Hugging Face's logo

Datasets:
Delta-Vector
/
Configs

Dataset card Files Community
Delta-Vector commited on
Commit
d1dd6f3
·
verified ·
1 Parent(s): 4f303b9

Update control14b.yaml

Browse files
Files changed (1) hide show
  1. control14b.yaml +144 -140
control14b.yaml CHANGED
@@ -1,140 +1,144 @@
1
- base_model: allura-org/TQ2.5-14B-Sugarquill-v1
2
- strict: false
3
-
4
- plugins:
5
- - axolotl.integrations.liger.LigerPlugin
6
- liger_rope: true
7
- liger_rms_norm: true
8
- liger_swiglu: true
9
- liger_fused_linear_cross_entropy: true
10
-
11
- # Output and HuggingFace
12
- hub_model_id: NewEden/control-14b-lora
13
- hf_use_auth_token: true
14
- hub_strategy: "all_checkpoints"
15
-
16
- wandb_project: huggingface
17
- wandb_entity:
18
- wandb_name: Control-14B
19
-
20
-
21
- chat_template: chatml
22
- group_by_length: false
23
- datasets:
24
- - path: Nitral-AI/Creative_Writing-ShareGPT
25
- type: chat_template
26
- roles_to_train: ["gpt"]
27
- field_messages: conversations
28
- message_field_role: from
29
- message_field_content: value
30
- train_on_eos: turn
31
- - path: Nitral-AI/ARES-ShareGPT
32
- type: chat_template
33
- chat_template: chatml
34
- roles_to_train: ["gpt"]
35
- field_messages: conversations
36
- message_field_role: from
37
- message_field_content: value
38
- train_on_eos: turn
39
- - path: NewEden/Claude-Instruct-5K
40
- type: chat_template
41
- chat_template: chatml
42
- roles_to_train: ["gpt"]
43
- field_messages: conversations
44
- message_field_role: from
45
- message_field_content: value
46
- train_on_eos: turn
47
- - path: NewEden/OpenCAI-ShareGPT
48
- type: chat_template
49
- roles_to_train: ["gpt"]
50
- field_messages: conversations
51
- message_field_role: from
52
- message_field_content: value
53
- train_on_eos: turn
54
- - path: NewEden/PIPPA-Mega-Filtered
55
- type: chat_template
56
- chat_template: chatml
57
- roles_to_train: ["gpt"]
58
- field_messages: conversations
59
- message_field_role: from
60
- message_field_content: value
61
- train_on_eos: turn
62
- - path: NewEden/Roleplay-Logs-Sharegpt-Ngram-cleaned
63
- type: chat_template
64
- chat_template: chatml
65
- roles_to_train: ["gpt"]
66
- field_messages: conversations
67
- message_field_role: from
68
- message_field_content: value
69
- train_on_eos: turn
70
- - path: Nitral-AI/Creative_Writing-ShareGPT
71
- type: chat_template
72
- chat_template: chatml
73
- roles_to_train: ["gpt"]
74
- field_messages: conversations
75
- message_field_role: from
76
- message_field_content: value
77
- train_on_eos: turn
78
-
79
- val_set_size: 0.01
80
- evals_per_epoch: 4
81
- eval_table_size:
82
- eval_max_new_tokens: 128
83
-
84
- sequence_len: 8192
85
- save_safetensors: true
86
- saves_per_epoch: 2
87
- logging_steps: 1
88
- special_tokens:
89
-
90
- # Quantization
91
- bf16: auto
92
- fp16:
93
- tf32: false
94
- ## For LoRA
95
- load_in_8bit: false
96
- load_in_4bit: True
97
-
98
- # LoRA
99
- peft_use_rslora: true
100
- adapter: qlora
101
- lora_model_dir:
102
- lora_r: 128
103
- lora_alpha: 16
104
- lora_dropout: 0.05
105
- lora_target_linear: true
106
- lora_fan_in_fan_out:
107
- lora_target_modules:
108
-
109
- ## if oom
110
- # lora_r: 64
111
- # lora_alpha: 32
112
- # lora_dropout: 0.1
113
-
114
-
115
- weight_decay: 0.02
116
- max_grad_norm: 1.0
117
-
118
- warmup_ratio: 0.05
119
- learning_rate: 0.00002
120
- lr_scheduler: cosine
121
- #lr_scheduler_kwargs:
122
- optimizer: paged_adamw_8bit # usually adamw_torch or paged_adamw_8bit
123
-
124
- ## Batch Size
125
- gradient_accumulation_steps: 8
126
- micro_batch_size: 1
127
- eval_batch_size: 1
128
-
129
- # Optimizations
130
- pad_to_sequence_len: true
131
- sample_packing: true
132
- eval_sample_packing: false
133
- flash_attention: true
134
- xformers_attention:
135
- gradient_checkpointing: "unsloth"
136
- gradient_checkpointing_kwargs:
137
- use_reentrant: true
138
- local_rank:
139
- early_stopping_patience:
140
- debug:
 
 
 
 
 
