AGENCYSCRIPT
CoursesEnterpriseBlog
👑FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

What RLHF Actually DoesWhy Not Just Fine-Tune?The Three-Stage PipelineStage 1: Supervised Fine-Tuning (SFT)Stage 2: Reward Model TrainingStage 3: Reinforcement Learning Fine-TuningPrerequisites: What You Actually NeedChoosing Your Starting ModelThe Right Tooling for a First RunDesigning Your First Annotation TaskA Realistic First-Project ScopeCommon Failure Modes and How to Avoid ThemFrequently Asked QuestionsDo I need to train my own reward model, or can I use an existing one?How much compute does a first RLHF experiment realistically require?Is RLHF still the dominant alignment technique, or has it been superseded?What's the minimum annotation budget for a meaningful experiment?Can agencies offer RLHF as a service to clients?How do I know if my RLHF run actually worked?Key Takeaways
Home/Blog/What Makes a Model Follow Instructions, Not Just Predict
General

What Makes a Model Follow Instructions, Not Just Predict

A

Agency Script Editorial

Editorial Team

·March 24, 2026·10 min read

Reinforcement learning from human feedback is the technique that turned capable-but-erratic language models into tools people actually trust. If you've used a modern AI assistant and noticed it follows instructions rather than just predicts the next token, RLHF is a large part of why. Understanding it isn't just academic — agencies and professionals who grasp how models get aligned to human preferences will make better product decisions, write better prompts, and spot failure modes before they cost clients money.

The barrier to entry is real but manageable. RLHF sits at the intersection of reinforcement learning, supervised fine-tuning, and human annotation design — none of which require a PhD to engage with meaningfully. What trips most people up isn't the math; it's approaching the technique without a clear mental model of what each stage does and why it exists. This article gives you that model, then maps the shortest credible path from zero to a first hands-on result.

One clarification upfront: "getting started" means different things at different levels. A solo practitioner might run an open-source RLHF pipeline on a small model and a few hundred labeled examples. An agency operator might be designing a feedback collection system to fine-tune a vendor model through an API. Both count. Both are addressed here.

What RLHF Actually Does

RLHF is a training method that uses human preferences — not just human-labeled categories — to shape a model's behavior. The distinction matters. Standard supervised fine-tuning tells a model "this output is correct." RLHF tells it "humans prefer this output over that one," which is a richer signal when correctness is hard to define objectively.

The goal is to produce a model whose outputs consistently reflect what a thoughtful human evaluator would rank as better: more helpful, more accurate, more appropriately cautious. In practice, this means teaching the model to optimize for a learned human preference, not a hand-coded rule.

Why Not Just Fine-Tune?

Supervised fine-tuning on high-quality examples works well when you have a clear right answer. But many real tasks — summarization, tone adjustment, advice-giving, creative work — don't have a single correct output. Two acceptable answers exist, one is better, and the judgment is contextual. RLHF is built for exactly that scenario. It learns the shape of "better" from comparative human judgments rather than from gold-standard labels.

The Three-Stage Pipeline

Every mainstream RLHF implementation follows the same core structure. Knowing each stage lets you diagnose problems and make informed trade-offs.

Stage 1: Supervised Fine-Tuning (SFT)

Start with a pretrained base model. Fine-tune it on a curated dataset of (prompt, ideal response) pairs. This stage produces a model that can follow instructions at a basic level — it's the scaffolding that makes later stages tractable.

Quality matters more than quantity here. A few thousand well-chosen examples will outperform tens of thousands of noisy ones. The SFT model is your starting point; its weaknesses will amplify through subsequent stages.

Stage 2: Reward Model Training

Collect human preference data: show annotators two or more model responses to the same prompt, have them rank or choose the better one. Train a separate model — the reward model (RM) — to predict those preferences. The RM's job is to output a scalar score representing how much a hypothetical human would prefer a given response.

This is where most first projects underestimate the work. Annotation guidelines need to be precise enough that two different annotators agree most of the time (inter-annotator agreement above 70–75% is a reasonable target for early experiments). Vague guidelines produce noisy reward models, and noisy reward models produce reward hacking — the policy model learns to game the score rather than genuinely improve.

Stage 3: Reinforcement Learning Fine-Tuning

Use the reward model's scores as the reward signal to fine-tune the SFT model via a policy gradient algorithm — almost universally PPO (Proximal Policy Optimization) in the literature. The model generates outputs, the RM scores them, and the policy updates to produce higher-scoring outputs over time.

