OpenMath-Reason

Optimizing Mathematical Reasoning in Tiny Language Models with LoRA

OpenMath-Reason is a lightweight mathematical reasoning Tiny Language Model (TLM) developed using LoRA-based fine-tuning on AMD Instinct MI300X GPUs with ROCm-enabled workflows.

This project benchmarks compact reasoning-oriented language models across GSM8K and OpenMathReasoning datasets and demonstrates that parameter-efficient fine-tuning can significantly improve mathematical reasoning performance in small models.

---

Features

• LoRA-based mathematical reasoning fine-tuning
• ROCm + AMD MI300X optimized workflows
• GSM8K benchmark evaluation
• OpenMathReasoning evaluation
• Tiny Language Model benchmarking
• Inference latency vs intelligence analysis
• Lightweight mathematical reasoning adaptation

---

Evaluated Models

Model	Parameters
Qwen2.5-Math-1.5B	1.5B
OpenMath-Reason	1.5B
Qwen3-1.7B	1.7B
Qwen3.5-2B-Base	2B
DeepSeek-R1-Distill-Qwen-1.5B	1.5B

---

Hardware

Experiments were conducted on:

• AMD Instinct MI300X
• ROCm 7.0
• PyTorch 2.6
• 192GB HBM3 GPU Memory

---

Datasets

GSM8K

Grade-school mathematical reasoning benchmark.

OpenMathReasoning

Chain-of-thought mathematical reasoning dataset.

---

Benchmark Results

Model	GSM8K (%)	OpenMath (%)	Combined (%)	Avg Time (s)
Qwen2.5-Math-1.5B	14.50	26.00	20.25	1.87
OpenMath-Reason	29.00	26.00	27.50	3.18
Qwen3-1.7B	6.00	17.00	11.50	2.34
Qwen3.5-2B-Base	20.50	24.00	22.25	2.86
DeepSeek-R1-Distill-Qwen-1.5B	5.50	14.00	9.75	1.95

---

Training

The model was fine-tuned using:

• LoRA (Low-Rank Adaptation)
• PEFT
• Hugging Face Transformers
• ROCm-enabled PyTorch

Training was conducted for 2 epochs using reduced subsets of GSM8K and OpenMathReasoning for rapid experimentation.

---

Running Evaluation

Example:

python benchmark_combined.py

---

Key Findings

• OpenMath-Reason achieved the highest combined benchmark accuracy.
• LoRA adaptation significantly improved GSM8K performance.
• ROCm workflows proved effective for compact reasoning-oriented AI experimentation.
• Tiny Language Models can achieve meaningful mathematical reasoning improvements without frontier-scale infrastructure.

---

Technologies Used

• ROCm
• AMD Instinct MI300X
• PyTorch
• Hugging Face Transformers
• PEFT
• LoRA
• Python
• GSM8K
• OpenMathReasoning

---

Future Work

• Larger-scale benchmark evaluation
• MATH and AIME integration
• Multi-stage reasoning fine-tuning
• Self-consistency decoding
• Distributed ROCm training

---

License

Apache 2.0