comethrusws commited on
Commit
5f90e2a
·
verified ·
1 Parent(s): a950fe3

Upload config.yaml with huggingface_hub

Browse files
Files changed (1) hide show
  1. config.yaml +56 -0
config.yaml ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ language: ne
3
+ sample_rate: 16000
4
+ hf_dataset_id: Firoj112/voxcpm-nepali-data
5
+ train_manifest: finetuning_data/manifests/train_manifest.jsonl
6
+ val_manifest: finetuning_data/manifests/val_manifest.jsonl
7
+ hardware:
8
+ gpu_count: 1
9
+ load_in_4bit: false
10
+ gradient_checkpointing: true
11
+ vram_gb_per_gpu: 24.0
12
+ lora:
13
+ alpha: 96
14
+ bias: none
15
+ dropout: 0.05
16
+ enabled: true
17
+ rank: 96
18
+ target_modules:
19
+ - q_proj
20
+ - v_proj
21
+ - k_proj
22
+ - o_proj
23
+ metadata:
24
+ config_type: lora
25
+ language: ne
26
+ model_version: '2'
27
+ model:
28
+ name_or_path: openbmb/VoxCPM2
29
+ optimize: false
30
+ version: '2'
31
+ training:
32
+ amp: bf16
33
+ bf16: true
34
+ dataloader_num_workers: 0
35
+ ddp_find_unused_parameters: false
36
+ epochs: 20
37
+ eval_steps: 500
38
+ fp16: false
39
+ gradient_accumulation_steps: 32
40
+ gradient_checkpointing: true
41
+ learning_rate: 0.00005
42
+ logging_steps: 50
43
+ max_grad_norm: 1.0
44
+ max_steps: 20000
45
+ output_dir: ./checkpoints/lora_nepali_v2
46
+ per_device_train_batch_size: 1
47
+ report_to:
48
+ - wandb
49
+ - tensorboard
50
+ save_steps: 500
51
+ warmup_steps: 500
52
+ lr_scheduler_type: cosine
53
+ wandb:
54
+ log_model: checkpoint
55
+ project: Vox-CPM-Nepali-Fine-Tune
56
+ watch: gradients