Views
No views yet
0.4.11base_model: Qwen/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: Jennny/math-conversations2
11 conversation: qwen-7b-chat
12 type: sharegpt
13 split: "train"
14 train_on_split: "train"
15
16warmup_ratio: 0.05
17val_set_size: 0.0
18output_dir: ./prm
19wandb_project: preference-models
20# wandb_entity: domain-generalization
21wandb_watch:
22wandb_name: "qwen-7b-bs32_lr2e-6_prm"
23wandb_log_model:
24
25train_on_inputs: false
26
27save_safetensors: true
28#noisy_embedding_alpha: 10.0 # default for sharegpt type
29dataset_prepared_path: ~/data/preference-models/last_run_prepared
30
31dataset_processes: 48
32#torch_compile: true
33sequence_len: 8192
34sample_packing: true
35pad_to_sequence_len: true
36
37trust_remote_code: True
38adapter:
39lora_model_dir:
40#lora_r: 32
41#lora_alpha: 16
42#lora_dropout: 0.05
43#lora_target_linear: true
44#lora_fan_in_fan_out:
45
46gradient_checkpointing: True
47
48#warmup_ratio: 0.1
49gradient_accumulation_steps: 4
50micro_batch_size: 1
51num_epochs: 1
52#max_steps: 10
53#optimizer: adamw_torch_fused
54optimizer: paged_adamw_32bit
55#lr_scheduler: constant_with_warmup
56lr_scheduler: cosine
57learning_rate: 2.0e-6
58
59weight_decay: 0.0
60max_grad_norm: 1.0
61
62group_by_length: false
63bf16: auto
64fp16: false
65tf32: true
66
67early_stopping_patience:
68local_rank:
69logging_steps: 2
70xformers_attention:
71flash_attention: true
72
73eval_steps:
74eval_table_size:
75eval_table_max_new_tokens:
76#save_steps: 100
77save_strategy: "epoch"
78save_total_limit: 4
79#save_safetensors: false
80debug:
81
82ddp: #true
83deepspeed: #deepspeed/zero1.json # multi-gpu only
84
85fsdp:
86fsdp_config:
87special_tokens:
88 pad_token: <|end_of_text|>
89