This post is based on google cloud course
Define LLM
- subset of DL
- LLM : large, general purpose language models can be pre-trained and then fine-tuned for specific purposes
trained for commen language problems .. and then tailored to solve specific problems
Large -> large training set and large number of parameters
General purpose -> commonality of human languages
Pre-trained an fine-tuned
Benefits?
single model can be used for different tasks
fine-tune process requires minimal field data
continously growing performance
ex) PaLM , GPT, LaMDA
LLL use cases
- Question Answering (QA)
Prompt tuning
Tuning
process of adapting a model to a new domain of custom use casses. by training new data
3 main kinds of LLM -> each requries different way of prompting
- Generic language model : predict next word ex) auto complete, search
-
Instruction Tuned : predict a response to the instructions ex) summarize, writing (generate poam...), keyword extraction
-
Dialog Tuned : have dialog by predicting next response.
ex) chat bot
==> task specific tuning can make LLMs more reliable !
fine-tuning
retrain the pre-trained model by weighting -> expensive
althernative?
PETM parameter-efficient tuning methods
method for tuning LLM on own data. The base model is not changed. few add-on layers are tuned. ex) Prompt Tuning
Gen AI development tools
