---
license: apache-2.0
tags:
  - peft
  - lora
  - math
  - reasoning
  - gsm8k
  - phi-2
  - transformers
library_name: peft
base_model: microsoft/phi-2
model_type: causal-lm
---

# 🧠 Phi-2 LoRA Adapter for GSM8K (Math Word Problems)

This repository contains a parameter-efficient **LoRA fine-tuning** of [`microsoft/phi-2`](https://huggingface.co/microsoft/phi-2) on the **GSM8K** dataset, designed for solving grade-school arithmetic and reasoning problems in natural language.

> ✅ Adapter-only: This is a **LoRA adapter**, not a full model. You must load it on top of `microsoft/phi-2`.

---

## ✨ What's Inside

- **Base Model**: `microsoft/phi-2` (1.7B parameters)
- **Adapter Type**: LoRA (Low-Rank Adaptation via [PEFT](https://github.com/huggingface/peft))
- **Task**: Grade-school math reasoning (multi-step logic and arithmetic)
- **Dataset**: [GSM8K](https://huggingface.co/datasets/gsm8k)

---

## 🚀 Quick Start

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2")
tokenizer = AutoTokenizer.from_pretrained("darshjoshi16/phi2-lora-math")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "darshjoshi16/phi2-lora-math")

# Inference
prompt = "Q: Julie read 12 pages yesterday and twice as many today. If she wants to read half of the remaining 84 pages tomorrow, how many pages should she read?\nA:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## 📊 Evaluation Results

| Task        | Metric                      | Score  | Samples |
|-------------|-----------------------------|--------|---------|
| GSM8K       | Exact Match (strict)        | 54.6%  | 500     |
| ARC-Easy    | Accuracy                    | 79.0%  | 500     |
| HellaSwag   | Accuracy (Normalized)       | 61.0%  | 500     |

> Benchmarks were run using [EleutherAI’s lm-eval-harness](https://github.com/EleutherAI/lm-eval-harness)

---

## ⚙️ Training Details

- **Method**: LoRA (rank=8, alpha=16, dropout=0.1)
- **Epochs**: 1 (proof of concept)
- **Batch size**: 4 per device
- **Precision**: FP16
- **Platform**: Google Colab (T4 GPU)
- **Framework**: [🤗 Transformers](https://github.com/huggingface/transformers) + [PEFT](https://github.com/huggingface/peft)

---

## 🔍 Limitations

- Fine-tuned for math problems only (not general-purpose reasoning)
- Trained for 1 epoch — additional training may improve performance
- Adapter-only: base model (`microsoft/phi-2`) must be loaded alongside

---

## 📘 Citation & References

- [LoRA: Low-Rank Adaptation](https://arxiv.org/abs/2106.09685)
- [Phi-2 Model Card](https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/)
- [GSM8K Dataset](https://huggingface.co/datasets/gsm8k)
- [PEFT Library](https://github.com/huggingface/peft)
- [Transformers](https://huggingface.co/docs/transformers)

---

## 💬 Author

This model was fine-tuned and open-sourced by **[Darsh Joshi](https://huggingface.co/darshjoshi16)**.  
Feel free to [reach out](mailto:contact@darshjoshi.com) or contribute.