A KL-divergence penalty is applied to prevent the policy from drifting so far from the SFT model that it produces fluent nonsense that happens to score well. This penalty is a tunable hyperparameter and a common source of instability. Too low and you get reward hacking; too high and the RL stage does almost nothing.

Prerequisites: What You Actually Need

Don't skip this section. Starting RLHF without the right foundations wastes weeks.

Conceptual prerequisites:

  • Solid grasp of how transformer-based language models work (tokenization, attention, next-token prediction)
  • Basic understanding of supervised learning — loss functions, gradient descent, overfitting
  • Familiarity with what fine-tuning does at a conceptual level

If any of those feel shaky, spend a week with A Framework for Machine Learning Basics before touching an RLHF codebase. You'll move faster, not slower.

Practical prerequisites:

  • Python fluency (not expertise — reading and modifying existing scripts is enough)
  • Access to a GPU environment (Google Colab Pro, a rented A100 for a few hours, or a local workstation with at least 16GB VRAM)
  • Familiarity with HuggingFace's transformers and datasets libraries

Data prerequisites:

  • At minimum 500–1,000 preference pairs for a toy experiment; 5,000–50,000 for anything production-adjacent
  • A clear annotation rubric before you collect a single label

Choosing Your Starting Model

For a first experiment, smaller is better. A 1–7B parameter open-weight model (LLaMA 3, Mistral, Phi-3) lets you iterate quickly on consumer hardware or inexpensive cloud compute. Don't start with a 70B model to "get better results" — you'll spend your first month fighting memory constraints and miss the conceptual learning.

Pick a model that already has a reasonable SFT checkpoint if you can. HuggingFace Hub hosts dozens of instruction-tuned variants of popular base models. Starting from one of those skips Stage 1 or dramatically shortens it, letting you focus on the reward model and RL stages where the interesting learning happens.

The Right Tooling for a First Run

Three libraries handle the heavy lifting for most first RLHF projects:

  • TRL (HuggingFace) — the most accessible entry point. Wraps PPO training with sensible defaults and integrates cleanly with the HuggingFace ecosystem. Start here.
  • OpenRLHF — better for larger models and multi-GPU setups once you've outgrown TRL's defaults.
  • Axolotl — primarily for SFT, but excellent for that stage's data formatting and training stability.

For the reward model specifically, the simplest approach for beginners is to take your SFT model, add a scalar head, and train it on preference pairs using a Bradley-Terry loss. TRL's RewardTrainer does exactly this with minimal boilerplate.

The Best Tools for Machine Learning Basics covers the broader ML tooling landscape if you need context on where these libraries sit relative to the ecosystem.

Designing Your First Annotation Task

This is where most practitioners underinvest. Your reward model is only as good as your preference data, and your preference data is only as good as your annotation design.

Effective annotation guidelines do four things:

  • Define the evaluation criteria explicitly (helpfulness, accuracy, safety, tone — pick two or three, not ten)
  • Give annotators concrete examples of borderline cases
  • Specify how to handle ties (force a choice, or allow ties — be consistent)
  • State what to do when both responses are bad (common in early experiments)

Run a calibration round: have annotators label the same 50 pairs independently, then compare. If agreement is below 65%, the guidelines are too vague. Fix them before scaling.

For agencies building client-specific models, this annotation design phase is actually a high-value consulting deliverable — it forces clients to articulate what "good" looks like in their domain, which they often haven't done explicitly before.

A Realistic First-Project Scope

Aim for this in your first 4–6 weeks:

  1. Choose a narrow task (summarization, customer email tone, Q&A over a specific corpus)
  2. Collect or generate 200–500 SFT examples; fine-tune a 1–3B model for a few hours
  3. Generate model outputs for 100–200 prompts; collect preference labels (yourself or a small team)
  4. Train a reward model; verify it correlates with human intuition on a held-out set
  5. Run PPO for a few hundred to a few thousand steps; compare outputs before and after

Don't aim for a production system. Aim for a system where you can point to a specific behavior that changed because of your preference data. That's the learning. The Case Study: Machine Learning Basics in Practice illustrates how this kind of narrow, concrete first project translates into broader applied competence.

Common Failure Modes and How to Avoid Them

Reward hacking: The policy learns to produce verbose, over-hedged, or sycophantic responses that score well but aren't actually better. Mitigation: tighten annotation guidelines, lower the KL penalty gradually, and evaluate outputs qualitatively throughout training — not just by reward score.

