LORA for Embedding layer token expansion ???

[chapter544]

Hi,
By now we may all have heard about the Llama model. However, it was train on a limited set of languages. I want to "transfer" the Llama knowledge to a new language that it wasn't trained on, say Korean. My thought is to expand the Llama token vocab (~32K) by adding Korean tokens, resulting a bigger token embedding map. Then, I will freeze all Llama layers except the word embedding layer and the LMHead layer, to train a CausalLM model on Korean text. I only have a 3090 GPU so I am thinking about peft/LORA.

My question: is there an example or guide to do this? Like setting LORA module targets to an Embedding Layer?

Thanks,

[EeyoreLee]

LoRa is not suitable for this situation.

[pacman100]

Hello @chapter544, that is a nice idea and as per what you have described it might work. You can make the embedding layers additionally trainable and add lora layers to the respective attention subblocks of Llama. Please refer this comment #334 (comment) on adding additional modules as trianable through LoRAConfig.

[chapter544]

Thank you very much for the info.

[pacman100]

Hello @chapter544, see this: #337 (comment)