import os
import glob
import json
from dotenv import load_dotenv
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, DataCollatorForLanguageModeling, LlamaTokenizerFast
from datasets import Dataset, load_dataset
import torch
load_dotenv()
MODEL_NAME = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" # 改为指定的模型
OUTPUT_DIR = "trained_model"
TRAIN_FILE = "train.jsonl"
MAX_LENGTH = 512
BATCH_SIZE = 8
EPOCHS = 3
def create_training_data(posts_dir):
all_texts = []
for lang_dir in os.listdir(posts_dir):
lang_path = os.path.join(posts_dir, lang_dir)
if not os.path.isdir(lang_path):
continue
for file_path in glob.glob(os.path.join(lang_path, "*.md")):
try:
with open(file_path, 'r', encoding='utf-8') as f:
content = f.read()
# 移除前言
content = content.split("---", 2)[-1].strip()
all_texts.append(content)
except Exception as e:
print(f"读取文件 {file_path} 时出错: {e}")
return all_texts
def prepare_dataset(texts, tokenizer):
encodings = tokenizer(texts, truncation=True, padding=True, max_length=MAX_LENGTH, return_tensors="pt")
return Dataset.from_dict(encodings)
def train_model(dataset, tokenizer):
training_args = TrainingArguments(
output_dir=OUTPUT_DIR,
overwrite_output_dir=True,
num_train_epochs=EPOCHS,
per_device_train_batch_size=BATCH_SIZE,
save_steps=10_000,
save_total_limit=2,
prediction_loss_only=True,
remove_unused_columns=False,
)
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, trust_remote_code=True)
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset,
data_collator=data_collator,
)
trainer.train()
trainer.save_model(OUTPUT_DIR)
def main():
posts_dir = "_posts"
texts = create_training_data(posts_dir)
tokenizer = LlamaTokenizerFast.from_pretrained(MODEL_NAME, trust_remote_code=True, use_fast=True)
tokenizer.pad_token = tokenizer.eos_token
dataset = prepare_dataset(texts, tokenizer)
train_model(dataset, tokenizer)
if __name__ == "__main__":
main()