MODULE m9 · 10.0 HOURS

Sample Human Review and Comparison

0% complete

LEARNING OBJECTIVES

Module objectives

  1. Understand the gap between automated RAG evaluation metrics and actual factual accuracy.
  2. Design a protocol for humans to review the Factual Consistency of model-generated answers.
  3. Identify the limitations of LLM evaluation and analyze the significance of synthetic data techniques such as TrueTeacher [S5].
  4. Learn how to systematically classify error types and manage reliability datasets.

Limitations of Automated Evaluation and the Necessity of Human Review

When evaluating the quality of Retrieval Augmented Generation (RAG) systems, tools like Ragas provide quantitative metrics quickly, but there are limitations in perfectly capturing subtle factual errors in model-generated answers. In particular, it is difficult to distinguish whether an LLM is reasoning within the knowledge or relying on learned data to generate hallucinations in complex contexts.

Factual Consistency Evaluation

Recent research utilizes Natural Language Inference (NLI) models or Large Language Models (LLMs) to evaluate the factuality of summaries or answers. However, existing methods rely on human-written summary datasets, and differences can arise from the characteristics of results actually generated by models [S5]. Approaches like TrueTeacher attempt to overcome these limitations by using LLMs to generate synthetic factuality evaluation data from model-generated data [S5].

Role of Human-in-the-Loop

No matter how advanced an automated evaluation pipeline becomes, human review is essential for final reliability verification. Human review performs the following roles:

  1. Verification of automated evaluation metrics: Identifying cases where a specific answer was evaluated as ‘relevant’ but is not actually factual.
  2. Hallucination type classification: Diagnosing the structural flaws of the system (Retrieval Error vs. Generation Model Error).
  3. Calibration of regression test sets: Continuously improving the quality of evaluation sets based on human-verified data.

WORKED EXAMPLES

Worked examples

  1. Example 1: When an automated evaluation metric (e.g., Faithfulness) results in a high score of 0.9, but human review shows the inclusion of content not in the retrieved document. Analysis: Classify as hallucination where the model used historical information contained in internal weights, not retrieved information, and record this in system error logs.
  2. Example 2: When designed so the system evaluates the factuality of answers itself using the TrueTeacher model. Some answers evaluated as 'true' by the LLM are sampled and investigated by humans to measure the error rate (False Positive) of the LLM evaluator and specify this in the evaluation report [S5].

LAB PROTOCOL

Sample Human Review and Error Analysis Execution

  1. 1

    Derive Faithfulness scores for 100 answers through an automated evaluation pipeline (Ragas, etc.).

  2. 2

    Create a review set by randomly extracting 10 of the lowest scores, 10 of the middle level, and 10 of the highest.

  3. 3

    Manually record the presence of 'Retrieval Omission', 'Information Distortion', or 'Hallucination Generation' while comparing answers, retrieved documents (Context), and Ground Truth.

  4. 4

    Perform correlation analysis by comparing recorded human judgment and automated evaluation scores.

Safety check
  • Always ensure that the dataset to be reviewed does not include actual personal information or sensitive private documents.
  • Safely store reviewed data in local storage and do not upload to unverified external APIs.

Lab deliverables

  • Error classification sheet (CSV/Excel) containing at least 30 human review records
  • Summary report analyzing the consistency between automated metrics and human evaluation

ASSIGNMENT

Writing RAG Reliability Analysis Report

Deliverables

Rubric

  • Were error types systematically classified?
  • Was the limitation of automated evaluation metrics logically described with concrete examples?
  • Was human review data appropriately utilized as a basis for reliability analysis?

KNOWLEDGE CHECK

Knowledge check

1What is the main reason why it is difficult to determine system reliability only with automated factuality evaluation metrics?
2What is the difference between the TrueTeacher method and existing summary dataset usage methods?

FIELD CHECK

Completion criteria

  • Human review logs for at least 30 data samples must be written.
  • A report including a comparative analysis between automated evaluation results and human review results must be submitted.
  • The current system's vulnerabilities must be clearly defined through error classification.

MODULE SOURCES

Module sources

  1. [2305.11171] TrueTeacher: Learning Factual Consistency ... TrueTeacher: Learning Factual Consistency Evaluation with ... TrueTeacher: Generating Synthetic Data for Factual ... (PDF) TrueTeacher: Learning Factual Consistency Evaluation ... (PDF) TrueTeacher: Learning Factual Consistency Evaluation ... [PDF] TrueTeacher: Learning Factual Consistency Evaluation ...arxiv.org · paper