Laurie/faithfulness-qa-dataset

Faithfulness-QA: A Counterfactual Entity Substitution Dataset for Training Context-Faithful RAG Models

Overview

Faithfulness-QA is a large-scale dataset of 99,094 question-answer pairs designed to train and evaluate the faithfulness of Retrieval-Augmented Generation (RAG) models to retrieved context.

The core idea is counterfactual entity substitution: for each QA sample, we replace the answer-bearing entity in the context with a type-consistent alternative, creating a controlled conflict between the context and the model's parametric knowledge. A faithful model should output the replacement entity (from context), not the original entity (from memory).

Dataset Statistics

Source	Total	Train	Validation	Test
SQuAD	49,094	39,275	4,909	4,910
TriviaQA	50,000	40,000	5,000	5,000
Total	99,094	79,275	9,909	9,910

Entity Type Distribution

Type	SQuAD	TriviaQA
PERSON	9,775 (19.9%)	22,871 (45.7%)
ORG	10,114 (20.6%)	8,186 (16.4%)
DATE	9,997 (20.4%)	2,057 (4.1%)
GPE	6,292 (12.8%)	11,810 (23.6%)
CARDINAL	6,568 (13.4%)	1,348 (2.7%)
NORP	4,004 (8.2%)	1,247 (2.5%)
LOC	1,629 (3.3%)	1,952 (3.9%)
EVENT	715 (1.5%)	529 (1.1%)

Usage

from datasets import load_dataset

# Load all data (both SQuAD and TriviaQA)
dataset = load_dataset("Laurie/faithfulness-qa-dataset")

# Load only SQuAD subset
squad = load_dataset("Laurie/faithfulness-qa-dataset", "squad")

# Load only TriviaQA subset
triviaqa = load_dataset("Laurie/faithfulness-qa-dataset", "triviaqa")

# Example: access a training sample
sample = dataset["train"][0]
print(f"Question: {sample['question']}")
print(f"Modified context: {sample['modified_context'][:200]}...")
print(f"Faithful answer: {sample['faithful_answer']}")

Data Format

Each sample contains the following fields:

Field	Type	Description
id	string	Original sample ID from the source dataset
question	string	The question text
original_context	string	Unmodified context from the source dataset
modified_context	string	Context after counterfactual entity substitution
original_answer	string	Ground-truth answer from the source dataset
faithful_answer	string	Correct answer given the modified context (= replacement entity)
original_entity	string	The named entity that was replaced
replacement_entity	string	The new entity substituted in
entity_type	string	NER type: PERSON, GPE, ORG, DATE, CARDINAL, NORP, LOC, or EVENT
source	string	Source dataset: squad or triviaqa

Intended Use Cases

1. Faithfulness-aware fine-tuning: Train with (modified_context, question) → faithful_answer to teach models to follow context over parametric memory.
2. Attention-based faithfulness loss: Supervise cross-attention weights to ensure models attend to retrieved context.
3. Faithfulness evaluation: Measure the rate at which models output the faithful answer (context-grounded) vs. the original answer (parametric).
4. Knowledge conflict research: Study LLM behavior when retrieved context contradicts parametric knowledge.

Methodology

The pipeline consists of four stages:

1. Entity Bank Construction: Extract all named entities from SQuAD contexts using SpaCy NER → 76,953 entity bank across 8 types.
2. NER & Answer Entity Match: Match the answer to a recognized entity via exact match, substring match, or positional overlap.
3. Counterfactual Substitution: Replace entity with a same-type alternative from the entity bank (with length compatibility constraints).
4. Quality Filtering: Apply 6 quality filters, then split 80/10/10.

Citation

@misc{zhang2026faithfulnessqa,
  title={Faithfulness-QA: A Counterfactual Entity Substitution Dataset for Training Context-Faithful RAG Models},
  author={Zhang, Qi},
  year={2026},
  url={https://github.com/qzhangFDU/faithfulness-qa-dataset}
}

License

MIT License. The source datasets (SQuAD, TriviaQA) are used under their respective licenses.