MagistrTheOne/SHUTEN-DOJI API & Inference Endpoint

Roadmap status

Table
Stage	Item	Status
Infra	H200 train → vLLM → LoRA serve	✅ Done
Data v1	bootstrap trajectories	✅ Done (legacy — do not use for SFT)
SFT v1	smoke / warm-start	✅ Done — planner failed (Step N poison)
SFT v2	50 Constitution gold examples	✅ Done
Eval A–H	side-by-side vs base Qwen	✅ Done — v2 0 poison, 7/8 struct wins
SFT v2.1	110–130 gold + reviewed LLM + failure cases	⬜ Next
Eval v2	20–30 held-out planner cases	⬜ Next
DPO	preference pairs (chosen vs rejected plans)	⬜ After eval v2 pass
IMPACT cluster	consequence prediction fine-tune	⬜ Phase 2
NATIVE MoE	custom NULLXES architecture	⬜ Phase 3

Base model

Table

Foundation	Qwen/Qwen3.6-27B
Method	QLoRA SFT (rank 64, alpha 128) → merged into full weights
Train data	50 Constitution ShareGPT examples (no bootstrap poison)
Checkpoint	`shuten-sft-h200-v2` on RunPod H200

Training metrics (SFT v2)

Table
Metric	Value
Train examples	50
Eval examples	10
Epochs	3
Train loss	1.03
Eval loss	0.49
Trainable params	41.9M LoRA (merged at export)

Eval metrics (A–H, side-by-side)

Table
Metric	qwen_base	SHUTEN v2
Poison (`Step N:`, `tool_use`, …)	0/8	0/8
Avg structure markers	1.5	4.12
Wins vs base (structure)	—	7/8
Avg output length	5139	4684

Verdict: Constitution SFT removes bootstrap action-trace failure. Content quality still MVP — v2.1 dataset iteration required before DPO.

Usage (vLLM — no LoRA adapter needed)

bash
python -m vllm.entrypoints.openai.api_server \
  --model NULLXES/SHUTEN-DOJI \
  --max-model-len 8192 \
  --dtype bfloat16 \
  --trust-remote-code \
  --language-model-only \
  --gdn-prefill-backend triton

bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "NULLXES/SHUTEN-DOJI",
    "messages": [
      {"role": "system", "content": "You are SHUTEN, strategic intelligence by NULLXES DAI. Reason: State → Causes → Options → Impact → Future State → Confidence."},
      {"role": "user", "content": "[SHUTEN business]\n\nWorld State:\nRevenue down 18%. Backlog up 42%.\n\nObjective:\nRestore EBITDA margin >12% in 90d.\n\nRequired Output:\nState → Causes → Options → Impact → Future State → Confidence"}
    ],
    "max_tokens": 1200,
    "temperature": 0.3
  }'

Limitations

MVP release — 50 training examples only
May still prefix with Qwen-style reasoning traces
Not trained for DPO / impact cluster yet
Requires ~54GB VRAM at bf16 (single H200 / A100 80GB)

Citation

bibtex
@misc{nullxes-shuten-doji-v2,
  title={NULLXES SHUTEN-DŌJI: Strategic Intelligence (Constitution SFT v2)},
  author={NULLXES DAI},
  year={2026},
  note={Merged Qwen3.6-27B + Constitution LoRA. MVP eval 7/8 struct wins, 0 poison.}
}

SHUTEN-DOJI

Get help setting up a custom Dedicated Endpoints.

README

Roadmap status

Base model

Training metrics (SFT v2)

Eval metrics (A–H, side-by-side)

Usage (vLLM — no LoRA adapter needed)

Limitations

Links

Citation

Explore FriendliAI today