hAIgopala › AI terms › Rapid Fire (3 terms per short) › Day 25

GRPO

Model CustomisationChapter 7: Speaking in the School's VoiceTerm 123 of 253

RL method (used by DeepSeek) that compares a group of sampled answers to compute rewards, no separate value model.

At Sunrise School

Gopal asks

How do reasoning models get better at fee maths?

Bit answers

GRPO: I try 8 solutions, they're scored against each other, and I learn from the best in the group.

In the real world

Grade 8 attempts relative to each other.

Watch the short

Short coming soonSeason 5 · Day 25Follow @haigopala

Taught together with

DPO (Direct Preference Optimisation) · Distillation

Practise it

Try this term and all 253 others in the free Model Training Arena.

Open the Arena →

Follow Bit daily: YouTube English · తెలుగు · हिंदी  |  Instagram English · తెలుగు · हिंदी