README.md

metadata

license: cc-by-4.0
language:
  - en
pretty_name: Ballast T2
size_categories:
  - 10K<n<100K
task_categories:
  - text-retrieval
tags:
  - knowledge-base
  - textbooks
  - openstax
  - rag
  - grounding
configs:
  - config_name: passages
    data_files: passages/tier=T2/lang=en/lic=by/rank_bucket=*/*.parquet
  - config_name: sources
    data_files: sources.parquet

⚓ Ballast T2

The textbook tier of the Ballast knowledge artifact: 61 OpenStax textbooks as passage chunks, 90,445 chunks across 9,018 linked sections (≈92 MB of text, 34 MB zstd parquet). Where T0 carries facts as triples and T1 carries encyclopedia prose, T2 carries pedagogical prose: definitions, worked explanations, and method text of the kind reference corpora state but never teach.

Layout

passages/tier=T2/lang=en/lic=by/rank_bucket=7/part-0.parquet
sources.parquet        # one row per book: name, homepage, pinned git SHA, license
licenses.parquet       # license registry the per-chunk lic key resolves to
manifest.json          # per-file counts, bytes, sha256, build provenance

passages schema matches T1 (qid, chunk_idx, title, heading, text; chunker t1-chunk-v1, ~1.2 KB atomic chunks). Two T1 conventions do not apply here: T2 v1 has no notability ranking; every chunk sits in the tail bucket (rank_bucket=7), so corpus levels are a no-op; and chunks are keyed to book sections rather than Wikidata entities.

SELECT title, heading, text
FROM read_parquet('ballast-t2/passages/**/*.parquet', hive_partitioning=true)
WHERE text ILIKE '%standard deviation%' LIMIT 5;

Run it locally (serving format)

serving/sqlite/ carries the same chunks as a single-level SQLite corpus for the ballast CLI (v0.3.0+): 25 MB download, 116 MB on disk. Documents are the 9,018 book modules; the name index is module titles plus per-module index terms, so lookup resolves questions like "what is marginal cost" to the right chapter. Entity-name lookup over textbook prose is functional but unbenchmarked.

uvx openballast pull --tier t2
uvx openballast serve --corpus t2     # proxy :11435 + MCP :11436

Serving ids are corpus-local (D…); provenance stays in this parquet distribution (chunk_id, pinned revision, source_url per chunk).

Provenance: pinned against a license flip

Every book is ingested from its official OpenStax source repository at a pinned git commit, recorded per book in sources.parquet (access_path = git:<repo>@<sha>, retrieved 2026-08-02). The pinning is load-bearing: parts of the OpenStax catalog moved from CC BY to CC BY-NC-SA in spring 2026, and aggregator license metadata did not always track the change. This distribution therefore trusts only the license text at the pinned commit of the official repo, never catalog metadata.

The research build behind this artifact contains a second cell: 9 post-flip titles (10,050 chunks) under CC BY-NC-SA 4.0. That cell is excluded from this distribution to keep the published artifact commercially safe under a single license. The full manifest documents both cells, so the exclusion is visible rather than silent.

Curation & processing

What was selected, what was done to it, and why, so this is auditable rather than "some textbook text":

1. Selection. The OpenStax catalog, and nothing else: every book with an official source repository under github.com/openstax at retrieval time (2026-08-02). No aggregators, no mirrors, no PDF scrapes: the official CNXML/collection source is the only ingestion path, because it is the only place license status is authoritative.
2. Pinning. Each repository is pinned to one commit (sources.parquet.access_path), chosen as the newest commit whose LICENSE grants CC BY 4.0. Books whose current license is CC BY-NC-SA are carried at their NC state, not silently rolled back; that cell is excluded from this distribution (see Provenance).
3. License resolution. Per-book license read from the repo's license file at the pinned commit under versioned rules (t1-license-rules-v2), never from catalog or aggregator metadata: the spring-2026 relicensing showed the metadata lags the repos.
4. Extraction & chunking. Book markup to plain prose with the same deliberately simple parser as T1 (t1-chunk-v1): boilerplate sections dropped, figures and math markup dropped, paragraphs merged greedily to a ≈1.2 KB atomic-chunk target, section heading carried per chunk.
5. Dedup. Exact-duplicate chunks removed within the tier (dedup_ledger.parquet records what, if anything, was suppressed: zero for this build).
6. Why this tier exists. The companion tiers carry facts (T0) and encyclopedia prose (T1). Pedagogical prose is a different knowledge shape (definitions and worked method text), and keeping it a separate, separately-licensed tier lets downstream users include or exclude it deliberately rather than inheriting it mixed in.

Limitations

• Pedagogical prose, verbatim. Exercises, figure references, and chapter cross-references survive in chunk text; figures themselves and math markup are dropped by the parser (t1-chunk-v1, the same deliberately simple parser as T1, with the same class of markup residue).
• No notability ranking (see above): download is all-or-nothing at 34 MB, which at this size is not a hardship.
• English only, one catalog. T2 v1 is OpenStax-only; the tier is designed to grow by adding pinned sources, not by re-crawling.

License & attribution

CC BY 4.0 (OpenStax, Rice University). Attribution resolves through each chunk's partition and title to the book row in sources.parquet, which carries the book's name, homepage, and pinned source commit. OpenStax is not affiliated with and does not endorse this distribution.

Companion artifacts: ballast-t0, ballast-t1, ballast-evalsets.

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