{"id":7097,"date":"2024-08-12T16:59:34","date_gmt":"2024-08-12T08:59:34","guid":{"rendered":"https:\/\/aict.nkust.edu.tw\/digitrans\/?p=7097"},"modified":"2024-12-13T21:10:12","modified_gmt":"2024-12-13T13:10:12","slug":"apple-%e7%9a%84-dclm-7b%ef%bc%9a%e8%a8%ad%e5%ae%9a%e3%80%81%e4%bd%bf%e7%94%a8%e7%af%84%e4%be%8b%e3%80%81%e5%be%ae%e8%aa%bf","status":"publish","type":"post","link":"https:\/\/aict.nkust.edu.tw\/digitrans\/?p=7097","title":{"rendered":"Apple \u7684 DCLM-7B\uff1a\u8a2d\u5b9a\u3001\u4f7f\u7528\u7bc4\u4f8b\u3001\u5fae\u8abf"},"content":{"rendered":"\n

2024-08-12 | Dimitri Didmanidze<\/p>\n\n\n\n

<\/p>\n\n\n\n

DCLM-7B\uff1a\u4e3b\u8981\u7279\u6027\u8207\u529f\u80fd<\/h2>\n\n\n\n

Apple \u7684\u6700\u65b0\u8ca2\u737b\uff0cDCLM-7B\uff08\u8a9e\u8a00\u6a21\u578b\u8cc7\u6599\u8a08\u7b97\uff09\u57fa\u790e\u6a21\u578b\uff0c\u4f5c\u70ba LLM \u9818\u57df\u7684\u4e00\u500b\u503c\u5f97\u6ce8\u610f\u7684\u88dc\u5145\u800c\u812b\u7a4e\u800c\u51fa\u3002\u8b93\u6211\u5011\u63a2\u8a0e\u4e00\u4e0b\u5b83\u7684\u4e3b\u8981\u529f\u80fd\u3002<\/p>\n\n\n\n

Decoder-only Transformer<\/strong><\/h3>\n\n\n\n

DCLM-7B<\/strong> \u6a21\u578b\u63a1\u7528\u4e86 decoder-only Transformer<\/strong> \u67b6\u69cb\uff0c\u9019\u7a2e\u8a2d\u8a08\u65b9\u5f0f\u8b93\u6a21\u578b\u4e00\u6b21\u9810\u6e2c\u4e00\u500b\u8a5e\u5143\uff0c\u4e26\u5c07\u6bcf\u6b21\u751f\u6210\u7684\u8a5e\u5143\u56de\u994b\u5230\u6a21\u578b\u4e2d\uff0c\u4f86\u751f\u6210\u4e0b\u4e00\u500b\u8a5e\u5143\u3002<\/p>\n\n\n\n

\u9019\u7a2e\u67b6\u69cb\u7d93\u904e\u512a\u5316\uff0c\u80fd\u5920\u751f\u6210\u9023\u8cab\u4e14\u7b26\u5408\u8a9e\u5883\u7684\u6587\u672c\uff0c\u975e\u5e38\u9069\u5408\u7528\u65bc\u5404\u7a2e\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u4efb\u52d9\u3002\u9019\u4e5f\u662f\u76ee\u524d\u6700\u5148\u9032\u7684\u6a21\u578b\uff08\u5982 ChatGPT<\/strong> \u548c GPT-4<\/strong>\uff09\u6240\u63a1\u7528\u7684\u67b6\u69cb\uff0c\u5c55\u793a\u4e86\u5b83\u5728\u7406\u89e3\u548c\u751f\u6210\u985e\u4f3c\u4eba\u985e\u6587\u672c\u65b9\u9762\u7684\u9ad8\u6548\u6027\u3002<\/p>\n\n\n\n

