Creating A Training Job

​

Initiate a New Training Job

• Navigate to the My Trainings section in the platform.
• Click on Add a Training Job and select Whisper as the model type.

​

Basic Details and Model Selection

• Provide a name for your experiment.
• Enter the Hugging Face model path of the base Whisper model you wish to fine-tune (for example, openai/whisper-tiny).

​

Dataset Configuration

Provide a dataset in one of three ways:

• AWS S3 — point to a dataset file already in an S3 bucket.
• Hugging Face — point to an existing Hugging Face dataset.
• File Upload — upload a dataset file directly.

Dataset Requirements

• Supported file formats are CSV and JSONL (not shown for Hugging Face — it’s inferred from the dataset).
• You need an audio column and a text column — e.g. audio, text.
• For CSV, map them under Audio Column / Text Column. For JSONL, map them under Audio Key / Text Key.
• For AWS S3 and File Upload, set Train Split Ratio / Eval Split Ratio if your data isn’t pre-split.
• For Hugging Face, use Train Split / Eval Split to pass an HF split expression instead (e.g. test.clean / test.other).
• For AWS S3, if the bucket isn’t publicly accessible, attach a secret with the right credentials.
• For File Upload, attaching a secret marks the dataset as private — it doesn’t change how or where the file is uploaded.

​

Select Training Parameters

• Choose the Training Type: SFT for a standard full fine-tune, or LoRA for a parameter-efficient adapter-based fine-tune.
• Choose the Language and the Task (Transcribe or Translate).

Adapter Configuration (LoRA only)

When Training Type is LoRA, an Adapter Configuration section appears:

• Rank and Alpha — integers between 8 and 256.
• Dropout — a value between 0.1 and 0.9.
• Targets — one or more layers to apply the adapter on (all-linear, q_proj, k_proj, v_proj, out_proj).

Under Advanced, optionally configure QLoRA-style quantization — BNB 4 Bit Quant Type and Use BNB 4 Bit Double Quant both require Quant Bits to be set first:

• Quant Bits — 4 or 8.
• BNB 4 Bit Quant Type — nf4 or fp4.
• Use BNB 4 Bit Double Quant — use nested quantization for extra memory savings.

Update the training parameters based on your requirements for the training job: Number of Epochs: Total number of times the model trains on the complete dataset. Learning Rate: Controls how much the model adjusts its weights during training. Train / Eval Batch Size: Number of samples processed together per training/evaluation step. Available GPU RAM dictates the maximum batch size. Gradient Accumulation Steps: Number of steps to accumulate gradients before updating model weights — useful for simulating a larger effective batch size on limited GPU memory.

​

Select Infrastructure

Choose the GPU type, node count, and GPUs per node for your training job. Training runs on Simplismart-managed cloud infrastructure.

​

Start and Monitor the Training Job

Once the configuration is updated, start the training job and monitor its progress in the Recent Jobs section of the UI. Keep track of metrics, logs, and status to ensure the training meets your requirements.

​

Download, Compile, and Deploy Your Model

Once your job succeeds, a Compile button appears on the job, next to a ⋮ menu with a Download option — Download fetches the trained model weights directly, and Compile takes you to the Add Model page with your fine-tuned model already pre-filled as the source. From there:

• Choose the optimization infrastructure and configuration for the model — refer to the Add Model guide for details on these options.
• Click Add Model to save it to your My Models section.
• Once saved, deploy it via the Simplismart Model Suite — see Creating a Deployment for the deployment steps.