NLP Terminology

• Zero-shot classification: Give labels and classify without retraining
• Few-shot classification: Give few examples of labelling and classify without retraining
• Summarisation
  • Extractive: Select representative pieces of text
  • Abstractive: Generate new text
• Sentiment Analysis
• Translation
• Question-Answers
• Foundation models: Trained on text generation tasks such as predicting the next token in a sequence
• Instruction-following models: Tuned to follow (almost) arbitrary instructions or prompts
• Prompt Hacking
  • Prompt injection: Adding malicious content
  • Prompt leaking: Extract sensitive information
  • Jailbreaking: Bypass moderation rule
• Searching and Sampling
  • Search: Given the tokens generated so far, pick the next most likely token in a "search."
    • Greedy search (default): Pick the single next most likely token in a greedy search.
    • Beam search: Greedy search can be extended via beam search, which searches down several sequence paths, via the parameter num_beams.
  • Sampling: Given the tokens generated so far, pick the next token by sampling from the predicted distribution of tokens.
    • Top-K sampling: The parameter top_k modifies sampling by limiting it to the k most likely tokens.
    • Top-p sampling: The parameter top_p modifies sampling by limiting it to the most likely tokens up to probability mass p.
  • For more background on search and sampling, see this Hugging Face blog post.

Vector Databases and Libraries

1. FAISS (video): Vector Library for generating embeddings and searching
2. Chroma db (video): Vector database, for generating embedding and searching. It can be combined with LLM to build a RAG
3. Pinecone (video): Cloud based vector database
4. Weaviate (video): Vector database

Multi-stage reasoning:

1. LangChain