Debug a broken GPT training notebook containing NaNs, a mask bug, and shifted labels

Hard

Research Engineering & Debugging5 / 8

Chapter 255 / 8

Research Engineering & Debugging

Debug a broken GPT training notebook containing NaNs, a mask bug, and shifted labels

Hard

Target interface

training_step(model, input_ids, attn_mask_bias)

Implement the function/class skeleton 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 training_step(model, input_ids, attn_mask_bias):
    raise NotImplementedError

⌘/Ctrl↵to submit

AI review

No review yet

Submit your solution for a verdict, a score out of 100 and concrete feedback. Any correct approach passes.

Chapter 255 / 8

Research Engineering & Debugging

Debug a broken GPT training notebook containing NaNs, a mask bug, and shifted labels

Hard

Target interface

training_step(model, input_ids, attn_mask_bias)

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

Hints0 / 2

Reference solution

Your own code stays in the editor.