Autoscientist
Create an AutoScientist run
Recommend hyperparameters
Retrieve an AutoScientist run
List AutoScientist runs
Cancel an AutoScientist run
Download the best trained model artifact
List available base models for training
ModelsExpand Collapse
AutoscientistRun object { id, dataset_id, model, 12 more }
When the run reached a terminal state. Null while it is still running.
Resolved target win rate for the AutoScientist loop, including the model-specific default.
training_method: "instruction" or "alignment"The objective this run is training. Resolved at creation — a requested alignment is downgraded to instruction when the dataset is not preference pairs, so read this back to confirm what actually runs.
The objective this run is training. Resolved at creation — a requested alignment is downgraded to instruction when the dataset is not preference pairs, so read this back to confirm what actually runs.
best_hyperparams: object { training_type, n_epochs, learning_rate, 16 more } Hyperparameters that produced the best iteration so far, and the ones the downloadable artifact was trained with. Null until an iteration completes.
Hyperparameters that produced the best iteration so far, and the ones the downloadable artifact was trained with. Null until an iteration completes.
training_type: optional "lora" or "full"Training strategy override. AutoScientist optimizes this by default; set it only when strictly necessary.
Training strategy override. AutoScientist optimizes this by default; set it only when strictly necessary.
batch_size: optional "max" or numberEither the literal ‘max’ or an integer of 1 or more. No fixed upper bound: the ceiling depends on the model and hardware, and a larger value is rejected at launch. Defaults to ‘max’.
Either the literal ‘max’ or an integer of 1 or more. No fixed upper bound: the ceiling depends on the model and hardware, and a larger value is rejected at launch. Defaults to ‘max’.
Must be 1× or 2× lora_r. Only checked when you supply lora_r in the same request; otherwise it is validated against the platform default.
‘all-linear’ or comma-separated module list (e.g. ‘q_proj,v_proj’).
Number of cosine decay cycles across training. Defaults to 0.5, a single half-cycle decay from the peak learning rate down to the floor.
Whether training loss is computed over the prompt tokens as well as the completion. true trains on the full sequence, false trains only on the completion. Omit to let the platform decide per example, which is the default.
Controls how much the model can deviate from the reference model. Lower = more aggressive updates toward preferred responses; higher = stays closer to reference behavior.
Normalizes log ratios by sample length during DPO loss calculation. Automatically set to true when simpo_gamma > 0.
column_mapping: object { prompt, completion, reasoning_trace, 2 more } Column mapping the run trained with, echoing what was sent on create. Null when the mapping was left to the platform to infer.
Column mapping the run trained with, echoing what was sent on create. Null when the mapping was left to the platform to infer.
AutoscientistRecommendHyperparamsResponse object { model, hyperparams }
Resolved base model id the recommendation is sized for: the value you sent, or the automatically selected model when model was omitted. Pass it back as model on POST /autoscientist to launch against the same model.
hyperparams: object { training_type, n_epochs, learning_rate, 16 more } Hyperparameters derived for the resolved model and effective dataset size. POST /autoscientist applies the same values when hyperparams is omitted. Overriding is not advised.
Hyperparameters derived for the resolved model and effective dataset size. POST /autoscientist applies the same values when hyperparams is omitted. Overriding is not advised.
training_type: optional "lora" or "full"Training strategy override. AutoScientist optimizes this by default; set it only when strictly necessary.
Training strategy override. AutoScientist optimizes this by default; set it only when strictly necessary.
batch_size: optional "max" or numberEither the literal ‘max’ or an integer of 1 or more. No fixed upper bound: the ceiling depends on the model and hardware, and a larger value is rejected at launch. Defaults to ‘max’.
Either the literal ‘max’ or an integer of 1 or more. No fixed upper bound: the ceiling depends on the model and hardware, and a larger value is rejected at launch. Defaults to ‘max’.
Must be 1× or 2× lora_r. Only checked when you supply lora_r in the same request; otherwise it is validated against the platform default.
‘all-linear’ or comma-separated module list (e.g. ‘q_proj,v_proj’).
Number of cosine decay cycles across training. Defaults to 0.5, a single half-cycle decay from the peak learning rate down to the floor.
Whether training loss is computed over the prompt tokens as well as the completion. true trains on the full sequence, false trains only on the completion. Omit to let the platform decide per example, which is the default.
Controls how much the model can deviate from the reference model. Lower = more aggressive updates toward preferred responses; higher = stays closer to reference behavior.
Normalizes log ratios by sample length during DPO loss calculation. Automatically set to true when simpo_gamma > 0.