Skip to main content

Evaluation

Training produces many checkpoints. Evaluation is how you decide which one to fine-tune from, and which one to ship.

Both tools live inside the training sections rather than as separate areas of the product: Pre-training and Fine-tuning each have their own Leaderboard tab, and benchmarks are created from and consumed by those leaderboards.

Leaderboard

The leaderboard ranks checkpoints across runs, so you can compare a checkpoint from one run against a checkpoint from another under the same measurement.

Two properties matter more than the ranking itself:

  • It is scoped to its section. The Pre-training leaderboard ranks Pre-training runs; the Fine-tuning leaderboard ranks fine-tuning runs. The two are optimizing for different objectives and are not comparable, which is why they are not mixed.
  • It ranks checkpoints, not runs. A run's best checkpoint is often not its last one. Selecting the final checkpoint by default is the single most common way to ship a worse model than you trained.

Use it to pick the base checkpoint for a fine-tuning run, and to pick the checkpoint you deploy.

Benchmarks

A benchmark fixes the thing you measure against, so that a comparison means something. Without one, two checkpoints evaluated on different data are two unrelated numbers.

You add checkpoints to a benchmark from the leaderboard, and the benchmark's results come back into the leaderboard as a comparable column.

Metrics

Results are reported against the metrics your Model Design enabled:

MetricReads as
accuracyShare of correct predictions
precisionOf the positives predicted, how many were right
recallOf the actual positives, how many were found
f1Harmonic mean of precision and recall
roc_aucRanking quality across all thresholds
log_lossPenalty on confident wrong answers
ksMaximum separation between the two distributions
giniDiscriminatory power, rescaled from roc_auc

Custom metrics defined in the design appear alongside these.

Metrics arrive from four sources, and the distinction matters when a number looks wrong: TRAIN (emitted during training), CHECKPOINT_LOSS and CHECKPOINT_VALIDATION (evaluated per checkpoint against the design's test split), and BENCHMARK (evaluated against a benchmark's fixed dataset).

Reading a comparison honestly

  • Compare on the metric the task is judged on, not on whichever metric moved most.
  • Hold the benchmark fixed. A metric that improved against different data has not improved.
  • Watch the validation curve, not just the final number. A checkpoint whose validation loss has started rising is memorizing, and the leaderboard's headline metric can still look fine.

Next steps