LingweiGu/NanoKnow_Benchmark

NanoKnow Benchmark Qrels

[Paper] [Code]

Pre-built relevance judgments (qrels) that partition SQuAD and Natural Questions (NQ) into supported and unsupported splits based on whether the answer appears in the nanochat pre-training corpus (karpathy/fineweb-edu-100b-shuffle).

These qrels are part of the NanoKnow project: github.com/castorini/NanoKnow

Splits

Dataset	Total Questions	Supported	Unsupported
SQuAD	10,570	7,560 (72%)	3,010 (28%)
NQ-Open	3,610	2,391 (66%)	1,219 (34%)

• Supported — The gold answer was found in the pre-training corpus and verified by an LLM judge. These questions test parametric knowledge.
• Unsupported — The gold answer does not appear in the pre-training corpus. These questions test the model's ability to generalize or rely on external knowledge (RAG).

Files

File	Description	Format
qrels/squad_supported.txt	SQuAD supported questions (7,560 questions, 145,918 verified docs)	qid, question, answer, doc_id, answer_offset
qrels/squad_unsupported.txt	SQuAD unsupported questions (3,010 questions)	qid, question, answer
qrels/nq_supported.txt	NQ supported questions (2,391 questions, 56,857 verified docs)	qid, question, answer, doc_id, answer_offset
qrels/nq_unsupported.txt	NQ unsupported questions (1,219 questions)	qid, question, answer

File Format

Supported qrels map each question to one or more pre-training documents that contain a verified answer:

qid, question, official_answer, doc_id, answer_offset

• doc_id: Document identifier in the format shard_XXXXX_YYYYY (shard number and row offset within the FineWeb-Edu parquet files).
• answer_offset: Character offset of the answer string within the document.

Unsupported qrels list questions whose answers were not found in the corpus:

qid, question, official_answer

Pipeline

The qrels were generated using a three-stage pipeline:

1. BM25 Retrieval — Search the corpus for the top-100 candidate documents per question using Pyserini.
2. Answer String Matching — Filter to documents containing the gold answer as a substring.
3. LLM Verification — Use Qwen/Qwen3-8B as a judge to filter out coincidental matches (e.g., "Paris" in a passage about Paris, Texas).

Usage

Download

huggingface-cli download LingweiGu/NanoKnow_Benchmark --repo-type dataset --local-dir ./nanoknow-benchmark

Load in Python

import csv

def load_supported_qrels(filepath):
    qrels = []
    with open(filepath) as f:
        for line in f:
            if line.startswith("#") or not line.strip():
                continue
            parts = [p.strip() for p in line.split(",")]
            qrels.append({
                "qid": int(parts[0]),
                "question": parts[1],
                "answer": parts[2],
                "doc_id": parts[3],
                "answer_offset": int(parts[4]),
            })
    return qrels

squad_supported = load_supported_qrels("nanoknow-benchmark/qrels/squad_supported.txt")
print(f"Loaded {len(squad_supported)} supported entries")

Related Resources

• Lucene Index: LingweiGu/NanoKnow-Fineweb-Edu-Index — The pre-built BM25 index over the FineWeb-Edu corpus (~326 GB) used to generate these qrels.
• Code: github.com/castorini/NanoKnow — Scripts to project new benchmarks, evaluate nanochat checkpoints, and analyze frequency effects.

Citation

@article{gu2026nanoknow,
  title={NanoKnow: How to Know What Your Language Model Knows},
  author={Gu, Lingwei and Jedidi, Nour and Lin, Jimmy},
  journal={arXiv preprint arXiv:2602.20122},
  year={2026}
}

License

Apache 2.0