Github repo for ICLR-2025 paper, Fine-tuning Large Language Models with Sparse Matrices
Python
26
26 commits
updated Feb 2, 2026
SMT: Fine-Tuning Large Language Models with Sparse Matrices
The implementation of SMT: Fine-Tuning Large Language Models with Sparse Matrices. This paper introduce a method for selecting sparse sub-matrices that aims to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational costs and memory costs. We explored both gradient-based and activation-based parameter selection methods to identify the most significant sub-matrices for downstream tasks, updating only these blocks during fine-tuning. In our experiments, we demonstrated that SMT consistently surpasses other PEFT baselines (e.g., LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67% compared to FT.
We implemented SMT in two frameworks: DeepSpeed and Hugging Face Trainer/PEFT. Instructions for setting up the environment, training, and evaluation can be found in the subfolders.
Obsrervation 1: Deepseek-R1-Distill-LLaMA8B model underperforms the base LLaMA-3-8B model on Commonsense Reasoning dataset without reasoning trace.
| DeepSeek-R1-Distill-LLaMA8B | BoolQ | PIQA | SIQA | HellaSwag | WinoGrande | ARC-e | ARC-c | OBQA | AVG |
|---|---|---|---|---|---|---|---|---|---|
| base | 53.9 | 50.0 | 37.4 | 23.4 | 25.3 | 30.6 | 28.0 | 23.4 | 34.0 |
| SMT(0.86%) | 70.6 | 66.4 | 77.8 | 62.4 | 84.6 | 60.9 | 53.2 | 72.6 | 68.6 |
| Full FT. | 71.0 | 66.2 | 76.5 | 62.2 | 85.4 | 61.8 | 52.8 | 72.6 | 68.6 |
| LLaMA3-8B Model | BoolQ | PIQA | SIQA | HellaSwag | WinoGrande | ARC-e | ARC-c | OBQA | AVG |
|---|---|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 75.7 | 88.4 | 81.4 | 96.2 | 88.2 | 92.7 | 83.2 | 88.6 | 86.8 |
Observation 2: Deepseek-R1-Distill-LLaMA8B model largely outperforms the base LLaMA-3-8B model on Math Reasoning dataset with reasoning trace.
| DeepSeek-R1-DistillLLaMA8B Model | GSM8k | SingleEq | SVAMP | MultiArith | AddSub | AQuA | AVG |
|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 60.8 | 92.5 | 70.6 | 95.3 | 87.3 | 31.2 | 73.0 |
| LLaMA3-8B Model | GSM8k | SingleEq | SVAMP | MultiArith | AddSub | AQuA | AVG |
|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 42.8 | 88.5 | 60.4 | 93.9 | 85.8 | 25.2 | 66.1 |
For Commonsense Reasoning downstream evaluation datasets, we use the processed dataset splits provided by the LLM-Adapters repository: https://github.com/AGI-Edgerunners/LLM-Adapters/tree/main/dataset. We follow the same dataset organization and evaluation settings as described in the LLM-Adapters paper.
If you find this repository or our paper useful in your research, please consider citing:
@inproceedings{he2025smt,
title={SMT: Fine-Tuning Large Language Models with Sparse Matrices},
author={He, Haoze and Li, Juncheng B and Jiang, Xuan and Miller, Heather},
booktitle={The Thirteenth International Conference on Learning Representations}
}
Github repo for ICLR-2025 paper, Fine-tuning Large Language Models with Sparse Matrices
Python
26
26 commits
updated Feb 2, 2026
SMT: Fine-Tuning Large Language Models with Sparse Matrices
The implementation of SMT: Fine-Tuning Large Language Models with Sparse Matrices. This paper introduce a method for selecting sparse sub-matrices that aims to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational costs and memory costs. We explored both gradient-based and activation-based parameter selection methods to identify the most significant sub-matrices for downstream tasks, updating only these blocks during fine-tuning. In our experiments, we demonstrated that SMT consistently surpasses other PEFT baselines (e.g., LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67% compared to FT.
We implemented SMT in two frameworks: DeepSpeed and Hugging Face Trainer/PEFT. Instructions for setting up the environment, training, and evaluation can be found in the subfolders.
Obsrervation 1: Deepseek-R1-Distill-LLaMA8B model underperforms the base LLaMA-3-8B model on Commonsense Reasoning dataset without reasoning trace.
| DeepSeek-R1-Distill-LLaMA8B | BoolQ | PIQA | SIQA | HellaSwag | WinoGrande | ARC-e | ARC-c | OBQA | AVG |
|---|---|---|---|---|---|---|---|---|---|
| base | 53.9 | 50.0 | 37.4 | 23.4 | 25.3 | 30.6 | 28.0 | 23.4 | 34.0 |
| SMT(0.86%) | 70.6 | 66.4 | 77.8 | 62.4 | 84.6 | 60.9 | 53.2 | 72.6 | 68.6 |
| Full FT. | 71.0 | 66.2 | 76.5 | 62.2 | 85.4 | 61.8 | 52.8 | 72.6 | 68.6 |
| LLaMA3-8B Model | BoolQ | PIQA | SIQA | HellaSwag | WinoGrande | ARC-e | ARC-c | OBQA | AVG |
|---|---|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 75.7 | 88.4 | 81.4 | 96.2 | 88.2 | 92.7 | 83.2 | 88.6 | 86.8 |
Observation 2: Deepseek-R1-Distill-LLaMA8B model largely outperforms the base LLaMA-3-8B model on Math Reasoning dataset with reasoning trace.
| DeepSeek-R1-DistillLLaMA8B Model | GSM8k | SingleEq | SVAMP | MultiArith | AddSub | AQuA | AVG |
|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 60.8 | 92.5 | 70.6 | 95.3 | 87.3 | 31.2 | 73.0 |
| LLaMA3-8B Model | GSM8k | SingleEq | SVAMP | MultiArith | AddSub | AQuA | AVG |
|---|---|---|---|---|---|---|---|
| SMT(0.71%) | 42.8 | 88.5 | 60.4 | 93.9 | 85.8 | 25.2 | 66.1 |
For Commonsense Reasoning downstream evaluation datasets, we use the processed dataset splits provided by the LLM-Adapters repository: https://github.com/AGI-Edgerunners/LLM-Adapters/tree/main/dataset. We follow the same dataset organization and evaluation settings as described in the LLM-Adapters paper.
If you find this repository or our paper useful in your research, please consider citing:
@inproceedings{he2025smt,
title={SMT: Fine-Tuning Large Language Models with Sparse Matrices},
author={He, Haoze and Li, Juncheng B and Jiang, Xuan and Miller, Heather},
booktitle={The Thirteenth International Conference on Learning Representations}
}