1
+ base_model: allura-org/TQ2.5-14B-Sugarquill-v1
2
+ strict: false
3
+
4
+ plugins:
5
+ - axolotl.integrations.liger.LigerPlugin
6
+ liger_rope: true
7
+ liger_rms_norm: true
8
+ liger_swiglu: true
9
+ liger_fused_linear_cross_entropy: true
10
+
11
+ # Output and HuggingFace
12
+ hub_model_id: NewEden/control-14b-lora
13
+ hf_use_auth_token: true
14
+ hub_strategy: "all_checkpoints"
15
+
16
+ wandb_project: huggingface
17
+ wandb_entity:
18
+ wandb_name: Control-14B
19
+
20
+
21
+ chat_template: chatml
22
+ group_by_length: false
23
+ datasets:
24
+ - path: Nitral-AI/Creative_Writing-ShareGPT
25
+ type: chat_template
26
+ roles_to_train: ["gpt"]
27
+ field_messages: conversations
28
+ message_field_role: from
29
+ message_field_content: value
30
+ train_on_eos: turn
31
+ - path: Nitral-AI/ARES-ShareGPT
32
+ type: chat_template
33
+ chat_template: chatml
34
+ roles_to_train: ["gpt"]
35
+ field_messages: conversations
36
+ message_field_role: from
37
+ message_field_content: value
38
+ train_on_eos: turn
39
+ - path: NewEden/Claude-Instruct-5K
40
+ type: chat_template
41
+ chat_template: chatml
42
+ roles_to_train: ["gpt"]
43
+ field_messages: conversations
44
+ message_field_role: from
45
+ message_field_content: value
46
+ train_on_eos: turn
47
+ - path: NewEden/OpenCAI-ShareGPT
48
+ type: chat_template
49
+ roles_to_train: ["gpt"]
50
+ field_messages: conversations
51
+ message_field_role: from
52
+ message_field_content: value
53
+ train_on_eos: turn
54
+ - path: NewEden/PIPPA-Mega-Filtered
55
+ type: chat_template
56
+ chat_template: chatml
57
+ roles_to_train: ["gpt"]
58
+ field_messages: conversations
59
+ message_field_role: from
60
+ message_field_content: value
61
+ train_on_eos: turn
62
+ - path: NewEden/Roleplay-Logs-Sharegpt-Ngram-cleaned
63
+ type: chat_template
64
+ chat_template: chatml
65
+ roles_to_train: ["gpt"]
66
+ field_messages: conversations
67
+ message_field_role: from
68
+ message_field_content: value
69
+ train_on_eos: turn
70
+ - path: Nitral-AI/Creative_Writing-ShareGPT
71
+ type: chat_template
72
+ chat_template: chatml
73
+ roles_to_train: ["gpt"]
74
+ field_messages: conversations
75
+ message_field_role: from
76
+ message_field_content: value
77
+ train_on_eos: turn
78
+
79
+ #val_set_size: 0.01
80
+ #evals_per_epoch: 1
81
+ # eval_table_size:
82
+ # eval_max_new_tokens: 128
83
+ num_epochs: 2
84
+
85
+ sequence_len: 8192
86
+ save_safetensors: true
87
+ saves_per_epoch: 2
88
+ logging_steps: 1
89
+ special_tokens:
90
+
91
+ # Quantization
92
+ bf16: auto
93
+ fp16:
94
+ tf32: false
95
+ ## For LoRA
96
+ load_in_8bit: false
97
+ load_in_4bit: True
98
+
99
+ # LoRA
100
+ peft_use_rslora: true
101
+ adapter: qlora
102
+ lora_model_dir:
103
+ lora_r: 128
104
+ lora_alpha: 16
105
+ lora_dropout: 0.05
106
+ lora_target_linear: true
107
+ lora_fan_in_fan_out:
108
+ lora_target_modules:
109
+
110
+ ## if oom
111
+ # lora_r: 64
112
+ # lora_alpha: 32
113
+ # lora_dropout: 0.1
114
+
115
+
116
+ weight_decay: 0.02
117
+ max_grad_norm: 1.0
118
+
119
+ warmup_ratio: 0.05
120
+ learning_rate: 0.00002
121
+ lr_scheduler: cosine
122
+ #lr_scheduler_kwargs:
123
+ optimizer: paged_adamw_8bit # usually adamw_torch or paged_adamw_8bit
124
+
125
+ ## Batch Size
126
+ gradient_accumulation_steps: 8
127
+ micro_batch_size: 1
128
+ eval_batch_size: 1
129
+
130
+ # Optimizations
131
+ pad_to_sequence_len: true
132
+ sample_packing: true
133
+ eval_sample_packing: false
134
+ flash_attention: true
135
+ xformers_attention:
136
+ gradient_checkpointing: "unsloth"
137
+ gradient_checkpointing_kwargs:
138
+ use_reentrant: true
139
+ local_rank:
140
+ early_stopping_patience:
141
+ debug:
142
+ special_tokens:
143
+ pad_token: <|endoftext|>
144
+ eos_token: <|im_end|>