File size: 2,940 Bytes
d8333a8 8da1f6a d8333a8 8da1f6a d8333a8 8da1f6a d8333a8 8da1f6a d8333a8 8da1f6a d8333a8 8da1f6a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 |
---
base_model: unsloth/phi-4-unsloth-bnb-4bit
tags:
- text-generation
- transformers
- unsloth
- llama
- trl
license: apache-2.0
language:
- en
---
# 🧮 Phi-4 Math Reasoning Model (LoRA Finetuned)
## 📌 Model Overview
This model is a **LoRA fine-tuned version** of [unsloth/phi-4-unsloth-bnb-4bit](https://huggingface.co/unsloth/phi-4-unsloth-bnb-4bit).
It has been fine-tuned specifically for **math reasoning tasks**, capable of solving step-by-step arithmetic, algebra, and logic problems.
The base model is **Phi-4**, a 14B-parameter LLaMA variant optimized with [Unsloth](https://github.com/unslothai/unsloth) for **2x faster training** using Hugging Face’s [TRL](https://huggingface.co/docs/trl) library.
This version uses **bnb-4bit quantization**, making it memory efficient and suitable for single-GPU setups such as **Tesla T4 (16GB)** or consumer GPUs.
---
## ⚡ Key Features
- 🧠 Fine-tuned for **math reasoning and step-by-step solutions**
- ⚡ Efficient: 4-bit quantized, runs on **a single GPU** or even CPU (slower)
- 🚀 Trained with **Unsloth + TRL** for fast and memory-efficient fine-tuning
- 📚 Based on **Phi-4 (14B LLaMA model)**
---
## 📥 Installation
Ensure you have the latest versions of the required libraries:
```bash
pip install unsloth transformers accelerate bitsandbytes
```
🖥️ Usage (Colab / Local GPU)
```bash
import torch
from unsloth import FastLanguageModel
from transformers import TextStreamer
# Load the LoRA fine-tuned model
model_name = "RobinMillford/phi-4-math-reasoning-lora"
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=model_name,
max_seq_length=2048,
dtype=torch.float16, # fp16 recommended for GPU
load_in_4bit=True, # load in 4-bit quantized mode
device_map="auto" # automatically place layers on GPU/CPU
)
# Prepare for inference
FastLanguageModel.for_inference(model)
# Example: Generate a step-by-step solution
streamer = TextStreamer(tokenizer)
inputs = tokenizer(
"Solve step by step: Q: What is 24 * 17 ? A:",
return_tensors="pt"
).to("cuda")
_ = model.generate(**inputs, streamer=streamer, max_new_tokens=500)
```
📊 Example Output
Prompt:
Solve step by step: Q: What is 45 + 67 ?
Response:
Step 1: Add the ones digits: 5 + 7 = 12. Write down 2 and carry over 1.
Step 2: Add the tens digits plus carry: 4 + 6 + 1 = 11.
Step 3: Combine the results: 112.
Answer: 112
⚠️ Disclaimer
This model is intended for research and educational purposes only.
It may not be fully accurate for complex math reasoning tasks. Always verify critical calculations independently.
## ❤️ Made With
- [Unsloth](https://github.com/unslothai/unsloth)
- [Transformers](https://huggingface.co/docs/transformers)
- [TRL](https://huggingface.co/docs/trl)
- [Kaggle Notebook](https://www.kaggle.com/code/yaminh/finetuning-a-llm-for-math-reasoning-sbs)
--- |