第 6 章 分类微调：替换输出头的任务适配

// include/finetuning.h —— create_balanced_dataset
std::vector<SpamItem> ham, spam;
for (const auto& it : data) (it.label == 1 ? spam : ham).push_back(it);
std::mt19937 rng(seed);                       // seed=123，固定随机种子
std::shuffle(ham.begin(), ham.end(), rng);
ham.resize(spam.size());                      // ham 下采样到 747

class SpamDataset {
public:
    SpamDataset(const std::string& csv_path, ch2::BpeTokenizer& tokenizer,
                int64_t max_length = -1, int64_t pad_token_id = 50256) {
        for (const auto& it : data_)
            encoded_texts_.push_back(tokenizer.encode(it.text));
        if (max_length < 0)
            max_length_ = longest_encoded_length();       // 以训练集最长者为基准
        else {
            max_length_ = max_length;                     // val/test 沿用训练集的长度
            for (auto& e : encoded_texts_)                // 超长截断对齐
                if ((int64_t)e.size() > max_length_) e.resize(max_length_);
        }
        for (auto& e : encoded_texts_)
            e.resize(max_length_, pad_token_id_);         // 尾部填充
    }
};

class GPTClassifierImpl : public torch::nn::Module {
public:
    GPTClassifierImpl(ch4::GPTModel& base, int64_t num_classes) {
        base_ = base;                                    // 引用同一模块
        register_module("base", base_);
        out_head_ = register_module("out_head",
            torch::nn::Linear(emb_dim, num_classes));    // 768 → 2

        for (auto& np : base_->named_parameters())       // 1. 全部冻结
            np.value().set_requires_grad(false);
        auto& last = base_->trf_blocks->at<ch4::TransformerBlockImpl>(
            base_->trf_blocks->size() - 1);
        for (auto& np : last.named_parameters())         // 2. 解冻最后一个块
            np.value().set_requires_grad(true);
        for (auto& np : base_->final_norm->named_parameters())
            np.value().set_requires_grad(true);          // 3. 解冻 final_norm
    }
    torch::Tensor forward(torch::Tensor x) {
        auto hidden = base_->forward_hidden(x);          // [b, n, 768]（不接 out_head）
        return out_head_->forward(hidden);               // [b, n, 2]
    }
};

inline torch::Tensor calc_loss_batch(const torch::Tensor& input_batch,
                                     const torch::Tensor& target_batch,
                                     GPTClassifier& model) {
    auto logits = model->forward(input_batch).index(
        {torch::indexing::Slice(), -1, torch::indexing::Slice()});  // [b, 2]
    return torch::nn::functional::cross_entropy(logits,
                                                target_batch.to(logits.device()));
}

torch::manual_seed(123);
torch::optim::AdamWOptions options(5e-5);
options.weight_decay(0.1);
auto optimizer = torch::optim::AdamW(classifier->trainable_parameters(), options);
ch6::train_classifier_simple(classifier, train_loader, val_loader, optimizer,
                             /*num_epochs=*/5, /*eval_freq=*/50, /*eval_iter=*/5);

./scripts/run.sh chapter06_finetuning

=== 6.2 准备数据集
总样本数: 5572
标签分布: ham = 4825, spam = 747
平衡后: ham = 747, spam = 747（总计 1494）
划分: train = 1045, val = 149, test = 300

=== 6.3 创建数据加载器
训练集 max_length = 120
批次数量: 130 training, 19 validation, 38 test

=== 6.4 初始化带有预训练权重的模型
生成文本: Every effort moves you toward finding an ideal new way to practice
指令输入生成（未微调，不遵循指令）: ...（续写文本，不会回答 yes/no）

=== 6.6 计算分类损失和准确率
Training accuracy: 46.25% | Validation accuracy: 45.00% | Test accuracy: 48.75%
Training loss: 2.453 | Validation loss: 2.583 | Test loss: 2.322

=== 6.7 在有监督数据上微调模型
Ep 1 (Step 50): Train loss ..., Val loss ...    ← 每 50 步评估一次损失
每轮末尾另打印一行准确率（分母 40 = eval_iter 5 × batch 8，2.5% 步进，
只能看趋势）：训练准确率从 60% 附近起步，第 5 轮升至 95% 附近，
验证集同步上行到 92% 附近

=== 6.8 使用大语言模型作为垃圾消息分类器
text_1: You are a winner you have been specially selected to receive $1000 cash ...
分类结果: spam
text_2: Hey, just wanted to check if we're still on for dinner tonight? ...
分类结果: not spam
已保存分类模型到 data/review_classifier.pth