apunkt/doc-to-lora
1
0
Fork 0
mirror of https://github.com/SakanaAI/doc-to-lora.git synced 2026-07-23 17:01:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

data + configs testing model capability

Browse source
This commit is contained in:
51616 2025-01-07 14:44:51 +00:00
parent c7f7a2f788
commit 2a4e5742ca
9 changed files with 442 additions and 128 deletions
Download patch file Download diff file Expand all files Collapse all files

130
configs/context_numbers_128.yaml
View file

@ -16,9 +16,9 @@ label_names: ["labels"]
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 64
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 50
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
@ -36,132 +36,15 @@ target_modules:
# data
train_ds_names:
- data/raw_datasets/context_numbers_2
- data/raw_datasets/context_numbers_3
- data/raw_datasets/context_numbers_4
- data/raw_datasets/context_numbers_5
- data/raw_datasets/context_numbers_6
- data/raw_datasets/context_numbers_7
- data/raw_datasets/context_numbers_8
- data/raw_datasets/context_numbers_9
- data/raw_datasets/context_numbers_10
- data/raw_datasets/context_numbers_11
- data/raw_datasets/context_numbers_12
- data/raw_datasets/context_numbers_13
- data/raw_datasets/context_numbers_14
- data/raw_datasets/context_numbers_15
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_17
- data/raw_datasets/context_numbers_18
- data/raw_datasets/context_numbers_19
- data/raw_datasets/context_numbers_20
- data/raw_datasets/context_numbers_21
- data/raw_datasets/context_numbers_22
- data/raw_datasets/context_numbers_23
- data/raw_datasets/context_numbers_24
- data/raw_datasets/context_numbers_25
- data/raw_datasets/context_numbers_26
- data/raw_datasets/context_numbers_27
- data/raw_datasets/context_numbers_28
- data/raw_datasets/context_numbers_29
- data/raw_datasets/context_numbers_30
- data/raw_datasets/context_numbers_31
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_33
- data/raw_datasets/context_numbers_34
- data/raw_datasets/context_numbers_35
- data/raw_datasets/context_numbers_36
- data/raw_datasets/context_numbers_37
- data/raw_datasets/context_numbers_38
- data/raw_datasets/context_numbers_39
- data/raw_datasets/context_numbers_40
- data/raw_datasets/context_numbers_41
- data/raw_datasets/context_numbers_42
- data/raw_datasets/context_numbers_43
- data/raw_datasets/context_numbers_44
- data/raw_datasets/context_numbers_45
- data/raw_datasets/context_numbers_46
- data/raw_datasets/context_numbers_47
- data/raw_datasets/context_numbers_48
- data/raw_datasets/context_numbers_49
- data/raw_datasets/context_numbers_50
- data/raw_datasets/context_numbers_51
- data/raw_datasets/context_numbers_52
- data/raw_datasets/context_numbers_53
- data/raw_datasets/context_numbers_54
- data/raw_datasets/context_numbers_55
- data/raw_datasets/context_numbers_56
- data/raw_datasets/context_numbers_57
- data/raw_datasets/context_numbers_58
- data/raw_datasets/context_numbers_59
- data/raw_datasets/context_numbers_60
- data/raw_datasets/context_numbers_61
- data/raw_datasets/context_numbers_62
- data/raw_datasets/context_numbers_63
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_65
- data/raw_datasets/context_numbers_66
- data/raw_datasets/context_numbers_67
- data/raw_datasets/context_numbers_68
- data/raw_datasets/context_numbers_69
- data/raw_datasets/context_numbers_70
- data/raw_datasets/context_numbers_71
- data/raw_datasets/context_numbers_72
- data/raw_datasets/context_numbers_73
- data/raw_datasets/context_numbers_74
- data/raw_datasets/context_numbers_75
- data/raw_datasets/context_numbers_76
- data/raw_datasets/context_numbers_77
- data/raw_datasets/context_numbers_78
- data/raw_datasets/context_numbers_79
- data/raw_datasets/context_numbers_80
- data/raw_datasets/context_numbers_81
- data/raw_datasets/context_numbers_82
- data/raw_datasets/context_numbers_83
- data/raw_datasets/context_numbers_84
- data/raw_datasets/context_numbers_85
- data/raw_datasets/context_numbers_86
- data/raw_datasets/context_numbers_87
- data/raw_datasets/context_numbers_88
- data/raw_datasets/context_numbers_89
- data/raw_datasets/context_numbers_90
- data/raw_datasets/context_numbers_91
- data/raw_datasets/context_numbers_92
- data/raw_datasets/context_numbers_93
- data/raw_datasets/context_numbers_94
- data/raw_datasets/context_numbers_95
- data/raw_datasets/context_numbers_96
- data/raw_datasets/context_numbers_97
- data/raw_datasets/context_numbers_98
- data/raw_datasets/context_numbers_99
- data/raw_datasets/context_numbers_100
- data/raw_datasets/context_numbers_101
- data/raw_datasets/context_numbers_102
- data/raw_datasets/context_numbers_103
- data/raw_datasets/context_numbers_104
- data/raw_datasets/context_numbers_105
- data/raw_datasets/context_numbers_106
- data/raw_datasets/context_numbers_107
- data/raw_datasets/context_numbers_108
- data/raw_datasets/context_numbers_109
- data/raw_datasets/context_numbers_110
- data/raw_datasets/context_numbers_111
- data/raw_datasets/context_numbers_112
- data/raw_datasets/context_numbers_113
- data/raw_datasets/context_numbers_114
- data/raw_datasets/context_numbers_115
- data/raw_datasets/context_numbers_116
- data/raw_datasets/context_numbers_117
- data/raw_datasets/context_numbers_118
- data/raw_datasets/context_numbers_119
- data/raw_datasets/context_numbers_120
- data/raw_datasets/context_numbers_121
- data/raw_datasets/context_numbers_122
- data/raw_datasets/context_numbers_123
- data/raw_datasets/context_numbers_124
- data/raw_datasets/context_numbers_125
- data/raw_datasets/context_numbers_126
- data/raw_datasets/context_numbers_127
- data/raw_datasets/context_numbers_128
val_ds_names:
@ -172,7 +55,8 @@ val_ds_names:
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_512
- data/raw_datasets/context_numbers_1024
- data/raw_datasets/context_numbers_2048
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

