jeiku commited on
Commit
3f205f0
·
verified ·
1 Parent(s): 011a0c9

Upload controlkto.yml

Browse files
Files changed (1) hide show
  1. controlkto.yml +90 -0
controlkto.yml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ base_model: Delta-Vector/Control-8B-V1.1
2
+ model_type: AutoModelForCausalLM
3
+ tokenizer_type: AutoTokenizer
4
+
5
+ load_in_8bit: false
6
+ load_in_4bit: false
7
+ strict: false
8
+
9
+ hub_model_id: jeiku/controlkto
10
+ hub_strategy: "all_checkpoints"
11
+ push_dataset_to_hub:
12
+ hf_use_auth_token: true
13
+
14
+ chat_template: llama3
15
+
16
+ rl: kto
17
+ rl_beta: 0.2
18
+ kto_desirable_weight: 0.2
19
+
20
+ datasets:
21
+ - path: anthracite-core/full-opus-chosen-hermes-rejected-kto-v1
22
+ type: llama3.argilla
23
+
24
+ shuffle_merged_datasets: true
25
+ val_set_size: 0.0
26
+ output_dir: ./outputs/out
27
+
28
+ adapter: lora
29
+ lora_model_dir:
30
+
31
+ lora_r: 32
32
+ lora_alpha: 64
33
+ lora_dropout: 0.05
34
+ lora_target_linear: true
35
+ lora_fan_in_fan_out:
36
+
37
+ sequence_len: 8192
38
+ sample_packing: false
39
+ eval_sample_packing: false
40
+ pad_to_sequence_len: false
41
+
42
+ wandb_project: controlkto
43
+ wandb_entity:
44
+ wandb_watch:
45
+ wandb_name: controlkto
46
+ wandb_log_model:
47
+
48
+ gradient_accumulation_steps: 16
49
+ micro_batch_size: 2
50
+ num_epochs: 2
51
+ max_steps: 500
52
+
53
+ optimizer: adamw_8bit
54
+ lr_scheduler: cosine
55
+ learning_rate: 0.0001
56
+ weight_decay: 0.05
57
+
58
+ train_on_inputs: false
59
+ group_by_length: false
60
+ bf16: auto
61
+ fp16:
62
+ tf32: true
63
+
64
+ gradient_checkpointing: true
65
+ gradient_checkpointing_kwargs:
66
+ use_reentrant: true
67
+ remove_unused_columns: false
68
+ early_stopping_patience:
69
+ resume_from_checkpoint:
70
+ local_rank:
71
+ logging_steps: 1
72
+ xformers_attention:
73
+ flash_attention: true
74
+
75
+ warmup_steps: 10
76
+ evals_per_epoch: 2
77
+ eval_table_size:
78
+ eval_max_new_tokens:
79
+ saves_per_epoch: 1
80
+
81
+ debug:
82
+ deepspeed:
83
+ fsdp:
84
+ fsdp_config:
85
+ fsdp:
86
+ fsdp_config:
87
+
88
+ special_tokens:
89
+ pad_token: <|finetune_right_pad_id|>
90
+ eos_token: <|eot_id|>