Skip to content
Adaption Adaption Adaption
SupportLogin
AutoScientist
Overview
Edit on GitHub

Overview

Understand what AutoScientist builds and how its iterative training loop works.

AutoScientist automates the research loop behind model training and alignment. You choose the outcome; AutoScientist optimizes the data and training recipe, trains a model, evaluates it, and repeats until the run reaches its goal or exhausts its iterations.

Start with a dataset that has completed an Adaptive Data run. If your source is already a training-ready prompt/completion dataset, you can instead run on non-adapted data.

Select a dataset for AutoScientist

What you can build

Section titled “What you can build”
  • Instruction datasets pair prompts with expected completions for supervised fine-tuning.
  • Preference-pair datasets compare chosen and rejected completions for Direct Preference Optimization (DPO).

Set training_method="instruction" for supervised fine-tuning or training_method="alignment" for the SFT → DPO pipeline when you create a run. Alignment requires a dataset prepared as preference_pairs; pass the method explicitly rather than relying on saved app settings.

How the loop works

Section titled “How the loop works”

Each run cycles through four stages:

  1. Data optimization adds optional domain-targeted and general-purpose rows to improve diversity and reduce overfitting.
  2. Training runs SFT or DPO with the selected model, LoRA or full fine-tuning, and the current hyperparameter recipe.
  3. Evaluation measures the iteration against target_win_rate and captures training diagnostics such as loss, learning rate, and gradient norm.
  4. Hyperparameter adjustment analyzes the evaluation and revises the recipe for the next iteration.

The loop stops when it reaches target_win_rate or completes max_iterations. A succeeded status can mean either condition, so inspect best_win_rate after completion.

Next steps

Section titled “Next steps”