Blog

Notes on adapters, fine-tuning, and running models on your own hardware.

Learning Rate Schedules for Adapter Training

How learning rate and schedule choice affect LoRA stability, convergence, and overfitting risk.

2024-12-02

QLoRA: 4-bit Training on Modest GPUs

Training LoRA on top of a 4-bit quantized base model to cut VRAM without hurting quality much.

2024-11-05

Merging a LoRA into the Base Model

When to bake an adapter into the base weights and when to keep them separate.

2024-10-11
Choosing Target Modules for LoRA Training cover

Choosing Target Modules for LoRA Training

How to decide which layers receive LoRA updates and why the choice changes training behavior and artifact size.

2024-09-23

Preparing Image Datasets for Style Adapters

How to curate image datasets for style LoRA training without polluting the signal with bad crops, bad captions, or bad balance.

2024-09-08

RAG or a Fine-Tuned Adapter: Choosing the Path

When retrieval augmentation is enough and when a small LoRA is the better tool.

2024-08-15

Captioning Image Datasets for Better LoRA Results

How to write image captions that help a LoRA learn the right visual concept without overfitting to generic tags.

2024-07-15

Train and Validation Splits That Catch Real Regressions

How to split datasets so validation actually measures generalization instead of leakage or duplicate memorization.

2024-07-01

Dataset Cleaning Checklist for Fine-Tuning

A practical checklist for cleaning instruction and chat datasets before a fine-tune job starts.

2024-06-10