vkshdev/rag-hallucination-benchmark

RAG Hallucination Benchmark

Context

Retrieval-Augmented Generation (RAG) is the industry standard for reducing LLM hallucinations, but detecting when a RAG system fails is a massive challenge. Most existing benchmarks focus only on massive Deep Learning models and lack tabular features. This dataset provides a clean, engineered setup to train models (from XGBoost to RoBERTa) to detect hallucinations, predict context faithfulness, and measure answer relevance.

Content

The dataset contains 30,000 synthetic RAG interactions covering:

• Text Data: User Prompts, Retrieved Contexts, and LLM Responses.
• Metadata: Model Names, Temperature Settings, and Vector Similarity Scores.
• Engineered Features: Lexical Overlap, Entity Match Scores, and Complexity Indices.
• Targets:
  • Binary Hallucination flags
  • Multiclass Hallucination Types
  • Regression targets for Faithfulness

Inspiration

This was inspired by the need for a practical ML benchmark where you can practice both feature engineering for tabular models (LightGBM, XGBoost) and text classification for NLP Transformers (BERT, DeBERTa) on the same dataset.

Quick Load

import pandas as pd

url = "https://huggingface.co/datasets/vkshdev/rag-hallucination-benchmark/raw/main/RAG_context_adherence_and_hallucination_benchmark.csv"
df = pd.read_csv(url)

print(df.head())