{"id":61,"date":"2026-07-18T12:54:21","date_gmt":"2026-07-18T12:54:21","guid":{"rendered":"https:\/\/www.52elong.com\/?p=61"},"modified":"2026-07-18T12:55:50","modified_gmt":"2026-07-18T12:55:50","slug":"training-the-1000-questions-model-on-googles-colab","status":"publish","type":"post","link":"https:\/\/www.52elong.com\/index.php\/2026\/07\/18\/training-the-1000-questions-model-on-googles-colab\/","title":{"rendered":"Training the Qwen model on Google’s Colab"},"content":{"rendered":"\n

1. \u5f3a\u5236\u5347\u7ea7\u51b2\u7a81\u7684\u5e95\u5c42\u4f9d\u8d56\uff0c\u5e76\u5b89\u88c5\u5fc5\u5907\u5e93<\/h1>\n\n\n\n

!pip install –upgrade torchao peft -q
!pip install transformers datasets accelerate tiktoken -q<\/p>\n\n\n\n

import
os
import
torch
from transformers import
AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from transformers import
DataCollatorForLanguageModeling
from datasets import
load_dataset
from peft import
LoraConfig, get_peft_model, TaskType<\/p>\n\n\n\n

2. \u68c0\u67e5\u8bbe\u5907<\/h1>\n\n\n\n

device =
“cuda” if torch.cuda.is_available() else “cpu”
print(
f”\u5f53\u524d\u4f7f\u7528\u7684\u8ba1\u7b97\u8bbe\u5907\u662f: {device}”
)<\/p>\n\n\n\n

3. \u52a0\u8f7d\u5b98\u65b9\u6a21\u578b\u4e0e\u5206\u8bcd\u5668<\/h1>\n\n\n\n

model_name =
“Qwen\/Qwen2.5-1.5B-Instruct”
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=
True
)
base_model = AutoModelForCausalLM.from_pretrained(
model_name,
trust_remote_code=
True
,
device_map=
“auto”
)<\/p>\n\n\n\n

4. \u914d\u7f6e LoRA \u53c2\u6570<\/h1>\n\n\n\n

peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
inference_mode=
False
,
r=
8, # LoRA \u7684\u79e9
lora_alpha=
32, # \u7f29\u653e\u7cfb\u6570
lora_dropout=
0.1
,
target_modules=[
“q_proj”, “v_proj”] # \u9488\u5bf9\u5343\u95ee\u6a21\u578b\u7684\u6ce8\u610f\u529b\u673a\u5236\u5173\u952e\u5c42\u8fdb\u884c\u5fae\u8c03
)<\/p>\n\n\n\n

\ud83d\udca1 \u8fd9\u6b21\u4f9d\u8d56\u66f4\u65b0\u540e\uff0c\u5305\u88c5 LoRA \u6a21\u578b\u5c06\u4e0d\u518d\u62a5\u9519<\/h1>\n\n\n\n

model = get_peft_model(base_model, peft_config)
print(
“\u53ef\u8bad\u7ec3\u7684\u6a21\u578b\u53c2\u6570\u5360\u6bd4\u8be6\u60c5\uff1a”
)
model.print_trainable_parameters()<\/p>\n\n\n\n

