This is meant to be used with JSON structured output with the following columns: reasoning and output.

Example output:

{
  "reasoning": "Reasoning string goes here.",
  "output": "Output string goes here."
}

If your choice of LLM inference backend supports JSON structured output, use it!

Shows excellent reasoning capabilities across a wide range of domains.

Use the following system prompt:

You are Starlette, a curious human being.

Example usage code (Python):

from ollama import AsyncClient
from pydantic import BaseModel
from typing import Dict, List

class ReasoningModel(BaseModel):
    reasoning: str
    output: str
    
class OllamaInference:
    def __init__(self, host: str, num_ctx: int, model: str):
        self.num_ctx = num_ctx
        self.model = model
        self.client = AsyncClient(host=host)
        
    async def chat(self, messages: List[Dict[str, str]]) -> ReasoningModel:
        response = await self.client.chat(
            messages=messages,
            model=self.model,
            keep_alive=-1,
            format=ReasoningModel.model_json_schema(),
            options={
                "num_ctx": self.num_ctx,
                "temperature": 0.2
            }
        )
        
        return ReasoningModel.model_validate_json(response.message.content)
    
if __name__ == "__main__":
    import asyncio
    service = OllamaInference(host="http://localhost:11434", num_ctx=32_000, model="starsnatched/Starlette-1") # My model supports up to 128k context
    messages = [
        {"content": "You are Starlette, a curious human being.", "role": "system"},
        {"content": "Explain what Rao-Blackwell theorem is.", "role": "user"},
    ]
    
    response = asyncio.run(service.chat(messages))
    print("Reasoning:", response.reasoning)
    print("Output:", response.output)