\u958b\u6e90<\/h3>\n\n\n\n

DCLM-7B \u578b\u865f\u53ef\u6839\u64daApple \u7bc4\u4f8b\u7a0b\u5f0f\u78bc\u6388\u6b0a\u9032\u884c\u7814\u7a76\u548c\u958b\u767c\u3002\u9019\u7a2e\u958b\u6e90\u65b9\u6cd5\u9f13\u52f5\u4eba\u5de5\u667a\u6167\u793e\u7fa4\u5167\u7684\u5ee3\u6cdb\u4f7f\u7528\u548c\u5354\u4f5c\u3002<\/p>\n\n\n\n

\u900f\u904e\u4f7f\u8a72\u6a21\u578b\u6613\u65bc\u4f7f\u7528\uff0cApple \u652f\u63f4\u4eba\u5de5\u667a\u6167\u7684\u6c11\u4e3b\u5316\uff0c\u8b93\u4e16\u754c\u5404\u5730\u7684\u7814\u7a76\u4eba\u54e1\u548c\u958b\u767c\u4eba\u54e1\u80fd\u5920\u5728\u57fa\u672c\u6a21\u578b\u7684\u57fa\u790e\u4e0a\u9032\u884c\u8a66\u9a57\u548c\u5efa\u69cb\u3002<\/p>\n\n\n\n

70 \u5104\u500b\u53c3\u6578<\/h3>\n\n\n\n

DCLM-7B\u6a21\u578b\u64c1\u670970\u5104\u500b\u53c3\u6578\uff0c\u5728\u6548\u80fd\u548c\u904b\u7b97\u6548\u7387\u4e4b\u9593\u53d6\u5f97\u4e86\u5e73\u8861\u3002<\/p>\n\n\n\n

\u9019\u7a2e\u5c3a\u5bf8\u4f7f\u5f97\u8a72\u6a21\u578b\u53ef\u4ee5\u5728\u5927\u591a\u6578\u9ad8 RAM\/VRAM \u8a2d\u5099\u548c\u96f2\u7aef\u5e73\u53f0\u4e0a\u904b\u884c\uff0c\u5f9e\u800c\u4f7f\u5176\u5177\u6709\u591a\u529f\u80fd\u6027\u4e26\u53ef\u7528\u65bc\u5404\u7a2e\u61c9\u7528\u7a0b\u5f0f\u3002\u5927\u91cf\u7684\u53c3\u6578\u4f7f\u6a21\u578b\u80fd\u5920\u6355\u6349\u8907\u96dc\u7684\u8a9e\u8a00\u6a21\u5f0f\uff0c\u5f9e\u800c\u589e\u5f37\u5176\u57f7\u884c\u5404\u7a2e\u4efb\u52d9\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n

\u5728\u6d77\u91cf\u8cc7\u6599\u96c6\u4e0a\u9032\u884c\u8a13\u7df4<\/h3>\n\n\n\n

\u8a72\u6a21\u578b\u5df2\u7d93\u5728\u5305\u542b 2.5 \u5146\u500b\u6a19\u8a18\u7684\u5ee3\u6cdb\u8cc7\u6599\u96c6\u4e0a\u9032\u884c\u4e86\u8a13\u7df4\uff0c\u70ba\u8655\u7406\u5404\u7a2e\u8a9e\u8a00\u4efb\u52d9\u63d0\u4f9b\u4e86\u5805\u5be6\u7684\u57fa\u790e\u3002\u9019\u4f7f\u5f97 DCLM-7B \u6a21\u578b\u80fd\u5920\u7406\u89e3\u4e26\u7522\u751f\u5177\u6709\u9ad8\u5ea6\u6e96\u78ba\u6027\u548c\u76f8\u95dc\u6027\u7684\u6587\u5b57\u3002\u6b64\u5916\uff0c\u9019\u4f7f\u5f97\u8a72\u6a21\u578b\u6210\u70ba\u7279\u5b9a\u4efb\u52d9\u5fae\u8abf\u7684\u826f\u597d\u9078\u64c7\uff0c\u56e0\u70ba\u5b83\u5c0d\u82f1\u8a9e\u6709\u5f37\u5927\u7684\u57fa\u790e\u7406\u89e3\u3002<\/p>\n\n\n\n

