TANG617 / Memorize

Local Language Model Assisted Word Memorization

MEMORIZE: Local Language Model Assisted Word Memorization

• Optimize the workflow personally, display any information you need
• Deploy locally without requirements of any API
• Choose from a variety of LLMs, including LLAMA2, Phi, etc
• Adapt for multi-language including but not limited to English or Chinese

Project Struct

• source: Source jsonfile of vocabulary
• tex: Generated tex file
• pdf: Generated pdf file
  • noline: No dividing line between 2 words
  • nochinese: No Chinese meaning of the word itself

PipeLine

• Use Ollama(https://ollama.com) as the backend of LLM, which is capable of multiple LLM and hardware acceleration (like CUDA). Also, it's able to deploy LLM remotely and use REST API to access.
  • As for me, I choose to deploy Llama-2-7b with RTX3080 CUDA. The performance is satisfied, each phase or sentence takes 3-5 seconds.
• Source of the IELTS/GRE vocabulary comes from the Internet.

Example

	Word itself Upper number indicates the level of difficulty Sentences corresponding to its difficulty Phases corresponding to its difficulty

Recommandation

• Use English description rather than Chinese translation
• Guess the accurate meaning by sentences and phases
• Validate the guessing meaning by taking more sentences and phases into account

TODO

• More features
  • mark the level of memorization
  • real-time optimization
• Convert each option to paramter without modifying the code
• Better prompts for LLM
• Store generated text to database rather than stream into .tex file ( reduce failure)
• Convert to Anki format