deepseek 1.5b离线模型微调

### 对 DeepSeek 1.5B 离线模型进行微调 为了对 DeepSeek 1.5B 离线模型进行微调，可以遵循以下方法： #### 准备环境 确保安装了必要的库和工具包。这通常涉及 Python 和 PyTorch 的安装以及 Hugging Face Transformers 库的引入。 ```bash pip install torch transformers datasets evaluate accelerate ``` #### 加载数据集 使用 `datasets` 库来加载所需的数据集，并对其进行预处理以便于后续训练过程中的应用[^2]。 ```python from datasets import load_dataset dataset = load_dataset('path_to_your_data') train_testvalid = dataset['train'].train_test_split(test_size=0.1, seed=42) tokenized_datasets = train_testvalid.map( tokenize_function, batched=True, num_proc=4, remove_columns=["text"] ) ``` #### 初始化模型与分词器 从本地路径加载已有的 DeepSeek 模型及其对应的分词器实例化对象。 ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_name_or_path = "path/to/deepseek-r1-1.5b" tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) model = AutoModelForCausalLM.from_pretrained(model_name_or_path) ``` #### 设置训练参数 定义用于控制训练行为的各种超参数配置项，比如批量大小、学习率等。 ```python batch_size = 8 learning_rate = 5e-5 weight_decay = 0.01 num_train_epochs = 3 gradient_accumulation_steps = 4 max_length = 512 lr_scheduler_type = "cosine" warmup_ratio = 0.1 output_dir = "./results" training_args = TrainingArguments( output_dir=output_dir, per_device_train_batch_size=batch_size, learning_rate=learning_rate, weight_decay=weight_decay, fp16=True, gradient_checkpointing=True, max_grad_norm=None, num_train_epochs=num_train_epochs, logging_strategy="steps", evaluation_strategy="epoch", save_strategy="epoch", warmup_ratio=warmup_ratio, lr_scheduler_type=lr_scheduler_type, report_to='tensorboard', push_to_hub=False, hub_model_id=model_name_or_path.split('/')[-1], gradient_accumulation_steps=gradient_accumulation_steps, optim="adamw_torch_fused", ddp_find_unused_parameters=False, group_by_length=True, do_train=True, do_eval=True, predict_with_generate=True, generation_max_length=max_length * 2, save_total_limit=1, disable_tqdm=False, dataloader_num_workers=4, load_best_model_at_end=True, metric_for_best_model="eval_loss", greater_is_better=False, label_names=['labels'] ) ``` #### 创建Trainer并启动训练 利用上述设置创建 Trainer 实例，并开始执行实际的微调操作。 ```python trainer = Trainer( model=model, args=training_args, tokenizer=tokenizer, data_collator=data_collator, compute_metrics=compute_metrics_fn, train_dataset=tokenized_datasets["train"], eval_dataset=tokenized_datasets["test"], ) trainer.train() ``` 完成以上步骤之后，即实现了基于特定任务需求下的 DeepSeek 1.5B 模型微调工作流。