Update README.md
Browse files
README.md
CHANGED
@@ -6,7 +6,6 @@ tags:
|
|
6 |
- unsloth
|
7 |
- llama
|
8 |
- trl
|
9 |
-
license: apache-2.0
|
10 |
language:
|
11 |
- en
|
12 |
- ja
|
@@ -24,27 +23,27 @@ This llama model was trained 2x faster with [Unsloth](https://github.com/unsloth
|
|
24 |
|
25 |
# 推論用コード
|
26 |
# Hugging Faceにアップロードしたモデルを用いてELYZA-tasks-100-TVの出力を得るためのコードです。 このコードで生成されたjsonlファイルは課題の成果として提出可能なフォーマットになっております。
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
|
29 |
-
!pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
|
30 |
-
!pip install --upgrade torch
|
31 |
-
!pip install --upgrade xformers
|
32 |
-
|
33 |
-
# Install Flash Attention 2 for softcapping support
|
34 |
-
import torch
|
35 |
-
if torch.cuda.get_device_capability()[0] >= 8:
|
36 |
-
!pip install --no-deps packaging ninja einops "flash-attn>=2.6.3"
|
37 |
-
|
38 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
39 |
from unsloth import FastLanguageModel
|
|
|
40 |
import torch
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
load_in_4bit = True
|
45 |
|
46 |
model_id = "llm-jp/llm-jp-3-13b"
|
47 |
-
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
model, tokenizer = FastLanguageModel.from_pretrained(
|
50 |
model_name=model_id,
|
@@ -53,203 +52,44 @@ model, tokenizer = FastLanguageModel.from_pretrained(
|
|
53 |
trust_remote_code=True,
|
54 |
)
|
55 |
|
56 |
-
|
57 |
-
model = FastLanguageModel.get_peft_model(
|
58 |
-
model,
|
59 |
-
r=32,
|
60 |
-
target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
|
61 |
-
"gate_proj", "up_proj", "down_proj"],
|
62 |
-
lora_alpha=32,
|
63 |
-
lora_dropout=0.05,
|
64 |
-
bias="none",
|
65 |
-
use_gradient_checkpointing="unsloth",
|
66 |
-
random_state=3407,
|
67 |
-
use_rslora=False,
|
68 |
-
loftq_config=None,
|
69 |
-
max_seq_length=max_seq_length,
|
70 |
-
)
|
71 |
-
|
72 |
-
# https://huggingface.co/settings/tokens
|
73 |
-
HF_TOKEN = "your-token" # @param {type:"string"}
|
74 |
-
|
75 |
-
from datasets import load_dataset, concatenate_datasets
|
76 |
-
|
77 |
-
# データセットのロード
|
78 |
-
ichikara_dataset = load_dataset("json", data_files="/content/ichikara-instruction-003-001-1.json")
|
79 |
-
elyza_dataset = load_dataset("elyza/ELYZA-tasks-100")
|
80 |
-
|
81 |
-
EOS_TOKEN = tokenizer.eos_token #
|
82 |
-
|
83 |
-
# 学習時のプロンプトフォーマットの定義
|
84 |
-
prompt = """### 指示
|
85 |
-
{}
|
86 |
-
### 回答
|
87 |
-
{}"""
|
88 |
-
|
89 |
-
"""
|
90 |
-
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
|
91 |
-
"""
|
92 |
-
def formatting_prompts_func(examples):
|
93 |
-
input = examples["text"]
|
94 |
-
output = examples["output"]
|
95 |
-
text = prompt.format(input, output) + EOS_TOKEN
|
96 |
-
return {"formatted_text": text}
|
97 |
-
|
98 |
-
# ichikara-instruction のデータフォーマット
|
99 |
-
ichikara_dataset = ichikara_dataset.map(
|
100 |
-
formatting_prompts_func,
|
101 |
-
num_proc=4,
|
102 |
-
)
|
103 |
-
|
104 |
-
# ELYZA-tasks-100 データセットのフォーマット関数
|
105 |
-
def elyza_formatting_prompts_func(examples):
|
106 |
-
input = examples["input"]
|
107 |
-
output = examples["output"]
|
108 |
-
text = prompt.format(input, output) + EOS_TOKEN
|
109 |
-
return {"formatted_text": text}
|
110 |
-
|
111 |
-
# ELYZA-tasks-100 のデータフォーマット
|
112 |
-
elyza_dataset = elyza_dataset.map(
|
113 |
-
elyza_formatting_prompts_func,
|
114 |
-
num_proc=4
|
