That's a really nice figure! I would say in addition to all the parameter-efficient finetuning techniques, one category that is missing is full finetuning. That is, just training the model as is without PEFT.
@rasbt
-
Community should adopt ‘finetuning’ over ‘fine-tuning’ terminology
By
–
That's a nice summary! Hah, one personal pet-peeve: I hope the community is moving from "fine-tuning" to "finetuning" similar to what happened to pretraining 😛
-
LIMA Dataset Support Added to Lit-GPT Framework
By
–
Finally got around adding LIMA to Lit-GPT (upon popular request). You can use it now to finetune any of the supported LLMs (Llama 2, Falcon, LongChat, … you name it). More usage info here: https://
github.com/Lightning-AI/l
it-gpt/blob/main/tutorials/prepare_dataset.md#lima
… -
100B Models Will Eventually Be Considered Small
By
–
> 100B (subset of the 500B) on top, that’s not really small One day it will be considered small 😛
-
Dataset Definition Determines Fine-tuning, Not Technique
By
–
Yeah, maybe it's the dataset that defines it, not the technique. I.e., finetuning is if it's using the dataset corresponding to the target task.
-
Questioning the terminology between training and finetuning
By
–
yes, fair, but don't we call this simply "more training"? I dunno, is this really "finetuning"?
-
Self-supervised Learning and Autoregressive Next-word Prediction Clarification
By
–
Self-supervised = autoregressive, next-word prediction like in pretraining? That's what I had in mind with "like pretraining" but in a shorter way to make it fit into the 260 character limit
-
Defining and Categorizing LLM Fine-tuning Approaches
By
–
How do we define and categorize LLM "finetuning" these days? There's 1) Supervised finetuning: to 1a) train an encoder-LLM for classification or 1b) train a decoder-LLM to generate a specific response text. [1/2]
-
Three types of fine-tuning approaches for language models
By
–
[2/2] 2) Instruction finetuning: usually RLHF to generate a specific response. (But can also refer to 1b depending on the technique) 3) Long-context finetuning: like pretraining but w longer contexts (as per Code Llama paper)?