62
configs/context_numbers_128_new.yaml Normal file
View file

@ -0,0 +1,62 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_2
- data/raw_datasets/context_numbers_4
- data/raw_datasets/context_numbers_8
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_48
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_80
- data/raw_datasets/context_numbers_96
- data/raw_datasets/context_numbers_112
- data/raw_datasets/context_numbers_128
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_512
- data/raw_datasets/context_numbers_1024
- data/raw_datasets/context_numbers_2048

53
configs/context_numbers_128_only.yaml Normal file
View file

@ -0,0 +1,53 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_128_big
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

11
configs/context_numbers_256_only.yaml
View file

@ -36,7 +36,7 @@ target_modules:
# data
train_ds_names:
- data/raw_datasets/context_numbers_256
- data/raw_datasets/context_numbers_256_big
val_ds_names:
- data/raw_datasets/context_numbers_16
@ -46,7 +46,8 @@ val_ds_names:
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_512
- data/raw_datasets/context_numbers_1024
- data/raw_datasets/context_numbers_2048
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

61
configs/context_numbers_32.yaml Normal file
View file

@ -0,0 +1,61 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_2
- data/raw_datasets/context_numbers_4
- data/raw_datasets/context_numbers_8
- data/raw_datasets/context_numbers_12
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_20
- data/raw_datasets/context_numbers_24
- data/raw_datasets/context_numbers_28
- data/raw_datasets/context_numbers_32
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

53
configs/context_numbers_32_only.yaml Normal file
View file

@ -0,0 +1,53 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_32_big
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

61
configs/context_numbers_64.yaml Normal file
View file

@ -0,0 +1,61 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_4
- data/raw_datasets/context_numbers_8
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_24
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_40
- data/raw_datasets/context_numbers_48
- data/raw_datasets/context_numbers_56
- data/raw_datasets/context_numbers_64
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256

53
configs/context_numbers_64_only.yaml Normal file
View file

@ -0,0 +1,53 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 1
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.00001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- data/raw_datasets/context_numbers_64_big
val_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_16
- data/raw_datasets/context_numbers_32
- data/raw_datasets/context_numbers_64
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_256