Introducing Ponder Training

Our goal is for everyone to have their own continually updating model. Ponder Training is our post-training platform for taking the next step toward that goal.

On our path toward general continual learning, we've found that what part of the system you update matters. Ponder Training lets you explore those updates along with us, from supervised fine-tuning and reinforcement learning to modifications that change selected weights, neurons, memory, or decision behavior.

Choose what you update

We call these modifications Mods. Alongside SFT and RL with LoRA, you can explore:

AlphaEdit for targeted edits to model weights.
Neuron training for learning updates to selected neurons.
Engrams for training added memory tables over a frozen base model.
Decision models for learning defined decisions and scoring the available choices.

Different updates need different data and evaluation. The model modifications guide explains what each method changes and which model and runtime combinations it supports. These methods are experimental; execution and serving availability depend on the selected configuration.

From datasets to training

These Mods share our continual learning pipeline. Upload datasets and graders, then run hosted automated diagnostics to identify problems and clean up your tasks before training.

Review the examples, check that your graders measure the behavior you want, and keep evaluation tasks separate from training data. From there, configure an experiment, run it on hosted infrastructure, and compare the resulting model with your starting point.

The Ponder Training Console showing a completed run, compute and cost, and the stages from baseline evaluation through checkpoint saving and evaluation.

An example run in the Console, with status, compute, cost, checkpoints, and evaluation evidence in one view.

The Ponder Training documentation walks through preparing data, running training, and reviewing the result. For repeated updates as new tasks arrive, follow the continual learning guide.

A desktop app and harness to carry with you

If you'd like to bring your own key and use our hosted infrastructure for training, we've built an open-source desktop app to help you view, manage, and build your datasets and training modules. It also doubles as a durable harness to carry with you through training, keeping your projects and training work together as you iterate.

You can also use the CLI and maintained skills from your own coding agent. Start with the desktop guide, set up Ponder Training in your harness, or explore the open-source repository.

Two practical measures of continual learning

We propose two simple metrics to measure progress toward a continually updating model:

Adaptation time: How long from your feedback until the improved model is live.
Repeat rate: How often you see a mistake you already fixed.

We want feedback to turn into useful improvements sooner, and corrections to carry forward into future work. These are the measures we want to improve as we build toward general continual learning.

Serve it your way

Host and serve supported model versions directly through the app, or download your weights and host them on your own infrastructure. The serving guide covers deployment and compatible runtimes.

Start experimenting with Ponder Training, or read the documentation to choose your first workflow.