Debug a broken GRPO implementation with wrong grouping, wrong masks, and wrong old-logprob

Hard

Correctness, Debugging, and Performance Experiments6 / 8

Chapter 266 / 8

Correctness, Debugging, and Performance Experiments

Debug a broken GRPO implementation with wrong grouping, wrong masks, and wrong old-logprob

Hard

Debug a broken GRPO implementation with wrong grouping, wrong masks, and wrong old-logprobs.

Target interface

grpo_loss(logits, old_logprobs, actions, group_ids, rewards, completion_mask, eps=0.2)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.

solution.pyPython

Local draft

import torch, torch.nn.functional as F

def grpo_loss(logits, old_logprobs, actions, group_ids, rewards, completion_mask, eps=0.2):
    raise NotImplementedError

Local draft

⌘/Ctrl↵to submit

AI review

Ready when you are

Submit your solution for a score and feedback on correctness, edge cases, and clarity. Any correct approach is accepted.

Chapter 266 / 8

Correctness, Debugging, and Performance Experiments

Debug a broken GRPO implementation with wrong grouping, wrong masks, and wrong old-logprob

Hard

Debug a broken GRPO implementation with wrong grouping, wrong masks, and wrong old-logprobs.

Target interface

grpo_loss(logits, old_logprobs, actions, group_ids, rewards, completion_mask, eps=0.2)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.