用法： 脑子想到中文 → 找到对应英文零件 → 用万能连接词拼起来。不需要翻译，只需要组装。
你脑子里想说的 直接写
因为... This is because... 所以... As a result, ... / This leads to... 但是... However, ... 相比之下... Unlike [A], [B]... / While [A]..., [B]... 我建议... I would suggest... because...
[YES/NO], [建议] is likely to [improve/not improve] validation accuracy.
The model is currently [overfitting/underfitting], as evidenced by
[说出你看到的证据].
[建议] [helps/does not help] because [一句话机制].
实战例子（你脑子里想的 → 写出来的）：
中文思维 英文输出
过拟合了，train高val低 "The model is overfitting, as the training accuracy (95%) is much higher than the validation accuracy (60%)." 加正则化能帮忙 "L2 regularisation is likely to help because it penalises large weights, constraining model complexity." 欠拟合不能加dropout "Dropout will NOT help because the model is underfitting — dropout further constrains an already limited model." 多训练几轮会更差 "More epochs will worsen overfitting because the model will continue to memorise training noise."
[概念] is [一句话定义].
It works by [机制].
This is [beneficial/important] because [为什么有用].
高频概念直接背：
中文 英文零件（直接写）
注意力机制 "Attention computes a weighted sum of values, where weights reflect the relevance of each input position." 遮蔽注意力 "Masking prevents each position from attending to future tokens, preserving the autoregressive property." 多头注意力 "Multi-head attention runs several attention functions in parallel, each focusing on different aspects of the input." 位置编码 "Positional encoding is needed because the Transformer processes all positions in parallel, losing inherent ordering." [CLS] token "[CLS] is a learnable token that aggregates information from all patches via attention for classification." Dropout "Dropout randomly deactivates neurons during training, forcing the network to learn redundant representations." Batch Norm "Batch normalisation normalises activations within each mini-batch, keeping gradients in a healthy range." L2 正则化 "L2 regularisation penalises large weights, encouraging simpler and more generalisable models." 梯度消失 "Gradients are multiplied through many layers. With sigmoid (max derivative 0.25), they shrink to near zero." 死亡ReLU "If a neuron consistently receives negative inputs, ReLU outputs zero and the neuron stops learning permanently." 动量 "Momentum maintains a running average of past gradients, smoothing updates and accelerating convergence." Adam优化器 "Adam combines momentum (past gradient direction) with adaptive per-parameter learning rates, making it effective for deep networks." RNN隐藏状态 "Each time step takes the current input AND the previous hidden state: h_t = f(W·h_{t-1} + U·x_t + b)." LSTM门控 "LSTM uses three gates (forget, input, output) to control information flow, solving the vanishing gradient problem." 残差连接 "Skip connections add the input directly to the output: y = F(x) + x, allowing gradients to flow through the shortcut."
The [loss curve / metric] shows that [你看到了什么].
This indicates [诊断].
[建议/解释] because [原因].
中文思维 英文零件
loss在震荡 "The loss curve oscillates and fails to converge." loss发散了 "The loss diverges, increasing over epochs." train和val之间有gap "There is a significant gap between training and validation loss." 两条线都很高 "Both training and validation loss remain high." 学习率太大 "This indicates the learning rate is too high, causing the optimisation to overshoot the minimum." 学习率太小 "This suggests the learning rate is too small, resulting in very slow convergence." 准确率高但recall低 "Despite high accuracy, the low recall indicates the model fails to identify most positive instances." 类别不平衡 "This is due to class imbalance — the model achieves high accuracy by simply predicting the majority class."
不需要造句，只需要写步骤。模板：
Conv layer:
Input: [H, W, C]
Output: [floor((H + 2p - f) / s) + 1, same for W, num_filters]
Pool layer:
Input: [H, W, C]
Output: [floor((H - f) / s) + 1, same for W, C] (depth unchanged)
Flatten: H × W × C = [answer]
While [A] [A的特点], [B] [B的特点].
The key advantage of [A] is [优势].
However, [A]'s limitation is [缺点], which [B] addresses by [B的解决方式].
中文思维 英文零件
RNN按顺序处理 "RNNs process tokens sequentially, naturally capturing order." Transformer并行 "Transformers process all positions in parallel using self-attention." RNN不能并行是缺点 "Sequential processing prevents parallelisation, making training slow for long sequences." Transformer用位置编码补顺序 "The Transformer compensates for the loss of order information by adding positional encoding." CNN有局部性 "CNNs have a strong inductive bias towards locality and translation invariance." ViT更灵活但需要大数据 "ViT makes fewer assumptions, offering more flexibility, but requires large-scale pretraining data."
你可能会写 改成
"The model performance is not good" "The model performs poorly on validation data" "It can make the model more better" "It is likely to improve generalisation" "The reason is because..." "This is because..." "prevent to overfit" "prevent overfitting" "the accuracy is high so the model is good" "despite high accuracy, the model may be ineffective due to class imbalance" "add more regularisation to make it good" "applying regularisation constrains model complexity, reducing overfitting"
第一步（诊断）：看 train acc 和 val acc → 判断过拟合/欠拟合
第二步（判断）：这个建议能不能解决诊断出的问题？
第三步（输出）：YES/NO + 用英文写原因
观察 → 诊断 → 建议 → 原因
Observe → Diagnose → Suggest → Explain
"The training accuracy is [X]% while validation accuracy is [Y]%." （观察）
→ "This indicates [overfitting/underfitting]." （诊断）
→ "[Suggestion] is [likely/unlikely] to help." （建议）
→ "This is because [mechanism]." （原因）
分值 目标字数 结构
1 分 1-2 句 结论 2 分 2-3 句 结论 + 原因 3-4 分 4-6 句 结论 + 诊断 + 原因 + 机制 5-6 分 1-2 段 完整分析（每个要点 2-3 句）