Glossary · Term

SFT

← all terms

Definition

Plain language

Training a model by showing it examples of correct answers and having it imitate them.

As stated in the literature

Supervised Fine-Tuning, a post-training stage in which a model is trained on labeled demonstrations via standard next-token prediction.

Also called: supervised fine-tuning

Why it matters: It's the simplest and cheapest way to inject new behavior into a base model, and it's almost always the first step of any post-training pipeline.

For example, to teach a model to refuse harmful requests politely, you fine-tune it on thousands of (harmful prompt, polite refusal) pairs.

Heard on the show

“First, supervised fine-tuning on all those traces — that installs the shape of good reasoning, the structure of the loop.”
Episode 187 — An 8-Billion Agent That Beats Models 80 Times Its Size By Looking Things Up

Mentioned in 21 episodes

  1. 187
    An 8-Billion Agent That Beats Models 80 Times Its Size By Looking Things Up
  2. 167
    How Teaching an AI to Predict, Not Act, Made It a Better Actor
  3. 166
    A Router That Beats the Frontier Models It Calls
  4. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
  5. 156
    Why More Human Demonstrations Made a Computer-Use Agent Worse
  6. 141
    How Two Tokens Reopened a Reasoning Method the Field Had Given Up On
  7. 099
    How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes
  8. 091
    When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning
  9. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
  10. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
  11. 080
    How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents
  12. 066
    Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer
  13. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
  14. 047
    When Agent Benchmarks Lie: The Harness Problem in Open-Source AI
  15. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
  16. 022
    Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap
  17. 021
    Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents
  18. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't
  19. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
  20. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training
  21. 003
    How to Pick the Best of Sixteen Coding Agent Rollouts

Related concepts

Related terms