I can't share the corpus itself — it's built from private traffic, and the imatrix is basically a fingerprint of it. But the corpus was never the secret. The shape was, and that part reproduces from your own data:
~525K tokens, 30% agentic / 30% code / 15% conversation / 12.5% math / 12.5% writing. The ratio is the product, not the provenance.
Three things that actually moved the needle:
Put your system prompt and tool definitions in it. In agentic use those tokens are in every single forward pass, and they're completely absent from a wikitext-style imatrix. It's the activation pattern the model always fires.
Go bigger than the usual ~41K tokens. At 10x that, mixed, the expert ranking gets much more stable.
Dedupe before you weight. Agentic logs repeat the same system prompt thousands of times and the imatrix collapses onto it. And keep ~10% general text as insurance — a corpus overfitted to your own traffic quietly costs general capability.
Why it's worth the trouble: same K, same bits, same disk, changing only the corpus took HumanEval from 89.6% to 95.1% and dropped fabrication on a knowledge probe from 58% to 17%. Which experts you keep matters more than how many.
One warning, because it cost me: don't rank two selections by retained routing mass. The metric is circular — each selection wins when scored against the corpus that produced it. Mine confidently predicted the better build would lose. Budget for a real benchmark.
