Sistem Umpan Balik Manusia dalam Proses

When working with Large Language Models (LLMs), sometimes their outputs aren't quite right. This is where Human-in-the-Loop (HITL) systems come in.

HITL means humans actively review, correct, or provide feedback on an LLM's output. This feedback then helps the model learn and improve over time, making it smarter and more reliable.

Even powerful LLMs have limitations. They can:

• Hallucinate: Make up facts.
• Be biased: Reflect biases from their training data.
• Lack up-to-date info: Not know about recent events.
• Misunderstand complex requests: Struggle with nuanced instructions.

Human oversight helps catch and fix these issues, ensuring higher quality and safer outputs.

A typical HITL system involves a few key parts:

1. LLM Generation: The model creates an output.
2. Human Review: A person evaluates the output.
3. Feedback Collection: The human's judgment is captured.
4. Model Improvement: This feedback is used to refine the LLM, either by adjusting prompts or fine-tuning the model itself.

It's a continuous cycle of creation, evaluation, and learning.

Human feedback can be broadly categorized into two types:

• Explicit Feedback: Direct ratings or corrections from users or annotators. Think "thumbs up/down" or editing a generated text.
• Implicit Feedback: Inferred from user behavior without direct input. For example, how long a user spends on a response or if they click on certain links.

Both types are valuable for different reasons!

Explicit feedback is clear and intentional. Common ways to collect it include:

• Thumbs Up/Down: Simple binary feedback on response quality.
• Star Ratings: A more granular scale (e.g., 1-5 stars).
• Direct Edits: Users correcting or rewriting parts of the LLM's output.
• Annotator Labels: Expert human annotators tagging specific issues like factual errors or toxicity.

This data directly tells you what worked and what didn't.

Implicit feedback is less direct but still powerful. It's about observing how users interact with LLM outputs:

• Session Duration: Users spending more time on a response might indicate engagement.
• Click-Through Rates: If an LLM suggests links, clicks show relevance.
• Search Refinements: If a user immediately rephrases their query, the initial response might have been poor.
• Copying Text: Users copying output might find it useful.

Analyzing these behaviors can reveal subtle preferences and issues.

Let's imagine a basic way to capture explicit feedback (like a "thumbs up" or "thumbs down") after an LLM generates a response. This simple Python code shows how you might structure the collection of feedback.

It doesn't call an LLM, but focuses on the feedback part.

Once feedback is collected, it's not just stored; it's used to make the LLM better:

• Prompt Refinement: If many users rate an output as "bad" for a specific prompt, you can adjust the prompt to be clearer or more specific.
• Data for Fine-tuning: High-quality human-corrected data can be used to fine-tune the LLM, teaching it directly from good examples.
• Reinforcement Learning from Human Feedback (RLHF): Feedback can be used to train a reward model, which then guides the LLM to generate preferred outputs.

This closes the loop and improves performance.

While HITL is powerful, it comes with challenges:

• Scalability: Manually reviewing every LLM output for a large application is impossible.
• Cost: Human annotators or user feedback mechanisms can be expensive to build and maintain.
• Bias in Feedback: Human reviewers can also introduce their own biases, which might then be reinforced in the model.
• Subjectivity: What one person considers a "good" response, another might not.

Careful design is needed to overcome these.

In this lesson, we explored Human-in-the-Loop (HITL) systems. We learned that HITL integrates human judgment to continuously improve LLM performance by addressing limitations like hallucinations and biases.

You now understand the components of a HITL system, the difference between explicit and implicit feedback, and how this feedback drives model improvement. You also saw a simple code example for collecting feedback and are aware of the challenges involved.

Keep exploring how to build robust LLM applications!