Upload kto-4b.yml
Browse files- kto-4b.yml +109 -0
kto-4b.yml
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: NewEden/Hamanasu-4B-R2
|
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: NewEden/KTO-4B
|
10 |
+
hub_strategy: "all_checkpoints"
|
11 |
+
push_dataset_to_hub:
|
12 |
+
hf_use_auth_token: true
|
13 |
+
|
14 |
+
chat_template: chatml
|
15 |
+
|
16 |
+
rl: kto
|
17 |
+
rl_beta: 0.2
|
18 |
+
kto_desirable_weight: 0.2
|
19 |
+
|
20 |
+
datasets:
|
21 |
+
- path: NewEden/Opus-accepted-hermes-rejected-shuffled
|
22 |
+
split: train
|
23 |
+
type: chatml.argilla
|
24 |
+
- path: NewEden/KTO-IF-Dans
|
25 |
+
split: train
|
26 |
+
type: chatml.argilla
|
27 |
+
- path: NewEden/KTO-IF-Dans
|
28 |
+
split: train
|
29 |
+
type: chatml.argilla
|
30 |
+
- path: NewEden/Purpura-Arkhaios-CC-KTO
|
31 |
+
split: train
|
32 |
+
type: chatml.argilla
|
33 |
+
dataset_prepared_path: last_run_prepared
|
34 |
+
|
35 |
+
shuffle_merged_datasets: true
|
36 |
+
val_set_size: 0.0
|
37 |
+
output_dir: ./outputs/out
|
38 |
+
|
39 |
+
adapter: lora
|
40 |
+
lora_model_dir:
|
41 |
+
|
42 |
+
lora_r: 64
|
43 |
+
lora_alpha: 32
|
44 |
+
lora_dropout: 0.0
|
45 |
+
lora_target_linear: true
|
46 |
+
lora_fan_in_fan_out:
|
47 |
+
lora_target_modules:
|
48 |
+
- gate_proj
|
49 |
+
- down_proj
|
50 |
+
- up_proj
|
51 |
+
- q_proj
|
52 |
+
- v_proj
|
53 |
+
- k_proj
|
54 |
+
- o_proj
|
55 |
+
|
56 |
+
sequence_len: 32768
|
57 |
+
sample_packing: false
|
58 |
+
eval_sample_packing: false
|
59 |
+
pad_to_sequence_len: false
|
60 |
+
|
61 |
+
wandb_project: tavbussy
|
62 |
+
wandb_entity:
|
63 |
+
wandb_watch:
|
64 |
+
wandb_name: kto-1
|
65 |
+
wandb_log_model:
|
66 |
+
|
67 |
+
gradient_accumulation_steps: 2
|
68 |
+
micro_batch_size: 4
|
69 |
+
num_epochs: 1
|
70 |
+
optimizer: paged_adamw_8bit
|
71 |
+
lr_scheduler:
|
72 |
+
learning_rate: 1e-6
|
73 |
+
max_grad_norm: 0.001
|
74 |
+
|
75 |
+
optimizer: paged_adamw_8bit
|
76 |
+
lr_scheduler: cosine
|
77 |
+
learning_rate: 1e-5
|
78 |
+
weight_decay: 0.02
|
79 |
+
|
80 |
+
train_on_inputs: false
|
81 |
+
group_by_length: false
|
82 |
+
bf16: auto
|
83 |
+
fp16:
|
84 |
+
tf32: true
|
85 |
+
|
86 |
+
gradient_checkpointing: true
|
87 |
+
remove_unused_columns: false
|
88 |
+
early_stopping_patience:
|
89 |
+
resume_from_checkpoint:
|
90 |
+
local_rank:
|
91 |
+
logging_steps: 1
|
92 |
+
xformers_attention:
|
93 |
+
flash_attention: true
|
94 |
+
|
95 |
+
warmup_steps: 35
|
96 |
+
evals_per_epoch: 2
|
97 |
+
eval_table_size:
|
98 |
+
eval_max_new_tokens:
|
99 |
+
saves_per_epoch: 2
|
100 |
+
|
101 |
+
debug:
|
102 |
+
deepspeed: ./deepspeed_configs/zero3_bf16.json
|
103 |
+
fsdp:
|
104 |
+
fsdp_config:
|
105 |
+
fsdp:
|
106 |
+
fsdp_config:
|
107 |
+
|
108 |
+
special_tokens:
|
109 |
+
pad_token: <|finetune_right_pad_id|>
|