5. \u51c6\u5907\u6570\u636e\u96c6<\/h1>\n\n\n\n

if not os.path.exists(“my_data.txt”
):
with open(“my_data.txt”, “w”, encoding=”utf-8″) as
f:
knowledge = [
“\u4eba\u5de5\u667a\u80fd\u662f\u5f15\u9886\u672a\u6765\u7684\u6218\u7565\u6027\u6280\u672f\uff0c\u6b63\u5728\u6df1\u523b\u6539\u53d8\u4eba\u7c7b\u793e\u4f1a\u7684\u751f\u4ea7\u4e0e\u751f\u6d3b\u65b9\u5f0f\u3002\\n”
,
“\u5927\u8bed\u8a00\u6a21\u578b\u901a\u8fc7\u5b66\u4e60\u6d77\u91cf\u6587\u672c\uff0c\u638c\u63e1\u4e86\u4eba\u7c7b\u8bed\u8a00\u7684\u8bed\u6cd5\u3001\u903b\u8f91\u548c\u4e30\u5bcc\u7684\u767e\u79d1\u77e5\u8bc6\u3002\\n”
,
“\u963f\u91cc\u5343\u95ee\uff08Qwen\uff09\u662f\u76ee\u524d\u5f00\u6e90\u793e\u533a\u4e2d\u5bf9\u4e2d\u6587\u652f\u6301\u6700\u597d\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e4b\u4e00\u3002\\n”
,
“\u5728\u672c\u5730\u6216\u8005\u4e91\u7aef\u5fae\u8c03\u5c0f\u53c2\u6570\u6a21\u578b\uff0c\u662f\u9ad8\u6027\u4ef7\u6bd4\u843d\u5730\u5782\u76f4\u9886\u57dfAI\u52a9\u624b\u7684\u6700\u4f73\u65b9\u6848\u3002\\n”
]
f.write(
“”.join(knowledge * 50
))<\/p>\n\n\n\n

dataset = load_dataset(
“text”, data_files={“train”: “my_data.txt”
})<\/p>\n\n\n\n

def tokenize_function(examples):
texts = []
for text in examples[“text”
]:
if len(text.strip()) == 0
:
continue
messages = [
{
“role”: “system”, “content”: “\u4f60\u662f\u4e00\u4e2a\u535a\u5b66\u3001\u4e13\u4e1a\u7684\u4e2d\u6587\u4eba\u5de5\u667a\u80fd\u52a9\u624b\u3002”
},
{
“role”: “user”, “content”: “\u8bf7\u544a\u8bc9\u6211\u5173\u4e8e\u5927\u6a21\u578b\u6216\u4eba\u5de5\u667a\u80fd\u7684\u884c\u4e1a\u77e5\u8bc6\u3002”
},
{
“role”: “assistant”, “content”
: text.strip()}
]
formatted_text = tokenizer.apply_chat_template(messages, tokenize=
False, add_generation_prompt=False
)
texts.append(formatted_text)<\/p>\n\n\n\n

