✳ BETWEEN QUESTIONS ML STUDY ATLAS

A QUESTION-LED LESSON

Preference Data: How Can We Learn from Comparing Answers?

Instead of writing an ideal answer, a reviewer can compare two candidate answers. These comparisons can become a learning signal.

What a record contains

A simple preference record includes a prompt, two responses and a preference label. The label reflects the annotation process and may be noisy, inconsistent or dependent on the instructions given to reviewers.

Modeling a comparison

One illustrative model assigns scores sA and sB and uses sigmoid(sA − sB) as the probability that A is preferred. Equal scores give 0.5. Scores here are arbitrary illustrative values, not measurements of quality. Make the two scores equal, then increase score A. The comparison changes because the score difference changes, not because either score proves factual correctness.

INTERACTIVE EXPLORATION

Compare two scores

Adjust score A while score B remains 0. The displayed sigmoid of their difference is an illustrative pairwise preference model.

How preferences are used

A learned reward model can use comparisons, and methods such as DPO use preference pairs in a different objective. Neither a single preference label nor a high model score guarantees that an answer is truthful or safe.

Check your understanding

If the two scores are equal, what does sigmoid(sA − sB) equal?

0.5, because the score difference is zero.

Keep exploring

Probability and expectation ↗Logistic regression ↗Train, validation and test ↗

Further reading

Source and further reading ↗