DeepSeek-R1-Interactive-Reasoning-finetune

DeepSeek-R1-Interactive-Reasoning (Reasoning + Clarification)

Model Summary

DeepSeek-R1-Interactive-Reasoning is a fine-tuned reasoning model based on DeepSeek-R1. Unlike standard reasoning models that only output final answers, this model explicitly integrates clarification tokens into its reasoning process:

• <ask> → requests missing context from the user.
• <question> → asks for specific factual input needed for reasoning.
• <doubt> → expresses uncertainty, highlighting assumptions or limitations.

This makes the model more interactive, transparent, and safe for ambiguous queries.

---

Intended Uses

• Interactive reasoning in education, research, planning, and assistance systems.
• Applications where ambiguity must be clarified before giving a final answer.
• Safer deployment in healthcare, legal, or customer support where incorrect assumptions can be harmful.

Not intended for:

• Replacing domain experts (medical, legal, financial advice).
• Unmoderated use where clarifications may be ignored.

---

Training Details

• Base Model: DeepSeek-R1
• Method: Supervised fine-tuning (LoRA/Full SFT)
• Dataset Size: ~5,000 curated examples
• Dataset Type: Reasoning traces with <ask>, <doubt>, <question> tags

---

Example Usage

first run ollama create modelname_of_your_choice -f ./Modelfile

import ollama

response = ollama.chat(
    model="<The Model Name you gave>",
    messages=[
        {"role": "user", "content": "I have fever and vomiting, what is happening to me?"}
    ]
)

print(response.message.content)

Output:

You are experiencing fever and vomiting, which could indicate many different conditions. 
<question>What is your temperature?</question>
(User responds: 102°F)
Okay, you have a high fever, and you are vomiting. These are classic symptoms of flu or other viral infections. 
<doubt>I am not a doctor, but the most common thing with fever and vomiting at 102°F is likely a viral infection.</doubt>
Final Answer: Most likely you have a viral infection such as the flu. However, consult a medical professional for an accurate diagnosis.

---

Another Example

User: "I want to plan an industrial visit with my engineering college."

Model Output:

Okay, to plan a successful industrial visit, I need some details. 
<question>What industry are you interested in? For example, IT, Automobile, Robotics, etc.</question>
(User responds: Robotics)
<ask>What is the budget for this visit? Including travel expenses, entrance fees and other costs?</ask>
(User responds: 20000 INR)
<ask>Which companies do you want to visit that are in the robotics space? Are they nearby or do you need to travel far?</ask>
(User responds: Boston Robotics Inc. in Bangalore)
Final Answer: Plan an industrial visit to Boston Robotics Inc, Bangalore with a budget of 20,000 INR, focusing on robotics.

---

Limitations

• May overuse clarification tokens even when unnecessary.
• Only trained on ~5k samples, so robustness on out-of-domain queries may be limited.
• Not a medical or legal advisor. Outputs should be used for informational purposes only.

---

Ethical Considerations

• Designed to reduce hallucination risk by prompting for clarification.
• Still requires human judgment before acting on outputs.
• Explicit <doubt> tokens highlight uncertainty — but users must not ignore them.

---

Future Work

• Expand dataset beyond 5k examples for stronger generalization.
• Collect preference data to fine-tune when clarifications are appropriate.
• Add multilingual clarification reasoning.

---

Citation

If you use this model, please cite:

@misc{deepseek-rc,
  title        = {DeepSeek-RC: Reasoning + Clarification Model},
  author       = {Your Name},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/RohitFrancis/DeepSeek-R1-Interactive-Reasoning-finetune/}}
}

---

Tags

reasoning clarification interactive ask-model cot

---

license: mit