Does training LLM on very high quality data might lead to bad performance on misspelled prompt?

Question

I've been wondering about the effects of training large language models (LLMs) like GPT on highly curated, clean datasets. If a model is trained predominantly on data with perfect grammar, spelling, and structure, would that make it less effective when handling user inputs that are less polished, such as prompts with misspellings, informal language, or typos?

It seems like there might be a trade-off here: focusing on high-quality data could help the model excel at generating accurate, professional responses but might also make it less adaptable to more "imperfect" language inputs.

Would such a model be more likely to struggle with understanding prompts that don’t match the formal language it was trained on? And if so, are there best practices for handling this, such as mixing in noisy or diverse data during training to improve robustness?

Answer 1

Most LLM's vast amount of balanced training data containing both high-quality formal inputs alongside informal ones allows them to create representations that capture the essence of various language styles. For instance, GPT-4 was trained on text from many different domains including Books and Academic Journals, Websites, Blogs, News Articles, Social Media and Forums. This balance helps it remain robust in understanding informal prompts but generally leads it to produce responses that are relevant and professional. The final optimization along with SFT including RLHF results in a model capable of interpreting informal language without being biased toward an informal or unstructured style in its outputs.

Furthermore byte pair encoding is the popular tokenization algorithm for most LLMs which breaks down words into common subwords and even individual characters if needed and thus helps capture the meaning of uncommon spellings which you're concerned about. For example, if a typo or slang term like "rel8ivity" appears, the model may still recognize "rel","8", and "ivity" as distinct tokens associated to "relativity" due to similar subword representations in LLM's enormous training corpus, most likely from "rel" and "ivity" tokens. This allows LLMs to further see through more minor informalities or typos and still associate the most probable relevant meanings.