Individual skill
rlhf-pipeline
Build the full reinforcement-learning-from-human-feedback stack — preference data, reward model, and policy optimization.
skills/rlhf-pipeline/SKILL.md
Source commit faf37c0859ebf627bfcb5225e8057d088691e0ae
What was checked
- Matches the observed repository HEAD
- Yes
- Scanner rules are current
- Yes
- Repository scan complete
- No
- Skill scan complete
- Yes
- Observed behavior
- Instructions only
- Last checked
- Nov 11, 2026
These are observations at the recorded commit, not a guarantee of safety. Review the source before running it.
Install this skill
Choose an agent in your Stack or copy a command below. Commands fetch the pinned commit and copy only this skill directory. Existing destinations are refused. Copying a command does not install anything.
An install command is unavailable: current rules, complete repository and skill scans, an eligible behavior level and a permitted license are required.
Repository context
- Repository stars (not skill ratings)
- 56
- Repository last updated
- Oct 10, 2026
- License
- MIT
Observed agent integration: No specific integration evidence. Standard SKILL.md installation does not imply verified compatibility.
- agent-skills Repository topic
- ai-agents Repository topic
- ai-agents-automation Repository topic
- ai-skills Repository topic
- awesome Repository topic
- awesome-list Repository topic
- chatgpt Repository topic
- claude-skills Repository topic
- meta-ai Repository topic
- muse-ai Repository topic
- data Derived category
- education Derived category
Files and excerpt
File metadata and permitted, redacted excerpts come from the recorded scan. Source links are pinned to that commit.
Browse files
SKILL.mdcore · 7,308 bytes
Read excerpt
## Overview
RLHF aligns a language model to human preferences in three stages. First, collect
preference data: human annotators rank or choose between pairs of model responses
to prompts. Second, train a reward model on these comparisons, typically with a
Bradley-Terry loss that predicts which response a human would prefer. Third,
optimize the policy — usually with PPO (proximal policy optimization) — to maximize
reward while staying close to the original model via a KL penalty, preventing
collapse into reward-hacking gibberish.
Variants replace pieces of this stack: DPO skips the reward model and optimizes
the policy directly on preference pairs; RLAIF uses AI-generated preferences;
online RLHF re-collects preferences as the policy changes. The classic three-stage
pipeline remains the reference design because each stage is independently