Overview
What this challenge is about.
You receive a small open-weights base model (around 7B), a previously trained reward model, and 5,000 prompts (no responses) for PPO rollouts. Run PPO with TRL's PPOTrainer for a fixed compute budget (24 GPU-hours). Save checkpoints every 500 steps. Evaluate each checkpoint on (a) reward-model score (training reward), (b) a held-out 200-prompt human-judgement set (preferences vs. the base model judged by 2 raters), and (c) reward-hacking diagnostics (response length growth, n-gram repetition, KL divergence from base). Recommend the best checkpoint and explain why it is not necessarily the highest-reward one.
The Brief
What you'll do, and what you'll demonstrate.
Run PPO RLHF and pick the best checkpoint via held-out human judgement plus reward-hacking diagnostics, not by training reward alone.
Earning criteria — what you'll demonstrate
- Run end-to-end PPO RLHF with a pretrained reward model
- Apply KL-divergence regularization to balance reward and base-model fidelity
- Detect reward-hacking via length, repetition, and KL diagnostics
- Choose the right checkpoint based on multi-metric trade-offs
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
ML Researcher
Running PPO RLHF end-to-end with reward-hacking analysis is the canonical post-training researcher job at AI labs and AI startups in 2024-25.
This challenge sharpens
- rlhf
- ppo
- reward-hacking
AI Safety Researcher
Reward-hacking detection and KL-control discipline are exactly the alignment-research skills safety teams hire for.
This challenge sharpens
- reward-hacking
- kl-control
- rlhf
Research Scientist
Multi-metric checkpoint selection plus methodology write-up is the rigor research-scientist roles look for.
This challenge sharpens
- evaluation
- rlhf
- ppo