大语言模型入门

• What is a language model?
• Base models
• Tokens
• The ULMFiT 3-step approach, arxiv.org/abs/1801.06…
  • Step 1: LM pre-training
  • Step 2: LM fine-tuning
  • Instruction tuning
    • an e.g. dataset: huggingface.co/datasets/Op…
  • Step 3: Classifier fine-tuning
    • RLHF and friends
• What GPT 4 can do
  • can't reason, arxiv.org/abs/2308.03…
  • chat.openai.com/share/4211a…
  • chat.openai.com/share/323bb…
  • chat.openai.com/share/ce2f8…
• What GPT 4 can’t do
  • hallucinations
  • It doesn’t know about itself. (Why not?)
  • It doesn’t know about URLs.
  • Knowledge cut-off.
  • Bad pattern recognize. chat.openai.com/share/3051f…
• Advanced data analysis
  • coder interpreter: chat.openai.com/share/143a0…
  • OCR: chat.openai.com/share/2bb6c…
  • Creating price table
• The OpenAI API
  • Python SDK basis
  • Create our code interpreter
• PyTorch and Huggingface
  • Your GPUs
  • Retrieval augmented generation
  • Private GPTs: github.com/h2oai/h2ogp…
• Fine-tuning
  • Llama.cpp, github.com/ggerganov/l…, llama-cpp-python
  • MLC, llm.mlc.ai/docs/

video link

notebook link