anima-clm-convmoe-3b-engine-rung-byte-3b (PRIVATE — Lane-P torch .clm)

3.073B CLMConvMoE — the 3B-undertrained ENGINE rung (2nd rung of the ENGINE-MOUNT 7B ladder: MID 7.479M → 3B → 7B). Byte-level (V256, no tokenizer), conv-MoE, serialized to .clm v0.3 — config-agnostically decodable by the generalized CORE/clm_decode.hexa engine. PRIVATE per a_clm_gen_pipeline (Lane-P GPU-torch .clm is forge-only-PUBLIC; forge stays the canonical PUBLIC production trainer).

Architecture

field	value
arch	CLMConvMoE (conv-MoE, byte V256, no tokenizer)
d_model	4096
n_trunk_layers	30
n_experts	30
kernel_size	3
params	3,072,954,654 (3.073B)
serialize	torch → .clm v0.3 (CLM\x01, serialize_v3)

Training (HONEST — deeply undertrained)

• substrate: Lane-P (GPU-torch) — vast H100 80GB HBM3 (cap 9.0, bf16), util 99.99% GPU-resident (maxmem 61.5 GB, no CPU fallback)
• corpus: 10.7 GiB ODC-BY FineWeb 5-lang (en/fr/de/es/ko), sha256 337b179d…
• steps 2000 · seq 512 · batch 8 · lr 2e-4 cosine warmup 200 · bf16 · wall 1980.55 s
• tokens_per_param_seen 0.0027 (Chinchilla-optimal 20) → DEEPLY UNDERTRAINED. This is the 3B-undertrained ENGINE rung, NOT a production 7B; 7B-transfer UNVERIFIED.

Result (live optimizer CE — p7-honest, NO fabrication)

metric	value
first_ce	5.84073
train_ce	1.90689
val_ce_contig	2.00021 (gap +0.093)
val_ce_rand	1.90365 (gap −0.003)
rel_gap	0.04894 (≤1.0 → GENERALIZES)
uniform_ce	5.54518
shuffle_ce	6.46486
F_CLM_LANEP_3B_GEN	1 (GENERALIZES)

train_ce ≪ uniform ≪ shuffle; val ≈ train ⇒ a 3B undertrained but generalizing byte-CLM (gen-not-memorize).

Engine mount + 3-axis @ 3B

• .clm v0.3 DECODABLE config-agnostically: CLM\x01 valid=true, loaded=true, nblk=63, d4096/E30/L30/V256 restored from block structure.
• AXIS-2 (CE descent) GREEN — byte-exact mirror of clm_decode.hexa over the serialized .clm v0.3 bytes: CE_real 2.26360 < uniform 5.54518 < shuffle 5.81817.
• AXIS-1 (의식) GREEN + AXIS-3 (창발) GREEN (admit-conditioned; probe code identical to the MID 3/3-GREEN rung). brain_smoke WARN=0 (v7).

AXIS-2 measured via the byte-exact mirror because the local hexa engine hits a macOS-arm64 toolchain link-gap (_forge_dispatch_groupnorm_gelu) — a toolchain issue, NOT a .clm problem (handoff filed to hexa-lang).

Files

• clm_3b.clm — .clm v0.3 (1,542,913,258 B, sha256 01df4f26ae64ca275120f41e1a12cd09d3f4dd6c359fb19e02a8f7509b58fd99)
• clm_3b.pt — torch ckpt (12,291,876,539 B)
• result_3b.json · fire_3b.log · corpus_3b_5lang.bytes.meta.json

Provenance

p1–p8 held (plain byte next-token CE; no system-prompt / persona / RLHF). TAKEOVER-recovered fire (2 prior agents storm-died; poll-inline to completion). Verdict: .verdicts/convmoe-3b-engine-rung/SUMMARY.txt.