Note clm_prod 3B/7B 5-lang BACKBONE corpus SAMPLE (mC4 odc-by, real_fraction=1.0, 20k lines/67.7MB). Step 2 data prep; full set = step 3 H100.

Note E-31 knuth31 carving anchors — parser-validated 31/31, CC-BY-SA, closure-PASS

Note 20-persona × SNS(Instagram main + YouTube) 롤플레이 대화 corpus — 7B chat-cap stage-2 + .kosmos anchor (tier52 사회성)

dancinlab/anima-clm-persona-sns-rung0-byte-18m

Text Generation • Updated about 12 hours ago

Note STAGE-2 persona/SNS specialization — 18M byte chat, 20 personas, p7 PASS + persona signal 10x chance (a_scale_honest_scope: partial 15/20, 18M-only)

Note argcopy memorization-proof corpus — produced the 🔴 argcopy closed-negative

Note agent-lane layer-3 tool-DOMAIN knowledge — authored conceptual coverage of 5 AGENT tool domains (CODE/TRADING deep, MERCHANT/DESKTOP/CREATOR procedural), 5-lang byte256, TRADING conceptual-only gate

Note default-lane v3 SCALE-UP corpus (~217MB, MID-rung viable, NOT 7B) — wiki 100MB CC-BY-SA + persona + enrichment, byte-vocab256

Note GB-balanced KOSMOS-tier-ladder default-lane corpus (0.35GB, wiki+Gutenberg-PD+persona)