metadata
tags:
- phi-2
- conversational-ai
- fine-tuned
license: mit
datasets:
- UltraChat
- Clocal_data
base_model: microsoft/phi-2
phi-2-chat
A fine-tuned conversational variant of Microsoft's Phi-2 (2.7B) optimized for dialogue tasks
Model Details
Base Model
microsoft/phi-2 (2.7B parameters, MIT License)
Training Data
- UltraChat (CC-BY-NC-4.0):
@misc{ultrachat,
title={UltraChat: A Large-Scale Auto-generated Multi-round Dialogue Dataset},
author={Ding et al.},
year={2023},
howpublished={\url{https://github.com/thunlp/UltraChat}}
}
- Custom synthetic data(Proprietary)
Fine-Tuning
- Objective: Instruction-following & conversational ability
- Framework: PyTorch + Transformers
- Context Window: 2048 tokens
Usage
Quick Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"Irfanuruchi/phi-2-chat",
trust_remote_code=True,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Irfanuruchi/phi-2-chat")
# Recommended prompt format:
input_text = "<|user|>Explain dark matter<|assistant|>"
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Limitations
- ** Licence Restrictions: License Restrictions: Non-commercial use applies to UltraChat-derived components (CC-BY-NC-4.0)
- ** Bias: May reflect biases in base model and training data
License
- Core Model: MIT (inherited from Phi-2)
- UltraChat Components: CC-BY-NC-4.0 (non-commercial clause applies)
- Custom Data: Proprietary
Citation
@misc{phi-2-chat,
author = {Irfan Uruchi},
title = {phi-2-chat: Fine-tuned Phi-2 for conversational AI},
year = {2025},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/Irfanuruchi/phi-2-chat}}
}
@misc{phi2,
title={Phi-2: The Surprisingly Capable Small Language Model},
author={Microsoft},
year={2023},
url={https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/}
}
Contact
For questions or issues, please open a discussion on the Hugging Face Hub.
Or you can do the same also in GitHub: