LetsFineTune

A collection of fine-tuned language models built with LoRA/QLoRA, Unsloth, and Hugging Face Transformers.

Each model gets its own training pipeline — from dataset preparation to evaluation — with reproducible configs and notebooks.

Current Models

Model	Base	Dataset	Method	Status
Qwen3.5-9B Reasoning	Qwen3.5-9B	OpenThoughts-114k	LoRA (bf16, r=32, alpha=64)	Training

Project Structure

LetsFineTune/
├── configs/           # Training configs (YAML)
├── scripts/           # Training, evaluation, and data prep scripts
├── notebooks/         # Jupyter notebooks for experimentation
├── src/               # Reusable Python modules
│   ├── data/          # Data loading and preprocessing
│   ├── training/      # Training loops and utilities
│   └── evaluation/    # Eval metrics and benchmarks
├── models/            # Saved checkpoints (gitignored)
├── results/           # Evaluation results and metrics
└── data/              # Datasets (gitignored, see download instructions)

Quick Start

# Set up environment
conda activate slm-finetune
pip install -r requirements.txt

# Run a fine-tuning job
python scripts/train.py --config configs/lora_config.yaml

# Evaluate a checkpoint
python scripts/evaluate.py --model models/<checkpoint> --dataset data/processed/eval.jsonl

Tech Stack

• Frameworks: PyTorch, Hugging Face Transformers, PEFT, TRL, Unsloth
• Training: LoRA/QLoRA, bf16 mixed precision, bitsandbytes quantization
• Data: Hugging Face Datasets, pandas, NumPy
• Evaluation: ROUGE, BLEU (NLTK), EvalPlus (HumanEval), GSM8K
• Tracking: Weights & Biases, TensorBoard
• Visualization: Matplotlib
• Compute: CUDA-enabled GPUs (Accelerate)
• Environment: Conda, Jupyter Notebooks
• Code Quality: Ruff, pytest

License

MIT