# start_training Submit a training job for a reference set: a LoRA, or a voice on a voice model. A voice is cloned from the set's audio clips, designed from a prompt, or imported from an existing ElevenLabs voice ID (voice-import). An import's run makes one test generation in the voice, and a voice ID ElevenLabs does not serve fails the run. A finished ElevenLabs voice is bound to every ElevenLabs model that speaks a voice; generate with it by passing the reference set in reference\_sets. Asynchronous: returns a training\_run\_id immediately. Errors: invalid-dataset (too few images / missing captions / no voice\_id on an import), unknown-base-model (id not trainable), unsupported-training-framework (framework not available for this base model), unsupported-training-parameter (a training\_parameters field the framework does not take), invalid-training-parameter (a training\_parameters value the model's trainer does not accept), model-disabled (base model disabled for this workspace), insufficient-balance (not enough Creative Units), and internal errors when the run fails to start. Other coded errors propagate unchanged. IMPORTANT: Call tool \`get\_instructions\` first for general platform guidance. ## Parameters | Parameter | Type | Default | Description | | -------------------------- | ------------------ | ------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | reference\_set\_idrequired | string (uuid) | — | The reference set (Style) to train a LoRA or a voice for. | | base\_model\_idrequired | string | — | Base model identifier from recommend\_training\_models\_for\_reference\_sets. | | primary\_audio\_file\_id | string (uuid) | — | The attached audio file to clone for a single-sample voice model such as Qwen 3 TTS or Seed Audio. Omit for other training models and when the reference set has only one audio clip. If it has multiple audio clips, ask the user which clip to use and pass that clip's file ID; never choose arbitrarily. | | training\_framework | enum | — | Training framework. OMIT this field for normal training — the server picks the correct default. Do NOT set it based on quality, speed, or 'best model' requests. Set 'ai-toolkit' ONLY if the user explicitly names ai-toolkit; in that case you MUST also set base\_model\_id to one of: 'flux-2-klein-4b', 'flux-2-klein-4b-edit', 'qwen-image', 'qwen-image-edit-2511', 'qwen-image-2512'. Never pass 'ai-toolkit' with any other base model. For a voice, pick by what the user has: 'voice-clone' (the default) clones the set's audio clips; 'voice-design' designs a voice from training\_parameters.training\_prompt; 'voice-import' adopts an existing ElevenLabs voice from training\_parameters.voice\_id, on base\_model\_id one of: 'eleven-labs-tts-v3', 'eleven-labs-tts-v4', 'eleven-labs-tts-v4-turbo'.One of: fal, ai-toolkit, voice-clone, voice-design, voice-import | | training\_parameters | TrainingOverrides | — | Settings for this run in place of the reference set's saved ones. Omit unless the user asks for specific values, or the voice framework needs one (voice\_id for voice-import, training\_prompt for voice-design). Each applies only where the run's framework offers it (repeats: kohya and ai-toolkit only); one the framework does not take is refused with unsupported-training-parameter, and a value outside what the model's trainer accepts with invalid-training-parameter. | | epochs | integer1–300 | — | Number of passes over the training images. | | repeats | integer1–50 | — | Times each image is seen within one epoch, so one epoch is (images x repeats) steps. Offered by the kohya and ai-toolkit frameworks only. | | learning\_rate | number≤0.01, >0 | — | Optimizer learning rate. | | lora\_rank | integer1–128 | — | LoRA rank (network dimension). | | training\_prompt | string1–1000 chars | — | A description of the voice to design: age, accent, tone, pacing. Offered by the voice-design framework only. | | voice\_id | string1–64 chars | — | The ElevenLabs voice ID to import: a voice from the public ElevenLabs voice library, or one made on Layer. A voice private to another ElevenLabs account can't be reached, and a voice ID that can't be used fails the run. Offered by the voice-import framework only. | | loudness | number\-1–1 | — | Loudness of the designed voice. Offered by ElevenLabs voice-design only. | | guidance\_scale | number0–100 | — | How closely the designed voice follows the prompt. Offered by ElevenLabs voice-design only. | | prompt\_strength | number0–1 | — | Weight of the prompt against a reference clip, when the reference set holds one. Offered by ElevenLabs voice-design only. | | workspace\_idrequired | string (uuid) | — | The user's workspace ID. | | project\_id | string (uuid) | — | Optional project to train in; use list\_projects to find project ids. The run follows that project's AI policy and the reference set is added to the project. |