CreateFineTuneJobRequestSchema

base_modelstringrequired

Foundation model to fine-tune. Use the model_name value from List base models.

Example: Llama-3.3-70B-Instruct

batch_sizeintegerrequired

Number of training examples processed per gradient update step. Recommended range: 1–16.

Default: 4

Example: 1

epochintegerrequired

Number of complete passes through the training data. Typical range: 1–10.

Default: 3

Example: 1

learning_ratenumberrequired

Step size for the optimizer. Typical range: 0.00005–0.001.

Default: 0.0004

Example: 0.0001

model_namestringrequired

Unique name for the resulting fine-tuned model. Used to reference the model in deployment and inference requests.

Example: example-ftmodel

custom_datasetstring

Name of a saved dataset to use as the training data source instead of tag-filtered logs.

Default: null

custom_logs_filenamestring

Filename of an uploaded custom logs file (from Upload custom logs file) to use as the training data source.

Default: null

is_sampleboolean

When true, runs the job using a small sample of the training data to quickly validate configuration before a full run.

Default: false

lora_alphainteger

LoRA scaling factor applied to adapter outputs. Typically set to ~2× lora_r.

Default: 16

lora_dropoutnumber

Dropout probability applied to LoRA adapter layers during training to prevent overfitting.

Possible values: >= 0 and <= 1

Default: 0.05

lora_rinteger

LoRA rank — controls the number of trainable parameters in each adapter layer. Higher rank = more capacity but more compute.

Default: 32

parent_model_idinteger

ID of an existing fine-tuned model to use as the starting point for incremental (continued) training.

Default: null

save_logs_with_tags[]

Tags used to filter training logs for this job. Only logs matching all specified tags are included.

Default: null

skip_logs_with_errorsboolean

When true, training log entries flagged as containing errors are excluded from the dataset.

Default: true

tagsstring[]

Filter training logs by these tags. Only logs tagged with one of these values are used for training.

Default: null

training_job_timeinteger

Maximum allowed wall-clock time for the training job in seconds. The job is stopped if it exceeds this limit.

Default: null

use_synthetic_databoolean

When true, includes synthetic data logs in the training dataset alongside real captured logs.

Default: true

warmup_stepsinteger

Number of gradient steps over which the learning rate is linearly warmed up from zero.

Default: 0

CreateFineTuneJobRequestSchema

{
  "base_model": "Llama-3.3-70B-Instruct",
  "model_name": "example-ftmodel",
  "batch_size": 1,
  "epoch": 1,
  "learning_rate": 0.0001
}