Views
No views yet
0.4.11base_model: /shared/ssd/models/Qwen2.5-7B-Instruct
2model_type: AutoModelForCausalLM
3tokenizer_type: AutoTokenizer
4
5load_in_8bit: false
6load_in_4bit: false
7strict: false
8
9datasets:
10 - path: nuojohnchen/judgelm-train-100k-pair-pm
11 type: sharegpt
12 conversation: chatml
13 split: "train"
14 train_on_split: "train"
15
16warmup_steps: 40
17val_set_size: 0.0
18output_dir: ./pm_models/qwen7b-it_bs128_lr1e-5
19#wandb_project: preference-models
20#wandb_entity: domain-generalization
21wandb_watch:
22wandb_name: "qwen-7b-it_bs128_lr1e-5"
23#_response_only
24wandb_log_model:
25
26train_on_inputs: false
27
28save_safetensors: true
29#noisy_embedding_alpha: 10.0 # default for sharegpt type
30dataset_prepared_path: ~/data/preference-models/last_run_prepared
31
32
33dataset_processes: 48
34#torch_compile: true
35sequence_len: 3072
36sample_packing: true
37pad_to_sequence_len: true
38
39trust_remote_code: True
40adapter:
41lora_model_dir:
42
43
44
45
46
47gradient_checkpointing: True
48
49#warmup_ratio: 0.1
50gradient_accumulation_steps: 16
51micro_batch_size: 1
52num_epochs: 1
53#max_steps: 10
54optimizer: adamw_torch_fused
55#optimizer: paged_adamw_32bit
56#lr_scheduler: constant_with_warmup
57lr_scheduler: cosine
58learning_rate: 5.0e-6
59
60weight_decay: 0.0
61max_grad_norm: 1.0
62
63
64group_by_length: false
65bf16: auto
66fp16: false
67tf32: true
68
69early_stopping_patience:
70local_rank:
71logging_steps: 2
72xformers_attention:
73flash_attention: true
74
75
76eval_steps:
77eval_table_size:
78eval_table_max_new_tokens:
79#save_steps: 100
80save_strategy: "epoch"
81save_total_limit: 1
82#save_safetensors: false
83debug:
84
85
86ddp: #true
87deepspeed: #deepspeed/zero1.json # multi-gpu only
88
89fsdp:
90fsdp_config:
91special_tokens:
92
93