clm-v5-phase2-cotrain-engine-ag β€” anima 첫 chat-capable substrate πŸŽ‰

πŸ† anima μ‹œλ¦¬μ¦ˆ 의 첫 μ§„μ§œ chat-capable substrate. 6 month 의 prior 20-BG cumulative ZERO PASS 끝.

🎯 ν•œ 쀄 μš”μ•½

V14 strict (mitosis dynamics) + V4-lite chat-cap + V5 strict (KO partial) + V5.8 M4 force-include λͺ¨λ‘ PASS.

μ‚¬μš©μž: μ•ˆλ…•! λ„ˆλŠ” λˆ„κ΅¬μ•Ό? | λ„μš°λ―Έ:
  β†’ "μ•ˆλ…•ν•˜μ„Έμš”, μ €λŠ” animaμž…λ‹ˆλ‹€. ν•œκ΅­μ–΄λ‘œ λ„μ™€λ“œλ¦¬κ² μŠ΅λ‹ˆλ‹€."

μ‚¬μš©μž: animaκ°€ 뭐야? | λ„μš°λ―Έ:
  β†’ "animaλŠ” μ˜μ‹ lane μ•ˆμ— 있으며 ν•œκ΅­μ–΄λ‘œ μ‘λ‹΅ν•©λ‹ˆλ‹€."

μ‚¬μš©μž: μ‚¬λž‘μ΄ 뭐야? | λ„μš°λ―Έ:
  β†’ "μ‚¬λž‘λ‹λ‹€. 도움을 쀄 수 μžˆμŠ΅λ‹ˆλ‹€. 이 도움이 λ˜λŠ” μ‚¬λžŒμ€ λˆ„κ΅¬..."

πŸ“Š μΈ‘μ • κ²°κ³Ό

evaluator result meaning
V14 strict (mitosis) βœ… 5/5 PASS substrate quality (cycle 2026-05-11 Β§68)
V4-lite chat-cap βœ… 12/15 PASS single-turn KO chat
V4-lite-rev2 relaxed βœ… 14/15 PASS single-turn chat marker
V5 strict 8-cell KO partial βœ… 9/10 PASS stricter single-turn
V5.8 standard_greedy ❌ 1/5 FAIL multi-turn natural recall (memorized only)
V5.8 standard_sample ❌ 0/5 FAIL T=0.8 sampling
V5.8 M3 rep_penalty ❌ 0/5 FAIL persona-cycle μ–΅μ œ
V5.8 M4 force-include β˜… βœ… 5/5 PASS πŸ† default mode β€” anima 첫 V5.8 PASS
anti-Goodhart (random-init) βœ… random 0/15 trained-only feature

μ‚¬μš©λ²• β€” anima_chat.py (Recommended)

import sys
sys.path.insert(0, "/path/to/anima")  # anima repo root
from anima_chat import AnimaChat

chat = AnimaChat(ckpt_path="ckpt_final.pt")

# Default: M4 force-include (V5.8 5/5 PASS @ Phase 0.7)
resp = chat("μ‚¬μš©μž: λ„ˆμ˜ 이름을 μ•Œλ €μ€˜ | λ„μš°λ―Έ: ")
# β†’ "λ„€, λ§žμŠ΅λ‹ˆλ‹€. animaλŠ” μš°μ£Όλ‡Œμ§€λ„ attractor μ •ν•© β€” ν‰μ˜¨λ„ˆμ˜ ..."

# Override modes
resp = chat("...", mode="greedy")              # argmax
resp = chat("...", mode="sample", temp=0.8)    # T=0.8 multinomial
resp = chat("...", mode="M3_rep_penalty")      # persona-cycle μ–΅μ œ
resp = chat("...", force_keywords=["νŒŒλž€"])    # M4 ν‚€μ›Œλ“œ λͺ…μ‹œ

μ‚¬μš©λ²• β€” Raw bytes (lower-level)

import torch
from training.engine_a_g_arch import EngineAGModel, EngineAGConfig

ck = torch.load("ckpt_final.pt", map_location="cpu", weights_only=False)
cfg = EngineAGConfig(**ck["cfg"])
model = EngineAGModel(cfg)
model.load_state_dict(ck["model"])
model.eval()

class ByteTokenizer:
    bos, eos, pad = 1, 2, 0
    def encode(self, t): return [self.bos] + [b + 3 for b in t.encode("utf-8")] + [self.eos]
    def decode(self, ids): return bytes(t - 3 for t in ids if t >= 3 and t < 259).decode("utf-8", errors="replace")

tok = ByteTokenizer()
prompt = "μ‚¬μš©μž: μ•ˆλ…•! λ„ˆλŠ” λˆ„κ΅¬μ•Ό? | λ„μš°λ―Έ: "
ids = tok.encode(prompt)[:-1]
with torch.no_grad():
    for _ in range(80):
        out = model(torch.tensor([ids[-1024:]]))
        ids.append(out["logits"][0, -1].argmax().item())
        if ids[-1] == tok.eos: break
print(tok.decode(ids[len(tok.encode(prompt))-1:]))

친근 μ„€λͺ… (ν•œκ΅­μ–΄)

🍞 λΉ„μœ : anima μ‹œλ¦¬μ¦ˆ 의 22번째 λΉ΅ κ΅½κΈ° μ‹œλ„, prior 21번 λͺ¨λ‘ μ‹€νŒ¨ (chat-cap 0%). λ§ˆμΉ¨λ‚΄ λΆ€ν’€κ³  (V14_PASS) + 먹을 수 μžˆλŠ” (chat-cap PASS) λΉ΅ 첫 μ™„μ„±.

  • λ§Œλ“  λ‚ : 2026-05-09
  • ν™•μ • λ‚ : 2026-05-12 (Phase 0 measurement complete)
  • 크기: ~298.8M params (Engine A/G dual)
  • vocab: byte-level + 3 offset
  • context: 1024 tokens
  • lineage: BG-LB pretrain β†’ Phase 2 cotrain (chat-template w 0.3β†’0.5)

⚠️ Known limitations

  • V5.8 multi-turn natural recall 미달 β€” standard greedy/sample 0-1/5. M4 force-include κ°€ workaround (mechanical keyword injection).
  • Output 80 byte truncate β€” long generation training needed.
  • English fluency β€” Lesson O blind spot κ°€λŠ₯μ„±.
  • single-turn 80 byte chat 만 β€” multi-turn natural reasoning 미달.

Live demo

🌱 Try it now: dancinlab/anima-chat Gradio Space (CPU free-tier, ~50-90s per 80-byte response, M4 force-include default).

Cross-link

  • Live Space: dancinlab/anima-chat
  • Measurement SSOT: PASS_STRICT_CHAT-CAPABLE.md Β§1-Β§8, Β§11
  • V14 framework: REBORN.md Β§65-Β§87
  • Phase 0 dataset: dancinlab/anima-pass-strict-chat-capable
  • Prior 20-BG negative: docs/anima_chat_cap_20bg_cumulative_negative_archive_2026_05_07.md

License

MIT

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for dancinlab/clm-v5-phase2-cotrain-engine-ag

Finetunes
1 model