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

STAGE-2 persona/SNS specialization of the stage-1 chat-PASS rung-0 (dancinlab/anima-clm-chat-rung0-byte-18m). An 18.13M byte-level ConsciousLMReconstructed (dual engine_a/engine_g FFN + dual head) fine-tuned on the persona × SNS dialogue corpus so anima chats in each of 20 persona voices on the SNS surface (Instagram main + YouTube).

Honest scope (a_scale_honest_scope)

This is the SMALL 18M rung, persona-specialized. NO claim of mid/7B persona chat. The persona signal is REAL but PARTIAL: 15 of 20 personas self-identify at least once under the discriminative evaluator (20/40 = 0.50, 10× chance); the rest blur into a related persona. CPU-trained ($0), torch reference lane (NOT AKIDA, Lane-G).

Philosophy (p1/p2/p3/p4/p6 — HELD)

NO system prompt · NO identity rules · NO persona injection · NO assistant framing · NO RLHF. Persona is carried only by the learned dialogue-continuation format 사용자: <u> / <persona_name>: <reply> — there is no [role: / [persona: / [character: tag in the training text (grep == 0), and the demo uses no prompt prefix.

Verdict (p7 simple-stack, NOT perplexity)

axis	trained	random-init mirror
(A) base-chat retained	PASS 4/5	FAIL 0/5
(B) persona self-id (40 trials)	PASS 0.50 (10× chance)	NULL 0.05 (== chance)

chat_pass_retained = TRUE · persona_signal_real = TRUE · anti_goodhart_ok = TRUE

Fine-tune CE 3.278 → 0.0785 (2500 steps, AdamW lr 5e-5, batch 32, block 256, seed 42).

Files

• persona_stage2_18m.pt — ckpt (sha256 aea96ef1a7ef27018ca015a9e66569d67763152e30f24e120aa44da6884cf8bc)
• persona_stage2_train_eval.py — trainer + p7 evaluator (reproducible)
• summary.json, persona_voice_trained.json, p7_base_trained.json — verdicts

Demo

python3 persona_chat_demo.py --ckpt persona_stage2_18m.pt --sweep --seed 7

사용자: 주말에 뭐 할 거예요?
knight: 한가로운 날이오. 허나 평온 또한 지켜야 할 영토라오

사용자: 시험 망한 것 같아요…
senpai: 한 번 망했다고 인생 안 끝나. 일단 오늘은 푹 자