RLHF: How ChatGPT Was Trained

RLHF: How ChatGPT Was Trained

Reinforcement Learning from Human Feedback (RLHF) is the technique behind ChatGPT, Claude, and other conversational AI systems. Instead of predicting next tokens (standard language modeling), RLHF trains models to optimize human preferences. A reward model learns to predict human preferences, then policy gradient methods optimize the language model against this reward.

The RLHF Pipeline: Three Stages

Stage 1: Supervised Fine-Tuning (SFT) - train on high-quality examples. Stage 2: Reward Modeling - train a model to predict human preference (which response is better?). Stage 3: RL - optimize language model using policy gradient to maximize predicted rewards. This pipeline transforms a base language model into one that follows instructions and responds to human preferences.

import numpy as np
import matplotlib.pyplot as plt

# Conceptual RLHF pipeline
print("RLHF Training Pipeline for Instruction-Following LLMs")
print("="*60)

print("
Stage 1: Supervised Fine-Tuning (SFT)")
print("-" * 60)
print("Input: Base LLM + High-quality (prompt, response) pairs")
print("Goal: Adapt model to follow instructions")
print("Method: Standard next-token prediction on curated data")
print("Result: SFT model that generates reasonable responses")

print("
Stage 2: Reward Modeling")
print("-" * 60)
print("Input: (prompt, response_A, response_B, preference)")
print("Goal: Learn to predict which response humans prefer")
print("Method: Binary classification - does model prefer A or B?")
print("Training data: ~5k-10k comparisons, ranked by annotators")
print("Result: Reward model r(prompt, response) -> scalar score")

print("
Stage 3: Reinforcement Learning")
print("-" * 60)
print("Input: SFT model + Reward model")
print("Goal: Optimize LLM to maximize predicted rewards")
print("Method: PPO (Proximal Policy Optimization)")
print("Loss: L = E[log π(response|prompt) × r(prompt, response)]")
print("- Subtract KL penalty to stay close to original model")
print("Result: RL-trained model that follows preferences")

# Illustration: reward landscape
fig, axes = plt.subplots(2, 2, figsize=(12, 10))

# Stage 1: SFT loss
iterations = np.arange(100)
sft_loss = 5.0 * np.exp(-iterations / 30) + 0.5
axes[0, 0].plot(iterations, sft_loss, linewidth=2, color='blue')
axes[0, 0].set_title('Stage 1: SFT - Supervised Fine-Tuning')
axes[0, 0].set_xlabel('Training Iteration')
axes[0, 0].set_ylabel('Loss (Cross-Entropy)')
axes[0, 0].grid(True, alpha=0.3)
axes[0, 0].text(50, 4, 'Minimize prediction loss
on curated examples', fontsize=10)

# Stage 2: Reward model accuracy
rm_acc = 0.55 + 0.35 * (1 - np.exp(-iterations / 20))
axes[0, 1].plot(iterations, rm_acc, linewidth=2, color='green')
axes[0, 1].axhline(0.7, color='red', linestyle='--', alpha=0.5, label='Target accuracy')
axes[0, 1].set_title('Stage 2: Reward Model Training')
axes[0, 1].set_xlabel('Training Iteration')
axes[0, 1].set_ylabel('Accuracy (Preference Prediction)')
axes[0, 1].set_ylim(0.5, 1.0)
axes[0, 1].grid(True, alpha=0.3)
axes[0, 1].legend()

# Stage 3: RL reward (with KL penalty)
rl_iterations = np.arange(200)
kl_weight = 0.5
reward_signal = 0.5 * (1 - np.exp(-rl_iterations / 50)) - kl_weight * 0.002 * rl_iterations
axes[1, 0].plot(rl_iterations, reward_signal, linewidth=2, color='orange')
axes[1, 0].axhline(0, color='black', linestyle='-', alpha=0.2)
axes[1, 0].set_title('Stage 3: RL Training with KL Penalty')
axes[1, 0].set_xlabel('RL Step')
axes[1, 0].set_ylabel('Objective (Reward - KL Penalty)')
axes[1, 0].grid(True, alpha=0.3)
axes[1, 0].text(100, 0.3, 'Maximize reward while
staying close to SFT', fontsize=10)

