Model modifications

Train weights, edit knowledge, or work with activations. Explore how each method changes a model.

Train on examples of the answers you want. SFT supplies the learning objective; LoRA updates a small adapter while the base weights stay frozen.

Reviewed examplesInputs and approved answers

Supervised trainingLearn from target-token loss

Trained adapterSave with its exact base model

Support conversation

Frozen base model

Train LoRA weights

Your answer format

Added adapter weights

ExampleTeach a support model to return your response format from reviewed conversations.

Ponder Training

Train, evaluate and improve a model from the same experiment setup.

Supervised fine-tuning (SFT) with LoRAContinual learning: Teach the model from reviewed examples. Save a trained adapter.
Reinforcement learning with LoRA (GRPO)Continual learning: Score groups of responses and learn from the rewards.
Continual learning: Schedule reviewed SFT batches or compatible GRPO updates, with review, spending limits and independent evaluation.

Training loopLoRA adapter

New outcomes. Next adapter.

Training requires the exact method, base and runtime to qualify together.

Nemotron Lightning 3.5 30B A3B

Available for hosted inference. This model is not a Ponder training base.

Work with agents beside your local projects, then turn a dataset into an evaluation plan without leaving the app.