\u4e0a\u4e0b\u6587\u8996\u7a97<\/h3>\n\n\n\n

\u57fa\u672c DCLM-7B \u6a21\u578b\u5177\u6709 2048 \u500b\u6a19\u8a18\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u9019\u4f7f\u5176\u80fd\u5920\u8655\u7406\u76f8\u5c0d\u8f03\u9577\u7684\u6587\u5b57\u5e8f\u5217\u3002\u5118\u7ba1\u4ee5\u4eca\u5929\u7684\u6a19\u6e96\u4f86\u770b\uff0c\u9019\u500b\u6578\u5b57\u76f8\u5c0d\u8f03\u5c0f\uff0c\u4f46\u860b\u679c\u4e5f\u767c\u5e03\u4e86\u5e36\u6709 8K \u4ee4\u724c\u4e0a\u4e0b\u6587\u8996\u7a97\u7684\u8b8a\u9ad4\u3002<\/p>\n\n\n\n

\u9019\u7a2e\u64f4\u5c55\u7684\u4e0a\u4e0b\u6587\u8996\u7a97\u70ba\u8655\u7406\u8f03\u9577\u7684\u8f38\u5165\u63d0\u4f9b\u4e86\u66f4\u5927\u7684\u9748\u6d3b\u6027\uff0c\u4f7f\u8a72\u6a21\u578b\u9069\u5408\u9700\u8981\u8655\u7406\u64f4\u5c55\u6587\u672c\u6216\u6587\u6a94\u7684\u61c9\u7528\u7a0b\u5e8f\uff0c\u4f8b\u5982\u6aa2\u7d22\u589e\u5f37\u751f\u6210\uff08RAG\uff09\u3002<\/p>\n\n\n

\n
\"\u4e3b\u8981\u7279\u6027\u548c\u529f\u80fd\"\/<\/figure>\n<\/div>\n\n\n

DCLM-7B \u5165\u9580<\/h2>\n\n\n\n

Apple \u5df2\u4f7f DCLM-7B \u578b\u865f\u8207Hugging Face\u7684transformers<\/code>\u7a0b\u5f0f\u5eab\u76f8\u5bb9\uff0c\u4f7f\u5176\u6613\u65bc\u5b58\u53d6\u548c\u4f7f\u7528\u3002<\/p>\n\n\n\n

\u60a8\u53ef\u4ee5\u5728Hugging Face\u4e0a\u627e\u5230\u8a72\u6a21\u578b\u7684\u7db2\u9801\uff0c\u4e26\u67e5\u770bGitHub \u5132\u5b58\u5eab\u4ee5\u7372\u53d6\u66f4\u591a\u8a73\u7d30\u8cc7\u8a0a\u3002\u8981\u4f7f\u7528\u548c\u5b58\u53d6\u6a21\u578b\uff0c\u6211\u5011\u9700\u8981\u5b89\u88dd\u8a72transformers<\/code>  \u5eab\uff1a<\/p>\n\n\n\n

