127 lines
5.7 KiB
Markdown
127 lines
5.7 KiB
Markdown
---
|
||
license: mit
|
||
language:
|
||
- en
|
||
base_model:
|
||
- microsoft/phi-4
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
tags:
|
||
- text-generation-inference
|
||
- llama
|
||
- phi3
|
||
- phi
|
||
---
|
||
|
||
|
||
# **Phi-4-QwQ [ Responsible Problem Solving & Advanced Reasoning ]**
|
||
|
||
`[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:
|
||
|
||
1. **Responsible Problem Solving**
|
||
- Breaking down complex problems into logical, actionable steps.
|
||
- Offering ethical, well-rounded solutions in academic and professional contexts.
|
||
|
||
2. **Advanced Reasoning Tasks**
|
||
- Excelling in mathematics, logic, and scientific reasoning.
|
||
- Providing detailed explanations and systematic answers.
|
||
|
||
3. **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.
|
||
|
||
4. **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.
|
||
|
||
5. **Customer Support & Dialogue Systems**
|
||
- Enabling chatbots and virtual assistants to provide accurate, helpful, and responsible responses.
|
||
- Enhancing customer service with reasoning-driven automation.
|
||
|
||
6. **Multilingual Capabilities**
|
||
- Supporting multilingual communication and content generation while maintaining contextual accuracy.
|
||
- Assisting in translations with a focus on retaining meaning and nuance.
|
||
|
||
7. **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:
|
||
|
||
1. **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.
|
||
|
||
2. **Contextual Interpretation**
|
||
- The model may occasionally misinterpret highly nuanced prompts or ambiguous contexts, leading to suboptimal responses.
|
||
|
||
3. **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.
|
||
|
||
4. **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.
|
||
|
||
5. **Computational Requirements**
|
||
- Deploying Phi-4-QwQ efficiently may require substantial computational resources, particularly for large-scale deployments or real-time applications.
|
||
|
||
6. **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.
|
||
|
||
7. **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. |