Orca 2 7B
Explanation Tuning for Reasoning

Cautious System Messages

During training, Microsoft Research used "Cautious System Messages" that instructed the teacher model (GPT-4) to use specific reasoning strategies for specific types of problems. The student model (Orca 2) then learned to internalize when each strategy is appropriate — without needing the system message at inference time.

Source: "Orca 2: Teaching Small Language Models How to Reason" — Mitra et al., November 2023 (arXiv:2311.11045)

Python API Integration

import requests
import json

def query_orca2(prompt: str, system: str = "") -> str:
    """Query Orca 2 via Ollama API."""
    response = requests.post(
        "http://localhost:11434/api/generate",
        json={
            "model": "orca2",
            "prompt": prompt,
            "system": system,
            "stream": False,
            "options": {
                "temperature": 0.7,
                "num_ctx": 4096
            }
        }
    )
    return response.json()["response"]

# Example: Reasoning task
answer = query_orca2(
    "A train travels 120 km in 2 hours. "
    "It then travels 90 km in 1.5 hours. "
    "What is the average speed for the entire journey?"
)
print(answer)

# Example: With system prompt for step-by-step reasoning
answer = query_orca2(
    "If a shirt costs $25 after a 20% discount, what was the original price?",
    system="Think step by step before giving the final answer."
)
print(answer)

Better Alternatives in 2026

For most use cases in 2026, Qwen 2.5 7B (ollama pull qwen2.5:7b) is the recommended replacement — it scores ~16 MMLU points higher, has 128K context, and uses Apache 2.0 license.