pip install transformers<\/code><\/pre>\n\n\n\n
\u6b64\u5916\uff0c\u6211\u5011\u9084\u9700\u8981\u5b89\u88ddopen_lm<\/code>\u6846\u67b6\uff1a<\/p>\n\n\n\n
pip install git+https:\/\/github.com\/mlfoundations\/open_lm.git<\/code><\/pre>\n\n\n\n
\u5168\u7cbe\u5ea6 DCLM-7B \u8eca\u578b\u76f8\u7576\u5927\uff0c\u7d04 27.5GB\uff0c\u9700\u8981\u5927\u91cf RAM \u6216 VRAM \u624d\u80fd\u904b\u4f5c\u3002\u60a8\u5c07\u9700\u8981\u4e00\u53f0\u9ad8\u968e\u96fb\u8166\u6216\u67d0\u7a2e\u96f2\u7aef\u74b0\u5883\u3002\u6211\u5c07\u4f7f\u7528 Google Colab \u7684\u9ad8\u7d1a\u8a02\u95b1\u7b46\u8a18\u578b\u96fb\u8166\uff0c\u914d\u5099 50GB RAM \u548c L4 GPU\u3002<\/p>\n\n\n\n
\u5b89\u88dd\u4e86\u6240\u6709\u5fc5\u8981\u7684\u5eab\u5f8c\uff0c\u6211\u5011\u5c31\u53ef\u4ee5\u958b\u59cb\u4f7f\u7528\u8a72\u6a21\u578b\u4e86\uff01<\/p>\n\n\n\n
DCLM-7B\uff1a\u7528\u6cd5\u7bc4\u4f8b<\/h2>\n\n\n\n
\u4f8b\u5982\uff0c\u6211\u5c07\u904b\u884c\u6a21\u578b\u7684 Huggingface \u7db2\u9801\u4e0a\u63d0\u4f9b\u7684\u57fa\u672c\u7bc4\u4f8b\u3002\u9996\u5148\uff0c\u6211\u5011\u5c0e\u5165\u6240\u6709\u5fc5\u9700\u7684\u5eab\uff1a<\/p>\n\n\n\n
from open_lm.hf import *
from transformers import AutoTokenizer, AutoModelForCausalLM<\/code><\/pre>\n\n\n\n
\u7136\u5f8c\uff0c\u6211\u5011\u9700\u8981\u4e0b\u8f09\u4e26\u521d\u59cb\u5316\u5206\u8a5e\u5668\u548c\u6a21\u578b\uff08\u8acb\u6ce8\u610f\uff0c\u5728\u672c\u4f8b\u4e2d\uff0c\u6211\u5011\u5728 CPU \u4e0a\u4ee5\u5168\u7cbe\u5ea6\u6d6e\u9ede\u6578\u904b\u884c\u6a21\u578b\uff09\uff1a<\/p>\n\n\n\n
tokenizer = AutoTokenizer.from_pretrained(\"apple\/DCLM-Baseline-7B\")
model = AutoModelForCausalLM.from_pretrained(\"apple\/DCLM-Baseline-7B\")<\/code><\/pre>\n\n\n\n
\u6700\u5f8c\uff0c\u6211\u5011\u904b\u884c\u7bc4\u4f8b\u63d0\u793a\uff1a<\/p>\n\n\n\n
inputs = tokenizer([\"Machine learning is\"], return_tensors=\"pt\")
gen_kwargs = {\"max_new_tokens\": 50, \"top_p\": 0.8, \"temperature\": 0.8, \"do_sample\": True, \"repetition_penalty\": 1.1}
output = model.generate(inputs['input_ids'], **gen_kwargs)
output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True)<\/code><\/pre>\n\n\n\n
\u6211\u5f97\u5230\u4ee5\u4e0b\u8f38\u51fa\uff1a<\/p>\n\n\n\n
[Machine learning is not the solution to everything, it just enables you to solve a problem that otherwise would have been impossible. The biggest challenge for me as a manager of an AI team was to identify those problems where machine learning can really add value and be successful.]<\/code><\/pre>\n\n\n\n
\u9032\u968e\u4f7f\u7528\u8207\u5fae\u8abf<\/h2>\n\n\n\n
