Implement a minimal PPO update for LLMs: ratio, clipped surrogate, value loss, per-token K

Hard

RLHF: Reward Models, PPO & GRPO4 / 7

Chapter 124 / 7

RLHF: Reward Models, PPO & GRPO

Implement a minimal PPO update for LLMs: ratio, clipped surrogate, value loss, per-token K

Hard

Implement a minimal PPO update for LLMs: ratio, clipped surrogate, value loss, per-token KL penalty to a reference. [OpenAI]

Target interface

ppo_step(logp_new, logp_old, logp_ref, values_new, returns, advantages, mask, eps=0.2, vf_coef=0.5, kl_coef=0.1)

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
import torch.nn.functional as F

def ppo_step(logp_new, logp_old, logp_ref, values_new, returns, advantages, mask, eps=0.2, vf_coef=0.5, kl_coef=0.1):
    raise NotImplementedError

def _m(x, mask):
    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 124 / 7

RLHF: Reward Models, PPO & GRPO

Implement a minimal PPO update for LLMs: ratio, clipped surrogate, value loss, per-token K

Hard

Implement a minimal PPO update for LLMs: ratio, clipped surrogate, value loss, per-token KL penalty to a reference. [OpenAI]

Target interface

ppo_step(logp_new, logp_old, logp_ref, values_new, returns, advantages, mask, eps=0.2, vf_coef=0.5, kl_coef=0.1)

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

Hints0 / 2

Reference solution

Your own code stays in the editor.