Training loss
—
LossSmoothedSelected base
See how our models learned.
Choose one of our models below.
Loading training record…
Pre-training
Learning general language from text. Explore loss and training performance over time.
Validation loss
—
Held-out dataSelected base
Selected base132,005
Base tokens0B
Base validation loss0
Training points—
Fine-tuning
Adapting the pretrained base to JSON extraction. Each stage shows its own task score and evaluation set.
Inherited checkpoint
Task score
- Whole request correct
- Output NLL
- Schema-valid JSON
- Requests