\u5fae\u8abf DCLM-7B \u6a21\u578b\u6709\u52a9\u65bc\u6839\u64da\u7279\u5b9a\u4efb\u52d9\u9032\u884c\u5b9a\u5236\uff0c\u5f9e\u800c\u589e\u5f37\u5176\u5728\u61c9\u7528\u4e2d\u7684\u6027\u80fd\u3002\u4e0d\u5e78\u7684\u662f\uff0cHuggingface \u7684\u51fd\u5f0f\u5eab\u4e0d\u652f\u63f4 DCLM-7B \u6a21\u578bpeft<\/code>\uff0c\u56e0\u6b64\u6211\u5011\u9700\u8981\u4f7f\u7528transformers<\/code>\u51fd\u5f0f\u5eab\u5c0d\u5176\u9032\u884c\u5fae\u8abf\u3002<\/p>\n\n\n\n
\u5982\u679c\u6c92\u6709LoRA\u9019\u6a23\u7684\u5de5\u5177\uff0c\u5fae\u8abf\u9019\u9ebc\u5927\u7684\u6a21\u578b\u9700\u8981\u5927\u91cf\u8cc7\u6e90\uff0c\u56e0\u70ba\u5b83\u57fa\u672c\u4e0a\u8207\u5f9e\u982d\u958b\u59cb\u8a13\u7df4\u5b83\u76f8\u540c\u3002\u56e0\u6b64\uff0c\u6211\u5c07\u5728\u9019\u88e1\u6982\u8ff0\u5fae\u8abf\u904e\u7a0b\uff0c\u800c\u4e0d\u6703\u5be6\u969b\u904b\u884c\u5b83\u4f86\u67e5\u770b\u7d50\u679c\u3002<\/p>\n\n\n\n
\u6e96\u5099\u8cc7\u6599\u96c6<\/h3>\n\n\n\n
\u8981\u4e0b\u8f09\u548c\u4f7f\u7528\u516c\u958b\u53ef\u7528\u7684\u8cc7\u6599\u96c6\uff0c\u6211\u5011\u5c07\u4f7f\u7528 Hugging Face \u7684datasets<\/code>\u5eab\u3002\u6211\u5011\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88dd\u5b83\uff1a<\/p>\n\n\n\n
pip install datasets<\/code><\/pre>\n\n\n\n
\u5b89\u88dd\u5f8c\uff0c\u6211\u5011\u532f\u5165\u4e26\u4f7f\u7528\u8a72load_dataset<\/code>\u51fd\u6578\uff0c\u5728\u672c\u4f8b\u4e2d\u6211\u5c07\u4f7f\u7528\u8cc7\u6599wikitext<\/code>\u96c6\uff1a<\/p>\n\n\n\n
from datasets import load_dataset
dataset = load_dataset('wikitext', 'wikitext-2-raw-v1')<\/code><\/pre>\n\n\n\n
\u73fe\u5728\uff0c\u6211\u5011\u9700\u8981\u5c0d\u8cc7\u6599\u96c6\u9032\u884c\u6a19\u8a18\uff1a<\/p>\n\n\n\n
def tokenize_function(examples):
\treturn tokenizer(examples['text'], padding='max_length', truncation=True, max_length=512)
tokenized_datasets = dataset.map(tokenize_function, batched=True)<\/code><\/pre>\n\n\n\n
\u73fe\u5728\uff0c\u6211\u5011\u6e96\u5099\u958b\u59cb\u5fae\u8abf\uff01<\/p>\n\n\n\n
\u5fae\u8abf<\/h3>\n\n\n\n
\u70ba\u4e86\u9032\u884c\u5fae\u8abf\uff0c\u6211\u5011\u9700\u8981\u532f\u5165\u4e26\u521d\u59cb\u5316\u5c0d\u8c61TrainingArguments<\/code>\uff0cTrainer<\/code>\u7136\u5f8c\u57f7\u884c\u8a72train()<\/code>\u51fd\u6578\u3002<\/p>\n\n\n\n
from transformers import TrainingArguments, Trainer
training_args = TrainingArguments(
\t    report_to = \"none\",
\t    output_dir=\".\/results\",
\t    evaluation_strategy=\"epoch\",
\t    learning_rate=2e-5,  # Controls how much to change the model weights during training
\t    per_device_train_batch_size=2,  # Number of samples per batch per device during training
\t    per_device_eval_batch_size=2,  # Number of samples per batch per device during evaluation
\t    num_train_epochs=3,  # Number of times the entire training dataset will be passed through the model
\t    weight_decay=0.01,  # Regularization technique to prevent overfitting
