_cortana_train_data.jsonl \
    --per_device_train_batch_size {batch_size} \
    --voice_prompt_drop_rate {drop_rate} \
    --gradient_accumulation_steps {grad_accum} \
    --learning_rate {lr} \
    --lora_r {lora_r} \
    --lora_alpha {lora_alpha} \
    --num_train_epochs {epochs} \
    --train_diffusion_head {train_diff} \
    --bf16 {bf16} \
    --gradient_clipping \
    --max_grad_norm {max_grad} \
    --gradient_checkpointing {grad_checkpoint} \
    --diffusion_loss_weight {diff_weight} \
    --ce_loss_weight {ce_weight} \
    --warmup_ratio {warmup} \
    --lr_scheduler_type {scheduler} \
    \
    --logging_steps 10 \
    --save_steps 1528 \
    \
    --report_to wandb \
    --remove_unused_columns False \
    --do_train \
    --ddpm_batch_mul 4 \
    --lora_target_modules q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj

</CODE>