psinger commited on
Commit
9ede93a
1 Parent(s): 3a52ff1

Upload cfg.yaml

Browse files
Files changed (1) hide show
  1. cfg.yaml +90 -0
cfg.yaml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.1
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: output
16
+ data_sample: 1.0
17
+ data_sample_choice:
18
+ - Train
19
+ - Validation
20
+ mask_prompt_labels: false
21
+ parent_id_column: parent_id
22
+ prompt_column:
23
+ - instruction
24
+ text_answer_separator: <|answer|>
25
+ text_prompt_start: <|prompt|>
26
+ train_dataframe: data/user/oasst/train_full_allrank.pq
27
+ validation_dataframe: data/user/oasst/gpt4_val_v0.csv
28
+ validation_size: 0.01
29
+ validation_strategy: custom
30
+ environment:
31
+ compile_model: false
32
+ find_unused_parameters: false
33
+ gpus:
34
+ - '0'
35
+ - '1'
36
+ - '2'
37
+ - '3'
38
+ mixed_precision: true
39
+ number_of_workers: 8
40
+ seed: -1
41
+ trust_remote_code: false
42
+ use_fsdp: false
43
+ experiment_name: h2ogpt-gm-oasst1-en-1024-20b
44
+ llm_backbone: EleutherAI/gpt-neox-20b
45
+ logging:
46
+ logger: Neptune
47
+ neptune_project: Zoo/h2o-llm
48
+ number_of_texts: 10
49
+ output_directory: output/user/h2ogpt-gm-oasst1-en-1024-20b/
50
+ prediction:
51
+ batch_size_inference: 1
52
+ do_sample: false
53
+ max_length_inference: 256
54
+ metric: GPT3.5
55
+ min_length_inference: 2
56
+ num_beams: 2
57
+ repetition_penalty: 1.2
58
+ stop_tokens: ''
59
+ temperature: 0.3
60
+ problem_type: text_causal_language_modeling
61
+ tokenizer:
62
+ add_prefix_space: false
63
+ add_prompt_answer_tokens: false
64
+ max_length: 1024
65
+ max_length_answer: 512
66
+ max_length_prompt: 512
67
+ padding_quantile: 1.0
68
+ training:
69
+ batch_size: 8
70
+ differential_learning_rate: 1.0e-05
71
+ differential_learning_rate_layers: []
72
+ drop_last_batch: true
73
+ epochs: 3
74
+ evaluate_before_training: true
75
+ evaluation_epochs: 0.5
76
+ grad_accumulation: 1
77
+ gradient_clip: 0.0
78
+ learning_rate: 0.0005
79
+ lora: true
80
+ lora_alpha: 33
81
+ lora_dropout: 0.05
82
+ lora_r: 16
83
+ lora_target_modules: ''
84
+ loss_function: CrossEntropy
85
+ optimizer: AdamW
86
+ save_best_checkpoint: false
87
+ schedule: Cosine
88
+ train_validation_data: false
89
+ warmup_epochs: 0.0
90
+ weight_decay: 0.0