Modelpublic

kuotient/Llama-3-8B-Instruct-vector-diff

sourceHugging Faceotherupdated 2y agoView on Hugging Face

4likes23downloads

Model Card

Llama-3 chat vector

—Update 0426: A small problem with the deployment of the model 'Llama-3-Seagull-Evo-8B', but we hope to have it back in good time!
—Update 0526: Check our newest EMM model, Alpha-Ko-8B-Instruct

This is 'modelified' version of chat vector from the paper Chat Vector: A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages. So this is not a model, its just weight diff, just for ease to use myself(or you too)!

What I understand here: 'Chat vector method' is a merging method that utilizes the difference between the base model, the continuously pre-trained (usually language transferred) model, and the chat model; so the recipe is

model(base) + weight_diff(continous pretrained) + weight_diff(instruct) or

model(base) + weight_diff(continous pretrained + fine-tuned) + weight_diff(instruct).

So before (my) initial purpose in comparing which method is better, llama3 → CP + chat vector → FT vs. llama3 → CP → FT + chat vector, it seems reasonable to compare it with other methods in Mergekit.

Model	Method	Kobest(f1)	Haerae(acc)
beomi/Llama-3-Open-Ko-8B-Instruct-preview	chat vector	0.4368	0.439
kuotient/Llama-3-Ko-8B-ties	Ties	0.4821	0.5160
kuotient/Llama-3-Ko-8B-dare-ties	Dare-ties	0.4950	0.5399
kuotient/Llama-3-Ko-8B-TA	Task Arithmetic(maybe...? not sure about this)	-
WIP	Model stock(I don't read this paper yet but still)	-
kuotient/Llama-3-Seagull-Evo-8B	Evolutionary Model Merging	0.6139	0.5344
---	---	---	---
meta-llama/Meta-Llama-3-8B	Base	-	-
meta-llama/Meta-Llama-3-8B-Instruct	-	0.4239	0.4931
beomi/Llama-3-Open-Ko-8B	Korean Base	0.4374	0.3813

All that aside, I'd like to thank @beomi for creating such an awesome korean-based model.