AutoScientist
AutoScientist co-optimizes your data and model training recipe automatically, self-improving across both until quality converges on the goal you describe. Where Adaptive Data shapes the inputs, AutoScientist shapes the model: it runs the full research loop end-to-end so you walk away with a model adapted to your goal, rather than a dataset you still need to train on yourself.
Getting Started
Section titled “Getting Started”Once you’ve launched AutoScientist, start by selecting a dataset you’ve already run through Adaptive Data.
Step 1 | Instruction Tuning
Section titled “Step 1 | Instruction Tuning”Do you have size constraints for the model you want to train?
- All sizes (no restrictions, best possible results)
- Tiny AutoScientist (restricts training to models under 10B parameters)
Step 2 | Column Training
Section titled “Step 2 | Column Training”Functionally the same concept as Adaptive Data’s column mapping: choose a Prompt column and Completion column.
The screen previews the result of the Original prompt & completion vs the Enhanced prompt, Enhanced completion. The option to select the original columns is also available.
Step 3 | Augmentation
Section titled “Step 3 | Augmentation”AutoScientist evaluates your dataset and can surface up to two independent, opt-in expansions, each priced and sized live as you move its slider:
- Domain dataset expansion — appears when you have fewer than 20,000 datapoints. You set a target datapoint count with a slider; the menu shows the resulting Credits required and Expanded dataset size live.
- Diversity dataset expansion — appears when the dataset isn’t diverse enough. Same slider-driven credits/row-count pattern as domain expansion, using general-purpose rather than domain-specific datapoints.
Step 4 | Recipe
Section titled “Step 4 | Recipe”AutoScientist proposes a full hyperparameter recipe as editable fields, including Algorithm, Epochs, Rank of lora, Model Name, Alpha, Target Layers, Optimizer, Warmup Ratio, and Gradient clipping. Every field is editable before you continue.
Step 5 | Confirmation & Training
Section titled “Step 5 | Confirmation & Training”Confirm your optimization choices — a raw JSON view of the exact configuration, with Copy and Edit controls, giving you a full technical read of what will actually run.
You are then met with a final order summary breaking cost down by component: domain dataset expansion, diversity dataset expansion, and AutoScientist training itself.
On your dashboard, under AutoScientist, you can track the progression of the run.
Step 6 | Understanding Results Once a run completes, the results page shows your trained model ID, with Weights export and AutoScientist Config.
Section titled “Step 6 | Understanding Results Once a run completes, the results page shows your trained model ID, with Weights export and AutoScientist Config.”The Training Winrates chart compares your base vs adapted dataset winrate.
For example: base 30 vs adapted 70
-The adapted model won roughly 70% of head-to-head comparisons against the original base model on your own dataset
Train/Eval Metrics
Shows standard training diagnostics over training steps
Common issues
Section titled “Common issues”- Flat or negative quality/win rate after training — use at least a 7–8B model, increase epochs, and consider the Augmentation step’s 20,000-datapoint threshold.
- No way to get the model onto Hugging Face — use the manual download → blank repo → web UI upload.
- Can’t tell if a job is running, queued, or stuck — check the AutoScientist tab for state and Measure tab for metrics; ask support to confirm status directly if truly uncertain after a long wait.
- Pass rows in Autoscientist without adapting it - you are not obliged to use enhanced rows, you can choose to use the original rows.