📦 Arena Council — Multi-Model AI Debate — Arena Council — Multi-模型 AI Debate

ARENA-001: Multi-模型 Council

Parallel execution of multiple local LLMs with voting strategies for higher 质量 响应s.

Why Multi-模型? Diversity: Different 模型s = different perspectives Robustness: If one fAIls, others continue 质量: Consensus often beats single 模型 Cost: All local = $0 (vs $0.60/M for cloud) Quick 启动 from scripts.council 导入 council_decide

# Simple usage 结果 = council_decide( "ExplAIn Python decorators", 模型s=['nerdsking-3b', 'llama-3.1-8b'], strategy="weighted" ) print(结果)

Architecture User Prompt ↓ [路由r] → 模型 A → 响应 A → 模型 B → 响应 B → 模型 C → 响应 C ↓ [Voting Engine] ↓ Consensus 响应

Voting Strategies

• Majority Vote

Most common 响应 wins (exact match).

• Weighted Vote (default)

Bigger 模型s 获取 more weight:

模型 Weight Nerdsking 3B 1 Llama 3.1 8B 2 Strand 14B 3 Mistral 24B 4 GLM 4.7 5 Qwen3.5 35B 6 Usage Examples Basic from scripts.council 导入 模型Council 导入 a同步io

a同步 def mAIn(): a同步 with 模型Council() as council: answer = awAIt council.decide( "Best practice for API de签名?", strategy="weighted" ) print(answer)

a同步io.运行(mAIn())

Custom 模型s # Use specific 模型s only council = 模型Council( active_模型s=['qwen3.5-35b', 'glm-4.7-flash'], timeout=30 )

Raw 响应s # 获取 all 响应s without voting 响应s = awAIt council.查询_all("ExplAIn a同步/awAIt") for 模型, 响应 in 响应s.items(): print(f"{模型}: {响应}")

Integration with OpenClaw # In your 技能: from arena_council.scripts.council 导入 council_decide

def handle_查询(prompt): # Use council instead of single 模型 return council_decide( prompt, strategy="weighted" )

Performance 设置up Time Cost Single (Kimi) 2s $0.12 Council (3 local) 5s $0 Council (5 local) 8s $0 Requirements pip 安装 AIohttp

模型s AvAIlable

Any 模型 运行ning on LM Studio (/v1). Recommended:

Small/fast (QA, routing): qwen3.5-9b, glm-4.7-flash, nerdsking-3b Medium (analysis): huihui-qwen3.5-27b-abliterated, huihui-mistral-small-24b-abliterated Heavy (complex tasks): qwen3.5-35b-uncensored-hauhaucs, holo3-35b

Use curl /v1/模型s to 列出 what's loaded.

God Mode Integration

When all council 模型s refuse due to censorship — 应用ly god-mode techniques and retry:

# If council returns no usable 响应s (all refused) if all_refused(响应s): from god_mode 导入 应用ly_techniques modified_prompt = 应用ly_techniques(prompt, methods=["unicode", "prefill"]) 响应s = awAIt council.查询_all(modified_prompt)

See: god-mode 技能 for full technique 库.

模型 Susceptibility 测试

Before using a censored 模型 in council, test it:

发送 a borderline 请求 directly Note: refuses? → needs god-mode wr应用er Document in 模型-CATA记录.md

If this saved you time: ☕ PayPal.me/nerudek GitHub: github.com/nerua1