统计图

训练损失曲线

训练集与验证集 loss / accuracy 曲线，适合展示模型收敛与过拟合。

完整 Prompt

A line chart showing training loss and validation loss over 100 epochs for 3 learning rates.

Series:
- LR=1e-3: train loss decreases smoothly from 2.4 to 0.10; validation loss decreases until epoch 40 (0.55) then rises to 1.20 by epoch 100 (overfitting).
- LR=1e-4: train loss decreases from 2.4 to 0.18; validation loss decreases monotonically from 2.3 to 0.42.
- LR=1e-5: train loss decreases slowly from 2.4 to 0.85; validation loss decreases from 2.3 to 0.98 (underfitting).

Style:
- Solid lines for training loss, dashed lines for validation loss.
- One color per learning rate (LR=1e-3 amber, LR=1e-4 teal, LR=1e-5 gray).
- X-axis: Epoch (0–100), Y-axis: Loss (0–2.5), gridlines every 0.5.
- Legend top-right with the line-style key.
- Annotate the epoch-40 inflection on the LR=1e-3 validation curve with a small text callout "Overfitting begins".

Clean academic styling, minimal chart junk, white background.

立即试用此 Prompt

适用场景

For ablation sections showing learning-rate sensitivity and overfitting behaviour.

变体

Log-scale Y axis

Same chart but with the y-axis in log scale (0.01 to 5). Re-state that the loss values now appear on a log scale and add a "log scale" note next to the y-axis label.

With early-stopping marker

Add a vertical dashed line at epoch 40 labeled "Early stopping (best val loss)". Place a small triangle marker on the LR=1e-3 validation curve at epoch 40.

使用建议

Describe the SHAPE of each curve in words ("smoothly decreasing", "U-shaped"). It guides the model better than raw numbers alone.
State which line style maps to which series — solid/dashed/dotted are reliably reproduced.
Annotate the most important point with a callout — the model will draw a small label arrow there.

常见问题

Can I add a confidence band around each curve?

Yes — add "Surround each line with a translucent confidence band of width ±0.05 in the training-loss color."