About the session
A technical and personal talk about training a LoRA adapter on years of Telegram conversations to see whether an LLM can learn a person's writing style, humor, recurring expressions, and group chat dynamics.
The talk would cover the full pipeline: exporting Telegram data, cleaning and structuring messy chat logs, preparing instruction-style datasets, choosing a base model, applying LoRA/QLoRA fine-tuning, running the training process, and comparing the fine-tuned model against the original base model.
Beyond the technical implementation, the talk also explores evaluation: how do you know if the model really sounds like you? I would show qualitative examples, failure modes, overfitting risks, privacy concerns, and the strange experience of interacting with an AI system trained on your own digital footprint.
The goal is to make LoRA fine-tuning concrete, practical, and memorable through a real personal use case rather than a generic benchmark.