115 |
-
)
|
116 |
-
|
117 |
-
from datasets import concatenate_datasets
|
118 |
-
|
119 |
-
# ichikara-instruction と ELYZA-tasks-100 を統合
|
120 |
-
combined_dataset = concatenate_datasets([
|
121 |
-
ichikara_dataset["train"],
|
122 |
-
elyza_dataset["test"]
|
123 |
-
])
|
124 |
-
|
125 |
-
# データ品質チェック
|
126 |
-
# 1. ランダムサンプルを確認
|
127 |
-
import random
|
128 |
-
sample_indices = random.sample(range(len(combined_dataset)), 10)
|
129 |
-
for idx in sample_indices:
|
130 |
-
print(combined_dataset[idx]["formatted_text"])
|
131 |
-
|
132 |
-
# 2. 自動検査ルール
|
133 |
-
# 短すぎるデータをチェック(Noneチェックを追加)
|
134 |
-
short_data = combined_dataset.filter(
|
135 |
-
lambda x: x["input"] is not None and x["output"] is not None and (len(x["input"]) < 5 or len(x["output"]) < 5)
|
136 |
-
)
|
137 |
-
print(f"\n短すぎるデータ数: {len(short_data)}")
|
138 |
-
|
139 |
-
# 指示と回答が同一のデータ(Noneチェックを追加)
|
140 |
-
duplicate_data = combined_dataset.filter(
|
141 |
-
lambda x: x["input"] is not None and x["output"] is not None and x["input"].strip() == x["output"].strip()
|
142 |
-
)
|
143 |
-
print(f"\n指示と回答が同一のデータ数: {len(duplicate_data)}")
|
144 |
-
|
145 |
-
# 問題のあるデータをフィルタリング(Noneチェックを追加)
|
146 |
-
filtered_dataset = combined_dataset.filter(
|
147 |
-
lambda x: x["input"] is not None and x["output"] is not None and len(x["input"]) > 5 and len(x["output"]) > 5 and x["input"].strip() != x["output"].strip()
|
148 |
-
)
|
149 |
-
|
150 |
-
print(f"元のデータ数: {len(combined_dataset)}")
|
151 |
-
print(f"フィルタリング後のデータ数: {len(filtered_dataset)}")
|
152 |
-
print(f"除外されたデータ数: {len(combined_dataset) - len(filtered_dataset)}")
|
153 |
-
|
154 |
-
# フィルタリング後のデータの例を確認
|
155 |
-
print(filtered_dataset[0])
|
156 |
-
|
157 |
-
"""
|
158 |
-
training_arguments: 学習の設定
|
159 |
-
"""
|
160 |
-
from trl import SFTTrainer
|
161 |
-
from transformers import TrainingArguments
|
162 |
-
from unsloth import is_bfloat16_supported
|
163 |
-
|
164 |
-
trainer = SFTTrainer(
|
165 |
-
model=model,
|
166 |
-
tokenizer=tokenizer,
|
167 |
-
train_dataset=filtered_dataset,
|
168 |
-
max_seq_length=max_seq_length,
|
169 |
-
dataset_text_field="formatted_text",
|
170 |
-
packing=False,
|
171 |
-
args=TrainingArguments(
|
172 |
-
per_device_train_batch_size=2,
|
173 |
-
gradient_accumulation_steps=4,
|
174 |
-
num_train_epochs=3,
|
175 |
-
logging_steps=10,
|
176 |
-
warmup_steps=10,
|
177 |
-
save_steps=50,
|
178 |
-
save_total_limit=2,
|
179 |
-
max_steps=200,
|
180 |
-
learning_rate=2e-4,
|
181 |
-
fp16=not is_bfloat16_supported(),
|
182 |
-
bf16=is_bfloat16_supported(),
|
183 |
-
group_by_length=True,
|
184 |
-
seed=3407,
|
185 |
-
output_dir="outputs",
|
186 |
-
report_to="none",
|
187 |
-
),
|
188 |
-
)
|
189 |
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
import json
|
194 |
-
from datasets import load_dataset
|
195 |
-
|
196 |
-
dataset = load_dataset("json", data_files="/content/elyza-tasks-100-TV_0.jsonl", split="train")
|
197 |
-
|
198 |
-
datasets = []
|
199 |
-
with open("/content/elyza-tasks-100-TV_0.jsonl", "r", encoding="utf-8") as f:
|
200 |
item = ""
|
201 |
for line in f:
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
|
208 |
-
|
209 |
-
import json
|
210 |
-
|
211 |
-
# 推論するためにモデルのモードを変更
|
212 |
FastLanguageModel.for_inference(model)
|
213 |
|
214 |
results = []
|
215 |
for dt in tqdm(datasets):
|
216 |
-
|
217 |
-
