Launch a Whisper Finetuning Job

❗️

Pre-requisites

Before you start, ensure that you meet the following prerequisites:

• Valid MonsterAPI Account: If you don’t have an account, Sign up here.
• Minimum 1,000 API Credits: If you haven’t purchased credits yet, Explore our plans.

Step-by-Step Guide

1. Open the Fine-Tuning Portal

   After logging into your account, navigate to the "Fine-Tuning" Portal from the left side menu. Click on the "Create New Finetuning Job" button and select “Finetune Whisper Model”.

   
   

2. Select a Whisper Model and Define Task and Language

   Choose a Whisper model from the dropdown menu that suits your needs. Options include models like whisper-large-v2, v3, tiny, medium, base, large, and distill whisper. Define the task as either Transcribe or Translate.

   
   

3. Select a Dataset

   • Option 1 - Use a Hugging Face Dataset:

     Provide the dataset name from Hugging Face. If access requires an HF key, supply it. If subsets are available, select them from the drop-down menu. Example datasets:

     • Mozilla Common Voice Dataset
     • English Dialects
     • Peoples Speech
     • Voxpopuli Dataset

   • Option 2 - Upload a Custom Dataset:
     Upload a zip file containing your custom dataset. If no subsets are available, the default subset will be selected.

📘 Convert ZIP of Audio Files to Parquet:
Use the provided Google Colab Notebook to convert your zip file into a Parquet dataset for MonsterAPI Whisper Finetuner. The notebook will handle the upload to Hugging Face and ensure compatibility.

4. Specify Hyperparameter Configuration

   Configure your hyper-parameters, such as epochs, learning rate, cutoff length, and warmup steps. These parameters are pre-filled based on your selected model but can be adjusted as needed. Provide HuggingFace and WandB credentials to upload the model and record training logs.

🚧

Incorrect hyperparameters can impact the fine-tuning process.

Click "Next" to review the summary page. Ensure all settings are correct and then submit your request. Your fine-tuning
job will start within a few minutes. Once it switches to the "IN PROGRESS" state, you can view job logs and metrics.

⚙️ Optional Settings

1. Track Your Fine-Tuning Job Using WandB:

   To track your job, add your WandB credentials in the third step:

   • WandB username
   • WandB key (get it here)
   • Project name (create a project in WandB if needed)
   • Run name (choose any name)

   Valid credentials will enable automatic tracking of metrics in your WandB project.

2. Upload Model Outputs to Hugging Face Repo:

   To store the fine-tuned model in Hugging Face, provide your credentials:

   • Hugging Face API Key (must have write access)
   • Hugging Face Repo Path
   
   

These credentials will allow the job to publish the fine-tuned model to your Hugging Face repository upon completion.