# Summary table
ax = axes[1, 1]
ax.axis('off')

summary_text = """
RLHF Key Concepts:

ADVANTAGE: r(s,a) = r_reward(s,a) - β × KL(π_RL || π_SFT)
  • r_reward: predicted human preference
  • KL: divergence from SFT model (prevents collapse)
  • β: regularization coefficient (~0.5)

PPO Update: θ ← θ + ∇ E[min(rt·Ât, clip(rt,1-ε,1+ε)·Ât)]
  • Ât: advantage estimate
  • rt: probability ratio (new/old)
  • clip: prevent too-large updates

Challenges:
  ✗ Reward model can be gamed (adversarial examples)
  ✗ Expensive: requires human annotations
  ✗ Training instability (RL is notoriously finicky)
  ✗ Potential value drift (model optimizes metric, not true goal)
"""

ax.text(0.05, 0.95, summary_text, transform=ax.transAxes,
        fontsize=10, verticalalignment='top', family='monospace')

plt.tight_layout()
plt.show()

Practical Implementation: PPO for Language Models

The RL stage uses Proximal Policy Optimization (PPO), which improves stability over pure policy gradient methods. PPO clips probability ratios to prevent too-large updates: Loss = -E[min(r_t × A_t, clip(r_t, 1-ε, 1+ε) × A_t)], where r_t is the probability ratio and A_t is advantage estimate.

import numpy as np
import torch
import torch.nn as nn

class LanguageModelPPO:
    """Simplified PPO for language model RL fine-tuning"""
    def __init__(self, model, reward_model, learning_rate=1e-6, epsilon=0.2):
        self.model = model
        self.reward_model = reward_model
        self.lr = learning_rate
        self.epsilon = epsilon  # Clipping range
        self.optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

    def compute_advantage(self, rewards, values, gamma=0.99):
        """Compute generalized advantage estimate"""
        advantages = []
        gae = 0
        for t in reversed(range(len(rewards))):
            delta = rewards[t] + gamma * (values[t+1] if t < len(values)-1 else 0) - values[t]
            gae = delta + gamma * 0.95 * gae  # gamma-weighted advantage
            advantages.insert(0, gae)
        return torch.tensor(advantages, dtype=torch.float32)

    def ppo_loss(self, log_probs_old, log_probs_new, advantages):
        """PPO clipped loss"""
        ratio = torch.exp(log_probs_new - log_probs_old)
        surr1 = ratio * advantages
        surr2 = torch.clamp(ratio, 1-self.epsilon, 1+self.epsilon) * advantages
        return -torch.min(surr1, surr2).mean()

    def update(self, prompts, responses, epochs=3):
        """RL update step"""
        for epoch in range(epochs):
            # Get model outputs
            log_probs_old = self.model.log_prob(prompts, responses)

            # Get reward predictions
            with torch.no_grad():
                rewards = self.reward_model(prompts, responses)
                values = self.model.value_estimate(prompts)

            # Compute advantages
            advantages = self.compute_advantage(rewards, values)

            # Get new log probs
            log_probs_new = self.model.log_prob(prompts, responses)

            # PPO loss
            loss = self.ppo_loss(log_probs_old, log_probs_new, advantages)

            # Optimize
            self.optimizer.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
            self.optimizer.step()

print("PPO for LLM RL Fine-tuning:")
print("- Generates responses using current policy")
print("- Rewards evaluated by reward model (trained in stage 2)")
print("- Clipped PPO prevents too-large policy updates")
print("- KL penalty keeps model close to SFT version")
print("- Multiple epochs on same batch for sample efficiency")

Challenges: Reward Hacking and Alignment

RLHF faces fundamental challenges. Reward models can have blind spots the policy exploits (adversarial examples). Models may achieve high reward through unintended means (reward hacking). Scaling RLHF beyond human feedback is an open problem—Constitutional AI and other approaches aim to improve scalability and robustness.

Key Takeaways

• RLHF has three stages: SFT (supervised fine-tuning), reward modeling, and RL optimization.
• Reward model predicts human preferences; RL optimizes language model to maximize rewards.
• PPO is the standard RL algorithm for language models; clipping prevents destabilizing updates.
• KL penalty keeps model close to SFT version, preventing reward specification gaming.
• Reward hacking (achieving high reward through unintended means) is an open challenge.
• RLHF is expensive; scaling and improving it remains an active research area.

