finetuning-scheduler

Overview

A PyTorch Lightning extension that provides flexible fine-tuning schedules for foundation model experimentation. It allows users to define and automate phase transitions during training, such as switching between frozen and unfrozen layers. The tool integrates directly into Lightning's training loop to manage schedule-driven parameter updates.

Best for

Best for
Researchers and engineers using PyTorch Lightning who need structured fine-tuning schedules for foundation models.

Use cases

• Defining multi-phase fine-tuning schedules with conditional transitions
• Automating layer freezing and unfreezing during model training
• Reproducing and comparing fine-tuning strategies across experiments

Notes

A PyTorch Lightning extension that provides flexible fine-tuning schedules for foundation model experimentation. It allows users to define and automate phase transitions during training, such as switching between frozen and unfrozen layers. The tool integrates directly into Lightning’s training loop to manage schedule-driven parameter updates.

69 stars on GitHub. Last updated 2026-01-26. Licensed Apache-2.0.

Use cases

• Defining multi-phase fine-tuning schedules with conditional transitions
• Automating layer freezing and unfreezing during model training
• Reproducing and comparing fine-tuning strategies across experiments

Pros

• Tight integration with PyTorch Lightning for minimal code changes
• Flexible schedule definitions support complex training strategies
• Open source with a permissive license

Cons

• Small community (69 GitHub stars) limits support and contributions
• Requires PyTorch Lightning as a dependency, not standalone
• Documentation and examples may be sparse for advanced use cases

Indexed from awesome-llmops and enriched against its public facts.