FineTuneLab

Unified LoRA/QLoRA Fine-Tuning Framework

FineTuneLab combines UnSloth's efficient LoRA training pipeline with woct0rdho's GGUF-over-LoRA training methods, enabling fine-tuning from both safetensor and GGUF model formats on a single consumer GPU (RTX 4070 Super, 16GB VRAM).

Quick Start

pip install -e ".[dev,gguf,pandas]"
finetune-lab train --model model.gguf --dataset data.jsonl --output output/ --lora-rank 16

Architecture

FineTuneLab uses a dual-path loading strategy:

Model Source	Training Path	Inference Path
.safetensors	Unsloth FastModel (frozen + LoRA)	Unsloth FastModel
.gguf	GGUF Dequant-JIT (on-demand) + Frozen LoRA	llama.cpp native kernels

Modules

• finetune_lab.config — Dataclass-based configuration (LoRA, Training, Model, Data)
• finetune_lab.data — Multi-format dataset loading (CSV, JSONL, Parquet, DataFrame, HF Datasets)
• finetune_lab.models — Model loaders: SafetensorsModelLoader, GGUFModelLoader, LoRAMerger
• finetune_lab.training — LoRATrainer orchestration (safetensors via SFTTrainer, GGUF via custom loop)
• finetune_lab.export — GGUF export utilities for adapters and merged models

Reference

• docs/research_notes.md — Detailed research notes on Bonsai 1-bit quantization, GGUF Over-LoRA, and Unsloth integration.

License

Apache 2.0