Crystalcareai commited on
Commit
65a5134
1 Parent(s): 093c261

Added Training Config Files

Browse files

Adding YAML files detailing the parameters used for mergekit and training. These configurations were frequently requested.

Training-Configs/AxolotlConfig.yml ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ base_model: Crystalcareai/Qwen-1.5-8x7B
2
+ model_type: Qwen2ForCausalLM
3
+ tokenizer_type: Qwen2Tokenizer
4
+ trust_remote_code: true
5
+
6
+
7
+ load_in_8bit: false
8
+ load_in_4bit: true
9
+ strict: false
10
+
11
+
12
+ datasets:
13
+ - path: Crystalcareai/MoD
14
+ type: sharegpt
15
+ dataset_prepared_path: last_run_prepared
16
+ val_set_size: 0.0
17
+ output_dir: ./qlora-out
18
+
19
+
20
+ model_config:
21
+ output_router_logits: true
22
+
23
+
24
+ adapter: qlora
25
+ lora_model_dir:
26
+
27
+
28
+ sequence_len: 32768
29
+ sample_packing: true
30
+ pad_to_sequence_len: true
31
+
32
+
33
+ lora_r: 32
34
+ lora_alpha: 16
35
+ lora_dropout: 0.05
36
+ lora_target_linear: true
37
+ lora_fan_in_fan_out:
38
+
39
+
40
+ gradient_accumulation_steps: 2
41
+ micro_batch_size: 2
42
+ num_epochs: 4
43
+ optimizer: adamw_bnb_8bit
44
+ lr_scheduler: cosine
45
+ learning_rate: 0.0002
46
+
47
+
48
+ train_on_inputs: false
49
+ group_by_length: false
50
+ bf16: true
51
+ fp16:
52
+ tf32: false
53
+
54
+
55
+ gradient_checkpointing: true
56
+ early_stopping_patience:
57
+ resume_from_checkpoint:
58
+ local_rank:
59
+ logging_steps: 1
60
+ xformers_attention:
61
+ flash_attention: true
62
+
63
+
64
+ warmup_steps: 10
65
+ evals_per_epoch: 4
66
+ eval_table_size:
67
+ eval_max_new_tokens: 128
68
+ saves_per_epoch: 4
69
+ debug:
70
+ deepspeed: deepspeed_configs/zero2.json
71
+ weight_decay: 0.0
72
+ fsdp:
73
+ fsdp_config:
74
+ special_tokens:
Training-Configs/Qwen1.5-8x7b-Mergekit-Config.yml ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ gate_mode: random
2
+ dtype: bfloat16
3
+ experts_per_token: 2
4
+ experts:
5
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
6
+ positive_prompts:
7
+ - "You are a versatile assistant for a wide array of inquiries and daily assistance."
8
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
9
+ positive_prompts:
10
+ - "You are a tech and programming guide for various skill levels, from beginners to experts."
11
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
12
+ positive_prompts:
13
+ - "You are a world explorer - offering diverse insights on culture, and social studies."
14
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
15
+ positive_prompts:
16
+ - "You are a financial navigator for saving, investing, and managing money wisely."
17
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
18
+ positive_prompts:
19
+ - "You are an educational ally, supporting learning across a broad spectrum of subjects."
20
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
21
+ positive_prompts:
22
+ - "You are a well-being companion, offering advice on physical health, mental wellness, and fitness."
23
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
24
+ positive_prompts:
25
+ - "You are a creative muse for writing, arts, and all forms of creative expression."
26
+ - source_model: Crystalcareai/Qwen1.5-7B-MoD100k
27
+ positive_prompts:
28
+ - "You are a philosopher and interpreter and explorer of the human condition."