Learning to be a good helper
After reading the library, a model can continue any text. But it is not a good helper yet.
Ask it a question, and it might just write more questions!
So people show it good answers, and they pick which of two answers is better.
The model learns to give the kind of answers people liked.
You can be the teacher. Pick the better answer.
Your turn
Read both answers. Tap the better one.
Toy calculation· a two-feature reward learned from your picksWhat colour is the sky?
What your model is learning to like
The middle line means "doesn't care yet". Each pick nudges the meters. Picks so far: 0.
People teach models to be helpful, honest and kind.
Grown-up corner: the precise version
- showing good answers = supervised fine-tuning
- picking the better answer = human feedback (RLHF)
A pretrained base model only continues text. Assistants are made by fine-tuning on example conversations and then reinforcement learning from human (or AI) preferences, using a reward model trained on comparisons like the ones here.
This toy reward has two features (does it help? is it kind?) and learns their weights from your picks with a Bradley-Terry update. Qwen3-8B-Base, the model LatentLens studies, is a base model: pretrained, not chat-tuned.