Ketak-ZoomRx commited on
Commit
a9ba1e0
1 Parent(s): 878b81b

Upload cfg.yaml

Browse files
Files changed (1) hide show
  1. cfg.yaml +117 -0
cfg.yaml ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ architecture:
2
+ backbone_dtype: float16
3
+ force_embedding_gradients: false
4
+ gradient_checkpointing: true
5
+ intermediate_dropout: 0.0
6
+ pretrained: true
7
+ pretrained_weights: ''
8
+ augmentation:
9
+ random_parent_probability: 0.0
10
+ skip_parent_probability: 0.0
11
+ token_mask_probability: 0.0
12
+ dataset:
13
+ add_eos_token_to_answer: true
14
+ add_eos_token_to_prompt: true
15
+ add_eos_token_to_system: true
16
+ answer_column: output
17
+ chatbot_author: H2O.ai
18
+ chatbot_name: h2oGPT
19
+ data_sample: 1.0
20
+ data_sample_choice:
21
+ - Train
22
+ - Validation
23
+ limit_chained_samples: false
24
+ mask_prompt_labels: true
25
+ parent_id_column: None
26
+ personalize: false
27
+ prompt_column:
28
+ - instruction
29
+ system_column: system
30
+ text_answer_separator: <|answer|>
31
+ text_prompt_start: <|prompt|>
32
+ text_system_start: <|system|>
33
+ train_dataframe: /app/h2o-llmstudio/data/user/Trial-Input - Prim_Llama/Trial-Input
34
+ - Prim_Llama.csv
35
+ validation_dataframe: None
36
+ validation_size: 0.01
37
+ validation_strategy: automatic
38
+ environment:
39
+ compile_model: false
40
+ find_unused_parameters: false
41
+ gpus:
42
+ - '0'
43
+ - '1'
44
+ huggingface_branch: main
45
+ mixed_precision: true
46
+ number_of_workers: 8
47
+ seed: -1
48
+ trust_remote_code: true
49
+ use_fsdp: false
50
+ experiment_name: Prim_Drug_Llama
51
+ llm_backbone: meta-llama/Llama-2-7b-chat-hf
52
+ logging:
53
+ logger: None
54
+ neptune_project: ''
55
+ number_of_texts: 10
56
+ output_directory: /app/h2o-llmstudio/output/user/Prim_Drug_Llama/
57
+ prediction:
58
+ batch_size_inference: 0
59
+ do_sample: false
60
+ max_length_inference: 256
61
+ metric: BLEU
62
+ metric_gpt_model: gpt-3.5-turbo-0301
63
+ min_length_inference: 2
64
+ num_beams: 1
65
+ num_history: 4
66
+ repetition_penalty: 1.2
67
+ stop_tokens: ''
68
+ temperature: 0.0
69
+ top_k: 0
70
+ top_p: 1.0
71
+ problem_type: text_causal_language_modeling
72
+ tokenizer:
73
+ add_prefix_space: false
74
+ add_prompt_answer_tokens: false
75
+ max_length: 2976
76
+ max_length_answer: 256
77
+ max_length_prompt: 2912
78
+ padding_quantile: 1.0
79
+ use_fast: true
80
+ training:
81
+ adaptive_kl_control: true
82
+ advantages_gamma: 0.99
83
+ advantages_lambda: 0.95
84
+ batch_size: 2
85
+ differential_learning_rate: 1.0e-05
86
+ differential_learning_rate_layers: []
87
+ drop_last_batch: true
88
+ epochs: 10
89
+ evaluate_before_training: false
90
+ evaluation_epochs: 1.0
91
+ grad_accumulation: 1
92
+ gradient_clip: 0.0
93
+ initial_kl_coefficient: 0.2
94
+ kl_horizon: 10000
95
+ kl_target: 6.0
96
+ learning_rate: 0.0015
97
+ lora: true
98
+ lora_alpha: 16
99
+ lora_dropout: 0.05
100
+ lora_r: 32
101
+ lora_target_modules: ''
102
+ loss_function: TokenAveragedCrossEntropy
103
+ offload_reward_model: false
104
+ optimizer: AdamW
105
+ ppo_batch_size: 1
106
+ ppo_clip_policy: 0.2
107
+ ppo_clip_value: 0.2
108
+ ppo_epochs: 4
109
+ ppo_generate_temperature: 1.0
110
+ reward_model: OpenAssistant/reward-model-deberta-v3-large-v2
111
+ save_best_checkpoint: false
112
+ scaling_factor_value_loss: 0.1
113
+ schedule: Cosine
114
+ train_validation_data: false
115
+ use_rlhf: false
116
+ warmup_epochs: 0.0
117
+ weight_decay: 0.0