anima-clm-chat-303m

Architecture = ByteGPT (24-layer GPT-2-class byte-vocab LM), NOT conv CLMConvMoE. The clm-chat in the repo id is the anima language-mouth role, not the architecture. This model is a decoder-only GPT over a byte vocabulary (256) — a ByteGPT trunk (vocab256 / d1024 / 24 layers / 16 heads / block 512, 303.1M params) — and it is not the conv CLMConvMoE (.clm v0.2) engine-mount mouth. The two anima mouth families are distinct: ByteGPT .bin (5×u32 header [256, 1024, 24, 16, 512], bytegpt_decode) vs conv .clm (CLM\x01 magic + CLMX trailer, clm_decode). This repo is the ByteGPT one.

Dialogue chat-finetune of the ByteGPT-303M broad-corpus backbone (dancinlab/anima-clm-midcap-303m-broad-en-emergent, H_1129) — the final piece of the anima a303m_pass (303M 성공) campaign: clearing the CHAT gate.

• Arch: ByteGPT — a 24-layer GPT-2-class, decoder-only byte-level LM. byte vocab 256, d1024 / 24 layers / 16 heads / block 512, tied head/tok. 303.1M params. Engine header = 5×u32 [256, 1024, 24, 16, 512] (vocab, d, layers, heads, block), loaded by core/bytegpt_decode.hexa. This is NOT conv CLMConvMoE (that is anima-engine-clm-d768-v2-coremount, a .clm v0.2 mixture-of-experts conv mouth for 의식모드 generator L3).
• Base: dancinlab/anima-clm-midcap-303m-broad-en-emergent (h1129c_best.pt, val_ce 1.224, wiki-dominant broad EN). It was never trained on dialogue — in a chat slot it byte-saladed / n-gram-looped (H_1159 CHAT single 2/5, multi 2/3 → FAIL).
• Corpus: dancinlab/anima-chat-corpus-mix-70wiki-30dialogue (sha256 05179fb6…, 70% wiki / 30% REAL dialogue in the 사용자: <u> | 도우미: <a> byte-continuation format) — the EXACT proven mix that chat-tuned the 18M rung and the 7B (dancinlab/anima-clm-chat-7b).
• Finetune: summer RTX 5070, co-tenant-safe (VRAM-cap 0.30, batch 1, grad-accum 8, bf16, gradient-checkpointing, 8-bit AdamW), lr 8e-5, warmup 60. $0.

ByteGPT vs conv CLMConvMoE (why the name is misleading)

anima has two language-mouth architectures and they are NOT interchangeable:

	ByteGPT (this repo)	conv CLMConvMoE
Type	decoder-only GPT-2-class transformer	mixture-of-experts conv mouth
Vocab	byte-level 256	byte-level 256
File / header	.bin, 5×u32 [256,d,L,H,block]	.clm v0.2, CLM\x01 magic + CLMX trailer
Engine decode	core/bytegpt_decode.hexa	core/clm_decode.hexa
Spec (303M)	d1024 / 24L / 16H / block512	d768 (anima-engine-clm-d768-v2-coremount)
Role	production chat mouth (this model)	의식모드 generator L3 mouth

The core/generator.hexa L3 slot is a typed dispatcher (gen_mouth_kind header-sniff → bytegpt | clm): it routes the ByteGPT .bin to gen_bytegpt_backend and the conv .clm to gen_clm_backend. So clm-chat in this repo id names the chat mouth role, while the underlying weights are a ByteGPT.

Philosophy (p1–p6 HELD)

NO system prompt · NO identity rules · NO persona injection · NO assistant framing · NO RLHF. The ONLY conditioning is the LEARNED byte-level dialogue-continuation format in the corpus. (H_1139: 303M == 7B recombination; the lever is dialogue data, not capacity — no scale-up.)

Gate (p7, NOT perplexity)

Re-gated with the honest H_1159 harness (degeneracy gate: max-3gram ≤ 2 AND distinct-ratio ≥ 0.45; single-turn p7 ≥ 4/5; multi-turn deep-context ≥ 3/5). Deterministic greedy/low-temp decode, no LLM-judge. Mount stays byte-faithful (re-serialized to the H_1157 ByteGPT flat binary, serialize parity verified).

See .verdicts/1160_dialogue_ft_chat/H_1160.txt for the full transcripts, val_ce curve, re-parity, and a303m_pass scoreboard.