Reasoning Datasets¶

Reasoning datasets include chain-of-thought traces that show how the model arrives at its answer.

When to Use¶

Training models to explain their thinking
Math, logic, or multi-step problems
Tasks requiring transparent decision-making
Improving model reliability through explicit reasoning

Configuration¶

config.yaml

topics:
  prompt: "Mathematical problem solving"
  mode: tree
  depth: 2
  degree: 3

generation:
  system_prompt: "Generate problems with step-by-step solutions."
  instructions: "Show clear reasoning before the final answer."

  conversation:
    type: cot
    reasoning_style: freetext

  llm:
    provider: "openai"
    model: "gpt-4o"

output:
  system_prompt: |
    You are an AI that explains its reasoning step-by-step.
  num_samples: 7
  batch_size: 1
  save_as: "reasoning-dataset.jsonl"

Key Settings

conversation.type: cot
conversation.reasoning_style: freetext

Output Format¶

Reasoning datasets include a reasoning field alongside messages:

Sample Output

reasoning-dataset.jsonl

{
    "messages": [
        {
            "content": "You are an AI that explains its reasoning step-by-step.\n",
            "role": "system"
        },
        {
            "content": "Can you create a reasoning problem that involves mathematical problem-solving and logical reasoning?",
            "role": "user"
        },
        {
            "content": "Sure, let's dive into a problem that requires a bit of logical reasoning and mathematical problem-solving.\n\n**Problem:**\n\nImagine you have a basket containing a mix of apples and oranges. The total number of fruits in the basket is 30. If the number of apples is twice the number of oranges, how many apples and oranges are there in the basket?\n\n**Reasoning:**\n\nHmm, let's think about this. We know the total number of fruits is 30...",
            "role": "assistant"
        }
    ],
    "reasoning": {
        "content": "Hmm, let's think about this. We know the total number of fruits is 30. So, if we let the number of oranges be x, then the number of apples would be 2x...",
        "style": "freetext"
    },
    "question": "Can you create a reasoning problem that involves mathematical problem-solving and logical reasoning?",
    "final_answer": "There are 20 apples and 10 oranges in the basket."
}

Reasoning Styles¶

DeepFabric supports one reasoning style for chain-of-thought:

Style	Description
`freetext`	Natural language explanation of thought process

The freetext style produces readable, conversational reasoning traces.

CLI Usage¶

Generate reasoning dataset

deepfabric generate \
  --topic-prompt "Logic puzzles" \
  --conversation-type cot \
  --reasoning-style freetext \
  --num-samples 1 \
  --batch-size 1 \
  --provider openai \
  --model gpt-4o \
  --output-save-as logic-dataset.jsonl

Training Considerations¶

Reasoning Placement

The reasoning field is stored separately from messages. During training, you can:

Include reasoning in the assistant's response (visible to users)
Use it as auxiliary training signal only
Format it as a special token section (e.g., <thinking>...</thinking>)