input_text = dt["input"]
|
218 |
-
|
219 |
-
# プロンプトを生成
|
220 |
-
prompt = f"### 指示\n{input_text}\n次の要件を満たしてください:\n1. 簡潔に回答する。\n2. 必要なら箇条書きを使用して要点を整理する。\n3. 指示された内容に忠実に答える。\n### 回答\n"
|
221 |
-
|
222 |
|
223 |
-
|
224 |
-
inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
|
225 |
|
226 |
-
|
227 |
-
outputs = model.generate(
|
228 |
-
**inputs,
|
229 |
-
max_new_tokens=512,
|
230 |
-
use_cache=True,
|
231 |
-
do_sample=False,
|
232 |
-
repetition_penalty=1.2,
|
233 |
-
)
|
234 |
-
prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
|
235 |
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
241 |
|
|
|
242 |
|
243 |
-
#
|
244 |
-
|
245 |
-
with open(
|
246 |
for result in results:
|
247 |
-
|
248 |
-
|
249 |
-
model.push_to_hub_merged(
|
250 |
-
new_model_id,
|
251 |
-
tokenizer=tokenizer,
|
252 |
-
save_method="lora",
|
253 |
-
token=HF_TOKEN,
|
254 |
-
private=True
|
255 |
-
)
|
|
|
6 |
- unsloth
|
7 |
- llama
|
8 |
- trl
|
|
|
9 |
language:
|
10 |
- en
|
11 |
- ja
|
|
|
23 |
|
24 |
# 推論用コード
|
25 |
# Hugging Faceにアップロードしたモデルを用いてELYZA-tasks-100-TVの出力を得るためのコードです。 このコードで生成されたjsonlファイルは課題の成果として提出可能なフォーマットになっております。
|
26 |
+
# セットアップ
|
27 |
+
!pip install unsloth
|
28 |
+
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
|
29 |
+
!pip install -U torch
|
30 |
+
!pip install -U peft
|
31 |
|
32 |
+
# モデル・トークナイザの読み込み
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
from unsloth import FastLanguageModel
|
34 |
+
from peft import PeftModel
|
35 |
import torch
|
36 |
+
import json
|
37 |
+
from tqdm import tqdm
|
38 |
+
import re
|
|
|
39 |
|
40 |
model_id = "llm-jp/llm-jp-3-13b"
|
41 |
+
adapter_id = "Gamoooo/llm-jp-3-13b-last"
|
42 |
+
|
43 |
+
HF_TOKEN = "your-token" #@param {type:"string"}
|
44 |
+
|
45 |
+
dtype = None
|
46 |
+
load_in_4bit = True
|
47 |
|
48 |
model, tokenizer = FastLanguageModel.from_pretrained(
|
49 |
model_name=model_id,
|
|
|
52 |
trust_remote_code=True,
|
53 |
)
|
54 |
|
55 |
+
model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
+
# 入力データの準備
|
58 |
+
datasets = []
|
59 |
+
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
item = ""
|
61 |
for line in f:
|
62 |
+
line = line.strip()
|
63 |
+
item += line
|
64 |
+
if item.endswith("}"):
|
65 |
+
datasets.append(json.loads(item))
|
66 |
+
item = ""
|
67 |
|
68 |
+
# 推論実行
|
|
|
|
|
|
|
69 |
FastLanguageModel.for_inference(model)
|
70 |
|
71 |
results = []
|
72 |
for dt in tqdm(datasets):
|
73 |
+
input = dt["input"]
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
+
prompt = f"### 指示\n{input}\n次の要件を満たしてください:\n1. 簡潔に回答する。\n2. 必要なら箇条書きを使用して要点を整理する。\n3. 指示された内容に忠実に答える。\n### 回答\n"
|
|
|
76 |
|
77 |
+
inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
+
outputs = model.generate(
|
80 |
+
**inputs,
|
81 |
+
max_new_tokens=512,
|
82 |
+
use_cache=True,
|
83 |
+
do_sample=False,
|
84 |
+
repetition_penalty=1.2,
|
85 |
+
)
|
86 |
+
prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
|
87 |
|
88 |
+
results.append({"task_id": dt["task_id"], "input": input, "output": prediction})
|
89 |
|
90 |
+
# 出力の保存
|
91 |
+
json_file_id = re.sub(".*/", "", adapter_id)
|
92 |
+
with open(f"/content/{json_file_id}_output.jsonl", 'w', encoding='utf-8') as f:
|
93 |
for result in results:
|
94 |
+
json.dump(result, f, ensure_ascii=False)
|
95 |
+
f.write('\n')
|
|
|
|
|
|
|
|
|
|
|
|
|
|