RAGTruth–LettuceDetect Evaluation

📌 Project Overview

This project presents an empirical evaluation and error analysis of LettuceDetect for RAG hallucination detection using the RAGTruth dataset.

The work focuses on understanding where LettuceDetect succeeds and fails, particularly through false-positive (FP) and false-negative (FN) analysis, and evaluates whether targeted rule-based improvements can reduce missed hallucinations.

🎯 Objectives

• Evaluate LettuceDetect against RAGTruth
• Analyze false positives and false negatives
• Identify common failure patterns in hallucination detection
• Categorize false-negative cases
• Develop a targeted rule-based improvement
• Evaluate whether the rule improves overall detection performance
• Apply regression checks to prevent precision and F1 degradation

🔬 Methodology

The evaluation was performed in multiple phases.

Phase 3 — Evaluation Corrections

The evaluation of LettuceDetect against RAGTruth was corrected by addressing:

• Incomplete gold annotations
• An overly strict span-matching threshold

These corrections primarily affected the false-positive side of the evaluation.

The original false negatives were then analyzed separately.

Phase 4 — False Negative Analysis

Phase 4 focuses on the false negatives that LettuceDetect missed.

A pool of 783 false-negative cases at a matching threshold of 0.1 was used, from which 100 cases were randomly sampled and manually categorized.

📊 False Negative Categories

The sampled false negatives were categorized into five groups:

Category	Count	Percentage
Attribute-related	30	30%
Subtle-phrasing	25	25%
Long-span	20	20%
Numeric-buried	17	17%
Other	8	8%

Key Finding

Attribute-related errors were the largest false-negative category (30%).

These involved incorrect claims about structured fields such as:

• Hours
• WiFi
• Parking
• Reservations
• Music

This pattern was also observed in the false-positive analysis, making attribute-related hallucinations the most evidence-backed target for a rule-based improvement.

📈 False Negative Breakdown

🛠️ Rule-Based Improvement

A rule-based post-processor called check_attributes() was developed to cross-check generated answers against structured source information.

The rule checks for:

• Attribute invention — claims about fields marked as unknown or None
• Hours mismatch — stated times that do not match the source's available opening or closing times

🔄 Rule Iterations

Multiple versions of the rule were tested using the same threshold=0.1 evaluation setup.

Version	Precision	Recall	F1	Result
Baseline	0.504	0.489	0.497	—
Full rule	0.372	0.552	0.445	Regression
Narrow rule	0.465	0.530	0.495	Slightly below baseline
Gap-fill rule	0.495	0.500	0.498	Marginal improvement

📊 Rule Iteration Results

Final Result

The final gap-fill approach produced a small improvement:

F1: 0.497 → 0.498

The rule was restricted to cases where LettuceDetect predicted nothing at all, allowing it to focus on detector blind spots instead of competing with existing predictions.

🔍 Error Analysis Findings

The analysis revealed several important failure modes.

1. Attribute-related hallucinations

Incorrect claims involving structured fields were the largest identified FN category.

2. Subtle phrasing

Some hallucinations were expressed naturally and fluently, without obvious linguistic signals.

3. Long spans

Long or multi-clause hallucinations can dilute model confidence and make detection more difficult.

4. Numeric information

Incorrect numbers embedded inside otherwise correct sentences were another common failure mode.

⚠️ Regression Analysis

The first rule-based version improved recall but significantly reduced precision.

The full rule produced:

• Precision: 0.372
• Recall: 0.552
• F1: 0.445

This demonstrated that increasing recall alone does not necessarily improve the overall detector.

A hedge-phrase filter and improvements to hours parsing recovered much of the lost precision.

The final gap-fill version produced the best overall result with:

• Precision: 0.495
• Recall: 0.500
• F1: 0.498

💡 Key Takeaways

• Attribute-related hallucinations were the largest identified FN category.
• Simple keyword-based rules can introduce many false positives.
• Rule-based improvements require careful regression testing.
• Improving recall without monitoring precision can make the overall system worse.
• Restricting the rule to detector blind spots produced a small net F1 improvement.
• The final improvement should be treated cautiously because the gain was only +0.001 F1.

🚀 Future Improvements

• Perform bootstrap resampling to test statistical significance
• Develop a more precise attribute-value extractor
• Compare generated attribute values directly against source values
• Improve detection of subtle-phrasing hallucinations
• Investigate long-span hallucinations
• Improve detection of numeric-buried hallucinations
• Evaluate the approach on a larger manually reviewed sample
• Explore more advanced model-based hallucination detection methods

🛠️ Technologies

• Python
• Jupyter Notebook
• Natural Language Processing (NLP)
• Retrieval-Augmented Generation (RAG)
• Large Language Models (LLMs)
• Hallucination Detection
• Error Analysis
• Rule-Based Classification
• RAGTruth
• LettuceDetect

📂 Project Structure

RAGTruth-LettuceDetect-Evaluation/
│
├── .gitignore
├── README.md
├── requirements.txt
│
├── RAGTruth_LettuceDetect_Evaluation.ipynb
│
├── phase3a_fp_adjudication_export.py
├── phase3b_boundary_diagnosis.py
├── phase3c_baseless_export.py
├── phase4a_fn_analysis.py
├── phase4b_attribute_rule.py
│
├── phase4_summary.md
├── phase4_fn_breakdown.png
└── phase4_rule_iterations.png