Super-hard

Build a toy post-training stack with intentional bugs, then write tests that catch every o

Research Engineering & Debugging · Problem 7 of 8

All problems

Chapter 17Research Engineering & Debugging

Build a toy post-training stack with intentional bugs, then write tests that catch every o

Super-hardProblem 7 / 8

Build a toy post-training stack with intentional bugs, then write tests that catch every one.

Implement the function/class skeleton in the editor. Any correct approach is accepted.

Hints

solution.pypython

local draft

import os, torch, torch.nn.functional as F

def seq_logprob(model, ids, prompt_len, mask_bug=False):
    raise NotImplementedError

def dpo_loss(policy, ref, chosen, rejected, prompt_len, beta=0.1):
    raise NotImplementedError

⌘/Ctrl + ↵ to submit

AI review

Ready when you are

Submit your solution and a structured review appears here — verdict, score, and concrete feedback. Any correct approach passes.

Chapter 17Research Engineering & Debugging

Build a toy post-training stack with intentional bugs, then write tests that catch every o

Super-hardProblem 7 / 8

Build a toy post-training stack with intentional bugs, then write tests that catch every one.

Implement the function/class skeleton in the editor. Any correct approach is accepted.

Hints