README.md

metadata

title: OpenBallast
emoji: 💻
colorFrom: green
colorTo: pink
sdk: static
pinned: true
license: cc-by-4.0
short_description: Just download more VRAM

⚓ OpenBallast

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.

Small models reason fine — they just don't know much. Parameters are the most expensive place to store facts (~100× more bytes per fact than a compressed triple corpus). OpenBallast builds ballasts: versioned, CC0, rank-quantized knowledge artifacts you pair with any local model — pick your knowledge level like you pick a GGUF quant.

Measured (two model families, 50,147 linked probes + a 43,137-probe hallucination suite):

• Raw, Gemma-4's 2B/4B/12B score 61/66/68%; given the same corpus to look facts up in, all three land at 87–91%. Replicates on Qwen3.5 (0.8B–9B: raw 32–54%, ballasted 77–83%, and the ballasted 4B beats the ballasted 9B). Size gaps are mostly memorization gaps.

• A 2B + a 470 MB file beats the 12B raw with a real, non-generative lookup in the loop; the parameter route costs ~19 GB. Hallucination on answerable questions: 0.24 → 0.07.

• Grounding cuts multi-hop hallucination 3–20× but raises fabrication on unanswerable questions — it fixes answerable questions, it does not teach abstention.

• Ballasted accuracy doubles as a 4-bit damage diagnostic: nf4 breaks some models and not others, unpredictably from size, and only the grounded score tells you which.

• Corpora tuned to a specific model's knowledge gaps lose to the one generic corpus at every size — a decisive negative result.

• 📦 ballast-t0 — the corpus, 8 quantization levels, CC0

• 🧪 ballast-evalsets — 50k recall probes + 43k hallucination probes

• 📖 Thesis + methodology + numbers

• ⚡ Live demo endpoint (MCP): https://mcp.openballast.org — runs on a $0/month stack