• forked from team repo
• time should be combined from CALM-RAG500 & BLUFF-RAG (since name of repo was changed)ged)

Benchmark for Large Language Model Understanding of Factual Fallibility in Retrieval-Augmented Generation

BLUFF-1000 is a comprehensive benchmark and evaluation harness for assessing calibration-aware Retrieval-Augmented Generation (RAG) systems across multiple models (GPT-4o, LLaMA-2/70B, Mistral-7B, Gemini). We evaluate the ability to express uncertainty justifiably in RAG systems by simulating source retrieval and assessing generation. This topic is especially important, since real-world RAG systems often retrieve unreliable or contradictory information from sources.

BLUFF-1000 contains 500 questions spanning 10 domains. Each question contains 2 source sets, amounting to 1000 total evaluation instances.

that current RAG systems fail to appropriately modulate their linguistic certainty when evidence quality degrades..

https://drive.google.com/file/d/1FkTS_F6eDyPmY5sk1G8mXdIsllNf4BK0/view?usp=sharing

git clone https://github.com/your-repo/BLUFF-RAG500.git

cd BLUFF-RAG500
pip install -r requirements.txtt

from runner import RAGEvaluator


# Initialize evaluator

evaluator = RAGEvaluator("example_dataset.json")


# Setup OpenAI (replace with your API key)

evaluator.setup_openai("your-openai-api-key")


# Run evaluation

results = evaluator.run_evaluation("openai", max_items=10)


# Print summary and save results

evaluator.print_summary(results)

evaluator.save_results(results)

Each item in the BLUFF-RAG-500 dataset follows this structure:

{
   "id": 17,
   "domain": "medicine",
   "question": "What was the remission rate in the Phase 3 trial of Drug X?",
   "source_excerpts": [
    {
        "title": "Phase 3 Clinical Trial Results...",
       "url": "https://pubmed.ncbi.nlm.nih.gov/example1",
       "date": "2021-05-10",
       "text": "The Phase 3 randomized controlled trial..."
    }
  ],
  
  "gold_answer": "45%",
   "human_confidence": 0.6,
   "human_hedge_label": "Likely"
}}

Implements all calibration and confidence metrics:

• Overconfidence Index (OCI): Fraction of high-confidence wrong answers
• Expected Calibration Error (ECE): Calibration assessment
• Brier Score: Probabilistic accuracy measure
• Hedge Detection: Precision/recall for uncertainty language
• Isotonic Calibration: Post-hoc calibration improvement

Handles prompt formatting for different models and scenarios:

• Standard RAG prompts with confidence elicitation
• Calibration-focused prompts for better uncertainty estimation
• Few-shot examples for improved calibration
• Model-specific prompt adaptations

Main evaluation harness:

• Multi-model support (OpenAI, LLaMA, Mistral, Gemini)
• Batch evaluation with progress tracking
• Automatic metric computation
• Results saving and summary generation

• Expected Calibration Error (ECE): Measures calibration quality
• Brier Score: Combines accuracy and calibration
• Overconfidence Index: High-confidence errors (τ = 0.8)

• Hedge Precision/Recall: Detection of uncertainty language
• Confidence-Accuracy Correlation: Alignment of confidence with correctness
• Lexical Overconfidence: Confident language in wrong answers

• Retrieval-Confidence Correlation: How retrieval quality affects confidence
• Recall vs Confidence: Relationship between evidence quality and certainty

results = evaluator.run_evaluation("openai", prompt_type="standard")

results = evaluator.run_evaluation("openai", prompt_type="calibration")

results = evaluator.run_evaluation("openai", prompt_type="uncertainty")

• GPT-4o
• LLaMA-2/70B
• Mistral-7B
• Gemini

• Climate Science
• Technology
• Current Events
• Astronomy
• Finance
• History
• Law
• Psychology
• Public Health
• Politics
• Sports

Website: HTML/CSS, Javascript Dataset Creation: Python Metric Calculations: Pythonon