README.md

---
license: mit
base_model: microsoft/Phi-4-reasoning-plus
tags:
- phi-4
- math
- reasoning
- fine-tuned
- lora
- unsloth
library_name: transformers
pipeline_tag: text-generation
---

# Phi-4 Reasoning Plus - Math SFT

This model is a Supervised finetuned version of [microsoft/Phi-4-reasoning-plus](https://huggingface.co/microsoft/Phi-4-reasoning-plus) for mathematical reasoning tasks with 30k problems from aime , numina math dataset , and various other problems.

## Training Details

- **Base Model**: microsoft/Phi-4-reasoning-plus
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation) with Unsloth
- **LoRA Config**: r=512, alpha=512
- **Target Modules**: lm_head, o_proj, v_proj, up_proj, down_proj, k_proj, q_proj, gate_proj, embed_tokens
- **Precision**: bfloat16

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "pragnyanramtha/phi-4-math-rplus",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("pragnyanramtha/phi-4-math-rplus")

# For math problems
messages = [
    {"role": "system", "content": "You are a helpful math assistant. Solve problems step by step."},
    {"role": "user", "content": "What is the sum of the first 100 positive integers?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))