OpenBallast/ballast-t0

⚓ Ballast T0

We measured how much of a bigger model's factual advantage is just memorized trivia, and whether you can buy that back with a file instead of with parameters. You can, and it's 40–100× cheaper per byte.

This is that file: the reference ballast (a versioned, rank-quantized knowledge artifact you pair with any language model). 25.4M entities, 197.4M cleaned Wikidata triples, 1.51 GB total, published in 8 nested rank buckets: pick your knowledge level like you pick a GGUF quant.

Measured effect (50,147 probes, 90.5% corpus-linked at full corpus; logprob choice scoring):

Gemma-4	raw	+ full ballast	hallucination
E2B	0.608	0.868	0.242 → 0.074
E4B	0.662	0.910	0.197 → 0.042
12B	0.683	0.910	0.207 → 0.049

Qwen3.5	raw	+ full ballast	hallucination
0.8B	0.324	0.784	0.601 → 0.114
2B	0.363	0.770	0.552 → 0.137
4B	0.434	0.831	0.474 → 0.095
9B	0.542	0.819	0.329 → 0.113

Protocol: these are 8-way multiple-choice scores read from the model's token probabilities (prediction = argmax length-normalized answer logprob), with abstention below 0.5 confidence. The hallucination column is the wrong-answer rate among non-abstentions under that protocol. Open-ended generative use scores lower across the board. See the retrieval results for generative numbers.

Both families, bf16, oracle-composed (perfect entity linking). The measured real-linker band realizes ≈63% of the corpus curves' gain. In the Qwen chart the 0.8B + corpus curve ends above every bare model in the family, including the 9B at 6× its footprint.

The pattern replicates across two unrelated families: raw floors spread, ballasted ceilings compress. The ballasted Qwen 4B beats the ballasted 9B (0.831 vs 0.819 on n = 50,147; protocol-specific, see the note above). With perfect entity resolution, +180 MB of ballast lifts Gemma E2B past the 12B's raw factual accuracy; with the shipped non-generative linker (which realizes about two thirds of that ideal gain), the crossing needs ≈470 MB. The exchange rate depends on what you charge the parameters: ≈100× against the ideal-resolution crossing at bf16, ≈40× at the realized crossing against the ≈19 GB bf16 parameter route, and ≈15× if the 12B is charged at Q4_K_M, the cheapest quant that leaves it intact.

Two more measured effects:

• Ballasted accuracy is a quantization-damage diagnostic. 4-bit (nf4) leaves Gemma E2B and 12B nearly intact but collapses E4B's ballasted ceiling from 0.910 to 0.650: when quantization breaks a model's ability to read evidence, no corpus buys it back, and the grounded score is what reveals it.

  

• This fixes answerable questions; it does not teach abstention. On 2-hop composition probes, grounding cuts hallucination 3–20×. On questions with no true answer (false premises, never-recorded facts), fabrication rises with corpus coverage: evidence reads as license to answer.

Full thesis, methodology, and caveats: github.com/OpenBallast/ballast, organized since 2026-08 as a slim overview plus per-topic results docs (equal-bytes, quantization, boost, hallucination, retrieval). The prose tier (full-body Wikipedia passages) now ships separately as ballast-t1.

Layout & quantization levels

Hive-partitioned Parquet; the partition key is the quantization axis. A level Lk = buckets 0..k. Download fewer bucket directories to hold a smaller corpus:

entities/tier=T0/lang=en/rank_bucket={0..7}/*.parquet
triples/tier=T0/lang=en/rank_bucket={0..7}/*.parquet
properties.parquet    # 13,704 property labels, shared by every level
manifest.json         # counts, sizes, hashes, build provenance

level	cumulative	entities
L0	36 MB	top 0.5%
L1	62 MB	top 1%
L2	107 MB	top 2%
L3	179 MB	top 4%
L4	288 MB	top 8%
L5	466 MB	top 16%
L6	756 MB	top 32%
L7	1,507 MB	all 25.4M

Ranking = composite of log(sitelinks) + log(claim count). Buckets nest: L3 is a byte prefix of L7. Truncation degrades honestly: an entity outside the level has no evidence; a triple whose object falls outside the level is dropped, never rendered as a bare Q-id.

