DATA_GUIDE.md

dsa data guide

Back to the dataset overview.

Folder layout

Each top-level dataset folder is a self-contained evaluation pack. There are no version directories in the current layout.

File	Purpose
pack.json	Entry point: identifies the database, case export, scoring policy, and their checksums.
cases.jsonl	Machine-readable questions, answer schemas, reference answers, and case metadata; one JSON object per line.
PROBLEMS.md	Human-readable problem catalog, including calculation conventions and answer fields.
ATTRIBUTION.md	Original sources, citations, licensing, and attribution.
database/<name>.duckdb	Ready-to-query database; its exact filename is in pack.json.
database/build_report.json	Dataset-specific build statistics and source provenance.
reference-queries.jsonl	Retained reference SQL for the cases recomputed during the revision. Not a complete reference implementation for every case.
changes.json	Per-case revision history and provisional difficulty labels.
source.json	Identity of the predecessor pack and unchanged source database.
validation.json	Export digests, case counts, and revision summary.
verification.json	Recorded schema checks, reference replays, and independent-check coverage.

Some files are dataset-specific:

• Online Retail II additionally has database-card.json, used by the chat application to describe the database.
• Online Retail II and EEA include database/schema.sql and database/views.sql, documenting the packaged relations. The DuckDB file is already built; these scripts need not be run to use it.
• EEA also includes database/source_manifest.json, listing the source-request archives and their identities. These archives are provenance references, not additional files bundled here.

The root .gitattributes file is repository storage configuration, not part of the evaluation contract.

JSON formats

All JSON and JSONL files are UTF-8. A .json file contains one JSON value; a .jsonl file contains one complete object per line and is not wrapped in an array. SHA-256 digests are lowercase hexadecimal strings. File paths in pack.json are relative to its dataset folder.

pack.json: evaluation manifest

The manifest uses the dsa-evaluation-pack/v1 contract. Its fields are:

Field	Type and meaning
format	String, exactly dsa-evaluation-pack/v1.
pack_id	String identifying the problem set.
version	Pack revision as a major.minor.patch string; currently 2.0.0. This is metadata, not a directory name.
license	License identifier, currently CC-BY-4.0.
database	Object with id, path, size_bytes, and sha256. Identifies the exact DuckDB file.
cases	Object with path, size_bytes, sha256, and case_count. Identifies the exact JSONL export.
scorer	Object selecting the answer-comparison policy.
provenance	Object containing source_datasets, an array of export identities with name, version, case_count, and export_sha256. These identify case exports, not the original real-world data providers.

All three current packs use {"name":"exact-json","version":"1"}. Answers are compared structurally, not by their display formatting: object-key order is irrelevant, but array order and numeric types matter. Follow each question's rounding and ordering requirements.

The loader also supports json-numeric-tolerance version 1, with numeric relative_tolerance and absolute_tolerance fields. That policy allows tolerance for expected floating-point values while preserving exact integer comparisons. It is not selected by the current three manifests; do not silently substitute it when reporting their results.

cases.jsonl: one evaluation case per line

Field	Type and meaning
case_id	String identifying the case within the pack.
case_version	String identifying the case revision.
question	Complete natural-language task, including calculation details and data conventions.
answer_schema	JSON Schema Draft 2020-12 object defining the required answer shape and constraints.
expected_answer	JSON value satisfying answer_schema; used only by the evaluator.
metadata	Object with string fields family (task category) and source_level (provisional difficulty).
derivation	Optional object {"format":"dsa-derivation/v1"} requesting a replay-verified derivation; absent from the current exports.

For example, a one-number answer may use this answer_schema:

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "properties": {"total_mwh": {"type": "number"}},
  "required": ["total_mwh"],
  "additionalProperties": false
}

Schema validation checks answer structure, not factual correctness. Evaluation additionally compares the answer with expected_answer using the manifest's scorer. The full task is in question; the short opening sentence alone can omit important constraints.

The authoritative validated manifest and case contracts are defined in dsa.pack. Each case carries its own answer schema rather than sharing a single universal answer format.

database-card.json: chat presentation

Currently provided only for Online Retail II, this object uses format: "dsa-database-card/v1" and contains:

• title and summary: strings describing the database;
• coverage: an object containing period_start and period_end strings, plus integer counts for transaction_lines, invoices, identified_customers, product_codes, and country_values;
• primary_relation: the schema-qualified name of the recommended table or view;
• relations: an array of objects with name, kind (table or view), row_count, description, and columns. Each column has string fields name, data_type, and description;
• example_questions: an array of strings for user orientation;
• analysis_notes: an array of model-facing analytical conventions, separated from the user-facing description.

This is a presentation document, not an evaluation-case export. Its current coverage fields are retail-specific. The validated contract is defined in the chat application's database-card module.

