Hyperparameters

AutoScientist proposes a complete training recipe before training, then evaluates and adjusts the recipe across iterations. For most runs, keep the recommended values unchanged. AutoScientist is optimized to select these values together and improve them based on measured results.

Changing individual hyperparameters is not recommended. An override can prevent AutoScientist from exploring the configuration it would otherwise select and can reduce training quality. Override a value only when you have a specific experimental or infrastructure constraint.

Hyperparameter reference

training_method is a top-level create argument (instruction or alignment). The remaining fields below are keys inside hyperparams.

General training

These hyperparameters apply to both SFT and DPO runs.

Hyperparameter	Definition	Valid Values
learning_rate	Peak learning rate for the optimizer. Controls the step size during weight updates.	Float in [1e-6, 5e-4]
n_epochs	Number of complete passes training makes over the dataset.	Integer, 1–5
batch_size	Number of samples per training step. In SFT, must be "max". In DPO, a power-of-2 integer or "max".	"max" or power-of-2 integer ≥ 1
lr_scheduler_type	How the learning rate changes during training. A linear schedule decays steadily; a cosine schedule slows its decay near the end.	"linear" | "cosine"
min_lr_ratio	Floor learning rate as a fraction of the peak. The learning rate will not decay below learning_rate × min_lr_ratio.	Float, 0.0–1.0
scheduler_num_cycles	Number or fraction of cosine decay cycles across training. Only affects cosine schedules.	Float > 0.0 (default 0.5)
warmup_ratio	Fraction of total training steps spent gradually increasing the learning rate from zero. Smooths the transition into the target learning rate.	Float, 0.0–1.0 (typically 0.03–0.1)
max_grad_norm	Maximum gradient norm for gradient clipping. Caps gradient magnitude to reduce exploding gradients and unstable updates. Set to 0.0 to disable.	Float ≥ 0.0
weight_decay	L2 regularization penalty. Discourages excessively large weights, which helps prevent overfitting and improves generalization to unseen data.	Float ≥ 0.0
training_type	Training strategy. "lora" inserts low-rank adapters; "full" updates all model weights. Full fine-tuning is only available for supported models.	"lora" | "full"

LoRA

These hyperparameters configure Low-Rank Adaptation and are only applicable when training_type is "lora". In alignment (SFT → DPO) pipelines, LoRA settings are frozen from the winning SFT iteration and cannot be changed during DPO.

Hyperparameter	Definition	Valid Values
lora_r	Rank of the LoRA adapter matrices. Controls adapter capacity — higher ranks can represent more complex adaptations but require more memory and compute.	Power-of-2 integer, 1–64
lora_alpha	Scaling factor for the LoRA adapter’s contribution to the base model. Must be 1× or 2× the value of lora_r.	Power-of-2 integer (1× or 2× lora_r)
lora_dropout	Dropout probability applied to LoRA layers during training. Adds regularization to reduce overfitting.	Float, 0.0–1.0 (typically 0.0–0.1)
lora_trainable_modules	Which model layers receive LoRA adapters. "all-linear" targets every linear layer; alternatively, specify a comma-separated list of module names.	"all-linear" or comma-separated list (e.g. "q_proj,v_proj,o_proj")

SFT specific

Hyperparameter	Definition	Valid Values
train_on_inputs	Whether training loss is computed over prompt tokens in addition to completion tokens. Setting to false masks the prompt so the model only learns to generate completions.	true | false

Alignment specific

These hyperparameters apply only to preference-based (DPO) training. They are inert for SFT runs.

Hyperparameter	Definition	Valid Values
dpo_beta	Controls how far the model is allowed to drift from the reference model. Lower values (around 0.1) update more aggressively toward the preferred output; higher values (around 0.7) stay closer to the reference.	Float, 0.05–0.9
dpo_normalize_logratios_by_length	Normalizes log ratios by sample length during loss calculation. Automatically set to true when simpo_gamma is above 0. Changing this flag alters the effective scale of dpo_beta; treat a change as a major intervention.	true | false
rpo_alpha	Weight of an auxiliary negative log-likelihood (NLL) loss on preferred responses, added on top of the standard DPO loss (Regularized Preference Optimization). Counteracts likelihood displacement — the failure mode where chosen-response probability drifts downward even as the reward margin grows. Set to 0.0 to disable. Cannot be combined with simpo_gamma.	Float ≥ 0.0 (typical 0.1–1.0)
simpo_gamma	Target reward margin from Simple Preference Optimization. Above 0.0, the objective becomes reference-free and length-normalized: the chosen response must beat the rejected response by at least this margin. Raise dpo_beta alongside it since SimPO typically calls for a substantially larger beta. Set to 0.0 to disable. Cannot be combined with rpo_alpha.	Float ≥ 0.0 (typical 0.3–1.6)

Review in the app

The proposed fields are editable, but the default workflow is to review and accept them:

1. Review the model, algorithm, and generated recipe.
2. Keep the recommended values unless a known constraint requires an override.
3. Inspect the final AutoScientist Config JSON before confirming.
4. Let AutoScientist evaluate and refine the recipe across iterations.
5. After training, compare win rate and diagnostics as described in Interpreting results.

Advanced: Apply API overrides

API overrides are an advanced escape hatch for controlled experiments or hard requirements. The SDK accepts the training objective through top-level training_method, the base model through top-level model, and recipe overrides through hyperparams. training_type is a field inside hyperparams, not a top-level argument.

If an override is necessary, avoid unsupported key names: copy a valid hyperparams object from the current app configuration or construct it from the current create API schema, save it as hyperparams.json, and pass it unchanged:

import json
from pathlib import Path

from adaption import Adaption

client = Adaption()

dataset_id = "dataset_abc123"
hyperparams = json.loads(Path("hyperparams.json").read_text())

run = client.autoscientist.create(
    dataset_id=dataset_id,
    training_method="alignment",
    hyperparams={
        **hyperparams,
        "training_type": "lora",
        "dpo_beta": 0.1,
    },
)
print(run.id, run.status)

Unset hyperparameter keys remain under platform control. To enforce a specific model, review the supported models, list the currently available IDs, and pass one through model:

response = client.autoscientist.list_models()
for model in response.models:
    print(model.id)

Prefer leaving model and hyperparams unset so AutoScientist can choose and iterate on them automatically. See Running AutoScientist for the complete request and status workflow.