Datasets:

tarekmasryo
/

rag-qa-logs-corpus-data

pretty_name: RAG QA Logs & Corpus (Synthetic)
license: cc-by-4.0
language:
  - en
size_categories:
  - 100K<n<1M
task_categories:
  - question-answering
  - tabular-classification
  - tabular-regression
tags:
  - rag
  - retrieval-augmented-generation
  - evaluation
  - hallucination
  - faithfulness
  - meta-modeling
  - risk-scoring
  - logs
  - telemetry
  - multi-table
  - synthetic-data
  - tabular
  - dataset
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/eval_runs.csv
  - config_name: eval_runs
    data_files:
      - split: train
        path: data/eval_runs.csv
  - config_name: retrieval_events
    data_files:
      - split: train
        path: data/rag_retrieval_events.csv
  - config_name: documents
    data_files:
      - split: train
        path: data/rag_corpus_documents.csv
  - config_name: chunks
    data_files:
      - split: train
        path: data/rag_corpus_chunks.csv
  - config_name: scenarios
    data_files:
      - split: train
        path: data/scenarios.csv
  - config_name: data_dictionary
    data_files:
      - split: train
        path: docs/data_dictionary.csv

🧠📚 RAG QA Logs & Corpus (Synthetic)

🧪 Multi-table synthetic RAG telemetry for quality, hallucinations, latency, and cost

A production-style, privacy-safe synthetic dataset that mimics telemetry exported from a real RAG system — from corpus → chunks → retrieval events → eval runs.

✅ Fully synthetic (no real users / orgs / PII).

⚡ Quick facts

Total rows: 103,255 across 6 linked tables
Labels (in eval_runs): is_correct, hallucination_flag, faithfulness_label
Join keys: doc_id, chunk_id, example_id, scenario_id, query_id, run_id
Splits: train, val, test (example-grain in eval_runs)
Retrieval strategies: dense, bm25, hybrid, dense_then_rerank, bm25_then_rerank

📦 What’s inside

data/eval_runs.csv — QA evaluation runs (labels + metrics)
data/rag_retrieval_events.csv — per-chunk retrieval telemetry (many rows per example)
data/rag_corpus_documents.csv — document metadata
data/rag_corpus_chunks.csv — chunk index + chunk text
data/scenarios.csv — scenario templates/use-cases
docs/data_dictionary.csv — schema documentation across tables

🔗 How to join

Documents → Chunks: doc_id
Eval runs → Retrieval events: example_id
Retrieval events → Chunks: chunk_id
Eval runs → Scenarios: scenario_id

🚀 How to use

🤗 Load from Hugging Face (recommended)

from datasets import load_dataset, get_dataset_config_names

repo_id = "tarekmasryo/rag-qa-logs-corpus-data"
print(get_dataset_config_names(repo_id))

runs   = load_dataset(repo_id, "eval_runs")["train"].to_pandas()
events = load_dataset(repo_id, "retrieval_events")["train"].to_pandas()
docs   = load_dataset(repo_id, "documents")["train"].to_pandas()
chunks = load_dataset(repo_id, "chunks")["train"].to_pandas()

💻 Local (after cloning)

git clone https://huggingface.co/datasets/tarekmasryo/rag-qa-logs-corpus-data
cd rag-qa-logs-corpus-data

import pandas as pd

runs   = pd.read_csv("data/eval_runs.csv")
events = pd.read_csv("data/rag_retrieval_events.csv")
docs   = pd.read_csv("data/rag_corpus_documents.csv")
chunks = pd.read_csv("data/rag_corpus_chunks.csv")

🧩 Example join (runs → events → chunks → docs)

Note: after joining runs → events, the result becomes retrieval-grain (rows ~= retrieval_events).

df = (
    runs
    .merge(events, on="example_id", how="left", suffixes=("_run", "_evt"))
    .merge(chunks[["chunk_id", "doc_id", "chunk_index"]], on="chunk_id", how="left")
    .merge(docs[["doc_id", "domain", "title", "source_type"]], on="doc_id", how="left")
)

print(df.shape)
print(runs["split"].value_counts(dropna=False))

🧾 Citation

Tarek Masryo. (2025). RAG QA Logs & Corpus (Synthetic).

📜 License

CC BY 4.0 (Attribution Required)