DeepSeek R1 Optimal Configuration Guide

# DeepSeek R1 Optimal Configuration

## Official Recommendations (from DeepSeek)
- **No system prompt**: Leave the system prompt empty for R1 models. All instructions go in the user message.
- **Temperature**: Set to 0.6 (recommended range: 0.5-0.7). Prevents endless repetitions or incoherent outputs.
- **Top-p**: Use 0.95
- **Thinking enforcement**: Prefix model output with "<think>\n" to ensure thorough reasoning

## What NOT to Do
- Do NOT use few-shot examples — this consistently degrades R1 performance
- Do NOT add chain-of-thought instructions — R1 already reasons internally
- Do NOT use complex, lengthy system prompts — keep instructions minimal and explicit
- Do NOT set temperature below 0.5 or above 0.7

## For Math Problems
Add this directive: "Reason step by step, and put your final answer within \boxed{}"

## For Validation
Use majority voting: generate multiple responses and select the most frequent answer.

## Key Differences from V3
- R1: No system prompt, user-prompt-only instructions, built-in reasoning
- V3: Supports system prompts, few-shot examples, function calling, JSON mode