anima-clm-default-lane-rung0-byte-18m

The first model ever trained on the anima DEFAULT-LANE unified corpus. A byte-level (vocab256) ConsciousLMReconstructed transformer (≈18.13M params) trained FROM SCRATCH on dancinlab/anima-corpus-5lang-unified-v2 — the unified 5-language corpus (en/fr/de/es/ko: ~40% wiki + ~40% persona/SNS + ~20% KOSMOS-grounded enrichment) that backs lane default. This is the default-lane LADDER rung-0 (a_scale_honest_scope).

Distinct from dancinlab/anima-clm-chat-rung0-byte-18m, which trained on an OLDER, smaller 70%-wiki/30%-dialogue mix. This rung-0 is the default lane's own first model.

arch (identical to the chat-rung0 sibling)

• ConsciousLMReconstructed: byte vocab256, d_model=384, 6 layers, 4 heads, block 256
• dual engine_a / engine_g FFN (out = engine_a(x) − engine_g(x)) + dual head_a / head_g
• params = 18,130,176 (~18.13M)

training (Lane G / GPU)

• substrate: pool host RTX 5070 (nvidia-smi 99% util / 250W / 2.6GB during the fire — GPU-resident, NO CPU fallback). torch-cuda REFERENCE lane for first-achievement, labeled honestly; forge-native (a_train_flame_forge) is the canonical production follow-on (NOT claimed done here). Substrate tag = GPU (a_lane_akida_gpu_split).
• 6000 steps · batch 32 · block 256 · AdamW lr 3e-4 cosine · warmup 300 · seed 42 · from-scratch
• train CE 5.7233 → 0.6983 (step 1 → 6000, ~431s wall)
• corpus sha256 550fed174d51be660810858e1e73e4590c21351b185ea22d0807403e120538ad (13,107,309 bytes)

verification (p7 simple-stack — NO perplexity verdict)

A STRICT structural real-text discriminator (C0-control-ratio < 0.02 + letter/space ratio ≥ 0.65 + no non-text run > 4 — pure structure, NOT perplexity). 5 mixed-language prompts in the corpus's own registers (user: dialogue across en/ko/fr + plain prose de/es).

• TRAINED: PASS 4/5 — coherent on-distribution multilingual text (Spanish/German prose, Korean persona dialogue; the one miss is an immediate-stop on the English dialogue seed).
• random-init mirror (anti-Goodhart): FAIL 0/5 — control-byte soup, rejected by the same evaluator. anti_goodhart_ok = TRUE, chat_pass = TRUE.

Note: an initial lenient evaluator scored the mirror 5/5 (Goodhart trap — str.isprintable() counts random Unicode/symbol soup as printable). The strict evaluator above is the honest one that actually separates trained text from random bytes; both are recorded in the repo verdict.

honest scope (a_scale_honest_scope)

🟢 coherent multilingual output at 18M-on-v2-default. Scope = small/toy byte rung; 18M only. Scale-transfer to mid / 7B is UNVERIFIED here (a later ladder rung, gated on this going 🟢). Not a production-scale claim. CE is ONE axis, not a perplexity truth (p7).

philosophy (p1·p2·p3·p4·p6 held)

NO system prompt, NO identity rules, NO persona injection, NO assistant framing, NO RLHF. The corpus turn structure is plain <speaker>: continuation with no role/persona tags (corpus tag-grep = 0). Capability is the learned byte distribution only.

files

• default_lane_rung0_18m.pt — torch ckpt (model_state + config + ce_log), sha256 4285bf35b002e1c2495587d38842fde0a2a6be4a40cb2eb513e7e762293fc476

Trainer: training/default_lane_rung0_train_eval.py (anima repo). Co-Authored-By Claude.