Revision and verification records

These are supporting authoring records, not additional model inputs or universally versioned runtime contracts. Their current shapes are:

File	Contents
reference-queries.jsonl	Objects with case_id, sql, origin, and array_key. The nullable array_key controls how query rows are wrapped into an answer object by the authoring code.
changes.json	Array of objects with case_id, previous_case_id, integer slot, action, task_type, analytical_operation, difficulty, difficulty_status, reference_status, and booleans answer_schema_changed and answer_changed. Slots are one-based positions in the revision.
source.json	Object with predecessor, database_relative_to_source_root, database_sha256, and status. predecessor contains format, repo_id, revision, path, and manifest_sha256, identifying the earlier pack at a pinned repository commit.
validation.json	Object with pack, version, cases, four digest fields (cases_sha256, database_sha256, manifest_sha256, reference_queries_sha256), counts (replaced, rewritten, recomputed_answers, pinned_predecessor_answers), and count mappings difficulty and types.
verification.json	Object with pack, cases_sha256, manifest_sha256, database_unchanged, production_loader, schemas_checked, reference_queries_replayed, and an array of independent_python_case_slots.

source.json is retained from authoring: its status may say unpublished-local-candidate, and its source path may contain an old version directory. These describe the historical build input, not the current publication status or download path. Use pack.json for current file locations.

The current revision records 244 replayed reference queries and 11 independent Python checks across the three packs. Another 56 answers were retained from pinned predecessors without new independent recomputation. The production_loader receipt records a local-transport check at authoring time; it is not a live service-health report. Difficulty labels are provisional, not calibrated model-performance measurements.

Database provenance records

database/build_report.json varies by dataset; it does not have a single shared schema:

• Online Retail II: source archive and SHA-256, build time and tool versions, worksheet/raw/analysis row counts, overlap checks, quality metrics, and build checks.
• SMARD: database identity and SHA-256, region and time coverage, interval count, licensing and attribution, and a source_files array. Each source-file record includes its identifier, request JSON string, response filename, retrieval time, byte count, row count, and SHA-256.
• EEA: format: "dsa-eea-air-quality-build-report/v1", database identity/size/SHA-256, source details, licensing, reporting-time coverage, cities, pollutants, sampling-point and relation counts, and a source_requests array containing archive/request identities, request JSON strings, checksums, and measurement-coverage statistics.

EEA's separate database/source_manifest.json uses format: "dsa-eea-source-manifest/v1", with retrieved_at_utc, endpoint, and entries. Each entry identifies a source request and archive using id, archive, request, city_id, city, source_city, country_code, start_utc_plus_1, end_utc_plus_1_exclusive, size_bytes, and sha256. Historical request strings are preserved as sent; reporting-time conventions are described below.

Working with the databases

The .duckdb files can be opened directly with DuckDB; using dsa is optional. Tables and views are organized into raw, analysis, and metadata schemas. Start with these relations:

Database	Recommended relations	Important conventions
Online Retail II	analysis.transaction_lines_nonoverlapping	One transaction line per row; 1,044,848 rows after removing cross-sheet overlap from 1,067,371 raw rows. Quantities are signed, prices are in GBP, and timestamps are timezone-naive. Missing identifiers and within-sheet duplicates remain. Follow the case's eligibility rules.
SMARD	analysis.quarter_hourly, analysis.daily	35,136 quarter-hour intervals in the 2024 local year. Use Europe/Berlin for calendar dates and UTC for unambiguous interval identities. Missing measurements are not imputed.
EEA	analysis.valid_measurements, analysis.city_hourly, analysis.city_daily, analysis.city_monthly, analysis.city_annual	Reporting timestamps use fixed UTC+1, not daylight-saving civil time. CO uses mg/m³; the other pollutants use µg/m³. Coverage varies, and missing observations are not imputed; Rome has no PM₁₀ or PM₂.₅ observations in this snapshot.

EEA's valid-measurement relation accepts validity codes 1–4. City-hour means average available valid sampling-point values; longer-period means weight each available city-hour mean equally. The original measurements and metadata remain available for coverage and validity audits.

For evaluation, keep expected_answer, reference SQL, and verification records out of the model's inputs. Supply the database, complete question, and answer schema; use the remaining material on the evaluator side. Public reference answers make this a transparent demo benchmark, not a contamination-resistant hidden test set.

Download one dataset with the standalone Hugging Face CLI:

uvx --from huggingface-hub hf download lschiemanowski/dsa-datasets \
  --repo-type dataset --include 'online-retail-ii/*' --local-dir data

For reproducible experiments, add --revision <commit-sha> and retain the selected commit together with the manifest digest. The main branch can change; a directory name alone is not a frozen benchmark identity. Existing commit-pinned historical packs remain available through repository history.