> ## Documentation Index > Fetch the complete documentation index at: https://docs.simplismart.ai/llms.txt > Use this file to discover all available pages before exploring further. # Start a new Whisper training job > Submit a new Whisper fine-tuning job with dataset, model, training, and infrastructure configuration. ## OpenAPI ````yaml openapi-specs/playground/whisper-training.yaml POST /api/job/whisper-training/ openapi: 3.1.0 info: title: Whisper Training API description: >- API endpoints for training, monitoring, and managing Whisper speech-to-text fine-tuning jobs version: 1.0.0 servers: - url: https://training-suite.app.simplismart.ai description: Training Suite Production Server security: - BearerAuth: [] tags: - name: Whisper Training description: Whisper fine-tuning job endpoints paths: /api/job/whisper-training/: post: tags: - Whisper Training summary: Start a new Whisper training job description: >- Submit a new Whisper fine-tuning job with dataset, model, training, and infrastructure configuration. operationId: createWhisperTrainingJob parameters: - in: header name: Authorization schema: type: string required: true description: Bearer token for authentication and authorization. example: Bearer requestBody: required: true content: multipart/form-data: schema: type: object properties: org: type: string description: Organization UUID the job belongs to. example: your_org_uuid experiment_name: type: string description: >- Name for this training run. A short random suffix is appended automatically, and it must be unique (after suffixing) within the org. example: whisper-hindi-run train_type: type: string description: Training methodology. example: finetune language: type: string description: Spoken language of the audio. enum: - English - Hindi - Bengali - Marathi - Telugu - Tamil - Gujarati - Urdu - Kannada - Malayalam - Punjabi - Sanskrit - Assamese example: Hindi task: type: string enum: - transcribe - translate example: transcribe num_train_epochs: type: integer description: Number of training epochs. example: 3 learning_rate: type: number description: Learning rate. example: 0.0001 per_device_train_batch_size: type: integer description: Training batch size per device. example: 8 per_device_eval_batch_size: type: integer description: Evaluation batch size per device. example: 8 gradient_accumulation_steps: type: integer description: Gradient accumulation steps. example: 1 model_details: type: object description: >- Model configuration. Sent as a JSON-encoded string within the multipart body. properties: base_model: type: string description: >- Hugging Face model path of the base Whisper model to fine-tune. example: openai/whisper-tiny secret_id: type: string description: >- Hugging Face secret UUID, for a private/gated base model. Omit for a public model. required: - base_model infrastructure_configurations: type: object description: >- Infrastructure configuration for the training job. Sent as a JSON-encoded string within the multipart body. properties: infrastructure_type: type: string enum: - SS description: Must be `SS`. GPU_type: type: string description: GPU type. example: h100 node_count: type: integer enum: - 1 - 2 - 4 - 8 description: Defaults to `1`. gpu_count_per_node: type: integer enum: - 1 - 2 - 4 - 8 description: Defaults to `1`. required: - infrastructure_type - GPU_type dataset_details: type: object description: >- Describes the dataset to create for this job. Sent as a JSON-encoded string within the multipart body. properties: datasource: type: string enum: - AWS S3 - Hugging Face - File Upload dataset_name: type: string description: >- Name for the dataset. A short random suffix is appended automatically and must be unique within the org. dataset_path: type: string description: >- Required for AWS S3/Hugging Face. Not used for File Upload — the file itself is sent via `dataset_file`. dataset_type: type: string enum: - csv - jsonl description: Not needed for Hugging Face — inferred from the dataset. audio_column_name: type: string description: Name of the audio column. Use for CSV datasets. text_column_name: type: string description: Name of the text column. Use for CSV datasets. audio_key_name: type: string description: Name of the audio key. Use for JSONL datasets. text_key_name: type: string description: Name of the text key. Use for JSONL datasets. secret: type: string description: >- Secret UUID for a private AWS S3 bucket, a gated Hugging Face dataset, or to mark a File Upload dataset as private. Omit for public sources. region: type: string description: AWS region, for AWS S3 sources. required: - datasource - dataset_name dataset_file: type: string format: binary description: >- The dataset file itself. Required when `dataset_details.datasource` is `"File Upload"`, and must be omitted for any other datasource. train_split: type: string description: >- Hugging Face split expression for the training set. Only meaningful when the dataset source is Hugging Face. Never send together with `split_type`/`train_split_ratio`/`eval_split_ratio`. example: '"test.clean"' eval_split: type: string description: >- Hugging Face split expression for the eval set. Same rule as `train_split`. example: '"test.other"' split_type: type: string enum: - random - stratified description: >- Required when the dataset source is AWS S3 or File Upload. Must not be sent for Hugging Face datasets. train_split_ratio: type: number description: >- Fraction of data used for training. Required together with `split_type`/`eval_split_ratio` for AWS S3/File Upload datasets. `train_split_ratio + eval_split_ratio` must equal `1`. example: 0.9 eval_split_ratio: type: number description: >- Fraction of data used for evaluation. Same rule as `train_split_ratio`. `0` is a valid value — send it explicitly, don't omit the field. example: 0.1 required: - org - experiment_name - train_type - language - task - num_train_epochs - learning_rate - per_device_train_batch_size - per_device_eval_batch_size - gradient_accumulation_steps - model_details - infrastructure_configurations - dataset_details encoding: model_details: contentType: application/json infrastructure_configurations: contentType: application/json dataset_details: contentType: application/json responses: '201': description: Whisper training job created successfully. content: application/json: schema: $ref: '#/components/schemas/WhisperTrainingJob' '400': description: >- Validation error — missing/invalid field, or a conditional rule above was violated. Body contains per-field error details. '401': description: Unauthorized access due to invalid or missing authorization token. '500': description: Internal server error while creating the training job. components: schemas: WhisperTrainingJob: type: object properties: experiment_name: type: string description: Training experiment name (including the auto-appended suffix). request_id: type: string description: Job's unique identifier. language: type: string description: Spoken language of the audio. status: type: string enum: - queued - initializing - running - succeeded - failed - cancelled created_at: type: string format: date-time securitySchemes: BearerAuth: type: http scheme: bearer bearerFormat: JWT description: JWT token for authentication ````