Schemas

entities: qid, label, description, aliases[], sitelinks, n_claims, rank triples: qid, pid, value_type (entity|time|quantity|string|coord), value

The partition keys (tier, lang, rank_bucket) are also materialized as columns inside every file, so load_dataset("OpenBallast/ballast-t0", "triples") keeps the level information: filter on rank_bucket <= k to reproduce a level after a full download. To download only a level, prune by path instead: DuckDB's hive_partitioning (below) reads only the buckets the query touches, and snapshot_download(..., allow_patterns=["triples/*/*/rank_bucket=[0-3]/*"]) fetches L3 and nothing deeper.

Excluded at publish: external identifiers, media references, URLs (bytes that ground nothing a language model can use).

Use it

DuckDB (the layout is built for it):

SELECT t.qid, p.label AS prop, t.value
FROM read_parquet('ballast-t0/triples/**/*.parquet', hive_partitioning=true) t
JOIN read_parquet('ballast-t0/properties.parquet') p USING (pid)
WHERE t.rank_bucket <= 3 AND t.qid = 'Q42';

Evidence rendering (the exact format the published numbers were measured under):

Facts about Douglas Adams:
- place of birth: Cambridge
- occupation: science fiction writer
...

Run it locally (serving format)

serving/sqlite/ carries the same corpus as zstd-compressed per-bucket SQLite, the format consumed by the ballast CLI. (The 1.51 GB above is the compressed parquet artifact; the serving format trades bytes for query speed: indexed SQLite runs larger on disk, per the table below.)

uvx openballast pull --level 3     # 275 MB download -> ~1 GB on disk
uvx openballast serve              # OpenAI grounding proxy :11435 + MCP :11436

level	download	on disk		level	download	on disk
L0	54 MB	0.21 GB		L4	444 MB	1.7 GB
L1	95 MB	0.36 GB		L5	721 MB	2.7 GB
L2	165 MB	0.62 GB		L6	1.2 GB	4.4 GB
L3	275 MB	1.0 GB		L7	2.3 GB	9.2 GB

Each bucket_k.sqlite holds that bucket's entities, names (normalized label/alias index), and triples; levels attach buckets 0..k, so upgrading a level downloads only the new buckets. Truncation semantics identical to the parquet artifact. Parquet remains the canonical/build format; the sqlite tree is a derived serving artifact.

Live demo endpoint

Levels L0–L5 are served free at https://mcp.openballast.org (MCP over streamable HTTP + plain GET), running entirely on a $0/month free-tier stack as an existence proof. Demo-grade: no auth, no SLA. Reference: docs/mcp.md.

curl "https://mcp.openballast.org/lookup?question=Where+was+Douglas+Adams+born%3F&level=5"

Curation & processing

This is not a Wikidata mirror; it is a curated, serving-shaped derivation. What was done to the dump, in order:

1. Inclusion. Every item with a resolvable English label (mul honored with English fallback). Wikimedia-internal pages (categories, templates, disambiguation scaffolding) filtered by instance of.
2. Statement filtering. External identifiers, media references, and URL datatypes are dropped at publish; they ground nothing a language model can answer with. What remains is the answerable slice: entity-, time-, quantity-, string-, and coordinate-valued statements.
3. Ranking. Each entity scored by a composite of log(sitelinks) + log(claim count) and assigned to 8 nested buckets at fixed population fractions. The rank axis is the product: it turns knowledge-per-byte into a download-time decision, which no raw dump or existing mirror offers.
4. Rendering contract. The published measurements were made against a fixed evidence rendering (the Facts about X: block above) over these exact bytes; truncation semantics (entity-outside-level, object-drop) are preserved end-to-end in every distribution format. That measurement anchoring (the corpus is the artifact the numbers were measured on, versioned together) is the second thing a mirror does not give you.

Limitations

• Coverage follows notability. The rank composite (sitelinks + claim count) concentrates the head buckets on globally notable, heavily interlinked entities; regional and non-English-world coverage thins earlier in the level ladder than English-world coverage does.
• Wikidata content ships as-is: vandalism that survived patrol, contested values, and personal data about living people (birth dates, family relations) are all present. CC0 licensing does not remove data-protection obligations for downstream use.
• Staleness horizon: facts reflect the 2026-07-20 dump; nothing after that date exists in the corpus. Rebuilds from newer dumps are versioned releases, not in-place updates.
• The measured numbers are protocol-specific (logprob multiple choice, abstention 0.5). Generative use is measured separately and scores lower. See the protocol note above.

Provenance & license

Built from the Wikidata JSON dump (wikidata-20260720; parameters and hashes in manifest.json). mul language code honored with English fallback; Wikimedia-internal pages filtered by instance of. Data: CC0 (Wikidata contributors). Companion eval sets: OpenBallast/ballast-evalsets.

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