Fine-Tuning
Adapting a pretrained model with further training — I fine-tuned a time-series foundation model on GPU; LLM methods like LoRA and DPO come from coursework.
Fine-tuning continues training a pretrained model on task-specific data. Modern variants include instruction tuning, parameter-efficient methods such as LoRA, and preference optimization such as RLHF and Direct Preference Optimization (DPO).
How I use it
- In practice — Real-time Data Systems. I fine-tuned IBM Granite TSPulse, a time-series foundation model, on GPU in Google Colab with PyTorch, running hyperparameter grids and evaluating it against baselines, including gradient-boosted trees from the scikit-learn ecosystem. The datasets are versioned on Hugging Face.
- Coursework — LLMs. My knowledge of LLM fine-tuning comes from the IBM Generative AI Engineering Professional Certificate professional certificate — in particular Generative AI Engineering with Transformers & LLMs and Generative AI Advanced Fine-Tuning for LLMs, which cover instruction tuning, LoRA and DPO with Hugging Face and PyTorch. I haven't shipped a fine-tuned language model; my LLM work sits at the application layer — RAG (Retrieval-Augmented Generation), Prompt Engineering and agent runtimes.
Why it matters
Knowing what fine-tuning can and cannot fix helps choose the right tool: retrieval for fresh knowledge, prompting for behaviour, and training only when neither is enough.