---
language:
  - en
license: mit
tags:
  - phi-3
  - distillation
  - knowledge-distillation
  - lora
  - code-generation
  - python
datasets:
  - Shuu12121/python-codesearch-dataset-open
model-index:
  - name: FStudent
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          type: custom
          name: Distillation Evaluation
        metrics:
          - name: Speedup Factor
            type: speedup
            value: 2.5x
            verified: false
---

# FStudent: Distilled Phi-3 Model

FStudent is a knowledge-distilled version of Microsoft's Phi-3-mini-4k-instruct model, trained through a comprehensive distillation pipeline that combines teacher-student learning with self-study mechanisms.

## Model Description

FStudent was created using a multi-stage distillation pipeline that transfers knowledge from a larger teacher model (Phi-4) to the smaller Phi-3-mini-4k-instruct model. The model was trained using LoRA adapters, which were then merged with the base model to create this standalone version.

### Training Data

The model was trained on a diverse set of data sources:

1. **PDF Documents**: Technical documentation and domain-specific knowledge
2. **Python Code Dataset**: Code examples from the [Shuu12121/python-codesearch-dataset-open](https://huggingface.co/datasets/Shuu12121/python-codesearch-dataset-open) dataset
3. **Teacher-Generated Examples**: High-quality examples generated by the Phi-4 teacher model

### Training Process

The distillation pipeline consisted of six sequential steps:

1. **Content Extraction & Enrichment**: PDF files were processed to extract and enrich text data
2. **Teacher Pair Generation**: Training pairs were generated using the Phi-4 teacher model
3. **Distillation Training**: The student model (Phi-3) was trained using LoRA adapters with the following parameters:
   - Learning rate: 1e-4
   - Batch size: 4
   - Gradient accumulation steps: 8
   - Mixed precision training
   - 4-bit quantization during training
4. **Model Merging**: The trained LoRA adapters were merged with the base Phi-3 model
5. **Student Self-Study**: The model performed self-directed learning on domain-specific content
6. **Model Evaluation**: The model was evaluated against the teacher model for performance

### Model Architecture

- **Base Model**: microsoft/Phi-3-mini-4k-instruct
- **Parameter-Efficient Fine-Tuning**: LoRA adapters (merged into this model)
- **Context Length**: 4K tokens
- **Architecture**: Transformer-based language model

## Intended Uses

This model is designed for:

- General text generation tasks
- Python code understanding and generation
- Technical documentation analysis
- Question answering on domain-specific topics

## Performance and Limitations

### Strengths

- Faster inference compared to larger models (approximately 2.5x speedup)
- Maintains much of the capability of the teacher model
- Enhanced code understanding due to training on Python code datasets
- Good performance on technical documentation analysis

### Limitations

- May not match the full capabilities of larger models on complex reasoning tasks
- Limited context window compared to some larger models
- Performance on specialized domains not covered in training data may be reduced

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("forge1825/FStudent")
tokenizer = AutoTokenizer.from_pretrained("forge1825/FStudent")

# Generate text
input_text = "Write a Python function to calculate the Fibonacci sequence:"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Quantized Usage

For more efficient inference, you can load the model with quantization:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# 4-bit quantization configuration
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)

# Load the model with quantization
model = AutoModelForCausalLM.from_pretrained(
    "forge1825/FStudent",
    device_map="auto",
    quantization_config=quantization_config
)
tokenizer = AutoTokenizer.from_pretrained("forge1825/FStudent")
```

## Training Details

- **Training Framework**: Hugging Face Transformers with PEFT
- **Optimizer**: AdamW
- **Learning Rate Schedule**: Linear warmup followed by linear decay
- **Training Hardware**: NVIDIA GPUs
- **Distillation Method**: Knowledge distillation with teacher-student architecture
- **Self-Study Mechanism**: Curiosity-driven exploration with hierarchical context

## Ethical Considerations

This model inherits the capabilities and limitations of its base model (Phi-3-mini-4k-instruct). While efforts have been made to ensure responsible behavior, the model may still:

- Generate incorrect or misleading information
- Produce biased content reflecting biases in the training data
- Create code that contains bugs or security vulnerabilities

Users should validate and review the model's outputs, especially for sensitive applications.

## Citation and Attribution

If you use this model in your research or applications, please cite:

```
@misc{forge1825_fstudent,
  author = {Forge1825},
  title = {FStudent: Distilled Phi-3 Model},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/forge1825/FStudent}}
}
```

## Acknowledgements

- Microsoft for the Phi-3-mini-4k-instruct base model
- Hugging Face for the infrastructure and tools
- The creators of the Python code dataset used in training