hirosan6595 commited on
Commit
e6e9f45
·
verified ·
1 Parent(s): d3e4d05

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +199 -0
README.md CHANGED
@@ -10,7 +10,206 @@ license: apache-2.0
10
  language:
11
  - en
12
  ---
 
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  # Uploaded model
15
 
16
  - **Developed by:** HiroSan6595
 
10
  language:
11
  - en
12
  ---
13
+ """python
14
 
15
+ !pip uninstall unsloth -y
16
+ !pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
17
+
18
+ !pip install --upgrade torch
19
+ !pip install --upgrade xformers
20
+
21
+ !pip install ipywidgets --upgrade
22
+
23
+ import torch
24
+ if torch.cuda.get_device_capability()[0] >= 8:
25
+ !pip install --no-deps packaging ninja einops "flash-attn>=2.6.3"
26
+
27
+ HF_TOKEN = "MY-TOKEN" #@param {type:"string"}
28
+
29
+ from unsloth import FastLanguageModel
30
+ import torch
31
+ max_seq_length = 1024 # 512->1024 unslothではRoPEをサポートしているのでコンテキスト長は自由に設定可能
32
+ dtype = None # Noneにしておけば自動で設定
33
+ load_in_4bit = True # 今回は13Bモデルを扱うためTrue
34
+
35
+ model_id = "llm-jp/llm-jp-3-13b"
36
+ new_model_id = "llm-jp-3-13b-it-f" #Fine-Tuningしたモデルにつけたい名前、it: Instruction Tuning
37
+
38
+ model, tokenizer = FastLanguageModel.from_pretrained(
39
+ model_name=model_id,
40
+ dtype=dtype,
41
+ load_in_4bit=load_in_4bit,
42
+ trust_remote_code=True,
43
+ )
44
+ model = FastLanguageModel.get_peft_model(
45
+ model,
46
+ r = 32,
47
+ target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
48
+ "gate_proj", "up_proj", "down_proj",],
49
+ lora_alpha = 32,
50
+ lora_dropout = 0.05,
51
+ bias = "none",
52
+ use_gradient_checkpointing = "unsloth",
53
+ random_state = 3407,
54
+ use_rslora = False,
55
+ loftq_config = None,
56
+ max_seq_length = max_seq_length,
57
+ )
58
+
59
+ from datasets import load_dataset, DatasetDict
60
+ dataset = load_dataset("DeL-TaiseiOzaki/Tengentoppa-sft-v1.0")
61
+ sampled_train = dataset["train"].shuffle(seed=42).select(range(5000))
62
+ dataset = DatasetDict({
63
+ "train": sampled_train
64
+ })
65
+ print(dataset)
66
+
67
+ prompt = """### 指示
68
+ {}
69
+ ### 回答
70
+ {}"""
71
+
72
+
73
+ """
74
+ formatting_prompts_func: 各データをプロンプトに合わせた形式に合わせる
75
+ """
76
+ EOS_TOKEN = tokenizer.eos_token # トークナイザーのEOSトークン(文末トークン)
77
+ def formatting_prompts_func(examples):
78
+ input = examples["instruction"] # 入力データ
79
+ output = examples["output"] # 出力データ
80
+ text = prompt.format(input, output) + EOS_TOKEN # プロンプトの作成
81
+ return { "formatted_text" : text, } # 新しいフィールド "formatted_text" を返す
82
+ pass
83
+
84
+ dataset = dataset.map(
85
+ formatting_prompts_func,
86
+ num_proc= 4, # 並列処理数を指定
87
+ )
88
+
89
+ dataset
90
+ print(dataset["train"]["formatted_text"][3501])
91
+
92
+ """
93
+ training_arguments: 学習の設定
94
+ - output_dir:
95
+ -トレーニング後のモデルを保存するディレクトリ
96
+ - per_device_train_batch_size:
97
+ - デバイスごとのトレーニングバッチサイズ
98
+ - per_device_eval_batch_size:
99
+ - デバイスごとの評価バッチサイズ
100
+ - gradient_accumulation_steps:
101
+ - 勾配を更新する前にステップを積み重ねる回数
102
+ - optim:
103
+ - オプティマイザの設定
104
+ - num_train_epochs:
105
+ - エポック数
106
+ - eval_strategy:
107
+ - 評価の戦略 ("no"/"steps"/"epoch")
108
+ - eval_steps:
109
+ - eval_strategyが"steps"のとき、評価を行うstep間隔
