Skip to main content

Flowcharts

The AI3 specific flowcharts can be found here, and the abbrevations used can be found below the flowcharts section.

Lookup DB​

  • The lookup database contains standard keys, called Phrase, and their Synonym. This is used to compare the keys from the JSON we receive in the subscriber and replace the keys to standard keys. This helps in efficiently using LLM models.
  • We create this in 2 ways:
    1. Old way: Call thesaurus.com to get the synonyms of keys found in the JSON and add any new keys not in the DB along with their synonyms.
    2. New way: Read the README provided along with the data and extract the keys as synonym, while the keys' description is the Phrase. And then calculate mean using the calculate_mean script.
    • TODO: One functionality to add is for new keys that are not in the table, we need to append them such that the keys are synonyms and their inference by LLM is the Phrase.

Fine-tuning LLaMA2 model​

This decsribes the porcess to finetune base LLaMA2 model from HuggingFace repository with checkpoints.

Standardization​

The below flow describes the flow for one JSON key in a message. This flow is recursive until we standardize all the keys.

Finetuned LLaMA2 model inference​

This describes the steps in order to use the finetuned model for text generation.

Subscriber​

This flowchart describes the flow of the subscriber we use to receive the data and perform operations like standardization and model inferencing.

Abbreviations​

  Abbreviation: Definition
  LLaMA: Llama-2-7b-chat-hf
  NIST: National Institute of Standards and Technology
  HF: Hugging Face
  LLM: Large Language Model
  DB: SQLite Database