Reading Loss Curves

~20 min · loss-curve, overfitting, underfitting, early-stopping

Level 0Observer

0 XP0/43 lessons0/11 achievements

0/120 XP to next level120 XP to go0% complete

Healthy training

Training loss decreases steadily.
Validation loss decreases, then plateaus.
Both curves stay relatively close.

Overfitting signs

Training loss keeps decreasing.
Validation loss starts rising.
The model memorizes training data instead of learning patterns.

Fixes: reduce epochs, increase dropout, add diverse training data, reduce rank (r), enable early stopping.

Underfitting signs

Both losses high and plateau early.
Model hasn't learned enough.

Fixes: increase epochs, increase learning rate, increase rank (r), more training data, check data quality.

Epochs vs steps

An epoch = one full pass through training data. With 1,000 examples and batch 8 → one epoch = 125 steps. Small datasets (< 1,000): 3–5 epochs. Large datasets (> 10,000): 1–2 epochs is often enough.

Code

Early stopping callback in TRL·python

from transformers import EarlyStoppingCallback
from trl import SFTTrainer

trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=dataset,
    eval_dataset=eval_dataset,
    processing_class=tokenizer,
    peft_config=peft_config,
    callbacks=[
        EarlyStoppingCallback(early_stopping_patience=3),
    ],
)
# Stops if validation loss doesn't improve for 3 evaluation rounds.

External links

Exercise

Find a published training run with a clear overfitting curve (W&B has many public runs). Identify the exact step where validation loss starts rising. Then look at training loss at the same step — what is the gap? This calibrates your eye for spotting overfitting in your own runs.