See how our models learned.
We’re opening up the recorded training history behind OpenSML, starting with JSONLite-175M.
Explore the observatory
Inspect the recorded data ↓
the purpose/
A closer look at the work behind the weights.
Building a model from scratch means making decisions about training, learning rates and which checkpoints to carry forward. We built the observatory to make that process easier to explore.
It brings the retained measurements into one place, so you can follow the training history and inspect the checkpoints behind a released model.
Replay pretraining.
Play through the recorded run, pause at a step and watch the curves unfold.
Inspect the details.
Explore loss, validation, throughput and learning rate. Expand charts and inspect recorded values.
Explore fine-tuning.
Move through five stages and inspect their checkpoint evaluations, with each metric and evaluation set identified.
Open to explore.
The observatory is a read-only record of past training, with pretraining and fine-tuning shown separately. No account or model download is needed.
Start with JSONLite-175M, press play, and take a closer look.