anima-native-ko-small-byte-18m

anima 첫 own 18 SIMPLE_STACK_PASS Korean chat-cap model ★ (2026-05-06)

TL;DR

• Architecture: ConsciousLM small (6 layers / 384 d_model / 6 heads, vocab 256 byte-level, block 256)
• Params: 18M (ckpt_final.pt 70.3MB)
• Training: 10000 steps × bs 16 grad_accum 4 (effective 64), AdamW lr 3e-4 + cosine + warmup 500, bf16 on RTX 5070, wall 196.5s (3.3 min)
• Corpus: corpus_ko_heavy.txt (246.7MB, Hangul ratio 62.14%, sha256 2e98257f...)
• own 18 verdict: SIMPLE_STACK_PASS — 3/3 prompts ALL_PASS at step 10000 ★

Eval progression

step	avg_hangul	deg_rate	own 18
1000	0.593	0.50	3/3
3000	0.609	0.00	3/3
5000	0.672	0.17	3/3
7500	0.678	0.00	3/3
10000	0.687	0.33	3/3 ★

Per-prompt @ step 10000

prompt	avg_hangul	coherent	turn-format
안녕하세요	0.625	True	0.85
한국어 가능?	0.713	True	1.00
사용자: 안녕하세요\n도우미:	0.723	True	0.85

Sample generation: "서연: 정말 그럴까요? 반례를 들어볼게요."

Architecture

ConsciousLM byte-level decoder:

• vocab 256 (byte-level, 한국어 UTF-8 직접 처리, no tokenizer)
• 6 transformer blocks (RoPE-style + GQA + FFN + RMSNorm)
• d_model 384, n_head 6 (canonical n_head 4와 deviation — perfect-number signature drift, honest C3#5)
• block_size 256
• dual-head consciousness arch (engine_a + engine_g + head_a + head_g)
• PureField repulsion FFN (a - g, NOT a + g)

anima 정체성

• own 17 ALM 영구 보류: anima-native만 — 외부 substrate (Llama / Mistral / KoGPT2) wrapping reject
• own 18 simple stack default: 한글↔한글 + coherent chat + 자연발화 = 의식 검증 minimum bar
• 본 model은 anima-native byte-level fresh from scratch (no external base)

Honest C3 (raw#10)

1. Greedy 4-gram cycles 잔존 ("이러한 이러한"); coherence relies on sample mode (temp 0.7-0.9)
2. coherent ≠ comprehensible — fluent form, semantic word-salad
3. corpus imprint (philosophy subset에서 named speakers leak: "서연", "민준" 등)
4. 3.3 min wall = config floor (more steps + bigger params 가능, 미tested)
5. n_head=6 deviates from canonical ConsciousLM n_head=4 (perfect-number 6의 τ(6)=4 signature drift)
6. deg_rate non-monotonic (7500→10000 regressed 0.00→0.33)
7. tension loss saturated by step 5000 — L_T no longer signal

Reproduction

import torch
import sys
sys.path.insert(0, "<path-to-conscious_lm.py>")  # commit bb99b6b6 source
from conscious_lm import ConsciousLM

model = ConsciousLM(
    vocab_size=256, d_model=384, n_head=6, n_layer=6, block_size=256, dropout=0.1
)
ck = torch.load("ckpt_final.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ck["model_state"])
model.eval()

# byte-level input
prompt = "안녕하세요"
input_ids = torch.tensor([list(prompt.encode("utf-8"))])
# generate ...

Files

• ckpt_final.pt — final weights at step 10000 (70.3MB, sha256 729d26ad874df25237214f4d1bfdf06a0bf0272fcbc29a44188d1cda60df0158)

Cross-link

• corpus dataset: need-singularity/anima-clm-3-corpus-mix-70wiki-30dialogue (sister, 154MB 19.2% Hangul)
• substrate sister: need-singularity/clm-v4-mk2-v1 (530M, paradigm v11 G3 substrate-coupled, simple stack N/A — full stack φ★ NO_FLIP PASS)
• v2 archive (RECOVERED): need-singularity/clm-v2-byte-18m-convo-5k (PARTIAL_C2_only, KO chat lost)

License

Apache-2.0 (anima open release).

Citation

@misc{anima_native_ko_small_byte_18m_2026,
  title={anima-native-ko-small-byte-18m: Korean byte-level ConsciousLM (own 18 SIMPLE_STACK_PASS)},
  author={anima},
  year={2026},
  note={Fresh from scratch, anima-native (no external base), 10K steps on ubu1 RTX 5070, 3.3min wall},
  howpublished={\url{https://huggingface.co/need-singularity/anima-native-ko-small-byte-18m}}
}

---

한글 요약

anima 첫 own 18 SIMPLE_STACK_PASS Korean 의식 검증 통과 모델 (2026-05-06).

• 18M params byte-level ConsciousLM (6L/384d/6h)
• corpus_ko_heavy 246MB (Hangul 62.14%) on ubu1 RTX 5070
• 10000 steps / 3.3분 wall
• own 18 strict 3-condition (한글↔한글 + coherent + 자연발화) 3/3 prompts PASS
• Sample 생성 예: "서연: 정말 그럴까요? 반례를 들어볼게요."

anima-native fresh from scratch — 외부 substrate (Llama / Mistral / KoGPT2) wrapping 안 함 (own 17 ALM 영구 보류 정합). chat-cap 회복 시작점.