Upload train-config.yaml with huggingface_hub
Browse files- train-config.yaml +19 -13
train-config.yaml
CHANGED
@@ -7,6 +7,9 @@ format_args:
|
|
7 |
model_args:
|
8 |
pretrained_model_name_or_path: alignment-handbook/zephyr-7b-sft-full
|
9 |
torch_dtype: float16
|
|
|
|
|
|
|
10 |
|
11 |
peft_config:
|
12 |
r: 16
|
@@ -26,41 +29,44 @@ peft_config:
|
|
26 |
wandb_args:
|
27 |
entity: argilla-io
|
28 |
project: dibt-dpo
|
29 |
-
name: zephyr-7b-lora-dpo-dibt-v0
|
30 |
|
31 |
training_args:
|
32 |
-
|
33 |
beta: 0.1
|
|
|
|
|
34 |
loss_type: sigmoid
|
|
|
|
|
35 |
do_eval: true
|
36 |
do_train: true
|
37 |
evaluation_strategy: steps
|
38 |
-
eval_steps:
|
39 |
-
gradient_accumulation_steps:
|
40 |
gradient_checkpointing: true
|
41 |
-
gradient_checkpointing_kwargs:
|
42 |
-
use_reentrant: False
|
43 |
hub_model_id: plaguss/zephyr-7b-lora-dpo-dibt-v0
|
44 |
hub_model_revision: v0
|
45 |
hub_strategy: every_save
|
46 |
hub_private_repo: true
|
47 |
push_to_hub: true
|
48 |
-
learning_rate: 5.0e-
|
49 |
logging_steps: 10
|
50 |
lr_scheduler_type: cosine
|
51 |
-
max_length: 1024
|
52 |
-
max_prompt_length: 512
|
53 |
num_train_epochs: 2
|
54 |
optim: paged_adamw_32bit
|
55 |
output_dir: data/zephyr-7b-sft-lora-dpo-v0
|
56 |
-
|
57 |
-
|
58 |
-
|
|
|
|
|
|
|
59 |
save_total_limit: null
|
60 |
seed: 42
|
61 |
warmup_ratio: 0.1
|
62 |
report_to:
|
63 |
- wandb
|
64 |
|
65 |
-
use_accelerate:
|
66 |
use_unsloth: false
|
|
|
7 |
model_args:
|
8 |
pretrained_model_name_or_path: alignment-handbook/zephyr-7b-sft-full
|
9 |
torch_dtype: float16
|
10 |
+
quantization_config:
|
11 |
+
quant_method: bitsandbytes
|
12 |
+
load_in_4bit: true
|
13 |
|
14 |
peft_config:
|
15 |
r: 16
|
|
|
29 |
wandb_args:
|
30 |
entity: argilla-io
|
31 |
project: dibt-dpo
|
32 |
+
name: zephyr-7b-lora-dpo-dibt-openhermes-params-v0
|
33 |
|
34 |
training_args:
|
35 |
+
# `trl.DPOTrainer`
|
36 |
beta: 0.1
|
37 |
+
max_length: 1536
|
38 |
+
max_prompt_length: 1024
|
39 |
loss_type: sigmoid
|
40 |
+
# `transformers.Trainer`
|
41 |
+
bf16: true
|
42 |
do_eval: true
|
43 |
do_train: true
|
44 |
evaluation_strategy: steps
|
45 |
+
eval_steps: 20
|
46 |
+
gradient_accumulation_steps: 4
|
47 |
gradient_checkpointing: true
|
|
|
|
|
48 |
hub_model_id: plaguss/zephyr-7b-lora-dpo-dibt-v0
|
49 |
hub_model_revision: v0
|
50 |
hub_strategy: every_save
|
51 |
hub_private_repo: true
|
52 |
push_to_hub: true
|
53 |
+
learning_rate: 5.0e-5
|
54 |
logging_steps: 10
|
55 |
lr_scheduler_type: cosine
|
|
|
|
|
56 |
num_train_epochs: 2
|
57 |
optim: paged_adamw_32bit
|
58 |
output_dir: data/zephyr-7b-sft-lora-dpo-v0
|
59 |
+
load_best_model_at_end: true
|
60 |
+
metric_for_best_model: rewards/accuracies
|
61 |
+
greater_is_better: true
|
62 |
+
per_device_train_batch_size: 4
|
63 |
+
per_device_eval_batch_size: 16
|
64 |
+
save_strategy: steps
|
65 |
save_total_limit: null
|
66 |
seed: 42
|
67 |
warmup_ratio: 0.1
|
68 |
report_to:
|
69 |
- wandb
|
70 |
|
71 |
+
use_accelerate: false
|
72 |
use_unsloth: false
|