---
license: gemma
base_model: google/gemma-4-12b-it
tags: [math, aime, competition-math, lora, sft, 1111AI]
pipeline_tag: text-generation
---

# 1111AI_MathR1-SFT

Gemma 4 12B supervised-fine-tuned on contamination-filtered olympiad CoT
for AIME 2026, by **mvbalaji / 1111AI_MathR1**.

## Results (AIME 2026, majority vote @ 4)

| Model | Score |
|-------|-------|
| Base Gemma 4 12B-IT | 80.0% |
| **+ SFT (this model)** | **20.0%** |
| Lift | **-60.0%** |

## Training
- LoRA r=128, alpha=256, 1 epochs
- Datasets: NuminaMath-CoT, DeepMath-103K, AIME 1983-2024, MATH
- 8-gram contamination filter vs AIME 2026 (threshold 0.15)
- Hardware: A100/H100 80GB

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = AutoModelForCausalLM.from_pretrained("mvbalaji/1111AI_MathR1-SFT", torch_dtype=torch.bfloat16, device_map="auto")
t = AutoTokenizer.from_pretrained("mvbalaji/1111AI_MathR1-SFT")
```