Upload folder using huggingface_hub
Browse files- configs/phi-lima.yml +76 -0
configs/phi-lima.yml
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: microsoft/phi-1_5
|
2 |
+
base_model_config: microsoft/phi-1_5
|
3 |
+
model_type: AutoModelForCausalLM
|
4 |
+
tokenizer_type: AutoTokenizer
|
5 |
+
is_llama_derived_model: false
|
6 |
+
trust_remote_code: true
|
7 |
+
|
8 |
+
load_in_8bit: false
|
9 |
+
load_in_4bit: true
|
10 |
+
strict: false
|
11 |
+
|
12 |
+
datasets:
|
13 |
+
- path: garage-bAInd/Open-Platypus
|
14 |
+
type: alpaca
|
15 |
+
|
16 |
+
dataset_prepared_path: last_run_prepared
|
17 |
+
val_set_size: 0.01
|
18 |
+
output_dir: ./phi-sft-out
|
19 |
+
|
20 |
+
sequence_len: 1024
|
21 |
+
sample_packing: false
|
22 |
+
pad_to_sequence_len:
|
23 |
+
|
24 |
+
adapter: qlora
|
25 |
+
lora_model_dir:
|
26 |
+
lora_r: 64
|
27 |
+
lora_alpha: 32
|
28 |
+
lora_dropout: 0.05
|
29 |
+
lora_target_linear: true
|
30 |
+
lora_fan_in_fan_out:
|
31 |
+
|
32 |
+
wandb_project: phi-1_5-sft
|
33 |
+
wandb_entity:
|
34 |
+
wandb_watch:
|
35 |
+
wandb_run_id:
|
36 |
+
wandb_log_model:
|
37 |
+
|
38 |
+
gradient_accumulation_steps: 1
|
39 |
+
micro_batch_size: 3
|
40 |
+
num_epochs: 2
|
41 |
+
optimizer: adamw_bnb_8bit
|
42 |
+
adam_beta2: 0.99
|
43 |
+
adam_epsilon: 0.00001
|
44 |
+
max_grad_norm: 1.0
|
45 |
+
lr_scheduler: cosine
|
46 |
+
learning_rate: 0.0002
|
47 |
+
|
48 |
+
train_on_inputs: false
|
49 |
+
group_by_length: true
|
50 |
+
bf16: true
|
51 |
+
fp16: false
|
52 |
+
tf32: true
|
53 |
+
|
54 |
+
gradient_checkpointing:
|
55 |
+
early_stopping_patience:
|
56 |
+
resume_from_checkpoint:
|
57 |
+
local_rank:
|
58 |
+
logging_steps: 1
|
59 |
+
xformers_attention:
|
60 |
+
flash_attention:
|
61 |
+
|
62 |
+
warmup_steps: 100
|
63 |
+
eval_steps: 0.05
|
64 |
+
save_steps:
|
65 |
+
debug:
|
66 |
+
deepspeed:
|
67 |
+
weight_decay: 0.01
|
68 |
+
fsdp:
|
69 |
+
fsdp_config:
|
70 |
+
resize_token_embeddings_to_32x: true
|
71 |
+
special_tokens:
|
72 |
+
bos_token: "<|endoftext|>"
|
73 |
+
eos_token: "<|endoftext|>"
|
74 |
+
unk_token: "<|endoftext|>"
|
75 |
+
pad_token: "<|pad|>"
|
76 |
+
|