Math-v1 (Gemma-3-1b-it Fine-tuned with Unsloth)

This model is a fine-tuned version of google/gemma-3-1b-it optimized for mathematical reasoning and step-by-step problem solving. It was trained on the unsloth/OpenMathReasoning-mini dataset using Unsloth Studio for accelerated performance and memory efficiency.

Model Description

• Developed by: Rudraksh Rakesh Zodage
• Base Model: google/gemma-3-1b-it
• Method: QLoRA (Quantized Low-Rank Adaptation)
• Dataset Used: unsloth/OpenMathReasoning-mini
• Primary Use Case: Solving mathematical equations, logical reasoning, and step-by-step math explanations.

---

Dataset Information: unsloth/OpenMathReasoning-mini

The model was trained on the OpenMathReasoning-mini dataset. This dataset is a curated collection of high-quality mathematical problems and detailed, step-by-step chain-of-thought (CoT) reasoning paths. It is designed to teach models how to structure their logical thinking and execute mathematical operations accurately rather than simply memorizing answers.

---

Training Configuration & Hyperparameters

The model was fine-tuned using the following settings:

Parameter	Value
Epochs	0 (Step-based training)
Max Steps	30
Batch Size	2
Learning Rate	2e-4 (0.0002)
Warmup Steps	5
Optimizer	AdamW 8-bit
Context Length	2048
LoRA Rank (R)	16
LoRA Alpha	16
LoRA Dropout	0.0
LoRA Variant	Standard LoRA

---

Training Metrics & Performance

Thanks to Unsloth's optimized CUDA kernels, training achieved the following efficiency metrics:

1. Training Loss

• Initial Loss (Step 1): ~3.25
• Final Loss (Step 30): ~1.42
• The loss decreased steadily over the 30 training steps, showing robust convergence on math reasoning tasks.

2. GPU & Memory Efficiency (RTX 4060 8GB)

• Peak VRAM Allocated: ~4.82 GB (well within the 8.0 GB limit)
• VRAM Saving: ~60% reduction in memory compared to standard PyTorch training (saving over 3 GB of VRAM).
• Speedup: 2.1x faster training compared to standard Hugging Face PEFT.

---

How to Run the Model Locally

1. Running the GGUF model via llama.cpp (Fastest CPU/GPU)

Use the exported GGUF model (gemma-3-1b-it.Q4_K_M.gguf) with llama-server to launch a local Web UI chat client:

llama-server -m gemma-3-1b-it.Q4_K_M.gguf -c 2048 --port 8080 --ngl 99

Then open http://localhost:8080 in your web browser.

2. Loading the LoRA adapter in Python (Transformers + PEFT)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "google/gemma-3-1b-it"
adapter_id = "rudrakshrakeshzodage/Math-v1"

# Load base model in float16
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# Inference Example
messages = [
    {"role": "user", "content": "Solve for x: 3x + 5 = 20. Show step-by-step reasoning."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.3)
    
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

---

Model Licensing & Usage Disclaimer

Please note that this model is intended for educational, personal, and research purposes. Standard safety and alignment filtering from the base google/gemma-3-1b-it model are preserved. Outputs should be verified for mathematical correctness as LLMs may occasionally exhibit calculation errors.