Shishir1807 commited on
Commit
d05d5ca
1 Parent(s): 2f81a98

Upload cfg.yaml

Browse files
Files changed (1) hide show
  1. cfg.yaml +116 -0
cfg.yaml ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: prompt
30
+ text_answer_separator: <|answer|>
31
+ text_prompt_start: <|prompt|>
32
+ text_system_start: <|system|>
33
+ train_dataframe: /app/h2o-llmstudio/data/user/Trial_M5/M5.csv
34
+ validation_dataframe: None
35
+ validation_size: 0.01
36
+ validation_strategy: automatic
37
+ environment:
38
+ compile_model: false
39
+ find_unused_parameters: false
40
+ gpus:
41
+ - '0'
42
+ - '1'
43
+ huggingface_branch: main
44
+ mixed_precision: true
45
+ number_of_workers: 8
46
+ seed: -1
47
+ trust_remote_code: true
48
+ use_fsdp: false
49
+ experiment_name: M5_llama
50
+ llm_backbone: meta-llama/Llama-2-7b-hf
51
+ logging:
52
+ logger: None
53
+ neptune_project: ''
54
+ number_of_texts: 10
55
+ output_directory: /app/h2o-llmstudio/output/user/M5_llama/
56
+ prediction:
57
+ batch_size_inference: 0
58
+ do_sample: false
59
+ max_length_inference: 256
60
+ metric: BLEU
61
+ metric_gpt_model: gpt-3.5-turbo-0301
62
+ min_length_inference: 2
63
+ num_beams: 1
64
+ num_history: 4
65
+ repetition_penalty: 1.2
66
+ stop_tokens: ''
67
+ temperature: 0.0
68
+ top_k: 0
69
+ top_p: 1.0
70
+ problem_type: text_causal_language_modeling
71
+ tokenizer:
72
+ add_prefix_space: false
73
+ add_prompt_answer_tokens: false
74
+ max_length: 800
75
+ max_length_answer: 32
76
+ max_length_prompt: 768
77
+ padding_quantile: 1.0
78
+ use_fast: true
79
+ training:
80
+ adaptive_kl_control: true
81
+ advantages_gamma: 0.99
82
+ advantages_lambda: 0.95
83
+ batch_size: 2
84
+ differential_learning_rate: 1.0e-05
85
+ differential_learning_rate_layers: []
86
+ drop_last_batch: true
87
+ epochs: 6
88
+ evaluate_before_training: false
89
+ evaluation_epochs: 1.0
90
+ grad_accumulation: 1
91
+ gradient_clip: 0.1
92
+ initial_kl_coefficient: 0.2
93
+ kl_horizon: 10000
94
+ kl_target: 6.0
95
+ learning_rate: 0.0015
96
+ lora: true
97
+ lora_alpha: 16
98
+ lora_dropout: 0.05
99
+ lora_r: 32
100
+ lora_target_modules: ''
101
+ loss_function: TokenAveragedCrossEntropy
102
+ offload_reward_model: false
103
+ optimizer: AdamW
104
+ ppo_batch_size: 1
105
+ ppo_clip_policy: 0.2
106
+ ppo_clip_value: 0.2
107
+ ppo_epochs: 4
108
+ ppo_generate_temperature: 1.0
109
+ reward_model: OpenAssistant/reward-model-deberta-v3-large-v2
110
+ save_best_checkpoint: false
111
+ scaling_factor_value_loss: 0.1
112
+ schedule: Cosine
113
+ train_validation_data: false
114
+ use_rlhf: false
115
+ warmup_epochs: 0.0
116
+ weight_decay: 0.0