Skip to content

Digital Hub

Introduction
Overview and architecture
Quick Start
Installation
Concepts and Tasks
Concepts and Tasks
Components
Components
- Landing Page
- Core UI
- CLI
- Kubernetes Resource Manager
- KubeAI and OpenWebUI
- Interactive Workspaces
  Interactive Workspaces
  - Code Toolbox
  - Community Workspaces
    Community Workspaces
    
    Dremio
    
    SQLPad
    
    Grafana
Scenarios and Tutorials
Scenarios and Tutorials
- Extract-Transform-Load
  Extract-Transform-Load
- DBT transformation
  DBT transformation
  - ETL with digitalhub-core and DBT scenario introduction
  - Workflow
- MLOps: Scikit-learn ML models
  MLOps: Scikit-learn ML models
- MLOps: MLflow ML models
  MLOps: MLflow ML models
  - MLFLow ML scenario introduction
  - Deploy and expose the model
- MLOps: Large Language Models
  MLOps: Large Language Models
  - Managing LLM Models
  - Managing LLM Models with KubeAI Runtime
- MLOps: Speech to Text
  MLOps: Speech to Text
  - Managing Speech-to-Text Models with KubeAI Runtime
- MLOps: Custom ML model training
  MLOps: Custom ML model training
  - Custom ML scenario introduction
  - Deploy and expose the model
- Dremio transformation
  Dremio transformation
  - Data transformation and usage with Dremio and Grafana
- API via PostgREST
  API via PostgREST
- Retrieval-Augmented Generation
  Retrieval-Augmented Generation
- Whisper fine-tuning
  Whisper fine-tuning
  - Fine-tuning speech-to-text Fine-tuning speech-to-text
    Table of contents
    
    Requirements
    
    Set-up
  - Create and process dataset
  - Fine-tuning
  - Convert to faster-whisper and test
- Flower Federated Learning
  Flower Federated Learning
  - Federated Learning with Flower
- Data validation
  Data validation
  - Data validation
Runtimes reference
Runtimes reference
SDK Reference
SDK Reference
- Python SDK Reference
CLI Reference
CLI Reference
- CLI commands

Fine-tuning speech-to-text

In this scenario, we fine-tune Whisper, a model for speech-to-text recognition.

Requirements

You'll need a HuggingFace token that has access to the voice library we will use. Enable your token to access this repository.

Create a workspace on Coder for Jupyter. When the workspace is up, access Jupyter and create a new notebook.

Set-up

Import the platform's library and create a project:

import digitalhub as dh

project = dh.get_or_create_project("whisper-fine-tuning")

Create a secret as follows, make sure you replace the value with the token that has access to the aforementioned repository:

project.new_secret(name="HF_TOKEN", secret_value="my-token")

The functions we will run use code from a Python file. Due to the many lines this file has, instead of presenting it in this documentation, we invite you to download it from the subfolder related to this scenario. Use the src directory and ensure it is at the same level of the notebook you're using.