README.md

metadata

license:
  - cc-by-sa-4.0
  - cc0-1.0
language:
  - en
pretty_name: Ballast T1
size_categories:
  - 10M<n<100M
task_categories:
  - text-retrieval
tags:
  - knowledge-base
  - wikipedia
  - rag
  - grounding
  - hallucination
configs:
  - config_name: passages
    data_files: passages/lic=by-sa/rank_bucket=*/*.parquet
  - config_name: statements
    data_files: statements/statements.parquet
  - config_name: labels
    data_files: statements/labels.parquet

⚓ Ballast T1

The prose tier of the Ballast knowledge artifact: full-body English Wikipedia as rank-quantized passage chunks. It spans 37.4M chunks across 7.0M articles (11.9 GB zstd parquet), in the same nested-bucket layout as Ballast T0. Where T0 carries facts as triples, T1 carries the paragraphs those facts live in. This is the tier for narrative and long-tail questions that triples cannot serve.

Measured context (all numbers, protocol, and caveats: results-retrieval.md): with Qwen3.5-9B-Base (int8) answering open-ended (≤24 tokens, normalized alias-containment grading) on 2,000 probes from SimpleQA + NQ-open, grounding against this corpus with a CPU-only retriever moves accuracy 0.1540 → 0.3315 (single-pass 0.2885; the two-pass support-check architecture adds the rest, and revalidated held-out on PopQA). At equal or less GPU memory, that beats a 12B model running alone (0.1565) by +17.5 points. These are generative string-match numbers: a different and harder instrument than the logprob multiple-choice scores on the T0 card; the two do not compare.

Layout

Hive-partitioned Parquet; the partition key is the corpus-quantization axis (same rank composite as T0: log sitelinks + log claim count). A level Lk = buckets 0..k. manifest.json at the repo root carries per-file counts, bytes, sha256 hashes, and build provenance.

passages/lic=by-sa/rank_bucket={0..7}/*.parquet
statements/statements.parquet     # 31.9M whitelisted Wikidata statements (CC0)
statements/labels.parquet         # 10.2M entity labels for rendering them
manifest.json

passages schema: qid, chunk_idx, title, heading, text (+ rank_bucket and lic from the path). Chunks are paragraph runs merged greedily to a ≈1.2 KB target and are atomic: serving stacks pack whole chunks or skip them, never truncate. This is the byte-identity condition the published measurements depend on.

bucket	chunks	articles	bytes	cumulative
0	1.96M	75K	662 MB	0.66 GB
1	0.76M	76K	258 MB	0.92 GB
2	1.46M	150K	582 MB	1.50 GB
3	2.25M	332K	784 MB	2.29 GB
4	3.28M	642K	1.00 GB	3.29 GB
5	4.69M	931K	1.40 GB	4.69 GB
6	7.04M	1.34M	2.13 GB	6.82 GB
7	16.0M	3.47M	5.05 GB	11.86 GB

Use it

SELECT chunk_idx, heading, text
FROM read_parquet('ballast-t1/passages/**/*.parquet', hive_partitioning=true)
WHERE rank_bucket <= 3 AND qid = 'Q42'
ORDER BY chunk_idx;

To download only a level: snapshot_download("OpenBallast/ballast-t1", repo_type="dataset", allow_patterns=["passages/*/rank_bucket=[0-3]/*", "manifest.json"]).

Run it locally (serving format)

serving/sqlite/ carries the same chunks (byte-identical to the parquet distribution) as per-bucket SQLite for the ballast CLI (v0.3.0+). The name index joins T0 labels and aliases by QID, with a normalized-title fallback for the 1.28M tail articles T0 never named. Entity-name lookup over prose is functional but unbenchmarked: the measured retrieval numbers above come from the two-pass research stack, which the CLI does not run yet.

uvx openballast pull --tier t1 --level 3   # 1.6 GB download, 8.7 GB on disk
uvx openballast serve --corpus t1          # proxy :11435 + MCP :11436

level	download	on disk		level	download	on disk
L0	0.52 GB	2.0 GB		L4	2.42 GB	12.1 GB
L1	0.71 GB	3.1 GB		L5	3.52 GB	16.7 GB
L2	1.09 GB	5.7 GB		L6	5.16 GB	23.5 GB
L3	1.64 GB	8.7 GB		L7	8.98 GB	40.5 GB

Sizes are measured from serving/sqlite/serving_manifest.json (per-file bytes + sha256), which pull also reads to print cost before downloading.

The statements config is the typed-fact sidecar (subject s, property prop, object o or literal o_lit), restricted to a ~70-property whitelist over English-Wikipedia entities, useful as a structured cross-check on prose answers.

Curation & processing

Not a Wikipedia mirror: a chunked, rank-quantized, license-cellular derivation of one. The pipeline, in order:

1. Scope. Article namespace only, from the pages-articles-multistream dump; redirects resolved (their titles feed the alias index rather than producing duplicate articles); Category/Portal/Template pages excluded. The build covers 99.4% of the dump's 7.54M articles.
2. Section filtering. References, external links, see-also, bibliographies, galleries and similar tail sections dropped by heading match before parsing.
3. Stripping & chunking (t1-chunk-v1). Wikitext stripped with a deliberately simple parser: templates, infoboxes, and tables are dropped, not rendered (cost measured and stated under Limitations); paragraphs merged greedily to a ≈1.2 KB atomic-chunk target, first chunk of each section prefixed Title — Heading: for self-containment.
4. Rank join. Articles joined to the T0 entity rank map by QID, so the prose tier shares the triple tier's notability axis and one entity spine; unmapped articles land in the tail bucket rather than being dropped. This cross-tier alignment (same QIDs, same buckets, prose + triples + statements composable) distinguishes this from existing passage dumps (KILT, wiki-dpr and kin are also years staler and single-granularity).
5. Dedup. Exact (qid, chunk_idx) duplicates removed across parse shards at packaging.
6. Statements sidecar. The CC0 statements config is Wikidata truthy statements restricted to a ≈70-property whitelist (media, sports, family, positions, taxa: the relation types the eval populations actually ask about), both endpoints restricted to English-Wikipedia entities, plus the labels needed to render them.

Limitations

• Raw encyclopedia prose, shipped verbatim. Chunks contain whatever English Wikipedia contained at the snapshot: biographical detail about living people, descriptions of violence and other sensitive topics, and any vandalism that survived patrol. Filter downstream to your use case.
• The parser is deliberately simple. Templates, tables, and most markup are dropped; occasional markup residue (HTML comments, entities) survives in chunk text. A raw-vs-stripped census priced the cost at 3.2 percentage points (NQ-open) / 4.4 (SimpleQA) of gold-answer capture versus a perfect parser. The gap is accepted deliberately; the residue is noise, not silent loss.
• Coverage follows notability, as in T0: head buckets skew to globally notable topics; articles without a mapped T0 entity land in the tail bucket (7).
• Staleness horizon: the enwiki-20260701 dump. Rebuilds are versioned releases.

Provenance & license

Parsed from the English Wikipedia pages-articles-multistream dump (enwiki-20260701): the same corpus build the published retrieval results were measured against, plus the statements sidecar. Ranking joins the T0 rank map.

Passages: CC BY-SA 4.0 (Wikipedia contributors). Attribution resolves through each chunk's title to the source article and its revision history; derivatives of the text carry share-alike. Statements + labels: CC0 (Wikidata contributors). The front-matter lists both licenses; each config's license is stated here and in manifest.json. Companion artifacts: ballast-t0, ballast-evalsets.

Questions and issues: github.com/OpenBallast/ballast/issues.