File size: 2,124 Bytes
3fbabe1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5704d69
36812a5
3fbabe1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
---
library_name: transformers
license: apache-2.0
basemodel: meta-llama/Meta-Llama-3-8B-Instruct
datasets:
- Saxo/total_ko_train_set_1_with_wiki_with_orca
language:
- ko
- en
pipeline_tag: text-generation
---

# Model Card for Model ID

<div align="center">
<img src="https://www.linkbricks.com/wp-content/uploads/2022/03/%E1%84%85%E1%85%B5%E1%86%BC%E1%84%8F%E1%85%B3%E1%84%87%E1%85%B3%E1%84%85%E1%85%B5%E1%86%A8%E1%84%89%E1%85%B3%E1%84%85%E1%85%A9%E1%84%80%E1%85%A9-2-1024x804.png" />
</div>


AI ์™€ ๋น…๋ฐ์ดํ„ฐ ๋ถ„์„ ์ „๋ฌธ ๊ธฐ์—…์ธ Linkbricks์˜ ๋ฐ์ดํ„ฐ์‚ฌ์ด์–ธํ‹ฐ์ŠคํŠธ์ธ ์ง€์œค์„ฑ ๋ฐ•์‚ฌ(Saxo)๊ฐ€ meta-llama/Meta-Llama-3-8B๋ฅผ ๋ฒ ์ด์Šค๋ชจ๋ธ๋กœ GCP์ƒ์˜ H100-80G 8๊ฐœ๋ฅผ ํ†ตํ•ด SFT-DPO ํ›ˆ๋ จ์„ ํ•œ(8000 Tokens) ํ•œ๊ธ€ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ.
ํ† ํฌ๋‚˜์ด์ €๋Š” ๋ผ๋งˆ3๋ž‘ ๋™์ผํ•˜๋ฉฐ ํ•œ๊ธ€ VOCA ํ™•์žฅ์€ ํ•˜์ง€ ์•Š์€ ๋ฒ„์ „ ์ž…๋‹ˆ๋‹ค.

Dr. Yunsung Ji (Saxo), a data scientist at Linkbricks, a company specializing in AI and big data analytics, trained the meta-llama/Meta-Llama-3-8B base model on 8 H100-60Gs on GCP for 4 hours of instructional training (8000 Tokens).
Accelerate, Deepspeed Zero-3 libraries were used. 

www.linkbricks.com, www.linkbricks.vc

## Configuration including BitsandBytes
---
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=False,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch_dtype
)


args = TrainingArguments(
    output_dir=project_name,
    run_name=run_name_str,
    overwrite_output_dir=True,
    num_train_epochs=20,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4, #1
    gradient_checkpointing=True,
    optim="paged_adamw_32bit",
    #optim="adamw_8bit",
    logging_steps=10,
    save_steps=100,
    save_strategy="epoch",
    learning_rate=2e-4, #2e-4
    weight_decay=0.01,
    max_grad_norm=1, #0.3
    max_steps=-1,
    warmup_ratio=0.1,
    group_by_length=False,
    fp16 = not torch.cuda.is_bf16_supported(),
    bf16 = torch.cuda.is_bf16_supported(),
    #fp16 = True, 
    lr_scheduler_type="cosine", #"constant",
    disable_tqdm=False,
    report_to='wandb',
    push_to_hub=False
)