Views
No views yet
0.8.0.dev01adapter: lora
2base_model: Qwen/Qwen2.5-7B-Instruct
3bf16: auto
4dataset_processes: 32
5datasets:
6- message_property_mappings:
7 content: content
8 role: role
9 path: tatsu-lab/alpaca
10 trust_remote_code: false
11 type: alpaca
12gradient_accumulation_steps: 4 # Increased to compensate for smaller batch size
13gradient_checkpointing: true # Enabled to save VRAM
14learning_rate: 0.0002
15lisa_layers_attribute: model.layers
16load_best_model_at_end: false
17load_in_4bit: true # Changed to 4-bit quantization for better memory efficiency
18load_in_8bit: false # Disabled 8-bit since we're using 4-bit
19lora_alpha: 16
20lora_dropout: 0.05
21lora_r: 8
22lora_target_modules:
23- q_proj
24- v_proj
25- k_proj
26- o_proj
27- gate_proj
28- down_proj
29- up_proj
30loraplus_lr_embedding: 1.0e-06
31lr_scheduler: cosine
32max_prompt_len: 512
33mean_resizing_embeddings: false
34micro_batch_size: 4 # Reduced from 16 to save memory
35num_epochs: 1.0
36optimizer: adamw_bnb_8bit
37output_dir: ./outputs/mymodel
38pretrain_multipack_attn: true
39pretrain_multipack_buffer_size: 10000
40qlora_sharded_model_loading: false
41ray_num_workers: 1
42resources_per_worker:
43 GPU: 1
44sample_packing_bin_size: 200
45sample_packing_group_size: 100000
46save_only_model: false
47save_safetensors: true
48sequence_len: 2048 # Reduced from 4096 to save memory
49shuffle_merged_datasets: true
50skip_prepare_dataset: false
51strict: false
52train_on_inputs: false
53trl:
54 log_completions: false
55 ref_model_mixup_alpha: 0.9
56 ref_model_sync_steps: 64
57 sync_ref_model: false
58 use_vllm: false
59 vllm_device: auto
60 vllm_dtype: auto
61 vllm_gpu_memory_utilization: 0.9
62use_ray: false
63val_set_size: 0.0
64weight_decay: 0.0
65