Dongwookss commited on
Commit
2deac7b
Β·
verified Β·
1 Parent(s): 7e6efcb

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +71 -2
README.md CHANGED
@@ -27,8 +27,77 @@ tags:
27
  - **SFT Trainer**λ₯Ό 톡해 ν›ˆλ ¨μ„ μ§„ν–‰
28
  - ν™œμš© 데이터
29
  - [q_a_korean_futsal](https://huggingface.co/datasets/Dongwookss/q_a_korean_futsal)
30
- - 말투 ν•™μŠ΅μ„ μœ„ν•΄ 'ν•΄μš”'체둜 λ³€ν™˜ν•˜κ³  λ‹΅λ³€ μ’…λ£Œμ‹œ 'μ–Όλ§ˆλ“ μ§€ λ¬Όμ–΄λ³΄μ„Έμš”~! ν’‹ν’‹~!'을 λ„£μ–΄ λͺ¨λΈ 컨셉을 μœ μ§€ν•˜μ˜€μŠ΅λ‹ˆλ‹€.
31
-
 
32
 
 
33
 
 
 
 
 
34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
27
  - **SFT Trainer**λ₯Ό 톡해 ν›ˆλ ¨μ„ μ§„ν–‰
28
  - ν™œμš© 데이터
29
  - [q_a_korean_futsal](https://huggingface.co/datasets/Dongwookss/q_a_korean_futsal)
30
+ - 말투 ν•™μŠ΅μ„ μœ„ν•΄ 'ν•΄μš”'체둜 λ³€ν™˜ν•˜κ³  인삿말을 λ„£μ–΄ λͺ¨λΈ 컨셉을 μœ μ§€ν•˜μ˜€μŠ΅λ‹ˆλ‹€.
31
+
32
+ - **Environment** : Colab ν™˜κ²½μ—μ„œ μ§„ν–‰ν•˜μ˜€μœΌλ©° L4 GPUλ₯Ό μ‚¬μš©ν•˜μ˜€μŠ΅λ‹ˆλ‹€.
33
 
34
+ ```python
35
 
36
+ from unsloth import FastLanguageModel
37
+ import torch
38
+ from trl import SFTTrainer
39
+ from transformers import TrainingArguments
40
 
41
+ max_seq_length = 256
42
+ dtype = None
43
+ load_in_4bit = False
44
+ model, tokenizer = FastLanguageModel.from_pretrained(
45
+ model_name="HuggingFaceH4/zephyr-7b-beta",
46
+ max_seq_length=max_seq_length,
47
+ dtype=dtype,
48
+ load_in_4bit=load_in_4bit,
49
+ #token = ,
50
+ )
51
+
52
+ model = FastLanguageModel.get_peft_model(
53
+ model,
54
+ r=32,
55
+ lora_alpha=64,
56
+ lora_dropout=0.05,
57
+ target_modules=[
58
+ "q_proj",
59
+ "k_proj",
60
+ "v_proj",
61
+ "o_proj",
62
+ "gate_proj",
63
+ "up_proj",
64
+ "down_proj",
65
+ ], # νƒ€κ²Ÿ λͺ¨λ“ˆ
66
+ bias="none",
67
+ use_gradient_checkpointing="unsloth",
68
+ random_state=123,
69
+ use_rslora=False,
70
+ loftq_config=None,
71
+ )
72
+
73
+ tokenizer.padding_side = "right"
74
+
75
+ trainer = SFTTrainer(
76
+ model=model,
77
+ tokenizer=tokenizer,
78
+ train_dataset=dataset,
79
+ dataset_text_field="text",
80
+ max_seq_length=max_seq_length,
81
+ dataset_num_proc=2,
82
+ packing=False,
83
+ args=TrainingArguments(
84
+ per_device_train_batch_size=20,
85
+ gradient_accumulation_steps=2,
86
+ warmup_steps=5,
87
+ num_train_epochs=3,
88
+ max_steps = 1761,
89
+ logging_steps = 10,
90
+ learning_rate=2e-5,
91
+ fp16=not torch.cuda.is_bf16_supported(),
92
+ bf16=torch.cuda.is_bf16_supported(),
93
+ optim="adamw_8bit",
94
+ weight_decay=0.01,
95
+ lr_scheduler_type="cosine",
96
+ seed=123,
97
+ output_dir="outputs",
98
+ ),
99
+ )
100
+
101
+ trainer.train()
102
+
103
+ ```