When to use each
If you haven’t decided what to train or how, start with Post-train a model. It covers deciding whether to train, evaluating, choosing a model and a method, and shipping.Loops
Use Loops to post-train a supported language model with LoRA, using reinforcement learning (RL) or supervised fine-tuning (SFT). You write the training loop in Python, and Baseten runs the trainer and sampler. Start with the Loops quickstart.Training Jobs
Use Training Jobs for full-parameter training, models Loops doesn’t support, or any workload that needs its own runtime. It’s an all-purpose tool: bring a container, framework, and training code, and choose the hardware. Start with the Jobs quickstart.RL Rollouts
Use RL Rollouts when you already run an RL trainer and rollout generation is the bottleneck. You own the training code, and Baseten’s inference platform scales your rollouts: your trainer publishes checkpoints to your bucket, and a rollout deployment loads each new policy while it keeps serving. Start with Deploy RL Rollouts.Applied Research
Work with Baseten Applied Research when you want help choosing a model, preparing training data, designing evaluations, or bringing a trained model to production.Next steps
Post-train a model
Decide, evaluate, choose a model and method, train, and ship.
Loops quickstart
Train with serverless Loops, sample from published weights, and save a checkpoint.
Jobs quickstart
Run your first training job and deploy the result.
Deploy RL Rollouts
Connect rollout serving to checkpoints from your trainer.
Evaluate a model
Define what a good answer is, and score models before and after training.