How do you add knowledge to LLMs?

Question

I recently heard an interesting comment from a gentleman on YouTube and it made sense instantly.

To paraphrase he explained that "fine-tuning" an LM is not necessarily adding knowledge to a model - it is rather teaching that model to do or suggest things in the style of the fine-tuning data.

So then my question is what if one does not want to change the model too much - just wants to make a model aware of a new set of data i.e.: a companies internal domain knowledge? Is there a recommended strategy?

Edit: My thought is some sort of distillation where you train a smaller or same model against itself - plus the additional data.

Answer 1

Have a look at llamaindex or langchain for information injection in the prompt of the LLM: https://docs.llamaindex.ai/en/latest/index.html

This retrieve-augment-generate (RAG) works without Finetuning/prolonged pre-training but comes with the cost of having to pay for the additional tokens in your "pre-promt" which is where you inject your new information

Also note that

1. pretraining or fine-tuning modifies the parameters $\theta$ of your neural net $f_\theta(x)$.
2. RAG on the other hand improves/modifies the input $x$ supplied to your model.

Answer 2

From LIMA: Less Is More for Alignment (2023) by Chunting Zhou et al.:

These results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

From Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? (2024) by Zorik Gekhman, Gal Yona, Roee Aharoni, Matan Eyal, Amir Feder, Roi Reichart, Jonathan Herzig:

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. Our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

Remains prolonged pre-training and adding information in the prompt, as Ggjj11 suggested.

Answer 3

The common way is through providing a context to a prompt. GPT type of bots' APIs allow for providing large amount of text for a context, e.g. 64K in a bot that I use. Also look up retrieval-augmented generation (RAG).

The way I use this in my teaching is by supplying the text of my slides and lecture notes to the bot when conversing with it. So that a bot can use the context to come up with better answers than a general bot.