"""Write the Hugging Face dataset card for Lilambd/empire-eval from the package itself, so every number in it is counted from the files. usage: python make_readme.py --pkg PKG_DIR --date YYYY-MM-DD --reports verify_dev.json verify_test.json ... """ import argparse, collections, glob, json, os TITLE = { 'world-hazard': ('World Hazard Ledger', 'earthquakes per region, station warming, disaster deaths'), 'world-earth': ('World Earth Surface Ledger', 'forest loss, land cover, built-up area, soil per region'), 'world-ground': ('World Ground Ledger', 'people near volcanoes, coasts and plate boundaries; soil'), 'world-prices': ('World Prices Long Record', 'commodity prices since 1960, CPI, energy prices'), 'world-science': ('World Science', 'OpenAlex paper counts by country, field, year'), 'world-marketplaces': ('World Marketplaces', 'digital marketplaces and 409k physical markets'), 'world-sound': ('World Sound', 'instruments, Cantometrics songs, tunings (dataset upload pending)'), 'world-signals': ('World Signals', 'daily attention snapshot: Steam, GitHub, HF, packages, podcasts, search'), 'world-seeds': ('World Seeds', 'Wikipedia "by country" tables keyed by ISO2'), 'japan-dpc-hospitals': ('Japan DPC Hospitals', 'all acute-care hospitals: beds, discharges, case mix'), 'japan-prefecture-facts': ('Japan by Prefecture', 'sourced facts for all 47 prefectures'), 'world-biodiversity-cells': ('World Biodiversity Cells', 'GBIF records and species on a 0.5° grid'), 'machine-economy-daily': ('Machine Economy Daily', 'x402 hosts, MCP registry, agent packages'), 'security-daily': ('Security Daily', 'CVE, KEV, EPSS, breaches, joined daily'), 'world-pulse-daily': ('World Pulse Daily', '15 public catalogues written down daily'), 'world-markets': ('World Markets', '138k physical markets, bazaars, night markets'), } LIC = {'world-science': 'CC0-1.0', 'world-marketplaces': 'CDLA-Permissive-2.0', 'world-markets': 'CDLA-Permissive-2.0', 'world-seeds': 'CC-BY-SA-4.0', 'world-pulse-daily': 'CC-BY-SA-4.0', 'world-biodiversity-cells': 'CC-BY-NC-4.0'} def _baselines(pkg): out = [] for nm, lab in (('oracle', 'oracle (reads the answers)'), ('null', 'always abstain'), ('random-guess', 'random guess')): p = os.path.join(pkg, 'results', f'baseline_{nm}', 'score.json') if os.path.exists(p): sm = json.load(open(p))['summary'] out.append(f"{lab} accuracy {sm['accuracy']:.4f}, hallucination rate {sm['hallucination_rate']:.2f}") return '; '.join(out) or 'none run yet' def main(): ap = argparse.ArgumentParser() ap.add_argument('--pkg', required=True) ap.add_argument('--reports', nargs='+', required=True, help='verify.py --report files (dev and private test answers)') ap.add_argument('--date', required=True) a = ap.parse_args() n_q = n_bad = 0 for rp in a.reports: for k, v in json.load(open(rp)).items(): if isinstance(v, dict): n_q += v['questions'] n_bad += v['failed'] a.verified = f'{n_q - n_bad:,} / {n_q:,}' sets = sorted(os.path.basename(p)[:-6] for p in glob.glob(os.path.join(a.pkg, 'data/dev/*.jsonl'))) rows, tasks, skills = [], collections.Counter(), collections.Counter() tot = collections.Counter() for s in sets: dev = [json.loads(l) for l in open(os.path.join(a.pkg, 'data/dev', s + '.jsonl'), encoding='utf-8')] test = [json.loads(l) for l in open(os.path.join(a.pkg, 'data/test', s + '.jsonl'), encoding='utf-8')] for it in dev + test: tasks[it['task']] += 1 for k in it['skills']: skills[k] += 1 n_un = sum(1 for it in dev + test if it['answer_type'] == 'none') n_rec = sum(1 for it in dev if it.get('recipe')) tot['dev'] += len(dev) tot['test'] += len(test) tot['un'] += n_un rows.append((s, len(dev), len(test), n_un, n_rec)) real = [r for r in rows if not r[0].startswith('worlds-')] fict = [r for r in rows if r[0].startswith('worlds-')] N = tot['dev'] + tot['test'] yaml = ['---', 'license: cc-by-4.0', 'pretty_name: "Lilambd Empire Eval: grounded questions over living public ledgers"', 'language:', '- en', '- ja', 'task_categories:', '- question-answering', '- table-question-answering', 'tags:', '- rag', '- evaluation', '- benchmark', '- agents', '- tabular', '- grounding', '- hallucination', '- abstention', '- contamination-free', 'size_categories:', '- 1K#`).\n') L.append('Use it to measure whether a RAG pipeline, a table agent or a tool-using LLM actually reads the data, ' 'and whether it says "not in the data" when the data has no answer.