Deep Dive: RLHF: How ChatGPT Was Trained

At this level, we stop simplifying and start engaging with the real complexity of RLHF: How ChatGPT Was Trained. In production systems at companies like Flipkart, Razorpay, or Swiggy — all Indian companies processing millions of transactions daily — the concepts in this chapter are not academic exercises. They are engineering decisions that affect system reliability, user experience, and ultimately, business success.

The Indian tech ecosystem is at an inflection point. With initiatives like Digital India and India Stack (Aadhaar, UPI, DigiLocker), the country has built technology infrastructure that is genuinely world-leading. Understanding the technical foundations behind these systems — which is what this chapter covers — positions you to contribute to the next generation of Indian technology innovation.

Whether you are preparing for JEE, GATE, campus placements, or building your own products, the depth of understanding we develop here will serve you well. Let us go beyond surface-level knowledge.

Transformer Architecture: The Engine Behind GPT and Modern AI

The Transformer architecture, introduced in the landmark 2017 paper "Attention Is All You Need," revolutionised NLP and eventually all of deep learning. Here is the core mechanism:

# Self-Attention Mechanism (simplified)
import numpy as np

def self_attention(Q, K, V, d_k):
    """
    Q (Query): What am I looking for?
    K (Key):   What do I contain?
    V (Value): What do I actually provide?
    d_k:       Dimension of keys (for scaling)
    """
    # Step 1: Compute attention scores
    scores = np.matmul(Q, K.T) / np.sqrt(d_k)

    # Step 2: Softmax to get probabilities
    attention_weights = softmax(scores)

    # Step 3: Weighted sum of values
    output = np.matmul(attention_weights, V)
    return output

# Multi-Head Attention: Run multiple attention heads in parallel
# Each head learns different relationships:
# Head 1: syntactic relationships (subject-verb agreement)
# Head 2: semantic relationships (word meanings)
# Head 3: positional relationships (word order)
# Head 4: coreference (pronoun → noun it refers to)

The key insight of self-attention is that every token can attend to every other token simultaneously (unlike RNNs which process sequentially). This parallelism enables efficient GPU training. The computational complexity is O(n²·d) where n is sequence length and d is dimension, which is why context windows are a major engineering challenge.

State-of-the-art developments include: sparse attention (reducing O(n²) to O(n·√n)), mixture of experts (MoE — activating only a subset of parameters per input), retrieval-augmented generation (RAG — grounding responses in external documents), and constitutional AI (alignment through principles rather than RLHF alone). Indian researchers at institutions like IIT Bombay, IISc Bangalore, and Microsoft Research India are actively contributing to these frontiers.

India's Scale Challenges: Engineering for 1.4 Billion

Building technology for India presents unique engineering challenges that make it one of the most interesting markets in the world. UPI handles 10 billion transactions per month — more than all credit card transactions in the US combined. Aadhaar authenticates 100 million identities daily. Jio's network serves 400 million subscribers across 22 telecom circles. Hotstar streamed IPL to 50 million concurrent viewers — a world record. Each of these systems must handle India's diversity: 22 official languages, 28 states with different regulations, massive urban-rural connectivity gaps, and price-sensitive users expecting everything to work on ₹7,000 smartphones over patchy 4G connections. This is why Indian engineers are globally respected — if you can build systems that work in India, they will work anywhere.

Advanced Algorithms: Dynamic Programming and Graph Theory

Dynamic Programming (DP) solves complex problems by breaking them into overlapping subproblems. This is a favourite in competitive programming and interviews:

# Longest Common Subsequence — classic DP problem
# Used in: diff tools, DNA sequence alignment, version control

def lcs(s1, s2):
    m, n = len(s1), len(s2)
    dp = [[0] * (n + 1) for _ in range(m + 1)]

    for i in range(1, m + 1):
        for j in range(1, n + 1):
            if s1[i-1] == s2[j-1]:
                dp[i][j] = dp[i-1][j-1] + 1
            else:
                dp[i][j] = max(dp[i-1][j], dp[i][j-1])

    return dp[m][n]

