Post-train a model

Go from a task to a deployed post-trained model: decide whether to train, define quality, choose a model and method, run, evaluate, and ship.

Post-training adapts an open model to a task you already run. Use an evaluation to test whether it improves quality or lowers serving cost or latency, compared with the model you use today. Baseten trains with Loops and serves the result on the same platform.

Define the evaluation before you train. Its quality rules help you select training data, and its scores tell you whether to deploy the resulting checkpoint.

​

The path

1. Decide whether to post-train

   Post-training pays off for a stable, well-defined task with a measured gap you can’t close with prompting or retrieval, at enough volume that serving cost or latency matters. See Decide whether to post-train.

2. Define what a good answer is

   Write quality rules, freeze a held-out set, and score the models you have today. Use the same rules to select training data and to decide what ships. See Evaluate a model.

3. Choose a base model

   Start from your constraints: license and publisher, serving latency and cost, and sequence length. Then baseline the candidates on your evaluation. See Choose a base model.

4. Choose SFT or RL

   Supervised fine-tuning (SFT) needs answers to imitate. Reinforcement learning (RL) needs a way to score outcomes. See Choose SFT or RL.

5. Prepare data or an environment

   For SFT, build conversations with targets you trust. For RL, write tasks, an agent, and a grader. See Prepare data for supervised fine-tuning or Design RL environments and rewards.

6. Run a first job

   Start small and confirm the run is learning before you scale. See Fine-tune tool calling with SFT or Train a model to count circles with RL. Both use models available on serverless Loops.

7. Ship it

   Check the run’s health, compare the checkpoint with its baseline on held-out data, deploy it, and roll it out behind the model it replaces. See Ship a post-trained model.

​

What has worked

Each of these published results started from a production task with a measurable definition of quality.

Task	What was trained	Reported result	Source
Structured extraction in a real-time voice pipeline	SFT on a 4B Qwen model, sized to fit the latency budget	99% accuracy on a key feature, 250 ms p90 latency down from 2.2 seconds, 60% lower inference cost	EliseAI
Clinical notes from clinician-patient conversations	A 27B open model fine-tuned for the task	60% better accuracy than the largest closed models tested, 50% to 80% lower inference cost, 2 to 3 times faster	Healthcare scribing
Emergency department summaries and charts	A specialist model trained on failures its evaluators found	Beats gemini-2.5-pro on summarization and chart generation, 6 to 8 times faster	Emergency medicine documentation
Long-horizon legal agents	A 27B open model post-trained with its agent harness in the loop	Moved into the closed-source frontier band on Harvey’s legal agent benchmark	Harvey

The examples in these docs follow the same patterns on public data: tool calling with SFT, an agent that edits files with RL, and a coding agent with RL.

​

Where training runs

Loops runs the trainer and sampler on Baseten. Your training script, or the Loops cookbook, can run locally or in a remote instance. See Choose where to run the Loops SDK. Loops picks an availability mode for each run, so you don’t manage infrastructure to get started. The serverless mode runs on shared capacity and is billed for tokens. The dedicated mode covers every supported model and lets you scale a run out with more trainer and sampler replicas.

Use Training Jobs instead when you need full-parameter training, a model Loops doesn’t support, or your own training framework in a container. Use RL Rollouts when you own the RL training code and want Baseten’s inference platform to scale its rollouts. Training on Baseten compares all three.

​

Next steps

Decide whether to post-train

Check the task, the gap, the data, and the economics before you train.

Evaluate a model

Define quality and freeze a held-out set before the first run.