Is it able to select languages after training for generating text?
Hello Google's team.
Your model is very good in my language - vietnamese. Thank you all a lot.
I see that the model is trained in over 140 languages. This one makes the vocabulary size is pretty big, over 260.000 tokens.
During training, no problem, I think so. Training with a pair of languages may take much effort, time and cost, not efficient.
However, I wonder whether there is any way to configure the model so that after training, we can select some specific languages.
Only choose the languages we need, instead of all language.
For example, I only need the model generates English and Vietnamese, no need Korean, Japan, Chinese,... through re-mapping token index, maybe.
This way may reduce the vocab_size, help the model faster a bit. I can't run a large model in a weak hardware. I guess many people is like me.
I'm not an expert in training LLM model, but I hope that my concern is meaningful.
Thanks.
Hi @DuongLeVan ,
Welcome to Gemma family of Google's open source models, Thank you so much for your kind words about the model's performance in Vietnamese We're thrilled to hear that it's working well for you.
Instead of modifying the core vocabulary structure, the AI community uses several powerful post-training techniques to achieve efficiency and deploy large models on resource-constrained devices.
- Quantization
- Parameter-Efficient Fine-Tuning (PEFT)
- Model Distillation
While your idea to isolate the English and Vietnamese parts of the vocabulary is clever, techniques like Quantization (for smaller memory/faster runtime) and PEFT (for specialization) runtime) and PEFT (for specialization) are the established and most practical solutions to run large LLMs efficiently on weaker hardware today.
Thanks.