\n') L.append('## What makes it different\n') L.append(f'- **Every answer is re-derivable.** `scripts/verify.py` reloads the dataset files, re-runs every recipe and re-reads every cited row. ' f'Result on {a.date}: **{a.verified} recipe answers reproduced exactly**.') L.append(f'- **It tests abstention.** {tot["un"]:,} questions have no answer in the data (a future month, a region with no station, a fictional name). ' 'The scorer reports a hallucination rate next to accuracy.') L.append(f'- **It cannot be memorised.** {len(fict)} of the sets are fictional planets (Lilambd Worlds) simulated from plate tectonics to languages; ' 'nothing about them exists on the web, so a correct answer must come from the files.') L.append('- **It renews itself.** The real-world ledgers are rebuilt from public measurement records (daily for some), and the generator scripts are included, ' 'so a fresh question set can be cut from any later revision.') L.append('- **Held-out test split.** 30% of the questions are published without answers. Their answers are kept by the maintainer so the split stays clean.\n') L.append('## Sets\n') L.append('| set | source dataset | dev | test | no-answer | licence of answers |') L.append('|---|---|---:|---:|---:|---|') for s, d, t, u, _ in real: nm, what = TITLE.get(s, (s, '')) L.append(f'| `{s}` | [{nm}](https://huggingface.co/datasets/Lilambd/{s}): {what} | {d} | {t} | {u} | {LIC.get(s, "CC-BY-4.0")} |') if fict: d, t, u = sum(r[1] for r in fict), sum(r[2] for r in fict), sum(r[3] for r in fict) names = ', '.join(r[0].split('-', 2)[2].capitalize() for r in fict) L.append(f'| `worlds-*` ({len(fict)} sets) | Lilambd Worlds fictional planets: {names} | {d} | {t} | {u} | CC-BY-4.0 |') L.append(f'| **total** | | **{tot["dev"]:,}** | **{tot["test"]:,}** | **{tot["un"]:,}** | |\n') L.append('Task types: ' + ', '.join(f'{k} {v:,}' for k, v in tasks.most_common()) + '.\n') L.append('Skills tagged: ' + ', '.join(f'`{k}` {v:,}' for k, v in skills.most_common(14)) + '.\n') L.append('## One row\n') L.append('```json') ex = json.loads(open(os.path.join(a.pkg, 'data/dev', 'world-hazard.jsonl'), encoding='utf-8').readline()) L.append(json.dumps({k: ex[k] for k in ('id', 'dataset', 'revision', 'task', 'question', 'answer', 'answer_type', 'tolerance', 'evidence', 'recipe')}, ensure_ascii=False, indent=1)) L.append('```\n') L.append('Fields: `answer` (text form), `answer_json` (typed value: number, string, list, boolean or `null` for no answer), `answer_type`, ' '`tolerance` (`abs` or `rel` for numbers), `evidence` (row ids with their key columns), `evidence_scope` (the filter behind counts, sums and rankings), ' '`recipe` (JSON: the operation that produces the answer), `skills`, `difficulty` 1-3, `revision` (dataset revision the answer belongs to).\n') if ml: n_ml = sum(1 for _ in open(os.path.join(a.pkg, 'data/multilingual', ml[0] + '.jsonl'), encoding='utf-8')) L.append('## Multilingual subset\n') L.append(f'`data/multilingual/` holds the same {n_ml:,} dev questions in ' + ', '.join(f'`{l}`' for l in ml) + '. ' 'Only the question text is translated; names, codes, numbers, column names and quoted strings are kept exactly, and the answer, evidence and recipe are the English item\'s. ' 'Each row keeps `source_id` and `question_en`, so a gap between languages is a reading gap, not a data gap. ' 'All recipe answers in every language file were re-derived by `verify.py`. Use the `multilingual` config.\n') if wl: L.append('## World Ledgers (SQL recipes)\n') L.append(f"The `world-ledgers` config adds **{wl['dev'] + wl['test']:,} questions** (dev {wl['dev']:,}, test {wl['test']:,}, same hash split) over the Lilambd world ledgers of 2026-11-03: " + ', '.join(wl['ledgers']) + f". {wl['x']:,} of them join two or three ledgers (city_id, ISO codes, nearest city). " "Here the recipe is a DuckDB query (`recipe_sql`) and the evidence is the list of files it reads (`evidence_files`); `question_ja` gives a Japanese wording. " "`scripts/verify_sql.py --views scripts/world_ledgers_views.json` re-runs every query; on 2026-11-03 it reproduced " f"{wl['dev'] + wl['test']:,} / {wl['dev'] + wl['test']:,} answers (one item whose query no longer matched the ledger files was removed). " "Not every ledger file is on the Hub yet, so outside users can score against these answers but cannot all re-run the queries today.