KavinduHansaka/phi4-mini-bnb-4bit

Phi-4-mini-reasoning (BitsAndBytes 4-bit NF4 Quantized)

This repository contains a 4-bit quantized version of microsoft/Phi-4-mini-reasoning, produced with BitsAndBytes via Hugging Face Transformers.
Quantization reduces VRAM usage while preserving most of the model’s reasoning capabilities.

---

Model Details

Model Description

• Developed by (base model): Microsoft
• Shared by (quantized version): KavinduHansaka
• Model type: Causal Language Model (decoder-only transformer)
• Context length: 128K
• Language(s): English
• License: MIT (inherited from base model)
• Finetuned from: microsoft/Phi-4-mini-reasoning

Model Sources

• Repository (quantized): KavinduHansaka/phi4-mini-bnb-4bit
• Repository (base model): microsoft/Phi-4-mini-reasoning

---

Uses

Direct Use

• Text and reasoning generation
• Educational and research experiments
• Running inference on lower-VRAM GPUs

Downstream Use

• Can be fine-tuned further for domain-specific reasoning tasks
• Integrated into chatbots, assistants, and research pipelines

Out-of-Scope Use

• Do not use for generating harmful, biased, or unsafe content
• Not recommended for high-accuracy production systems without further testing

---

Bias, Risks, and Limitations

• As with the base model, it may produce biased or incorrect content.
• Quantization may reduce numerical precision, which can slightly affect reasoning quality.
• Long-context reasoning (128k tokens) may still be resource-intensive.

Recommendations

• Apply appropriate safety filters before deploying in production.
• Be aware that outputs are not guaranteed to be factually correct.

---

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

model_id = "KavinduHansaka/phi4-mini-bnb-4bit"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.bfloat16,
    quantization_config=bnb_config
)

inputs = tokenizer("Explain why the sky is blue in simple terms.", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Training Details

This model inherits training data from microsoft/Phi-4-mini-reasoning. No additional fine-tuning was done.

• Quantization method: BitsAndBytes 4-bit (NF4, double quantization)
• Precision: bfloat16 compute
• Original precision: fp16

---

Technical Specifications

• Architecture: Decoder-only transformer
• Parameters: Same as Phi-4-mini-reasoning
• Quantization: 4-bit NF4

---

Citation

If you use this quantized model, please also cite the original Microsoft release:

@misc{microsoft2025phi4mini,
  title={Phi-4-mini-reasoning},
  author={Microsoft},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/microsoft/Phi-4-mini-reasoning}}
}

---

Model Card Authors

• Quantized version shared by KavinduHansaka
• Base model by Microsoft

---

Model Card Contact

• For issues/questions with this quantized release: open a discussion on KavinduHansaka/phi4-mini-bnb-4bit.
• For base model details: see microsoft/Phi-4-mini-reasoning.