return tokenizer(texts, truncation=True, max_length=512<\/code><\/pre>\n\n\n\n
)<\/p>\n\n\n\n
tokenized_datasets = dataset.
map(tokenize_function, batched=True, remove_columns=[“text”
])
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=
False
)<\/p>\n\n\n\n
6. \u914d\u7f6e\u7a33\u5065\u7684\u5fae\u8c03\u53c2\u6570\uff08\u4e0d\u5f00\u542f fp16 \u907f\u5f00 T4 \u7684\u6df7\u5408\u7cbe\u5ea6\u65ad\u8a00\u9519\u8bef\uff09<\/h1>\n\n\n\n
training_args = TrainingArguments(
output_dir=
“.\/qwen_lora_output”
,
num_train_epochs=
3
,
per_device_train_batch_size=
4
,
gradient_accumulation_steps=
4
,
save_steps=
100
,
logging_steps=
20
,
learning_rate=
1e-4
,
weight_decay=
0.01
,
fp16=
False
,
optim=
“adamw_torch”
,
report_to=
“none”
)<\/p>\n\n\n\n
7. \u542f\u52a8\u5fae\u8c03<\/h1>\n\n\n\n
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=tokenized_datasets[
“train”
],
)<\/p>\n\n\n\n
print(
“\ud83d\ude80 \u4f9d\u8d56\u4fee\u590d\u5b8c\u6bd5\uff01\u5f00\u59cb\u901a\u8fc7 LoRA \u5728 T4 GPU \u4e0a\u7a33\u5065\u5fae\u8c03\u963f\u91cc\u5343\u95ee\u2026”
)
trainer.train()<\/p>\n\n\n\n
8. \u4fdd\u5b58 LoRA \u6743\u91cd\u4e0e\u5206\u8bcd\u5668<\/h1>\n\n\n\n
trainer.save_model(
“.\/my_tuned_qwen”
)
tokenizer.save_pretrained(
“.\/my_tuned_qwen”
)
print(
“\ud83c\udf89 \u606d\u559c\uff01\u5343\u95ee\u5927\u6a21\u578b\u5df2\u6210\u529f\u8de8\u8d8a\u6240\u6709\u73af\u5883\u6697\u5751\uff0c\u5b8c\u6210\u5fae\u8c03\u5e76\u6210\u529f\u4fdd\u5b58\uff01”
)<\/p>\n\n\n\n
\u52a0\u8f7d\u5e76\u6d4b\u8bd5\u5fae\u8c03\u540e\u7684\u6a21\u578b\uff1f<\/h1>\n\n\n\n
import
torch
from transformers import
AutoModelForCausalLM, AutoTokenizer
from peft import
PeftModel<\/p>\n\n\n\n
model_name =
“Qwen\/Qwen2.5-1.5B-Instruct”
lora_path =
“.\/my_tuned_qwen” # \u4f60\u521a\u521a\u4fdd\u5b58\u7684\u5fae\u8c03\u6743\u91cd\u8def\u5f84<\/p>\n\n\n\n
print(
“\u6b63\u5728\u52a0\u8f7d\u57fa\u7840\u6a21\u578b\u4e0e\u5fae\u8c03\u540e\u7684 LoRA \u6743\u91cd\u2026”
)<\/p>\n\n\n\n
1. \u52a0\u8f7d\u57fa\u7840\u6a21\u578b<\/h1>\n\n\n\n
base_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,<\/p>\n\n\n\n
\u63a8\u7406\u65f6\u4f7f\u7528 fp16 \u901f\u5ea6\u66f4\u5feb\uff0c\u4e14 T4 \u663e\u5b58\u5b8c\u5168\u591f\u7528<\/h1>\n\n\n\n
device_map=<\/code><\/pre>\n\n\n\n
“auto”
)<\/p>\n\n\n\n
2. \u53e0\u52a0\u4e0a\u5c42 LoRA \u6743\u91cd<\/h1>\n\n\n\n
model = PeftModel.from_pretrained(base_model, lora_path)
tokenizer = AutoTokenizer.from_pretrained(lora_path)<\/p>\n\n\n\n
3. \u51c6\u5907\u6d4b\u8bd5\u5bf9\u8bdd\uff08\u6d4b\u8bd5\u5b83\u662f\u5426\u8bb0\u4f4f\u4e86\u6ce8\u5165\u7684\u884c\u4e1a\u77e5\u8bc6\uff09<\/h1>\n\n\n\n
messages = [
{
“role”: “system”, “content”: “\u4f60\u662f\u4e00\u4e2a\u535a\u5b66\u3001\u4e13\u4e1a\u7684\u4e2d\u6587\u4eba\u5de5\u667a\u80fd\u52a9\u624b\u3002”
},
{
“role”: “user”, “content”: “\u8bf7\u544a\u8bc9\u6211\u5173\u4e8e\u5927\u6a21\u578b\u6216\u4eba\u5de5\u667a\u80fd\u7684\u884c\u4e1a\u77e5\u8bc6\u3002”
}
]<\/p>\n\n\n\n
<\/p>\n\n\n\n
\u8fd0\u884c\u7ed3\u679c<\/h1>\n\n\n\n
\ud83e\udd16 AI \u7684\u56de\u7b54\uff1a
\u4f60\u597d\uff0c\u5f88\u9ad8\u5174\u4e3a\u60a8\u670d\u52a1\u3002<\/p>\n\n\n\n
<\/p>\n","protected":false},"excerpt":{"rendered":"
1. \u5f3a\u5236\u5347\u7ea7\u51b2\u7a81\u7684\u5e95\u5c42\u4f9d\u8d56\uff0c\u5e76\u5b89\u88c5\u5fc5\u5907\u5e93 !pip install –upgrade torch […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12,1],"tags":[],"class_list":["post-61","post","type-post","status-publish","format-standard","hentry","category-ai0","category-article0"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/posts\/61","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/comments?post=61"}],"version-history":[{"count":2,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/posts\/61\/revisions"}],"predecessor-version":[{"id":63,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/posts\/61\/revisions\/63"}],"wp:attachment":[{"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/media?parent=61"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/categories?post=61"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.52elong.com\/index.php\/wp-json\/wp\/v2\/tags?post=61"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}