Creating a job

Fine-tuning jobs are created against the API at https://gateway.omnia-voice.com/v1 with your workspace API key. Launching a job is an admin-gated action.

1. Upload a training file

Fine-tuning takes a training file (JSONL). Upload it with POST /v1/fine_tuning/files:

curl https://gateway.omnia-voice.com/v1/fine_tuning/files \
  -H "Authorization: Bearer $OMNIA_API_KEY" \
  -F "file=@training.jsonl"

import requests

BASE = "https://gateway.omnia-voice.com/v1"
headers = {"Authorization": f"Bearer {OMNIA_API_KEY}"}

with open("training.jsonl", "rb") as f:
    file = requests.post(
        f"{BASE}/fine_tuning/files",
        headers=headers,
        files={"file": f},
    ).json()

The response includes the file’s id, which you reference as trainingFileId. You can also prepare data as a managed dataset; see Datasets & formats.

2. Create the job

Create a job with POST /v1/fine_tuning/jobs. Only baseModel is required; you supply training data as a trainingFileId (or inline training).

curl https://gateway.omnia-voice.com/v1/fine_tuning/jobs \
  -H "Authorization: Bearer $OMNIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "baseModel": "meta-llama/Llama-3.1-8B-Instruct",
    "name": "support-tone-v1",
    "suffix": "support-v1",
    "trainingFileId": "<file-id>",
    "validationFileId": "<validation-file-id>",
    "method": "supervised",
    "seed": 42,
    "hyperparameters": {
      "n_epochs": 3,
      "learning_rate": 0.0001,
      "batch_size": 8,
      "lora": true,
      "lora_r": 16,
      "lora_alpha": 32
    }
  }'

import requests

BASE = "https://gateway.omnia-voice.com/v1"
headers = {
    "Authorization": f"Bearer {OMNIA_API_KEY}",
    "Content-Type": "application/json",
}

job = requests.post(
    f"{BASE}/fine_tuning/jobs",
    headers=headers,
    json={
        "baseModel": "meta-llama/Llama-3.1-8B-Instruct",
        "name": "support-tone-v1",
        "suffix": "support-v1",
        "trainingFileId": "<file-id>",
        "validationFileId": "<validation-file-id>",
        "method": "supervised",
        "seed": 42,
        "hyperparameters": {
            "n_epochs": 3,
            "learning_rate": 0.0001,
            "batch_size": 8,
            "lora": True,
            "lora_r": 16,
            "lora_alpha": 32,
        },
    },
).json()

A successful create returns the new job’s id:

{ "id": "<job-id>" }

Poll the job with GET /v1/fine_tuning/jobs/{id} to follow its status (queued → running → succeeded/failed/cancelled), events, and, once finished, its trained_tokens and the resulting model.

Fields

Field	Required	Description
`baseModel`	Yes	The base model to fine-tune (a catalog id).
`training`	No	Inline training data.
`validation`	No	Inline validation data.
`trainingFileId`	No	Id of an uploaded training file.
`validationFileId`	No	Id of a held-out file for validation metrics.
`name`	No	Human-readable name for the job.
`suffix`	No	Label appended to the resulting model name.
`seed`	No	For reproducibility.
`method`	No	`supervised` (LoRA or full) or `spec-draft`.
`hyperparameters`	No	Supervised training settings (see below).
`specDraftHyperparameters`	No	Settings for the `spec-draft` method.
`integrations`	No	External integrations for the job.

Supervised hyperparameters

All supervised hyperparameters are optional with sensible defaults; you only set what you want to override.

Training schedule

Field	Description
`n_epochs`	Number of passes over the training data.
`learning_rate`	Optimizer learning rate.
`batch_size`	Examples per training step.
`warmup_ratio`	Fraction of steps used to warm up the learning rate.
`weight_decay`	L2 regularization strength.
`max_grad_norm`	Gradient-clipping threshold.

Sequence handling

Field	Description
`context_length`	Maximum sequence length for training.
`packing`	Pack multiple short examples into one sequence.

LoRA

Field	Description
`lora`	`true` for LoRA adapter training, `false` for full fine-tuning.
`lora_r`	LoRA rank.
`lora_alpha`	LoRA alpha (scaling).
`lora_dropout`	LoRA dropout.

3. Track the job

List jobs with GET /v1/fine_tuning/jobs, and poll a single job with GET /v1/fine_tuning/jobs/{id} to watch its status and events:

curl https://gateway.omnia-voice.com/v1/fine_tuning/jobs/{id} \
  -H "Authorization: Bearer $OMNIA_API_KEY"

import requests

BASE = "https://gateway.omnia-voice.com/v1"
headers = {"Authorization": f"Bearer {OMNIA_API_KEY}"}

job = requests.get(f"{BASE}/fine_tuning/jobs/{id}", headers=headers).json()

Status progresses through queued → running → succeeded (or failed / cancelled). Events and checkpoints are available while the job runs. Cancel or delete a job with DELETE /v1/fine_tuning/jobs/{id}.

Billing happens once, on completion, from the trained-token count. Because trained_tokens isn’t known until the run finishes, the create-time gate uses an estimated-token budget: a workspace can’t launch training it can’t pay for. A job that fails or is cancelled is not billed.

Deploy the model

Once the job succeeds, deploy it to a dedicated endpoint.

Get started

Capture

Observe

Grade & Judge

Evaluate

Improve

Run at scale

Platform

Creating a job

1. Upload a training file

2. Create the job

Fields

Supervised hyperparameters

3. Track the job

Next

Deploy the model

​1. Upload a training file

​2. Create the job

​Fields

​Supervised hyperparameters

​3. Track the job

​Next

Deploy the model

1. Upload a training file

2. Create the job

Fields

Supervised hyperparameters

3. Track the job

Next