Fine-tune the LLM for mnemonics

[StephanAkkerman]

1. Description:

   • Problem:
     Currently we use a very general LLM for generating the mnemonics. However, we are using 1 template for input sentences and want 1 sentence as output. Maybe we can make it more efficient by finetuning the LLM.

   • Solution:
     Look into the process of finetuning LLMs and see if it's possible for our use case.

   • Prerequisites:
     [List any requirements or dependencies needed before starting.]

2. Tasks:

   • Research finetuning of LLM
   • Find out if this works for us
   • Train the model (or add a new issue for this)

3. Additional context
   Add any other context or screenshots about the feature request here.

[StephanAkkerman]

https://www.reddit.com/r/LocalLLaMA/comments/1fm59kg/how_do_you_actually_finetune_a_llm_on_your_own/?rdt=61769

[TimKoornstra]

Or use LORA

[StephanAkkerman]

https://www.reddit.com/r/MachineLearning/s/Vj23oXeGuu

[StephanAkkerman]

Like at https://huggingface.co/spaces/k-mktr/gpu-poor-llm-arena for some more models

[StephanAkkerman]

Simple guide on how to implement LORAs with huggingface: https://medium.com/@kram254/train-large-language-models-with-lora-and-huggingface-d04be693dd7a

[StephanAkkerman]

Could make our own dataset scraping sources like:

• https://forum.artofmemory.com/t/mnemonics-for-spanish-vocabulary/34816 from https://artofmemory.com/language-learning/
• https://mnemonicdictionary.com/wordlist/GREwordlist

[StephanAkkerman]

Plan for generating the dataset:

• Scrape sources for mnemonics
• Ask SOTA model to create the sentence
• Save this dataset