# Dijkstra's Shortest Path — used by Google Maps!
import heapq

def dijkstra(graph, start):
    dist = {node: float('inf') for node in graph}
    dist[start] = 0
    pq = [(0, start)]  # (distance, node)

    while pq:
        d, u = heapq.heappop(pq)
        if d > dist[u]:
            continue
        for v, weight in graph[u]:
            if dist[u] + weight < dist[v]:
                dist[v] = dist[u] + weight
                heapq.heappush(pq, (dist[v], v))

    return dist

# Real use: Google Maps finding shortest route from
# Connaught Place to India Gate, considering traffic weights

Dijkstra's algorithm is how mapping applications find optimal routes. When you ask Google Maps to navigate from Mumbai to Pune, it models the road network as a weighted graph (intersections are nodes, roads are edges, travel time is weight) and runs a variant of Dijkstra's algorithm. Indian highways, city roads, and even railway networks can all be modelled this way. IRCTC's route optimisation for trains across 13,000+ stations uses graph algorithms at its core.

Research Frontiers and Open Problems in RLHF: How ChatGPT Was Trained

Beyond production engineering, rlhf: how chatgpt was trained connects to active research frontiers where fundamental questions remain open. These are problems where your generation of computer scientists will make breakthroughs.

Quantum computing threatens to upend many of our assumptions. Shor's algorithm can factor large numbers efficiently on a quantum computer, which would break RSA encryption — the foundation of internet security. Post-quantum cryptography is an active research area, with NIST standardising new algorithms (CRYSTALS-Kyber, CRYSTALS-Dilithium) that resist quantum attacks. Indian researchers at IISER, IISc, and TIFR are contributing to both quantum computing hardware and post-quantum cryptographic algorithms.

AI safety and alignment is another frontier with direct connections to rlhf: how chatgpt was trained. As AI systems become more capable, ensuring they behave as intended becomes critical. This involves formal verification (mathematically proving system properties), interpretability (understanding WHY a model makes certain decisions), and robustness (ensuring models do not fail catastrophically on edge cases). The Alignment Research Center and organisations like Anthropic are working on these problems, and Indian researchers are increasingly contributing.

Edge computing and the Internet of Things present new challenges: billions of devices with limited compute and connectivity. India's smart city initiatives and agricultural IoT deployments (soil sensors, weather stations, drone imaging) require algorithms that work with intermittent connectivity, limited battery, and constrained memory. This is fundamentally different from cloud computing and requires rethinking many assumptions.

Finally, the ethical dimensions: facial recognition in public spaces (deployed in several Indian cities), algorithmic bias in loan approvals and hiring, deepfakes in political campaigns, and data sovereignty questions about where Indian citizens' data should be stored. These are not just technical problems — they require CS expertise combined with ethics, law, and social science. The best engineers of the future will be those who understand both the technical implementation AND the societal implications. Your study of rlhf: how chatgpt was trained is one step on that path.

Key Vocabulary

Here are important terms from this chapter that you should know:

🏗️ Architecture Challenge

Design the backend for India's election results system. Requirements: 10 lakh (1 million) polling booths reporting simultaneously, results must be accurate (no double-counting), real-time aggregation at constituency and state levels, public dashboard handling 100 million concurrent users, and complete audit trail. Consider: How do you ensure exactly-once delivery of results? (idempotency keys) How do you aggregate in real-time? (stream processing with Apache Flink) How do you serve 100M users? (CDN + read replicas + edge computing) How do you prevent tampering? (digital signatures + blockchain audit log) This is the kind of system design problem that separates senior engineers from staff engineers.

The Frontier

You now have a deep understanding of rlhf: how chatgpt was trained — deep enough to apply it in production systems, discuss tradeoffs in system design interviews, and build upon it for research or entrepreneurship. But technology never stands still. The concepts in this chapter will evolve: quantum computing may change our assumptions about complexity, new architectures may replace current paradigms, and AI may automate parts of what engineers do today.

What will NOT change is the ability to think clearly about complex systems, to reason about tradeoffs, to learn quickly and adapt. These meta-skills are what truly matter. India's position in global technology is only growing stronger — from the India Stack to ISRO to the startup ecosystem to open-source contributions. You are part of this story. What you build next is up to you.

Crafted for Class 10–12 • Frontier AI • Aligned with NEP 2020 & CBSE Curriculum