Update README.md
Browse files
README.md
CHANGED
@@ -1,2 +1,63 @@
|
|
1 |
|
2 |
์ด๋์ ๋ ์ด์ ๋ ๋ง์ ์กฐ๊ธ ์์ ๋ฃ๋๋ค.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
|
2 |
์ด๋์ ๋ ์ด์ ๋ ๋ง์ ์กฐ๊ธ ์์ ๋ฃ๋๋ค.
|
3 |
+
|
4 |
+
### 8bit ํ์ต ๊ธฐ์ค
|
5 |
+
|
6 |
+
{
|
7 |
+
"_name_or_path": "saltlux/Ko-Llama3-Luxia-8B",
|
8 |
+
"architectures": [
|
9 |
+
"LlamaForCausalLM"
|
10 |
+
],
|
11 |
+
"attention_bias": false,
|
12 |
+
"attention_dropout": 0.0,
|
13 |
+
"bos_token_id": 128000,
|
14 |
+
"eos_token_id": 128001,
|
15 |
+
"hidden_act": "silu",
|
16 |
+
"hidden_size": 4096,
|
17 |
+
"initializer_range": 0.02,
|
18 |
+
"intermediate_size": 14336,
|
19 |
+
"max_position_embeddings": 8192,
|
20 |
+
"model_type": "llama",
|
21 |
+
"num_attention_heads": 32,
|
22 |
+
"num_hidden_layers": 32,
|
23 |
+
"num_key_value_heads": 8,
|
24 |
+
"pretraining_tp": 1,
|
25 |
+
"rms_norm_eps": 1e-05,
|
26 |
+
"rope_scaling": null,
|
27 |
+
"rope_theta": 500000.0,
|
28 |
+
"tie_word_embeddings": false,
|
29 |
+
"torch_dtype": "float16",
|
30 |
+
"transformers_version": "4.39.1",
|
31 |
+
"use_cache": false,
|
32 |
+
"vocab_size": 145792
|
33 |
+
}
|
34 |
+
|
35 |
+
batch_size = 16
|
36 |
+
num_epochs = 1
|
37 |
+
micro_batch = 1
|
38 |
+
gradient_accumulation_steps = batch_size
|
39 |
+
|
40 |
+
### ํ๋ จ ๋ฐฉ๋ฒ์ ๋ํ ํ์ดํผ ํ๋ผ๋ฏธํฐ
|
41 |
+
cutoff_len = model.config.hidden_size
|
42 |
+
lr_scheduler = 'cosine'
|
43 |
+
warmup_ratio = 0.06 # warmup_steps = 100
|
44 |
+
learning_rate = 2e-4
|
45 |
+
optimizer = 'adamw_torch'
|
46 |
+
weight_decay = 0.01
|
47 |
+
max_grad_norm = 0.8 # ๋ชจ๋ธ์ด ๋๋ฌด ๊ณผ์ ํฉ๋๋๊ฑฐ๊ฐ๊ฑฐ๋ ์ด์ํ ๋ฐฉํฅ์ผ๋ก ํ์ต์ด ๋๋๊ฑฐ๊ฐ์ผ๋ฉด ๊ฐ์ ์ค์ฌ๋ณด์.
|
48 |
+
|
49 |
+
### LoRA config
|
50 |
+
lora_r = 16
|
51 |
+
lora_alpha = 16
|
52 |
+
lora_dropout = 0.05
|
53 |
+
lora_target_modules = ["gate_proj", "down_proj", "up_proj"]
|
54 |
+
|
55 |
+
### Tokenizer์์ ๋์ค๋ input๊ฐ ์ค์ ์ต์
|
56 |
+
train_on_inputs = False
|
57 |
+
add_eos_token = True
|
58 |
+
|
59 |
+
val_data = None
|
60 |
+
|
61 |
+
# Others
|
62 |
+
resume_from_checkpoint = False # !! ๋ง์ฝ ๋ชจ๋ธ์ ์ด์ด์ ํ๋ จํ๊ณ ์ถ๋ค๋ฉด, './custom_LLM/checkpoint-[xxx]'์ ๊ฐ์ด ํ์ผ ๊ฒฝ๋ก๋ฅผ ์
๋ ฅํด์ผ ํฉ๋๋ค!
|
63 |
+
output_dir = './custom_LLM'
|