LLM-Fine-Tuning
A WhatsApp chat-style fine-tune that turns years of personal messages into a small LLM that texts like you, a hands-on tour through the real constraints of consumer-GPU training.
Can a model learn your texting personality?
The idea was simple: take years of WhatsApp messages, turn them into conversation pairs, fine-tune a small language model on them, and see if it can respond in your style, casual grammar, verbal tics, and all.
Export, preprocess into pairs, fine-tune, generate.
The pipeline exports chat history, runs a preprocessing script that cleans the junk and builds message→response pairs, fine-tunes a small model on those pairs with LoRA, and produces a model that reads a message and replies in-style.
The README is honest about the journey: an ambitious Phi-3.5 attempt hit out-of-memory on a free Colab T4 and an Unsloth/torchvision compatibility wall, before settling on regular HuggingFace transformers with TinyLlama, a realistic lesson in matching model size to available hardware.