\t)
trainer = Trainer(
\t    model=model,
\t    args=training_args,
\t    train_dataset=tokenized_datasets['train'],
\t    eval_dataset=tokenized_datasets['test'],
\t    data_collator=data_collator,
\t    tokenizer=tokenizer,
\t)
trainer.train()<\/code><\/pre>\n\n\n\n
\u7d50\u8ad6<\/h2>\n\n\n\n
\u7e3d\u9ad4\u800c\u8a00\uff0cApple \u7684 DCLM-7B \u662f\u958b\u6e90\u8a9e\u8a00\u6a21\u578b\u9818\u57df\u7684\u91cd\u8981\u88dc\u5145\uff0c\u70ba\u7814\u7a76\u4eba\u54e1\u548c\u958b\u767c\u4eba\u54e1\u63d0\u4f9b\u4e86\u57f7\u884c\u5404\u7a2e NLP \u4efb\u52d9\u7684\u5f37\u5927\u5de5\u5177\u3002<\/p>\n\n\n\n
\u4f5c\u70ba\u50c5\u89e3\u78bc\u5668\u7684 Transformer \u6a21\u578b\uff0c\u5b83\u91dd\u5c0d\u6587\u5b57\u751f\u6210\u9032\u884c\u4e86\u6700\u4f73\u5316\uff0c\u63d0\u4f9b\u9023\u8cab\u4e14\u4e0a\u4e0b\u6587\u76f8\u95dc\u7684\u8f38\u51fa\u3002\u8a72\u6a21\u578b\u6839\u64da\u860b\u679c\u5b78\u8853\u8edf\u9ad4\u6388\u6b0a\u5354\u8b70\u63d0\u4f9b\uff0c\u9032\u4e00\u6b65\u9f13\u52f5\u4eba\u5de5\u667a\u6167\u793e\u7fa4\u7684\u5354\u4f5c\u548c\u5275\u65b0\u3002<\/p>\n\n\n\n
\u8cc7\u6599\u4f86\u6e90: https:\/\/www.datacamp.com\/tutorial\/apple-dclm-7b<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
2024-08-12 | Dimitri Didmanidze DCLM-7B\uff1a\u4e3b\u8981\u7279\u6027\u8207\u529f\u80fd Apple \u7684\u6700\u65b0\u8ca2\u737b\uff0cDCLM-7B\uff08\u8a9e\u8a00\u6a21\u578b\u8cc7\u6599\u8a08\u7b97\uff09\u57fa\u790e\u6a21\u578b\uff0c\u4f5c\u70ba LLM \u9818…<\/p>\n","protected":false},"author":4,"featured_media":7098,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[579,4],"tags":[40],"class_list":["post-7097","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-579","category-industry-news","tag-40"],"gutentor_comment":0,"jetpack_featured_media_url":"https:\/\/i0.wp.com\/aict.nkust.edu.tw\/digitrans\/wp-content\/uploads\/2024\/10\/%E8%9E%A2%E5%B9%95%E6%93%B7%E5%8F%96%E7%95%AB%E9%9D%A2-2024-10-04-170352.png?fit=688%2C199&ssl=1","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/posts\/7097","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7097"}],"version-history":[{"count":2,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/posts\/7097\/revisions"}],"predecessor-version":[{"id":7123,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/posts\/7097\/revisions\/7123"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=\/wp\/v2\/media\/7098"}],"wp:attachment":[{"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7097"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7097"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aict.nkust.edu.tw\/digitrans\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7097"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}