Fine-tuning
Taking a pre-trained model and continuing to train it on a narrower dataset so its behaviour shifts towards that domain.
Last updated
Fine-tuning takes a pre-trained model and continues training it on a narrower, usually smaller dataset so its behaviour shifts towards that domain. It is cheaper than training from scratch but still requires labelled examples, compute, and careful evaluation.
Why it matters for your research. Fine-tuned models are sometimes held up as the answer to “our sources are too specialised for general LLMs”. They can help — but they reproduce whatever biases are in the labels you provide, and most DH projects lack the labelled data required. Off-the-shelf models, carefully chosen, are usually the better starting point.
In Archēglyph. Not used. Our extractive pipeline does not need a fine-tuned model; we pick embedding models and OCR engines that behave well on prose and swap them out as the state of the art moves.
Not to be confused with. Prompting — no training, you change behaviour by changing the text sent in. And RAG — no training, you add a retrieval step at inference.