[Phi-4-QwQ finetuned] from Microsoft's Phi-4 is a state-of-the-art open model developed with a focus on responsible problem solving and advanced reasoning capabilities. Built upon a diverse blend of synthetic datasets, carefully filtered public domain websites, and high-quality academic books and Q&A datasets, Phi-4-QwQ ensures that small, capable models are trained with datasets of exceptional depth and precision.

Phi-4-QwQ adopts a robust safety post-training approach using open-source and in-house synthetic datasets. This involves a combination of SFT (Supervised Fine-Tuning) and iterative DPO (Direct Preference Optimization) techniques, ensuring helpful and harmless outputs across various safety categories.

Dataset Info

Phi-4-QwQ is fine-tuned on a carefully curated synthetic dataset generated using an advanced pipeline optimized for Chain of Thought (CoT) reasoning and Responsible Problem Breakdown (RPB) methodologies. This ensures that the model excels at:

Logical reasoning
Step-by-step problem-solving
Breaking down complex tasks into manageable parts

The dataset also emphasizes responsible decision-making and fairness in generating solutions.

Run with Transformers

python

# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("prithivMLmods/Phi-4-QwQ")
model = AutoModelForCausalLM.from_pretrained(
    "prithivMLmods/Phi-4-QwQ",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

input_text = "Explain the concept of black holes."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids, max_new_tokens=64)
print(tokenizer.decode(outputs[0]))

For chat-style interactions, use tokenizer.apply_chat_template:

python

messages = [
    {"role": "user", "content": "Explain the concept of black holes."},
]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True).to("cuda")

outputs = model.generate(**input_ids, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Intended Use

Phi-4-QwQ is tailored for a wide range of applications, especially those involving advanced reasoning, multilingual capabilities, and responsible problem-solving. Its primary use cases include:

Responsible Problem Solving
- Breaking down complex problems into logical, actionable steps.
- Offering ethical, well-rounded solutions in academic and professional contexts.
Advanced Reasoning Tasks
- Excelling in mathematics, logic, and scientific reasoning.
- Providing detailed explanations and systematic answers.
Content Generation
- Assisting in generating high-quality content for various domains, including creative writing and technical documentation.
- Supporting marketers, writers, and educators with detailed and well-structured outputs.
Educational Support
- Acting as a virtual tutor for students by generating practice questions, answers, and detailed explanations.
- Helping educators design learning material that promotes critical thinking and step-by-step problem-solving.
Customer Support & Dialogue Systems
- Enabling chatbots and virtual assistants to provide accurate, helpful, and responsible responses.
- Enhancing customer service with reasoning-driven automation.
Multilingual Capabilities
- Supporting multilingual communication and content generation while maintaining contextual accuracy.
- Assisting in translations with a focus on retaining meaning and nuance.
Safety-Critical Applications
- Ensuring safe and harmless outputs, making it suitable for sensitive domains.
- Providing aligned interactions with human oversight for critical systems.

Limitations

Despite its strengths, Phi-4-QwQ has some limitations that users should be aware of:

Bias and Fairness
- While great effort has been made to minimize biases, users should critically assess the model’s output in sensitive scenarios to avoid unintended bias.
Contextual Interpretation
- The model may occasionally misinterpret highly nuanced prompts or ambiguous contexts, leading to suboptimal responses.
Knowledge Cutoff
- Phi-4-QwQ’s knowledge is static and based on the data available at the time of training. It does not include real-time updates or information on recent developments.
Safety and Harmlessness
- Despite post-training safety alignment, inappropriate or harmful outputs may still occur. Continuous monitoring and human oversight are advised when using the model in critical contexts.
Computational Requirements
- Deploying Phi-4-QwQ efficiently may require substantial computational resources, particularly for large-scale deployments or real-time applications.
Ethical Considerations
- Users are responsible for ensuring that the model is not employed for malicious purposes, such as spreading misinformation, generating harmful content, or facilitating unethical behavior.
Domain-Specific Expertise
- While the model is versatile, it may not perform optimally in highly specialized domains (e.g., law, medicine, finance) without further domain-specific fine-tuning.