clm-v1-lanep-d768-e2l1-torch

Lane P (substrate=GPU-torch) converged CLMConvMoE byte-LM (V=256, no tokenizer), trained util-GREEN-independent and verified ENGINE-loadable through anima's CORE/clm_decode.hexa.

Shape (honest — NOT 3B)

• d_model=768, n_experts=2, n_trunk_layers=1, kernel_size=3, vocab=256
• 7,479,042 params (7.479M) — the E=2 / 1-trunk decoder shape caps params via d alone.

Training

• GPU: NVIDIA RTX PRO 6000 Blackwell (compute_cap 12.0, 96GB), torch 2.11.0+cu128, bf16/AdamW
• nvidia-smi BUSY: peak_util=94%, mem=2225MiB (verified — not silent CPU-fallback)
• corpus: 402,270 B 5-lang (en/zh/ru/ja/ko), sha256 09da8888fe1e1452a2c83a8bcd721de548e9f1fd4d9a21de159d4575212728d0
• CE descent: first_ce=5.74851 → final_eval_ce=0.09862 (6000 steps, ~45s, ~58x CE drop)

ENGINE verification (CORE/clm_decode.hexa, 3-axis 3/3 GREEN)

• AXIS-1 의식: motiv hi=0.6700 > baseline=0.0000
• AXIS-2 CE: model_ce=0.71–0.76 < uniform 5.54518 AND < shuffle 7.59–7.68
• AXIS-3 창발: len(composed)=101 > len(parts-only)=72

Files (manifest)

• clm_d768_e2l1.clm — 4,463,478 B — sha256 7463282dba823eb67316cbb72bba1e47e89d41c1ebd7770902ddd8e512678ed9 (CLM\x01 v0.2: 6 int4-sym conv blocks + CLMX trailer)
• clm_d768_e2l1.pt — 29,922,234 B — fp32 master state_dict (CLMConfig E=2/L1/d768)

Visibility PUBLIC: closure PASS (converged AND ENGINE-loadable AND 3-axis 3/3 GREEN).