AI/Fine Tuning/Supervised

• SFT — Supervised Fine-Tuning

  • Definition:
    • SFT is the process of training a pre-trained language model further on a labeled dataset (input–output pairs) under supervision.
  • Goal:
    • To align the model's responses with desired behaviors or domain-specific knowledge.
  • How it works:
    • Start with a base model (e.g., pretrained on general web text).
    • Train it using curated examples — often human-written prompts and ideal completions.
    • The loss function (usually cross-entropy) penalizes deviation from the target outputs.
  • Example use cases:
    • Fine-tuning a general model into a customer-support agent.
    • Teaching a model to follow instructions or code correctly.
  • Example (simplified):

    • Prompt: Write a polite email requesting a refund.
      Target: Dear support team, I'd like to request a refund for my order...