Create Model

POST

v1/models

Create Model allows you to create a copy of an existing model (called the parent model). Once created, your new model is independent and can be customized through training. You can experiment with different training parameters, datasets, and fine-tuning techniques without any risk to the original parent model.

Note: You can only train models that you have created yourself. The standard base models (like Llama-3.1-8b or DeepSeek-R1-8b) are read-only, so you must first create a copy of these models before you can train them.

Parameters

name:required string

The name for your new model. This will be used to identify the model in API calls and the model list.

Note: Model names must be unique within your account.

description:required string

A description of what this model is intended for or what makes it special. This helps you organize and identify your models.

parent_model:required string

The name of the parent model to base this new model on. This must be an existing model available in your model list.

Common parent models include:

Llama-3.1-8b - Meta's Llama 3.1 model
DeepSeek-R1-8b - DeepSeek's R1 model
Any model you have previously created

rank:optional number

The rank parameter for LoRA (Low-Rank Adaptation) fine-tuning. This controls the dimensionality of the low-rank matrices used in the adaptation.

Lower values (4-16) - Fewer parameters, faster training, less memory usage, but potentially less expressive
Higher values (32-128) - More parameters, slower training, more memory usage, but potentially more expressive

Note: If not specified, the model will use a default rank value based on the parent model architecture.

alpha:optional number

The alpha parameter for LoRA scaling. This controls the scaling factor applied to the LoRA weights.

Typically set as a multiple of rank (e.g., if rank=16, alpha=32). A common rule of thumb is alpha = 2 * rank.

Note: If not specified, the model will use a default alpha value based on the rank.

dropout:optional number

The dropout rate applied during training. Dropout helps prevent overfitting by randomly setting a fraction of input units to 0 during training.

0.0 - No dropout (may lead to overfitting)
0.1 - Light regularization (recommended for most cases)
0.2-0.5 - Stronger regularization (use when overfitting is a concern)

Note: If not specified, the model will use a default dropout value.

learning_rate:optional number

The initial learning rate for training this model. This can be overridden when you actually train the model, but setting it here provides a default value.

See the Train endpoint documentation for more details on learning rates.

Note: If not specified, the model will use a default learning rate based on the parent model.

Output

ModelInfo

Contains information about a model, including its configuration, training parameters, and metadata.

name:required string

Model name, this is the name given to to model.

id:required string

Model id, this is a id generated when the model is created.

description:required string

Model description, set when the model was created.

read_only:required boolean

Is the model readonly, only the models that you have created can be trained(that is not read only)

parent_model:optional string

The name of the parent model, this is the base model that the model is based on.

revision:optional number

The revision number of the current model. It start with 0 and is increase by 1 for every training.

last_modified:optional number

Time stamp of the latest training.

default_model:optional boolean

The default model the one selected at startup.

temperature:required number

Default temperature setting, this is the value user in getResponse if no other value if specified.

Controls the randomness and creativity of the model's responses. Lower values make output more focused and deterministic, while higher values increase randomness and creativity.

0.0 - Most deterministic, repeatable responses
1.0 - Balanced creativity and coherence (recommended for most use cases)
2.0 - Maximum randomness and creativity

Note: For tasks requiring consistency (like data extraction or classification), use lower values (0.0-0.3). For creative tasks (like brainstorming or storytelling), higher values (0.7-1.5) work better.

top_p:required number

Default temperature setting, this is the value user in getResponse if no other value if specified.

Also known as "nucleus sampling," this parameter controls the diversity of responses by limiting the model to consider only the most probable tokens whose cumulative probability reaches the specified threshold.

0.1 - Very focused, only highly probable tokens
0.5 - Moderately diverse output
1.0 - Considers all tokens based on their probability

Note: It's generally recommended to adjust either temperature OR top_p, but not both simultaneously. When top_p is less than 1.0, the model samples from the smallest set of tokens whose cumulative probability exceeds the threshold.

learning_rate:required number

Default value for the learningRate used in Training.

learning_steps:required number

Default value for the learningSteps used in Training.

beta:required number

Default value for the beta used in Training.

epsilon:required number

Default value for the epsilon used in Training.

API Reference

Create Model

Parameters

Output

ModelInfo

Create Model

Response