110
+ - logging_strategy:
111
+ - ログ記録の戦略
112
+ - logging_steps:
113
+ - ログを出力するステップ間隔
114
+ - warmup_steps:
115
+ - 学習率のウォームアップステップ数
116
+ - save_steps:
117
+ - モデルを保存するステップ間隔
118
+ - save_total_limit:
119
+ - 保存しておくcheckpointの数
120
+ - max_steps:
121
+ - トレーニングの最大ステップ数
122
+ - learning_rate:
123
+ - 学習率
124
+ - fp16:
125
+ - 16bit浮動小数点の使用設定(第8回演習を参考にすると良いです)
126
+ - bf16:
127
+ - BFloat16の使用設定
128
+ - group_by_length:
129
+ - 入力シーケンスの長さによりバッチをグループ化 (トレーニングの効率化)
130
+ - report_to:
131
+ - ログの送信先 ("wandb"/"tensorboard"など)
132
+ """
133
+ from trl import SFTTrainer
134
+ from transformers import TrainingArguments
135
+ from unsloth import is_bfloat16_supported
136
+
137
+ trainer = SFTTrainer(
138
+ model = model,
139
+ tokenizer = tokenizer,
140
+ train_dataset=dataset["train"],
141
+ max_seq_length = max_seq_length,
142
+ dataset_text_field="formatted_text",
143
+ packing = False,
144
+ args = TrainingArguments(
145
+ per_device_train_batch_size = 2, #
146
+ gradient_accumulation_steps = 4, #
147
+ num_train_epochs = 1, #
148
+ logging_steps = 10,
149
+ warmup_steps = 10,
150
+ save_steps=100,
151
+ save_total_limit=2,
152
+ max_steps=-1,
153
+ learning_rate = 2e-4,
154
+ fp16 = not is_bfloat16_supported(),
155
+ bf16 = is_bfloat16_supported(),
156
+ group_by_length=True,
157
+ seed = 3407,
158
+ output_dir = "outputs",
159
+ report_to = "none",
160
+ ),
161
+ )
162
+
163
+ #@title 現在のメモリ使用量を表示
164
+ gpu_stats = torch.cuda.get_device_properties(0)
165
+ start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
166
+ max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
167
+ print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
168
+ print(f"{start_gpu_memory} GB of memory reserved.")
169
+
170
+ #@title 学習実行
171
+ trainer_stats = trainer.train()
172
+ import json
173
+ datasets = []
174
+ with open("/content/drive/MyDrive/Student_LLM/05FinalReport/elyza-tasks-100-TV_0.jsonl", "r") as f:
175
+ item = ""
176
+ for line in f:
177
+ line = line.strip()
178
+ item += line
179
+ if item.endswith("}"):
180
+ datasets.append(json.loads(item))
181
+ item = ""
182
+ from tqdm import tqdm
183
+ FastLanguageModel.for_inference(model)
184
+
185
+ results = []
186
+ for dt in tqdm(datasets):
187
+ input = dt["input"]
188
+
189
+ prompt = f"""### 指示\n{input}\n### 回答\n"""
190
+
191
+ inputs = tokenizer([prompt], return_tensors = "pt").to(model.device)
192
+
193
+ outputs = model.generate(**inputs, max_new_tokens = 512, use_cache = True, do_sample=False, repetition_penalty=1.2)
194
+ prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
195
+
196
+ results.append({"task_id": dt["task_id"], "input": input, "output": prediction})
197
+
198
+ with open(f"{new_model_id}_output.jsonl", 'w', encoding='utf-8') as f:
199
+ for result in results:
200
+ json.dump(result, f, ensure_ascii=False)
201
+ f.write('\n')
202
+
203
+ https://docs.unsloth.ai/basics/saving-and-using-models
204
+
205
+ model.push_to_hub_merged(
206
+ new_model_id+"_lora",
207
+ tokenizer=tokenizer,
208
+ save_method="lora",
209
+ token=HF_TOKEN,
210
+ private=True
211
+ )
212
+ """
213
  # Uploaded model
214
 
215
  - **Developed by:** HiroSan6595