Implement DAPO on a toy RLVR task: clip-higher, dynamic sampling (drop all-correct/all-wro

Super-hard

RLHF & Preference Optimization7 / 7

Chapter 127 / 7

RLHF & Preference Optimization

Implement DAPO on a toy RLVR task: clip-higher, dynamic sampling (drop all-correct/all-wro

Super-hard

Implement DAPO on a toy RLVR task: clip-higher, dynamic sampling (drop all-correct/all-wrong groups), token-level loss, and overlong reward shaping. [NVIDIA]

Target interface

overlong_shaping(reward, length, soft_cap, hard_cap, penalty=1.0)

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

def overlong_shaping(reward, length, soft_cap, hard_cap, penalty=1.0):
    raise NotImplementedError

def dapo_loss(logp_new, logp_old, rewards, lengths, mask, G, eps_low=0.2, eps_high=0.28, soft_cap=4, hard_cap=8):
    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 127 / 7

RLHF & Preference Optimization

Implement DAPO on a toy RLVR task: clip-higher, dynamic sampling (drop all-correct/all-wro

Super-hard

Implement DAPO on a toy RLVR task: clip-higher, dynamic sampling (drop all-correct/all-wrong groups), token-level loss, and overlong reward shaping. [NVIDIA]

Target interface

overlong_shaping(reward, length, soft_cap, hard_cap, penalty=1.0)

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

Hints0 / 2

Reference solution

Your own code stays in the editor.