---
library_name: transformers
tags:
- llmcompressor
- quantization
- wint4
---

# Phi-4-mini-instruct-WINT4

This model is a 4-bit quantized version of [microsoft/Phi-4-mini-instruct](https://huggingface.co/microsoft/Phi-4-mini-instruct) "using the [llmcompressor](https://github.com/neuralmagic/llmcompressor) library.

## Quantization Details

*   **Base Model:** [microsoft/Phi-4-mini-instruct](https://huggingface.co/microsoft/Phi-4-mini-instruct)
*   **Quantization Library:** `llmcompressor`
*   **Quantization Method:** Weight-only 4-bit int (WINT4)
*   **Quantization Recipe:**
```yaml
quant_stage:
      quant_modifiers:
        QuantizationModifier:
          ignore: [lm_head]
          config_groups:
            group_0:
              weights: {num_bits: 4, type: int, symmetric: true, strategy: channel, dynamic: false}
              targets: [Linear]
```

## Evaluation Results

The following table shows the evaluation results on various benchmarks compared to the baseline (non-quantized) model.

| Task             | Baseline Metric (10.0% Threshold) | Quantized Metric | Metric Type         |
|------------------|-------------------------------------------------------|------------------|---------------------|
| winogrande       | 0.7545                                              | 0.6985           | acc,none            |

## How to Use

You can load the quantized model and tokenizer using the `transformers` library:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NoorNizar/Phi-4-mini-instruct-WINT4"

model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Example usage (replace with your specific task)
prompt = "Hello, world!"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Disclaimer

This model was quantized automatically using a script. Performance and behavior might differ slightly from the original base model.