Join our growing team. See the job board →
Talk to us →

Why human feedback still matters for fine-tuning

Models learn from examples, but what counts as a better answer is a human judgement. A look at how expert comparisons turn taste and domain knowledge into a training signal.

By Faisal Kawar · · 2 min read

Two document outlines side by side on a green grid, the right one highlighted with a check mark

After a model has read most of the internet, it can write almost anything. What it cannot know on its own is which answer a careful person would prefer. That judgement has to come from people.

This is the idea behind human feedback, and it is still the most reliable way to steer a model toward the behaviour you want.

Better is a judgement, not a fact#

Ask two experts which of two answers is better and you will often get the same pick for different reasons. One liked the structure. Another spotted a subtle error in the third line.

A model cannot derive that from raw text. It needs many examples where a person looked at two outputs and chose, ideally with a short reason.

How the comparison works#

The process is simple to describe and hard to do well.

  1. Show a reviewer a prompt and two or more model answers
  2. Ask them to rank the answers, using written guidelines
  3. Collect a short note on why, especially for close calls
  4. Use the rankings as the training signal for fine-tuning

The quality of the guidelines and of the reviewers decides the quality of the result. A rushed comparison teaches the model to be confidently average.

Why domain experts change the outcome#

For general chat, a well-trained reviewer is enough. For code, medicine, law or finance, the person ranking must be able to tell a correct answer from a convincing one.

That is why we use vetted domain experts for high-value areas such as coding and data science. Their rankings carry information a generalist cannot add. See how this works in our human feedback service.

Keep the loop honest#

Two habits keep feedback useful:

  • Measure agreement. If reviewers often disagree, the guidelines are unclear, not the reviewers
  • Test on fresh cases. Check the tuned model on questions it never saw, with the same experts judging

Human feedback is not a replacement for good data or good evaluation. It is the part that tells the model what good looks like. Teams that treat it as a craft, with clear guidelines and real expertise, get models that behave the way their users expect.

← All postsView as Markdown

Keep reading

A target with green hit marks next to four bars measured against a pass line on a green grid
· 2 min read

How to evaluate an LLM beyond the benchmark

Public benchmarks tell you how a model does on someone else's questions. Here is how to build tests from your own work, probe for safety gaps, and decide when a model is ready to ship.

A central agent box connected by lines to seven smaller agent boxes on a green grid
· 2 min read

What an AI agent needs before it ships

A working demo is not a finished agent. Here are the four things we check, from a clear task to a way to know when it breaks, before an agent goes in front of real users.

Want to talk about your AI project?

Talk to us →See our work →