Annotation drift: Annotators' standards shift over time, especially across sessions. Mitigation: re-run calibration pairs periodically; include "anchor" examples with known preferred labels in every batch.

SFT model too weak: If the base model can't follow basic instructions before RLHF, the RL stage will be chaotic. Mitigation: don't shortcut Stage 1. Verify SFT quality on a held-out prompt set before proceeding.

Treating reward score as ground truth: The reward model is a proxy for human preference, not a substitute. A rising reward score that doesn't correspond to visible quality improvement is a red flag, not a success metric. Always maintain a qualitative review loop.

Understanding these trade-offs before you hit them is exactly the kind of judgment the Machine Learning Basics: Trade-offs, Options, and How to Decide framework is designed to build.

Frequently Asked Questions

Do I need to train my own reward model, or can I use an existing one?

For learning purposes, training your own is worth the effort — it's where the conceptual understanding solidifies. For production use, API-based preference scoring (from models like GPT-4 used as a judge) is a legitimate shortcut, especially for smaller teams. The tradeoff is cost, latency, and reduced control over what the reward signal actually measures.

How much compute does a first RLHF experiment realistically require?

A toy experiment with a 1–3B model can run on a single A100 (40GB) in a few hours, costing $10–$30 in cloud compute. Scaling to 7B models and meaningful dataset sizes typically requires 8–16 hours on similar hardware. Budget for iteration — you'll rerun experiments as you tighten annotation guidelines and tune hyperparameters.

Is RLHF still the dominant alignment technique, or has it been superseded?

RLHF remains the most widely deployed alignment technique in production systems. Direct Preference Optimization (DPO) has emerged as a simpler alternative that skips the explicit reward model and RL loop, training directly on preference pairs. DPO is worth learning alongside RLHF — it's easier to implement and often competitive on narrow tasks, though RLHF gives finer-grained control for complex behavior shaping.

What's the minimum annotation budget for a meaningful experiment?

Five hundred preference pairs is a workable floor for demonstrating that the reward model learned something real. Below that, it's hard to separate signal from noise. For a reward model you'd actually trust in a production-adjacent setting, aim for 5,000–10,000 high-quality pairs at minimum.

Can agencies offer RLHF as a service to clients?

Yes, and it's becoming a differentiated offering. The annotation design, data collection infrastructure, and evaluation pipeline are where the agency value lives — not the raw model training, which is increasingly commoditized. Clients who need domain-specific behavior (legal tone, brand voice, compliance guardrails) are willing to pay for a credibly designed preference dataset and the fine-tuning process around it.

How do I know if my RLHF run actually worked?

Define success criteria before you start: a specific behavior that should increase (following formatting instructions, declining off-topic requests, matching a tone rubric) and a specific behavior that should decrease (hallucination on a test set, excessive hedging, policy violations). Evaluate on these criteria before and after the RL stage. A rising reward score with no observable behavioral change means something went wrong.

Key Takeaways

  • RLHF trains models on human preferences, not just correct labels — making it the right tool when "better" is contextual and hard to define as a single right answer.
  • The three-stage pipeline (SFT → Reward Model → RL Fine-Tuning) must be understood as a whole; weaknesses in any stage compound forward.
  • Annotation design is the highest-leverage investment in a first project — vague guidelines produce noisy reward models that undermine everything downstream.
  • Start small: a 1–3B model, a narrow task, and 500–1,000 preference pairs will teach you more than a poorly scoped large experiment.
  • TRL is the right entry-point library for most practitioners; pair it with a HuggingFace instruction-tuned checkpoint to skip or shorten Stage 1.
  • Reward hacking is the most common failure mode; always maintain qualitative evaluation alongside reward score tracking.
  • DPO is a credible alternative to full RLHF for simpler use cases — learn both and choose based on the control-complexity tradeoff your project actually needs.
  • For agencies, the annotation pipeline and evaluation design are the differentiated deliverables, not the model training itself.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Prompt Quality Decides Whether AI Earns Its Keep

Prompt quality is the single biggest variable in whether AI delivers real work or expensive noise. The model matters, the platform matters — but the prompt you write determines whether you get a first

A
Agency Script Editorial
June 1, 2026·10 min read
General

Counting the Real Cost of Every Token You Send

Tokens and context windows sit at the intersection of AI capability and operational cost—yet most business cases treat them as technical footnotes. That's a mistake that costs real money. Every time y

A
Agency Script Editorial
June 1, 2026·10 min read
General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification