When fine-tuning an LLM in Prem Studio, you’ll need to decide between reasoning and non-reasoning approaches. This choice significantly impacts your model’s behavior, training time, and performance on different types of tasks.

​
What’s the Difference?

Reasoning Fine-Tuning teaches models to show their thinking process step-by-step before arriving at an answer. The model learns to break down complex problems, consider multiple perspectives, and explain its reasoning. Non-Reasoning Fine-Tuning focuses on direct input-output mapping. The model learns to provide answers quickly without showing intermediate steps or explanations.

​
When to Use Each Approach

ScenarioReasoningNon-ReasoningExample Use Case
Complex problem-solving 🧮✅❌Mathematical word problems, multi-step analysis
Fast response times needed ⚡❌✅Chatbots, real-time translation, autocomplete
Transparency required 🔍✅❌Medical diagnosis support, legal research
Simple classification tasks 🏷️❌✅Sentiment analysis, content moderation
Educational applications 📚✅❌Tutoring systems, homework help
High-volume API calls 📈❌✅Content generation, summarization at scale
Debugging model decisions 🔧✅❌Understanding why a model made specific choices
Creative writing ✍️❌✅Story generation, marketing copy

​
When to Choose Non-Reasoning

1

Speed-Critical Applications

Choose non-reasoning when response time is crucial:
  • Real-time chat: Customer support bots, conversational AI
  • High-throughput processing: Batch content generation, data labeling
  • Interactive applications: Autocomplete, instant search suggestions
2

Simple, Direct Tasks

When the task has a clear input-output relationship:
  • Classification: Sentiment analysis, topic categorization
  • Format conversion: JSON to text, data transformation
  • Pattern matching: Named entity recognition, keyword extraction
Non-reasoning fine-tuning typically converges faster and requires less computational resources.
3

Creative or Stylistic Tasks

When the process matters less than the final output:
  • Creative writing and content generation
  • Style transfer and tone adjustment
  • Language translation where fluency matters more than showing steps

​
When to Choose Reasoning

1

Complex Multi-Step Tasks

Choose reasoning when your task requires breaking down problems into smaller steps:
  • Mathematical problems: “Solve this equation step by step”
  • Analysis tasks: “Analyze this business case and recommend actions”
  • Research questions: “Compare these theories and explain the differences”
Reasoning fine-tuning typically takes longer but produces more explainable and trustworthy outputs for complex tasks.
2

High-Stakes Decisions

When accuracy and explainability matter more than speed:
  • Medical or legal applications where decisions need justification
  • Financial analysis where reasoning must be auditable
  • Educational tools where learning the process is important
Use reasoning fine-tuning only when you can provide high-quality training data with step-by-step explanations.
3

Debugging and Interpretability

When you need to understand why a model made specific decisions:
  • Model behavior analysis
  • Identifying bias or errors in reasoning
  • Building trust with end users who need to understand outputs

​
Hybrid Approach: When You’re Unsure

If you’re uncertain which approach to use, consider these strategies:
  1. Start with Non-Reasoning for faster iteration and baseline performance
  2. Test with Reasoning if initial results lack the depth or accuracy you need
  3. Use both approaches for different parts of your application (reasoning for complex queries, non-reasoning for simple ones)

​
Implementation in Prem Studio

1

Select Your Fine-Tuning Type

When creating a fine-tuning job in Prem Studio:
  • Choose “Reasoning” for tasks requiring step-by-step thinking
  • Choose “Non-Reasoning” for direct input-output mapping
Choosing Fine-Tuning Type
2

After this it's all the same

For Reasoning: You have only two models to choose from. Those are:
  • Qwen 2.5 7B reasoning
  • Qwen 2.5 3B reasoning
Reasoning ModelsFor Non-Reasoning: You have a lot of models to choose from, ranging from Qwen, Gemma, Llama, Phi models.Reasoning Models

​
Output from reasoning vs non-reasoning models

The output from reasoning models is a bit different from the output from non-reasoning models. In reasoning models, it will first show it’s thought process under <think> </think> tag, and then it will show the final answer under <answer> </answer> tag. Here is an example: GIF showing reasoning model output in Prem Studio