Large Language Models (LLMs)

  • — What is a large language model?
    • — AI models that deal with language information trained huge billions of parameters.
    • — Transformers with selt attention is the most successful technique
  • — How they work?
    • — Basically matrix multiplication
    • — Turn tokens (split words) into vectors, that has meaning in high dimensional space
      • — Word embedding
    • — Process those vectors through different neural network layers
    • — Logits to output text
      • — Have to output one token at a time
    • — Pretraining and fine-tuning
      • — Tranined to predict one word given a list of word
      • — Then some of the layers can be changed and trained with smaller data for different use cases
    • — Can be part of a multimodal neural network
  • — Important concepts
    • — Tokenization
      • — Token to ID (number) mapping
      • — The map kept in memory as the same one is used for input and outputs
    • — Embedding
      • — Token ID to vector where the vectors have meaning
        • — token("queen") - token("female") + token("male") = token("king")
        • — token("Paris") - token("France") + token("Italy") = token("Rome")
    • — Attention mechanism
      • — Given a list of tokens, generating a table (like a heatamp) of which
    • — Transformers
      • — Use words in the context, to enrich each token with more meaning
        • — Done in many iterations
      • — KQV
        • — For each token you get key, query and value transformation
        • — To find the attention score for a token, you get the key for that word and query for the other word and get the product of the vectors
        • — Attention scores can be made to weights by normalization
        • — To transform the word, multiple all the token embeddings in the context with the attention score for that and add them up
      • — Multi-head attention
        • — Multiple attention head. Another dimension is added.
  • — Types of LLMs
    • — Translators
      • — String of tokens as input and string of tokens as output
      • — Encode with in one step with transformers, and decode until "endotext" token
    • — Generators
      • — Given a starting piece, predict the next token.
        • — Done recursively to generate long text
    • — Classifiers
      • — Like normal feed forward networks on top of attention head
    • — Chatbots
      • — Make a next word predictor
      • — Then fine tune with question and answer as one single text
      • — The whole conversation is one file that is recursively put to generate one token at a time
  • — Pretraining
    • — Take a long corpus of text
      • — [a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q]
    • — Make list of input and output for the context length
      • — [[a, b, c, d], [b, c, d, e]], [[e, f, g, h], [f, g, h, i]]
    • — Get [X, y] pairs for neural network
      • — [[a, *, *, *], b], [[a, b, *, *], c], ....
        • —
          • — is a padding to have causal attention
    • — Calculate loss with something like categorical crossentropy
    • — Backpropogate loss and adjust weights
    • — The model then can predict the next word give a list of word
  • — Finetuning for
  • — Famous models
    • — Word2Vec
    • — BART
    • — GPT
  • — Older ideas
    • — RNN
    • — LLM