Balkondekar Rohan Ramesh commited on
Commit
5f6cc3c
1 Parent(s): d0b6f09

Upload cfg.yaml

Browse files
Files changed (1) hide show
  1. cfg.yaml +112 -0
cfg.yaml ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ answer_column: "Answer\r\r"
16
+ chatbot_author: H2O.ai
17
+ chatbot_name: h2oGPT
18
+ data_sample: 1.0
19
+ data_sample_choice:
20
+ - Train
21
+ - Validation
22
+ limit_chained_samples: false
23
+ mask_prompt_labels: true
24
+ parent_id_column: None
25
+ personalize: false
26
+ prompt_column:
27
+ - Context
28
+ text_answer_separator: <|answer|>
29
+ text_prompt_start: <|prompt|>
30
+ train_dataframe: /home/ubuntu/h2o-llmstudio/data/user/QnA_with_context/latest_finetune.csv
31
+ validation_dataframe: None
32
+ validation_size: 0.01
33
+ validation_strategy: automatic
34
+ environment:
35
+ compile_model: false
36
+ find_unused_parameters: false
37
+ gpus:
38
+ - '0'
39
+ huggingface_branch: main
40
+ mixed_precision: true
41
+ number_of_workers: 4
42
+ seed: -1
43
+ trust_remote_code: true
44
+ use_fsdp: false
45
+ experiment_name: QnA-with-context
46
+ llm_backbone: lmsys/vicuna-7b-v1.3
47
+ logging:
48
+ logger: None
49
+ neptune_project: ''
50
+ number_of_texts: 10
51
+ output_directory: /home/ubuntu/h2o-llmstudio/output/user/QnA-with-context/
52
+ prediction:
53
+ batch_size_inference: 0
54
+ do_sample: false
55
+ max_length_inference: 256
56
+ metric: BLEU
57
+ metric_gpt_model: gpt-3.5-turbo-0301
58
+ min_length_inference: 2
59
+ num_beams: 1
60
+ num_history: 4
61
+ repetition_penalty: 1.2
62
+ stop_tokens: ''
63
+ temperature: 0.3
64
+ top_k: 0
65
+ top_p: 1.0
66
+ problem_type: text_causal_language_modeling
67
+ tokenizer:
68
+ add_prefix_space: false
69
+ add_prompt_answer_tokens: false
70
+ max_length: 512
71
+ max_length_answer: 256
72
+ max_length_prompt: 256
73
+ padding_quantile: 1.0
74
+ use_fast: true
75
+ training:
76
+ adaptive_kl_control: true
77
+ advantages_gamma: 0.99
78
+ advantages_lambda: 0.95
79
+ batch_size: 2
80
+ differential_learning_rate: 1.0e-05
81
+ differential_learning_rate_layers: []
82
+ drop_last_batch: true
83
+ epochs: 4
84
+ evaluate_before_training: false
85
+ evaluation_epochs: 1.0
86
+ grad_accumulation: 1
87
+ gradient_clip: 0.0
88
+ initial_kl_coefficient: 0.2
89
+ kl_horizon: 10000
90
+ kl_target: 6.0
91
+ learning_rate: 0.0001
92
+ lora: true
93
+ lora_alpha: 16
94
+ lora_dropout: 0.05
95
+ lora_r: 4
96
+ lora_target_modules: ''
97
+ loss_function: TokenAveragedCrossEntropy
98
+ offload_reward_model: false
99
+ optimizer: AdamW
100
+ ppo_batch_size: 1
101
+ ppo_clip_policy: 0.2
102
+ ppo_clip_value: 0.2
103
+ ppo_epochs: 4
104
+ ppo_generate_temperature: 1.0
105
+ reward_model: OpenAssistant/reward-model-deberta-v3-large-v2
106
+ save_best_checkpoint: false
107
+ scaling_factor_value_loss: 0.1
108
+ schedule: Cosine
109
+ train_validation_data: false
110
+ use_rlhf: false
111
+ warmup_epochs: 0.0
112
+ weight_decay: 0.0