that's nowadays literally a multi-million dollar question . Btw there are lots of (linear) approximations that came and went in the past few years, but it's still more or less the same olde because it just works so well!
@rasbt
-
Default approach in DPO experiments without extensive personal testing
By
–
I think it's a good default, but I am not an expert and haven't run (that) many DPO experiments myself.
-
Exercise: Convert Code to OCaml with LLM Assistance
By
–
Optional reader exercise: convert the code to OCaml . (Bonus points for training an LLM first for assisted coding )
-
DPO Preference Tuning from Scratch Notebook Resource
By
–
Ah yeah, and the notebook if you prefer a from-scratch approach: https://
github.com/rasbt/LLMs-fro
m-scratch/blob/main/ch07/04_preference-tuning-with-dpo/dpo-from-scratch.ipynb
… -
DPO from Scratch Notebook and Tulu 3 Length-Normalized Version
By
–
Nice, I should double check that to see if there's something worthwhile adding to my DPO from scratch notebook!
Btw one thing that intrigued me from the Tulu 3 paper is the length-normalized version https://
arxiv.org/abs/2501.00656 -
Congratulations and anticipation for Llama 4 release
By
–
Exciting! Big congrats!! (And I can't wait for Llama 4, make it awesome pls!)
-
Character-level tokenization inefficiency and token limit constraints
By
–
You could 100% do that. Actually, I have a paragraph on that in the post: the prob is that if you have a text of 100 characters, that would be 100 tokens (instead of ~20-30 tokens). In other words, it would be wasteful because you won't be able to input longer texts into the LLM.
-
Reader Project Training Tokenizer from Scratch on Nepali
By
–
And here's a great reader project who trained a tokenizer from scratch on Nepali:
-
Skeptical Take on Recent Research Paper Aging
By
–
I am tempted…But we will see if this ages well (just sayin https://
arxiv.org/abs/2412.09871 )