anacodicai-labs/covidqa-bench

COVID-QA Retrieval Benchmark — parsed with the AnacodicAI guidelines pipeline

A public retrieval / RAG benchmark over biomedical COVID-19 papers. Questions and gold answers come from deepset COVID-QA (Apache-2.0); the papers were fetched as PDFs and parsed with the same production pipeline as the AnacodicAI guidelines-generator and specialistRAG systems — so a number here reflects the real pipeline, not a toy one.

• 1,637 papers total: 137 with questions + 1,500 topic-matched decoys (hard negatives, no questions)
• 1,852 questions with marked gold answer spans
• Every answer is a span of its article, so standard IR metrics apply: Recall@k, nDCG@k, context recall/precision, faithfulness.

What's in here

File	What it is	Start here if you are…
gold/questions.jsonl	question, gold_answer, paper_id, gold_chunk_ids, match (1,852 rows)	computing metrics (everyone)
chunks.jsonl	64,965 production chunks for all 1,637 papers — chunk_id, paper_slug, chunk_type, section_path, pages, text, embed_text, …	scoring retrieval (fixed chunking)
parsed_questions_137.tar.zst	the 137 papers that have questions (.docling.json + assets) — 1.7 GB download, ~2.6 GB unpacked	comparing chunkers — start here (Task 1)
parsed_decoys_1500.tar.zst	the 1,500 decoy papers (no questions), same format — 17 GB download	add these for the hard setting (Task 2)
parsed_all_1637.tar.zst	both of the above in one archive — 18 GB download, 32 GB unpacked	if you want everything at once
manifest.csv	the 147 COVID-QA articles: PMC ID, DOI, licence, fetch status	provenance
manifest_distractors.csv	the decoys: PMC ID, DOI, title, licence, fetch status	provenance

All three archives unpack to the same layout, 02_parsed/covidqa/<paper>/, so parsed_questions_137 + parsed_decoys_1500 = parsed_all_1637.

Production chunker (identical to guidelines-generator / specialistRAG): Docling HybridChunker · 512 tokens · tiktoken for text-embedding-3-large · repeat_table_header · merge_peers · always_emit_headings · contextualize.

Two scoring rules that matter

1. gold_chunk_ids fit the 512-token production chunks only. If you re-chunk (256, 1024, a different splitter), score "does a top-k chunk contain the gold answer text" with a text match — not chunk IDs.
2. match = none flags answers the parse lost (~1%). Report it; part of it is COVID-QA's own OCR noise (e.g. 32uC for 32 °C).

Quick start

from huggingface_hub import hf_hub_download
import json

qpath = hf_hub_download("anacodicai-labs/covidqa-bench", "gold/questions.jsonl", repo_type="dataset")
cpath = hf_hub_download("anacodicai-labs/covidqa-bench", "chunks.jsonl", repo_type="dataset")

questions = [json.loads(l) for l in open(qpath)]
chunks    = [json.loads(l) for l in open(cpath)]

Unpack the parsed papers (for re-chunking) — start with the 137 question papers:

hf download anacodicai-labs/covidqa-bench parsed_questions_137.tar.zst --repo-type dataset --local-dir .
tar -I zstd -xf parsed_questions_137.tar.zst          # → 02_parsed/covidqa/<paper>/
# later, for the hard setting, add the decoys into the same folder:
hf download anacodicai-labs/covidqa-bench parsed_decoys_1500.tar.zst --repo-type dataset --local-dir .
tar -I zstd -xf parsed_decoys_1500.tar.zst

Licence & attribution

• Questions/answers: deepset COVID-QA, Apache-2.0.
• Papers: CC0 / CC BY only, each under its own licence — see manifest.csv (question papers) and manifest_distractors.csv (decoys). 10 COVID-QA articles without a redistributable PDF were excluded; retracted papers were skipped.
• Source acknowledgement: NIH NLM NCBI PubMed Central (PMC) Article Datasets, accessed via https://registry.opendata.aws/ncbi-pmc. This dataset is not endorsed by NLM / NIH.

Maintainers

Curated by Rashan (@rashanbu) at AnacodicAI Labs.