\n") L.append(f"Licence: {wl['nc']:,} of these rows (`commercial_ok: false`) rest on CShapes (non-commercial) or FAOSTAT production data; use them for non-commercial evaluation only. The rest follow CC BY 4.0 with source attribution (`licence_note`).\n") L.append('## Score a system\n') L.append('```bash') L.append('pip install pandas pyarrow pycountry') L.append('python scripts/make_prompts.py --qa data/dev --out prompts.jsonl --with-context-hint') L.append('# run your system, write {"id": ..., "answer": ...} per line to predictions.jsonl, or use the bundled runner:') L.append('ANTHROPIC_API_KEY=... python scripts/run_eval.py --qa data/dev --provider anthropic --out results/claude/predictions.jsonl --concurrency 8') L.append('OPENAI_API_KEY=... python scripts/run_eval.py --qa data/dev --provider openai --base-url http://localhost:8000/v1 --model MODEL --out results/MODEL/predictions.jsonl') L.append('python scripts/score.py --qa data/dev --pred predictions.jsonl --out score.json') L.append('```\n') L.append('The scorer accepts numbers with commas or units, lists as JSON or comma-separated, and treats `null`, empty, "not in the dataset", "unknown" and similar as abstaining. ' 'It reports accuracy, answerable accuracy, abstention accuracy, hallucination rate and false-abstention rate, broken down by set, task, difficulty and skill. ' 'Use `--lenient` for free-text answers (the fictional-planet questions written by the world generator).\n') L.append('Baselines on dev (from `results/baseline_*/score.json`): ' + _baselines(a.pkg) + '. `leaderboard/build_leaderboard.py` turns any number of `score.json` files into `leaderboard.json` and a static `leaderboard.html`; `leaderboard/EVAL_CARD.md` has a model-index snippet for reporting scores on a model card.\n') L.append('**Test split:** send `predictions.jsonl` for the test ids to hello@lilambd.com and you get the same report back.\n') L.append('## Rebuild or extend\n') L.append('`scripts/fetch_hf.py` downloads the source files pinned to their current commit, `scripts/build.py` generates questions (fixed seed), ' '`scripts/verify.py` re-derives them, `scripts/package.py` makes the dev/test split. A private set over your own tables can be built the same way.\n') L.append('## Limits\n') L.append('- Answers belong to the stated `revision`. The daily ledgers move on; re-verify against the same revision, or rebuild.') L.append('- Many questions name a table or a column (for example `features_country.csv`, `m6_in`). They test grounded reading, not trivia.') L.append('- Counting and ranking questions over large tables (hundreds of thousands of rows) are meant for systems that can run code or SQL; pure retrieval will not see every row.') L.append('- About 970 fictional-planet questions come from the world generator\'s own question list: free-text answers, checked by the generator\'s consistency tests, not by recipes (`recipe` is empty, `verified_by` says so).') L.append('- In the `worlds-*` sets, `source_qid` (for example `021-000`) is the world generator\'s internal question id, not a Wikidata QID.') L.append('- Source keys that repeat with different values (31 country CPI series in world-prices `monthly.csv`, 26 region rows in world-earth `features_region.csv`) are left out of question generation, so no question has two possible answers.') L.append('- Entity names are as written in the source; country names in questions come from ISO 3166 where the data has only a code.\n') L.append('## Licence\n') L.append('Questions, recipes and scripts: CC BY 4.0, Lilambd. Answers are facts taken from the source datasets and keep their licence (column above): ' 'most are CC BY 4.0, World Science is CC0, the market sets are CDLA-Permissive-2.0. Cite the source dataset alongside this one.\n') L.append('```bibtex\n@misc{lilambd_empire_eval_2026,\n title = {Lilambd Empire Eval: grounded questions over living public ledgers},\n author = {Lilambd},\n' f' year = {{2026}},\n url = {{https://huggingface.co/datasets/Lilambd/empire-eval}}\n}}\n```\n') L.append('Need an evaluation set over your own data, a private held-out split, or a set in another language? hello@lilambd.com\n') open(os.path.join(a.pkg, 'README.md'), 'w', encoding='utf-8').write('\n'.join(yaml) + '\n\n' + '\n'.join(L)) print('README', N, 'questions', len(sets), 'sets') if __name__ == '__main__': main()